Bump the crate version to 0.7.2. (#2517)

* Add the SigLIP model.

* Add more to the forward pass of the vision model.

* Complete the forward pass.

* Add the siglip example.

* Fix.

* Another fix.

* Get everything in place.

* Add a readme.

.cargo/config.toml

@@ -1,8 +1,8 @@
-[target.x86_64-unknown-linux-gnu]
-rustflags = ["-C", "target-cpu=native"]
-
-[target.aarch64-apple-darwin]
+[build]
 rustflags = ["-C", "target-cpu=native"]
 
 [target.wasm32-unknown-unknown]
 rustflags = ["-C", "target-feature=+simd128"]
+
+[target.x86_64-apple-darwin]
+rustflags = ["-C", "target-feature=-avx,-avx2"]

.github/dependabot.yml

@@ -0,0 +1,7 @@
+version: 2
+updates:
+  - package-ecosystem: "cargo"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    open-pull-requests-limit: 5

.github/workflows/ci_cuda.yaml

@@ -5,47 +5,16 @@ on:
   pull_request:
 
 jobs:
-  start-runner:
-    name: Start self-hosted EC2 runner
-    runs-on: ubuntu-latest
-    env:
-      AWS_REGION: us-east-1
-      EC2_AMI_ID: ami-03cfed9ea28f4b002
-      EC2_INSTANCE_TYPE: g5.xlarge
-      EC2_SUBNET_ID: subnet-931b34f5,subnet-ecb993cd,subnet-943dc2d8,subnet-45371f1a,subnet-ee93e0df,subnet-fddc3dfc
-      EC2_SECURITY_GROUP: sg-030175c435ac141d6
-    outputs:
-      label: ${{ steps.start-ec2-runner.outputs.label }}
-      ec2-instance-id: ${{ steps.start-ec2-runner.outputs.ec2-instance-id }}
-    steps:
-      - name: Configure AWS credentials
-        uses: aws-actions/configure-aws-credentials@v1
-        with:
-          aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
-          aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
-          aws-region: ${{ env.AWS_REGION }}
-      - name: Start EC2 runner
-        id: start-ec2-runner
-        uses: philschmid/philschmid-ec2-github-runner@main
-        with:
-          mode: start
-          github-token: ${{ secrets.GH_PERSONAL_ACCESS_TOKEN }}
-          ec2-image-id: ${{ env.EC2_AMI_ID }}
-          ec2-instance-type: ${{ env.EC2_INSTANCE_TYPE }}
-          subnet-id: ${{ env.EC2_SUBNET_ID }}
-          security-group-id: ${{ env.EC2_SECURITY_GROUP }}
-          aws-resource-tags: > # optional, requires additional permissions
-            [
-              {"Key": "Name", "Value": "ec2-tgi-github-runner"},
-              {"Key": "GitHubRepository", "Value": "${{ github.repository }}"}
-            ]
-
   test-cuda:
     concurrency:
       group: ${{ github.workflow }}-${{ github.job }}-${{ github.head_ref || github.run_id }}
       cancel-in-progress: true
-    needs: start-runner # required to start the main job when the runner is ready
-    runs-on: ${{ needs.start-runner.outputs.label }} # run the job on the newly created runner
+    runs-on:
+      group: aws-g4dn-2xlarge
+    container:
+      image: nvidia/cuda:12.3.1-devel-ubuntu22.04
+      options: --gpus 0 
+    if: ${{ github.event.pull_request.head.repo.full_name == github.event.pull_request.base.repo.full_name }}
     permissions:
       contents: write
       packages: write
@@ -56,32 +25,10 @@ jobs:
     steps:
       - name: Checkout repository
         uses: actions/checkout@v3
+      - name: Install dependencies
+        run: apt-get update && apt install curl build-essential libssl-dev protobuf-compiler pkg-config -y
       - name: Install Rust Stable
-        run: curl https://sh.rustup.rs -sSf | sh -s -- -y
+        uses: actions-rust-lang/setup-rust-toolchain@v1
       - uses: Swatinem/rust-cache@v2
-      - run: apt update -y && apt install libssl-dev -y
       - name: Test (cuda)
-        run: PATH=$PATH:/usr/local/cuda-11.8/bin/ /root/.cargo/bin/cargo test --features cuda
-  stop-runner:
-    name: Stop self-hosted EC2 runner
-    needs:
-      - start-runner
-      - test-cuda
-    runs-on: ubuntu-latest
-    env:
-      AWS_REGION: us-east-1
-    if: ${{ always() }} # required to stop the runner even if the error happened in the previous jobs
-    steps:
-      - name: Configure AWS credentials
-        uses: aws-actions/configure-aws-credentials@v1
-        with:
-          aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
-          aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
-          aws-region: ${{ env.AWS_REGION }}
-      - name: Stop EC2 runner
-        uses: philschmid/philschmid-ec2-github-runner@main
-        with:
-          mode: stop
-          github-token: ${{ secrets.GH_PERSONAL_ACCESS_TOKEN }}
-          label: ${{ needs.start-runner.outputs.label }}
-          ec2-instance-id: ${{ needs.start-runner.outputs.ec2-instance-id }}
+        run: cargo test --features cuda

.github/workflows/python.yml

@@ -0,0 +1,68 @@
+name: PyO3-CI
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+    paths:
+      - candle-pyo3/**
+  pull_request:
+    paths:
+      - candle-pyo3/**
+
+jobs:
+  build_and_test:
+    name: Check everything builds & tests
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest] # For now, only test on Linux
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Install Rust
+        uses: actions-rs/toolchain@v1
+        with:
+          toolchain: stable
+
+      - name: Install Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: 3.11
+          architecture: "x64"
+
+      - name: Cache Cargo Registry
+        uses: actions/cache@v1
+        with:
+          path: ~/.cargo/registry
+          key: ${{ runner.os }}-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
+
+      - name: Install Protoc
+        uses: arduino/setup-protoc@v2
+        with:
+            version: "25.0"
+            repo-token: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Install
+        working-directory: ./candle-pyo3
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          pip install -U pip
+          pip install pytest maturin black
+          python -m maturin develop -r --features onnx
+
+      - name: Check style
+        working-directory: ./candle-pyo3
+        run: |
+          source .env/bin/activate
+          python stub.py --check
+          black --check .
+
+      - name: Run tests
+        working-directory: ./candle-pyo3
+        run: |
+          source .env/bin/activate
+          python -m pytest -s -v tests

.github/workflows/rust-ci.yml

@@ -1,6 +1,6 @@
-on: 
+on:
   push:
-    branches: 
+    branches:
       - main
   pull_request:
 
@@ -15,7 +15,7 @@ jobs:
         os: [ubuntu-latest, windows-latest, macOS-latest]
         rust: [stable]
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal
@@ -34,7 +34,7 @@ jobs:
         os: [ubuntu-latest, windows-latest, macOS-latest]
         rust: [stable]
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal
@@ -49,7 +49,7 @@ jobs:
     name: Rustfmt
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal
@@ -65,7 +65,7 @@ jobs:
     name: Clippy
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal

.github/workflows/trufflehog.yml

@@ -0,0 +1,15 @@
+on:
+  push:
+
+name: Secret Leaks
+
+jobs:
+  trufflehog:
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+      with:
+        fetch-depth: 0
+    - name: Secret Scanning
+      uses: trufflesecurity/trufflehog@main

.gitignore

@@ -9,6 +9,10 @@ target/
 # More information here https://doc.rust-lang.org/cargo/guide/cargo-toml-vs-cargo-lock.html
 Cargo.lock
 
+# editor config
+.helix
+.vscode
+
 # These are backup files generated by rustfmt
 **/*.rs.bk
 
@@ -23,9 +27,22 @@ flamegraph.svg
 *.dylib
 *.so
 *.swp
+*.swo
 trace-*.json
 
+candle-wasm-examples/*/build
 candle-wasm-examples/*/*.bin
-candle-wasm-examples/*/*.wav
-candle-wasm-examples/*/*.safetensors
+candle-wasm-examples/*/*.jpeg
+candle-wasm-examples/*/audios/*.wav
+candle-wasm-examples/**/*.safetensors
+candle-wasm-examples/**/*.gguf
 candle-wasm-examples/*/package-lock.json
+candle-wasm-examples/**/config*.json
+.DS_Store
+.idea/*
+__pycache__
+out.safetensors
+out.wav
+bria.mp3
+bria.safetensors
+bria.wav

.vscode/settings.json

@@ -0,0 +1,11 @@
+{
+    "[python]": {
+        "editor.defaultFormatter": "ms-python.black-formatter"
+    },
+    "python.formatting.provider": "none",
+    "python.testing.pytestArgs": [
+        "candle-pyo3"
+    ],
+    "python.testing.unittestEnabled": false,
+    "python.testing.pytestEnabled": true
+}

CHANGELOG.md

@@ -0,0 +1,113 @@
+# Changelog
+This documents the main changes to the `candle` crate.
+
+## v0.3.1 - Unreleased
+
+### Added
+
+### Modified
+
+## v0.3.0 - 2023-10-01
+
+### Added
+
+- Added the Mistral 7b v0.1 model
+  [983](https://github.com/huggingface/candle/pull/983).
+- Quantized version of the Mistral model
+  [1009](https://github.com/huggingface/candle/pull/1009).
+- Add the gelu-erf op and activation function
+  [969](https://github.com/huggingface/candle/pull/969).
+- Add the mixformer/phi-v1.5 model
+  [930](https://github.com/huggingface/candle/pull/930).
+- Add the sclice-scatter op
+  [927](https://github.com/huggingface/candle/pull/927).
+- Add the Wuerstchen diffusion model
+  [911](https://github.com/huggingface/candle/pull/911).
+
+### Modified
+
+- Support for simd128 intrinsics in some quantized vecdots
+  [982](https://github.com/huggingface/candle/pull/982).
+- Optimize the index-select cuda kernel
+  [976](https://github.com/huggingface/candle/pull/976).
+- Self-contained safetensor wrappers
+  [946](https://github.com/huggingface/candle/pull/946).
+
+## v0.2.2 - 2023-09-18
+
+### Added
+- Support for `top_p` sampling
+  [819](https://github.com/huggingface/candle/pull/819).
+- T5 model including decoding
+  [864](https://github.com/huggingface/candle/pull/864).
+- 1-d upsampling
+  [839](https://github.com/huggingface/candle/pull/839).
+
+### Modified
+- Bugfix for conv2d
+  [820](https://github.com/huggingface/candle/pull/820).
+- Support tensor based indexing using `.i`
+  [842](https://github.com/huggingface/candle/pull/842).
+
+## v0.2.1 - 2023-09-11
+
+### Added
+- Add some RNNs (GRU and LSTM) in `candle-nn`
+  [674](https://github.com/huggingface/candle/pull/674),
+  [688](https://github.com/huggingface/candle/pull/688).
+- gguf v2 support
+  [725](https://github.com/huggingface/candle/pull/725).
+- Quantized llama example in Python using the pyo3 api
+  [716](https://github.com/huggingface/candle/pull/716).
+- `candle-nn` layer for conv2d-transposed
+  [760](https://github.com/huggingface/candle/pull/760).
+- Add the Segment-Anything Model (SAM) as an example
+  [773](https://github.com/huggingface/candle/pull/773).
+- TinyViT backbone for the segment anything example
+  [787](https://github.com/huggingface/candle/pull/787).
+- Shape with holes support
+  [770](https://github.com/huggingface/candle/pull/770).
+
+### Modified
+- Dilations are now supported in conv-transpose2d.
+  [671](https://github.com/huggingface/candle/pull/671).
+- Interactive mode for the quantized model
+  [690](https://github.com/huggingface/candle/pull/690).
+- Faster softmax operation
+  [747](https://github.com/huggingface/candle/pull/747).
+- Faster convolution operations on CPU and CUDA via im2col
+  [802](https://github.com/huggingface/candle/pull/802).
+- Moving some models to a more central location
+  [796](https://github.com/huggingface/candle/pull/796).
+
+## v0.2.0 - 2023-08-30
+
+### Added
+- Add the powf op
+  [664](https://github.com/huggingface/candle/pull/664).
+- Stable Diffusion XL support
+  [647](https://github.com/huggingface/candle/pull/647).
+- Add the conv-transpose2d op
+  [635](https://github.com/huggingface/candle/pull/635).
+- Refactor the VarBuilder api
+  [627](https://github.com/huggingface/candle/pull/627).
+- Add some quantization command
+  [625](https://github.com/huggingface/candle/pull/625).
+- Support more quantized types, e.g. Q2K, Q4K, Q5K...
+  [586](https://github.com/huggingface/candle/pull/586).
+- Add pose estimation to the yolo example
+  [589](https://github.com/huggingface/candle/pull/589).
+- Api to write GGUF files
+  [585](https://github.com/huggingface/candle/pull/585).
+- Support more quantization types
+  [580](https://github.com/huggingface/candle/pull/580).
+- Add EfficientNet as an example Computer Vision model
+  [572](https://github.com/huggingface/candle/pull/572).
+- Add a group parameter to convolutions
+  [566](https://github.com/huggingface/candle/pull/566).
+- New dtype: int64
+  [563](https://github.com/huggingface/candle/pull/563).
+- Handling of the GGUF file format.
+  [559](https://github.com/huggingface/candle/pull/559).
+
+## v0.1.2 - 2023-08-21

Cargo.toml

@@ -3,19 +3,24 @@ members = [
     "candle-core",
     "candle-datasets",
     "candle-examples",
+    "candle-book",
     "candle-nn",
     "candle-pyo3",
     "candle-transformers",
-    "candle-wasm-examples/llama2-c",
-    "candle-wasm-examples/whisper",
+    "candle-wasm-examples/*",
+    "candle-wasm-tests",
+    "tensor-tools",
 ]
 exclude = [
-    "candle-flash-attn",
-    "candle-kernels",
+   "candle-flash-attn",
+   "candle-kernels",
+   "candle-metal-kernels",
+   "candle-onnx",
 ]
+resolver = "2"
 
 [workspace.package]
-version = "0.1.1"
+version = "0.7.2"
 edition = "2021"
 description = "Minimalist ML framework."
 repository = "https://github.com/huggingface/candle"
@@ -24,35 +29,50 @@ categories = ["science"]
 license = "MIT OR Apache-2.0"
 
 [workspace.dependencies]
+ab_glyph = "0.2.23"
 accelerate-src = { version = "0.3.2" }
 anyhow = { version = "1", features = ["backtrace"] }
 byteorder = "1.4.3"
+candle = { path = "./candle-core", package = "candle-core", version = "0.7.2" }
+candle-datasets = { path = "./candle-datasets", version = "0.7.2" }
+candle-flash-attn = { path = "./candle-flash-attn", version = "0.7.2" }
+candle-kernels = { path = "./candle-kernels", version = "0.7.2" }
+candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.7.2" }
+candle-nn = { path = "./candle-nn", version = "0.7.2" }
+candle-onnx = { path = "./candle-onnx", version = "0.7.2" }
+candle-transformers = { path = "./candle-transformers", version = "0.7.2" }
 clap = { version = "4.2.4", features = ["derive"] }
-cudarc = { version = "0.9.14", features = ["f16"] }
-# TODO: Switch back to the official gemm implementation once it has caught up.
-gemm = { version = "0.15.6", package = "candle-gemm" }
-hf-hub = "0.2.0"
+criterion = { version = "0.5.1", default-features=false }
+cudarc = { version = "0.12.1", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16", "cuda-version-from-build-system", "dynamic-linking"], default-features=false }
+fancy-regex = "0.13.0"
+gemm = { version = "0.17.0", features = ["wasm-simd128-enable"] }
+hf-hub = "0.3.0"
 half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
-image = { version = "0.24.7", default-features = false, features = ["jpeg", "png"] }
+hound = "3.5.1"
+image = { version = "0.25.2", default-features = false, features = ["jpeg", "png"] }
+imageproc = { version = "0.24.0", default-features = false }
 intel-mkl-src = { version = "0.8.1", features = ["mkl-static-lp64-iomp"] }
 libc = { version = "0.2.147" }
 log = "0.4"
-memmap2 = "0.7.1"
+memmap2 = { version = "0.9.3", features = ["stable_deref_trait"] }
 num_cpus = "1.15.0"
 num-traits = "0.2.15"
+parquet = { version = "51.0.0" }
 rand = "0.8.5"
 rand_distr = "0.4.3"
 rayon = "1.7.0"
-safetensors = "0.3.1"
+safetensors = "0.4.1"
 serde = { version = "1.0.171", features = ["derive"] }
+serde_plain = "1.0.2"
 serde_json = "1.0.99"
 thiserror = "1"
-tokenizers = { version = "0.13.4", default-features = false }
+tokenizers = { version = "0.19.1", default-features = false }
 tracing = "0.1.37"
 tracing-chrome = "0.7.1"
 tracing-subscriber = "0.3.7"
-wav = "1.0.0"
-zip = { version = "0.6.6", default-features = false }
+yoke = { version = "0.7.2", features = ["derive"] }
+zip = { version = "1.1.1", default-features = false }
+metal = { version = "0.27.0", features = ["mps"]}
 
 [profile.release-with-debug]
 inherits = "release"

Makefile

@@ -1,3 +1,5 @@
+.PHONY: clean-ptx clean test
+
 clean-ptx:
 	find target -name "*.ptx" -type f -delete
 	echo "" > candle-kernels/src/lib.rs
@@ -11,8 +13,4 @@ clean:
 test:
 	cargo test
 
-pyo3-test:
-	cargo build --profile=release-with-debug --package candle-pyo3
-	python3 candle-pyo3/test.py
-
 all: test

README.md

@@ -1,5 +1,5 @@
 # candle
-[![discord server](https://dcbadge.vercel.app/api/server/hugging-face-879548962464493619)](https://discord.com/channels/879548962464493619/1136218819447238726)
+[![discord server](https://dcbadge.vercel.app/api/server/hugging-face-879548962464493619)](https://discord.gg/hugging-face-879548962464493619)
 [![Latest version](https://img.shields.io/crates/v/candle-core.svg)](https://crates.io/crates/candle-core)
 [![Documentation](https://docs.rs/candle-core/badge.svg)](https://docs.rs/candle-core)
 ![License](https://img.shields.io/crates/l/candle-core.svg)
@@ -7,57 +7,190 @@
 Candle is a minimalist ML framework for Rust with a focus on performance (including GPU support) 
 and ease of use. Try our online demos: 
 [whisper](https://huggingface.co/spaces/lmz/candle-whisper),
-[llama2](https://huggingface.co/spaces/lmz/candle-llama2).
+[LLaMA2](https://huggingface.co/spaces/lmz/candle-llama2),
+[T5](https://huggingface.co/spaces/radames/Candle-T5-Generation-Wasm),
+[yolo](https://huggingface.co/spaces/lmz/candle-yolo),
+[Segment
+Anything](https://huggingface.co/spaces/radames/candle-segment-anything-wasm).
 
+## Get started
+
+Make sure that you have [`candle-core`](https://github.com/huggingface/candle/tree/main/candle-core) correctly installed as described in [**Installation**](https://huggingface.github.io/candle/guide/installation.html).
+
+Let's see how to run a simple matrix multiplication.
+Write the following to your `myapp/src/main.rs` file:
 ```rust
-let a = Tensor::randn(0f32, 1., (2, 3), &Device::Cpu)?;
-let b = Tensor::randn(0f32, 1., (3, 4), &Device::Cpu)?;
+use candle_core::{Device, Tensor};
 
-let c = a.matmul(&b)?;
-println!("{c}");
+fn main() -> Result<(), Box<dyn std::error::Error>> {
+    let device = Device::Cpu;
+
+    let a = Tensor::randn(0f32, 1., (2, 3), &device)?;
+    let b = Tensor::randn(0f32, 1., (3, 4), &device)?;
+
+    let c = a.matmul(&b)?;
+    println!("{c}");
+    Ok(())
+}
 ```
 
+`cargo run` should display a tensor of shape `Tensor[[2, 4], f32]`.
+
+
+Having installed `candle` with Cuda support, simply define the `device` to be on GPU:
+
+```diff
+- let device = Device::Cpu;
++ let device = Device::new_cuda(0)?;
+```
+
+For more advanced examples, please have a look at the following section.
+
 ## Check out our examples
 
-Check out our [examples](./candle-examples/examples/):
+These online demos run entirely in your browser:
+- [yolo](https://huggingface.co/spaces/lmz/candle-yolo): pose estimation and
+  object recognition.
+- [whisper](https://huggingface.co/spaces/lmz/candle-whisper): speech recognition.
+- [LLaMA2](https://huggingface.co/spaces/lmz/candle-llama2): text generation.
+- [T5](https://huggingface.co/spaces/radames/Candle-T5-Generation-Wasm): text generation.
+- [Phi-1.5, and Phi-2](https://huggingface.co/spaces/radames/Candle-Phi-1.5-Wasm): text generation.
+- [Segment Anything Model](https://huggingface.co/spaces/radames/candle-segment-anything-wasm): Image segmentation.
+- [BLIP](https://huggingface.co/spaces/radames/Candle-BLIP-Image-Captioning): image captioning.
 
-- [Whisper](./candle-examples/examples/whisper/): speech recognition model.
-- [Llama and Llama-v2](./candle-examples/examples/llama/): general LLM.
+We also provide a some command line based examples using state of the art models:
+
+- [LLaMA v1, v2, and v3](./candle-examples/examples/llama/): general LLM, includes
+  the SOLAR-10.7B variant.
 - [Falcon](./candle-examples/examples/falcon/): general LLM.
-- [Bert](./candle-examples/examples/bert/): useful for sentence embeddings.
-- [StarCoder](./candle-examples/examples/bigcode/): LLM specialized to code
-  generation.
+- [Codegeex4](./candle-examples/examples/codegeex4-9b/): Code completion,code interpreter,web search,fuction calling,repository-level
+- [GLM4](./candle-examples/examples/glm4/): Open Multilingual Multimodal Chat LMs by THUDM
+- [Gemma v1 and v2](./candle-examples/examples/gemma/): 2b and 7b+/9b general LLMs from Google Deepmind.
+- [RecurrentGemma](./candle-examples/examples/recurrent-gemma/): 2b and 7b
+  Griffin based models from Google that mix attention with a RNN like state.
+- [Phi-1, Phi-1.5, Phi-2, and Phi-3](./candle-examples/examples/phi/): 1.3b,
+  2.7b, and 3.8b general LLMs with performance on par with 7b models.
+- [StableLM-3B-4E1T](./candle-examples/examples/stable-lm/): a 3b general LLM
+  pre-trained on 1T tokens of English and code datasets. Also supports
+  StableLM-2, a 1.6b LLM trained on 2T tokens, as well as the code variants.
+- [Mamba](./candle-examples/examples/mamba/): an inference only
+  implementation of the Mamba state space model.
+- [Mistral7b-v0.1](./candle-examples/examples/mistral/): a 7b general LLM with
+  better performance than all publicly available 13b models as of 2023-09-28.
+- [Mixtral8x7b-v0.1](./candle-examples/examples/mixtral/): a sparse mixture of
+  experts 8x7b general LLM with better performance than a Llama 2 70B model with
+  much faster inference.
+- [StarCoder](./candle-examples/examples/bigcode/) and
+  [StarCoder2](./candle-examples/examples/starcoder2/): LLM specialized to code generation.
+- [Qwen1.5](./candle-examples/examples/qwen/): Bilingual (English/Chinese) LLMs.
+- [RWKV v5 and v6](./candle-examples/examples/rwkv/): An RNN with transformer level LLM
+  performance.
+- [Replit-code-v1.5](./candle-examples/examples/replit-code/): a 3.3b LLM specialized for code completion.
+- [Yi-6B / Yi-34B](./candle-examples/examples/yi/): two bilingual
+  (English/Chinese) general LLMs with 6b and 34b parameters.
+- [Quantized LLaMA](./candle-examples/examples/quantized/): quantized version of
+  the LLaMA model using the same quantization techniques as
+  [llama.cpp](https://github.com/ggerganov/llama.cpp).
+
+<img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/quantized/assets/aoc.gif" width="600">
+  
 - [Stable Diffusion](./candle-examples/examples/stable-diffusion/): text to
-  image generative model, yet to be optimized.
+  image generative model, support for the 1.5, 2.1, SDXL 1.0 and Turbo versions.
 
-Run them using the following commands:
+<img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg" width="200">
+
+- [Wuerstchen](./candle-examples/examples/wuerstchen/): another text to
+  image generative model.
+
+<img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/wuerstchen/assets/cat.jpg" width="200">
+
+- [yolo-v3](./candle-examples/examples/yolo-v3/) and
+  [yolo-v8](./candle-examples/examples/yolo-v8/): object detection and pose
+  estimation models.
+
+<img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/yolo-v8/assets/bike.od.jpg" width="200"><img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/yolo-v8/assets/bike.pose.jpg" width="200">
+- [segment-anything](./candle-examples/examples/segment-anything/): image
+  segmentation model with prompt.
+
+<img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/segment-anything/assets/sam_merged.jpg" width="200">
+
+- [SegFormer](./candle-examples/examples/segformer/): transformer based semantic segmentation model.
+- [Whisper](./candle-examples/examples/whisper/): speech recognition model.
+- [EnCodec](./candle-examples/examples/encodec/): high-quality audio compression
+  model using residual vector quantization.
+- [MetaVoice](./candle-examples/examples/metavoice/): foundational model for
+  text-to-speech.
+- [Parler-TTS](./candle-examples/examples/parler-tts/): large text-to-speech
+  model.
+- [T5](./candle-examples/examples/t5), [Bert](./candle-examples/examples/bert/),
+  [JinaBert](./candle-examples/examples/jina-bert/) : useful for sentence embeddings.
+- [DINOv2](./candle-examples/examples/dinov2/): computer vision model trained
+  using self-supervision (can be used for imagenet classification, depth
+  evaluation, segmentation).
+- [VGG](./candle-examples/examples/vgg/),
+  [RepVGG](./candle-examples/examples/repvgg): computer vision models.
+- [BLIP](./candle-examples/examples/blip/): image to text model, can be used to
+  generate captions for an image.
+- [CLIP](./candle-examples/examples/clip/): multi-model vision and language
+  model.
+- [TrOCR](./candle-examples/examples/trocr/): a transformer OCR model, with
+  dedicated submodels for hand-writing and printed recognition.
+- [Marian-MT](./candle-examples/examples/marian-mt/): neural machine translation
+  model, generates the translated text from the input text.
+- [Moondream](./candle-examples/examples/moondream/): tiny computer-vision model 
+  that can answer real-world questions about images.
+
+Run them using commands like:
 ```
-cargo run --example whisper --release
-cargo run --example llama --release
-cargo run --example falcon --release
-cargo run --example bert --release
-cargo run --example bigcode --release
-cargo run --example stable-diffusion --release --features image -- --prompt "a rusty robot holding a fire torch"
+cargo run --example quantized --release
 ```
 
-In order to use **CUDA** add `--features cuda` to the example command line.
+In order to use **CUDA** add `--features cuda` to the example command line. If
+you have cuDNN installed, use `--features cudnn` for even more speedups.
 
 There are also some wasm examples for whisper and
 [llama2.c](https://github.com/karpathy/llama2.c). You can either build them with
 `trunk` or try them online:
 [whisper](https://huggingface.co/spaces/lmz/candle-whisper),
-[llama2](https://huggingface.co/spaces/lmz/candle-llama2).
+[llama2](https://huggingface.co/spaces/lmz/candle-llama2),
+[T5](https://huggingface.co/spaces/radames/Candle-T5-Generation-Wasm),
+[Phi-1.5, and Phi-2](https://huggingface.co/spaces/radames/Candle-Phi-1.5-Wasm),
+[Segment Anything Model](https://huggingface.co/spaces/radames/candle-segment-anything-wasm).
 
-For llama2, run the following command to retrieve the weight files and start a
+For LLaMA2, run the following command to retrieve the weight files and start a
 test server:
 ```bash
 cd candle-wasm-examples/llama2-c
 wget https://huggingface.co/spaces/lmz/candle-llama2/resolve/main/model.bin
 wget https://huggingface.co/spaces/lmz/candle-llama2/resolve/main/tokenizer.json
-trunk serve --release --public-url /candle-llama2/ --port 8081
+trunk serve --release --port 8081
 ```
 And then head over to
-[http://localhost:8081/candle-llama2](http://localhost:8081/candle-llama2).
+[http://localhost:8081/](http://localhost:8081/).
+
+<!--- ANCHOR: useful_libraries --->
+
+## Useful External Resources
+- [`candle-tutorial`](https://github.com/ToluClassics/candle-tutorial): A
+  very detailed tutorial showing how to convert a PyTorch model to Candle.
+- [`candle-lora`](https://github.com/EricLBuehler/candle-lora): Efficient and
+  ergonomic LoRA implementation for Candle. `candle-lora` has      
+  out-of-the-box LoRA support for many models from Candle, which can be found
+  [here](https://github.com/EricLBuehler/candle-lora/tree/master/candle-lora-transformers/examples).
+- [`optimisers`](https://github.com/KGrewal1/optimisers): A collection of optimisers
+  including SGD with momentum, AdaGrad, AdaDelta, AdaMax, NAdam, RAdam, and RMSprop.
+- [`candle-vllm`](https://github.com/EricLBuehler/candle-vllm): Efficient platform for inference and
+  serving local LLMs including an OpenAI compatible API server.
+- [`candle-ext`](https://github.com/mokeyish/candle-ext): An extension library to Candle that provides PyTorch functions not currently available in Candle.
+- [`candle-coursera-ml`](https://github.com/vishpat/candle-coursera-ml): Implementation of ML algorithms from Coursera's [Machine Learning Specialization](https://www.coursera.org/specializations/machine-learning-introduction) course.
+- [`kalosm`](https://github.com/floneum/floneum/tree/master/interfaces/kalosm): A multi-modal meta-framework in Rust for interfacing with local pre-trained models with support for controlled generation, custom samplers, in-memory vector databases, audio transcription, and more.
+- [`candle-sampling`](https://github.com/EricLBuehler/candle-sampling): Sampling techniques for Candle.
+- [`gpt-from-scratch-rs`](https://github.com/jeroenvlek/gpt-from-scratch-rs): A port of Andrej Karpathy's _Let's build GPT_ tutorial on YouTube showcasing the Candle API on a toy problem.
+- [`candle-einops`](https://github.com/tomsanbear/candle-einops): A pure rust implementation of the python [einops](https://github.com/arogozhnikov/einops) library.
+
+If you have an addition to this list, please submit a pull request.
+
+<!--- ANCHOR_END: useful_libraries --->
 
 <!--- ANCHOR: features --->
 
@@ -70,11 +203,51 @@ And then head over to
     - Optimized CPU backend with optional MKL support for x86 and Accelerate for macs.
     - CUDA backend for efficiently running on GPUs, multiple GPU distribution via NCCL.
     - WASM support, run your models in a browser.
-- Model support out of the box.
-    - LLMs: Llama v1 and v2, Falcon, StarCoder.
-    - Whisper.
-    - Stable Diffusion.
+- Included models.
+    - Language Models.
+        - LLaMA v1, v2, and v3 with variants such as SOLAR-10.7B.
+        - Falcon.
+        - StarCoder, StarCoder2.
+        - Phi 1, 1.5, 2, and 3.
+        - Mamba, Minimal Mamba
+        - Gemma v1 2b and 7b+, v2 2b and 9b.
+        - Mistral 7b v0.1.
+        - Mixtral 8x7b v0.1.
+        - StableLM-3B-4E1T, StableLM-2-1.6B, Stable-Code-3B.
+        - Replit-code-v1.5-3B.
+        - Bert.
+        - Yi-6B and Yi-34B.
+        - Qwen1.5, Qwen1.5 MoE.
+        - RWKV v5 and v6.
+    - Quantized LLMs.
+        - Llama 7b, 13b, 70b, as well as the chat and code variants.
+        - Mistral 7b, and 7b instruct.
+        - Mixtral 8x7b.
+        - Zephyr 7b a and b (Mistral-7b based).
+        - OpenChat 3.5 (Mistral-7b based).
+    - Text to text.
+        - T5 and its variants: FlanT5, UL2, MADLAD400 (translation), CoEdit (Grammar correction).
+        - Marian MT (Machine Translation).
+    - Text to image.
+        - Stable Diffusion v1.5, v2.1, XL v1.0.
+        - Wurstchen v2.
+    - Image to text.
+        - BLIP.
+        - TrOCR.
+    - Audio.
+        - Whisper, multi-lingual speech-to-text.
+        - EnCodec, audio compression model.
+        - MetaVoice-1B, text-to-speech model.
+        - Parler-TTS, text-to-speech model.
+    - Computer Vision Models.
+        - DINOv2, ConvMixer, EfficientNet, ResNet, ViT, VGG, RepVGG, ConvNeXT,
+          ConvNeXTv2, MobileOne, EfficientVit (MSRA), MobileNetv4, Hiera, FastViT.
+        - yolo-v3, yolo-v8.
+        - Segment-Anything Model (SAM).
+        - SegFormer.
+- File formats: load models from safetensors, npz, ggml, or PyTorch files.
 - Serverless (on CPU), small and fast deployments.
+- Quantization support using the llama.cpp quantized types.
 
 <!--- ANCHOR_END: features --->
 
@@ -91,7 +264,7 @@ Cheatsheet:
 | Operations | `tensor.view((2, 2))`                    | `tensor.reshape((2, 2))?`                                        |
 | Operations | `a.matmul(b)`                            | `a.matmul(&b)?`                                                  |
 | Arithmetic | `a + b`                                  | `&a + &b`                                                        |
-| Device     | `tensor.to(device="cuda")`               | `tensor.to_device(&Device::Cuda(0))?`                            |
+| Device     | `tensor.to(device="cuda")`               | `tensor.to_device(&Device::new_cuda(0)?)?`                            |
 | Dtype      | `tensor.to(dtype=torch.float16)`         | `tensor.to_dtype(&DType::F16)?`                                  |
 | Saving     | `torch.save({"A": A}, "model.bin")`      | `candle::safetensors::save(&HashMap::from([("A", A)]), "model.safetensors")?` |
 | Loading    | `weights = torch.load("model.bin")`      | `candle::safetensors::load("model.safetensors", &device)`        |
@@ -108,6 +281,7 @@ Cheatsheet:
 - [candle-datasets](./candle-datasets/): Datasets and data loaders.
 - [candle-transformers](./candle-transformers): transformers-related utilities.
 - [candle-flash-attn](./candle-flash-attn): Flash attention v2 layer.
+- [candle-onnx](./candle-onnx/): ONNX model evaluation.
 
 ## FAQ
 
@@ -144,35 +318,105 @@ Finally, Rust is cool! A lot of the HF ecosystem already has Rust crates, like [
 #### Missing symbols when compiling with the mkl feature.
 
 If you get some missing symbols when compiling binaries/tests using the mkl
-features, e.g.:
+or accelerate features, e.g. for mkl you get:
 ```
   = note: /usr/bin/ld: (....o): in function `blas::sgemm':
           .../blas-0.22.0/src/lib.rs:1944: undefined reference to `sgemm_' collect2: error: ld returned 1 exit status
 
   = note: some `extern` functions couldn't be found; some native libraries may need to be installed or have their path specified
   = note: use the `-l` flag to specify native libraries to link
-  = note: use the `cargo:rustc-link-lib` directive to specify the native libraries to link with Cargo (see https://doc.rust-lang.org/cargo/reference/build-scripts.html#cargorustc-link-libkindname)
+  = note: use the `cargo:rustc-link-lib` directive to specify the native libraries to link with Cargo
+```
+or for accelerate:
+```
+Undefined symbols for architecture arm64:
+            "_dgemm_", referenced from:
+                candle_core::accelerate::dgemm::h1b71a038552bcabe in libcandle_core...
+            "_sgemm_", referenced from:
+                candle_core::accelerate::sgemm::h2cf21c592cba3c47 in libcandle_core...
+          ld: symbol(s) not found for architecture arm64
 ```
 
 This is likely due to a missing linker flag that was needed to enable the mkl library. You
-can try adding the following at the top of your binary:
-```
+can try adding the following for mkl at the top of your binary:
+```rust
 extern crate intel_mkl_src;
 ```
+or for accelerate:
+```rust
+extern crate accelerate_src;
+```
 
-#### Cannot run llama example : access to source requires login credentials
+#### Cannot run the LLaMA examples: access to source requires login credentials
 
 ```
 Error: request error: https://huggingface.co/meta-llama/Llama-2-7b-hf/resolve/main/tokenizer.json: status code 401
 ```
 
-This is likely because you're not permissioned for the llama-v2 model. To fix
-this, you have to register on the huggingface-hub, accept the [llama-v2 model
+This is likely because you're not permissioned for the LLaMA-v2 model. To fix
+this, you have to register on the huggingface-hub, accept the [LLaMA-v2 model
 conditions](https://huggingface.co/meta-llama/Llama-2-7b-hf), and set up your
 authentication token. See issue
 [#350](https://github.com/huggingface/candle/issues/350) for more details.
 
+#### Missing cute/cutlass headers when compiling flash-attn
+
+```
+  In file included from kernels/flash_fwd_launch_template.h:11:0,
+                   from kernels/flash_fwd_hdim224_fp16_sm80.cu:5:
+  kernels/flash_fwd_kernel.h:8:10: fatal error: cute/algorithm/copy.hpp: No such file or directory
+   #include <cute/algorithm/copy.hpp>
+            ^~~~~~~~~~~~~~~~~~~~~~~~~
+  compilation terminated.
+  Error: nvcc error while compiling:
+```
+[cutlass](https://github.com/NVIDIA/cutlass) is provided as a git submodule so you may want to run the following command to check it in properly.
+```bash
+git submodule update --init
+```
+
+#### Compiling with flash-attention fails
+
+```
+/usr/include/c++/11/bits/std_function.h:530:146: error: parameter packs not expanded with ‘...’:
+```
+
+This is a bug in gcc-11 triggered by the Cuda compiler. To fix this, install a different, supported gcc version - for example gcc-10, and specify the path to the compiler in the NVCC_CCBIN environment variable.
+```
+env NVCC_CCBIN=/usr/lib/gcc/x86_64-linux-gnu/10 cargo ...
+```
+
+#### Linking error on windows when running rustdoc or mdbook tests
+
+```
+Couldn't compile the test.
+---- .\candle-book\src\inference\hub.md - Using_the_hub::Using_in_a_real_model_ (line 50) stdout ----
+error: linking with `link.exe` failed: exit code: 1181
+//very long chain of linking
+ = note: LINK : fatal error LNK1181: cannot open input file 'windows.0.48.5.lib'
+```
+
+Make sure you link all native libraries that might be located outside a project target, e.g., to run mdbook tests, you should run:
+
+```
+mdbook test candle-book -L .\target\debug\deps\ `
+-L native=$env:USERPROFILE\.cargo\registry\src\index.crates.io-6f17d22bba15001f\windows_x86_64_msvc-0.42.2\lib `
+-L native=$env:USERPROFILE\.cargo\registry\src\index.crates.io-6f17d22bba15001f\windows_x86_64_msvc-0.48.5\lib
+```
+
+#### Extremely slow model load time with WSL
+
+This may be caused by the models being loaded from `/mnt/c`, more details on
+[stackoverflow](https://stackoverflow.com/questions/68972448/why-is-wsl-extremely-slow-when-compared-with-native-windows-npm-yarn-processing).
+
 #### Tracking down errors
 
 You can set `RUST_BACKTRACE=1` to be provided with backtraces when a candle
 error is generated.
+
+#### CudaRC error
+
+If you encounter an error like this one `called `Result::unwrap()` on an `Err` value: LoadLibraryExW { source: Os { code: 126, kind: Uncategorized, message: "The specified module could not be found." } }` on windows. To fix copy and rename these 3 files (make sure they are in path). The paths depend on your cuda version.
+`c:\Windows\System32\nvcuda.dll` -> `cuda.dll`
+`c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin\cublas64_12.dll` -> `cublas.dll`
+`c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin\curand64_10.dll` -> `curand.dll`

candle-book/Cargo.toml

@@ -0,0 +1,48 @@
+[package]
+name = "candle-book"
+version.workspace = true
+edition.workspace = true
+description.workspace = true
+repository.workspace = true
+keywords.workspace = true
+categories.workspace = true
+license.workspace = true
+readme = "README.md"
+
+[dependencies]
+accelerate-src = { workspace = true, optional = true }
+candle = { workspace = true }
+candle-datasets = { workspace = true }
+candle-nn = { workspace = true }
+candle-transformers = { workspace = true }
+candle-flash-attn = { workspace = true, optional = true }
+safetensors = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+num-traits = { workspace = true }
+intel-mkl-src = { workspace = true, optional = true }
+cudarc = { workspace = true, optional = true }
+half = { workspace = true, optional = true }
+image = { workspace = true, optional = true }
+anyhow = { workspace = true }
+tokio = "1.29.1"
+
+[dev-dependencies]
+byteorder = { workspace = true }
+hf-hub = { workspace = true, features=["tokio"]}
+clap = { workspace = true }
+memmap2 = { workspace = true }
+rand = { workspace = true }
+tokenizers = { workspace = true, features = ["onig"] }
+tracing = { workspace = true }
+tracing-chrome = { workspace = true }
+tracing-subscriber = { workspace = true }
+# Necessary to disambiguate with tokio in wasm examples which are 1.28.1
+parquet = { workspace = true }
+image = { workspace = true }
+
+[build-dependencies]
+anyhow = { workspace = true }
+
+[features]
+default = []

candle-book/src/SUMMARY.md

@@ -10,9 +10,14 @@
 
 # Reference Guide
 
-- [Running a model](inference/README.md)
+- [Running a model](inference/inference.md)
     - [Using the hub](inference/hub.md)
-- [Error management]()
+- [Error management](error_manage.md)
+- [Training](training/training.md)
+    - [Simplified](training/simplified.md)
+    - [MNIST](training/mnist.md)
+    - [Fine-tuning]()
+    - [Serialization]()
 - [Advanced Cuda usage]()
     - [Writing a custom kernel]()
     - [Porting a custom kernel]()
@@ -21,7 +26,3 @@
     - [Creating a WASM app]()
     - [Creating a REST api webserver]()
     - [Creating a desktop Tauri app]()
-- [Training]()
-    - [MNIST]()
-    - [Fine-tuning]()
-    - [Serialization]()

candle-book/src/error_manage.md

@@ -29,7 +29,7 @@ After adding `RUST_BACKTRACE=1`:
 Error: WithBacktrace { inner: ShapeMismatchBinaryOp { lhs: [1, 784], rhs: [1, 784], op: "matmul" }, backtrace: Backtrace [{ fn: "candle::error::Error::bt", file: "/home/nicolas/.cargo/git/checkouts/candle-5bb8ef7e0626d693/f291065/candle-core/src/error.rs", line: 200 }, { fn: "candle::tensor::Tensor::matmul", file: "/home/nicolas/.cargo/git/checkouts/candle-5bb8ef7e0626d693/f291065/candle-core/src/tensor.rs", line: 816 }, { fn: "myapp::main", file: "./src/main.rs", line: 29 }, { fn: "core::ops::function::FnOnce::call_once", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/core/src/ops/function.rs", line: 250 }, { fn: "std::sys_common::backtrace::__rust_begin_short_backtrace", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/sys_common/backtrace.rs", line: 135 }, { fn: "std::rt::lang_start::{{closure}}", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/rt.rs", line: 166 }, { fn: "core::ops::function::impls::<impl core::ops::function::FnOnce<A> for &F>::call_once", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/core/src/ops/function.rs", line: 284 }, { fn: "std::panicking::try::do_call", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/panicking.rs", line: 500 }, { fn: "std::panicking::try", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/panicking.rs", line: 464 }, { fn: "std::panic::catch_unwind", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/panic.rs", line: 142 }, { fn: "std::rt::lang_start_internal::{{closure}}", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/rt.rs", line: 148 }, { fn: "std::panicking::try::do_call", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/panicking.rs", line: 500 }, { fn: "std::panicking::try", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/panicking.rs", line: 464 }, { fn: "std::panic::catch_unwind", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/panic.rs", line: 142 }, { fn: "std::rt::lang_start_internal", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/rt.rs", line: 148 }, { fn: "std::rt::lang_start", file: "/rustc/8ede3aae28fe6e4d52b38157d7bfe0d3bceef225/library/std/src/rt.rs", line: 165 }, { fn: "main" }, { fn: "__libc_start_main" }, { fn: "_start" }] }
 ```
 
-Not super pretty at the moment, but we can see error occured on `{ fn: "myapp::main", file: "./src/main.rs", line: 29 }`
+Not super pretty at the moment, but we can see error occurred on `{ fn: "myapp::main", file: "./src/main.rs", line: 29 }`
 
 
 Another thing to note, is that since Rust is compiled it is not necessarily as easy to recover proper stacktraces

candle-book/src/guide/hello_world.md

@@ -6,7 +6,7 @@ Open `src/main.rs` and fill in this content:
 
 ```rust
 # extern crate candle_core;
-use candle_core::{DType, Device, Result, Tensor};
+use candle_core::{Device, Result, Tensor};
 
 struct Model {
     first: Tensor,
@@ -25,11 +25,11 @@ fn main() -> Result<()> {
     // Use Device::new_cuda(0)?; to use the GPU.
     let device = Device::Cpu;
 
-    let first = Tensor::zeros((784, 100), DType::F32, &device)?;
-    let second = Tensor::zeros((100, 10), DType::F32, &device)?;
+    let first = Tensor::randn(0f32, 1.0, (784, 100), &device)?;
+    let second = Tensor::randn(0f32, 1.0, (100, 10), &device)?;
     let model = Model { first, second };
 
-    let dummy_image = Tensor::zeros((1, 784), DType::F32, &device)?;
+    let dummy_image = Tensor::randn(0f32, 1.0, (1, 784), &device)?;
 
     let digit = model.forward(&dummy_image)?;
     println!("Digit {digit:?} digit");
@@ -50,7 +50,7 @@ the classical `Linear` layer. We can do as such
 
 ```rust
 # extern crate candle_core;
-# use candle_core::{DType, Device, Result, Tensor};
+# use candle_core::{Device, Result, Tensor};
 struct Linear{
     weight: Tensor,
     bias: Tensor,
@@ -80,7 +80,7 @@ This will change the model running code into a new function
 
 ```rust
 # extern crate candle_core;
-# use candle_core::{DType, Device, Result, Tensor};
+# use candle_core::{Device, Result, Tensor};
 # struct Linear{
 #     weight: Tensor,
 #     bias: Tensor,
@@ -110,15 +110,15 @@ fn main() -> Result<()> {
     let device = Device::cuda_if_available(0)?;
 
     // Creating a dummy model
-    let weight = Tensor::zeros((784, 100), DType::F32, &device)?;
-    let bias = Tensor::zeros((100, ), DType::F32, &device)?;
+    let weight = Tensor::randn(0f32, 1.0, (784, 100), &device)?;
+    let bias = Tensor::randn(0f32, 1.0, (100, ), &device)?;
     let first = Linear{weight, bias};
-    let weight = Tensor::zeros((100, 10), DType::F32, &device)?;
-    let bias = Tensor::zeros((10, ), DType::F32, &device)?;
+    let weight = Tensor::randn(0f32, 1.0, (100, 10), &device)?;
+    let bias = Tensor::randn(0f32, 1.0, (10, ), &device)?;
     let second = Linear{weight, bias};
     let model = Model { first, second };
 
-    let dummy_image = Tensor::zeros((1, 784), DType::F32, &device)?;
+    let dummy_image = Tensor::randn(0f32, 1.0, (1, 784), &device)?;
 
     // Inference on the model
     let digit = model.forward(&dummy_image)?;
@@ -146,8 +146,8 @@ And rewrite our examples using it
 ```rust
 # extern crate candle_core;
 # extern crate candle_nn;
-use candle_core::{DType, Device, Result, Tensor};
-use candle_nn::Linear;
+use candle_core::{Device, Result, Tensor};
+use candle_nn::{Linear, Module};
 
 struct Model {
     first: Linear,
@@ -167,15 +167,15 @@ fn main() -> Result<()> {
     let device = Device::Cpu;
 
     // This has changed (784, 100) -> (100, 784) !
-    let weight = Tensor::zeros((100, 784), DType::F32, &device)?;
-    let bias = Tensor::zeros((100, ), DType::F32, &device)?;
+    let weight = Tensor::randn(0f32, 1.0, (100, 784), &device)?;
+    let bias = Tensor::randn(0f32, 1.0, (100, ), &device)?;
     let first = Linear::new(weight, Some(bias));
-    let weight = Tensor::zeros((10, 100), DType::F32, &device)?;
-    let bias = Tensor::zeros((10, ), DType::F32, &device)?;
+    let weight = Tensor::randn(0f32, 1.0, (10, 100), &device)?;
+    let bias = Tensor::randn(0f32, 1.0, (10, ), &device)?;
     let second = Linear::new(weight, Some(bias));
     let model = Model { first, second };
 
-    let dummy_image = Tensor::zeros((1, 784), DType::F32, &device)?;
+    let dummy_image = Tensor::randn(0f32, 1.0, (1, 784), &device)?;
 
     let digit = model.forward(&dummy_image)?;
     println!("Digit {digit:?} digit");
@@ -188,8 +188,8 @@ Feel free to modify this example to use `Conv2d` to create a classical convnet i
 
 Now that we have the running dummy code we can get to more advanced topics:
 
-- [For PyTorch users](./guide/cheatsheet.md)
-- [Running existing models](./inference/README.md)
-- [Training models](./training/README.md)
+- [For PyTorch users](../guide/cheatsheet.md)
+- [Running existing models](../inference/inference.md)
+- [Training models](../training/training.md)
 
 

candle-book/src/guide/installation.md

@@ -1,6 +1,46 @@
 # Installation
 
-Start by creating a new app:
+**With Cuda support**:
+
+1. First, make sure that Cuda is correctly installed.
+- `nvcc --version` should print information about your Cuda compiler driver.
+- `nvidia-smi --query-gpu=compute_cap --format=csv` should print your GPUs compute capability, e.g. something
+like:
+
+```bash
+compute_cap
+8.9
+```
+
+You can also compile the Cuda kernels for a specific compute cap using the 
+`CUDA_COMPUTE_CAP=<compute cap>` environment variable.
+
+If any of the above commands errors out, please make sure to update your Cuda version.
+
+2. Create a new app and add [`candle-core`](https://github.com/huggingface/candle/tree/main/candle-core) with Cuda support.
+
+Start by creating a new cargo:
+
+```bash
+cargo new myapp
+cd myapp
+```
+
+Make sure to add the `candle-core` crate with the cuda feature:
+
+```bash
+cargo add --git https://github.com/huggingface/candle.git candle-core --features "cuda"
+```
+
+Run `cargo build` to make sure everything can be correctly built.
+
+```bash
+cargo build
+```
+
+**Without Cuda support**:
+
+Create a new app and add [`candle-core`](https://github.com/huggingface/candle/tree/main/candle-core) as follows:
 
 ```bash
 cargo new myapp
@@ -8,17 +48,12 @@ cd myapp
 cargo add --git https://github.com/huggingface/candle.git candle-core
 ```
 
-At this point, candle will be built **without** CUDA support.
-To get CUDA support use the `cuda` feature
-```bash
-cargo add --git https://github.com/huggingface/candle.git candle-core --features cuda
-```
-
-You can check everything works properly:
+Finally, run `cargo build` to make sure everything can be correctly built.
 
 ```bash
 cargo build
 ```
 
+**With mkl support**
 
 You can also see the `mkl` feature which could be interesting to get faster inference on CPU. [Using mkl](./advanced/mkl.md)

candle-book/src/inference/hub.md

@@ -39,7 +39,7 @@ cargo add hf-hub --features tokio
 ```rust,ignore
 # This is tested directly in examples crate because it needs external dependencies unfortunately:
 # See [this](https://github.com/rust-lang/mdBook/issues/706)
-{{#include ../../../candle-examples/src/lib.rs:book_hub_1}}
+{{#include ../lib.rs:book_hub_1}}
 ```
 
 
@@ -58,7 +58,7 @@ Now that we have our weights, we can use them in our bert architecture:
 # 
 # let weights = repo.get("model.safetensors").unwrap();
 use candle_core::{Device, Tensor, DType};
-use candle_nn::Linear;
+use candle_nn::{Linear, Module};
 
 let weights = candle_core::safetensors::load(weights, &Device::Cpu).unwrap();
 
@@ -81,7 +81,7 @@ For more efficient loading, instead of reading the file, you could use [`memmap2
 and will definitely be slower on network mounted disk, because it will issue more read calls.
 
 ```rust,ignore
-{{#include ../../../candle-examples/src/lib.rs:book_hub_2}}
+{{#include ../lib.rs:book_hub_2}}
 ```
 
 **Note**: This operation is **unsafe**. [See the safety notice](https://docs.rs/memmap2/latest/memmap2/struct.Mmap.html#safety).
@@ -100,5 +100,5 @@ cargo add safetensors
 
 
 ```rust,ignore
-{{#include ../../../candle-examples/src/lib.rs:book_hub_3}}
+{{#include ../lib.rs:book_hub_3}}
 ```

candle-book/src/lib.rs

@@ -0,0 +1,199 @@
+#[cfg(test)]
+pub mod simplified;
+
+#[cfg(test)]
+mod tests {
+    use anyhow::Result;
+    use candle::{DType, Device, Tensor};
+    use parquet::file::reader::SerializedFileReader;
+
+    // NOTE: Waiting on https://github.com/rust-lang/mdBook/pull/1856
+    #[rustfmt::skip]
+    #[tokio::test]
+    async fn book_hub_1() {
+// ANCHOR: book_hub_1
+use candle::Device;
+use hf_hub::api::tokio::Api;
+
+let api = Api::new().unwrap();
+let repo = api.model("bert-base-uncased".to_string());
+
+let weights_filename = repo.get("model.safetensors").await.unwrap();
+
+let weights = candle::safetensors::load(weights_filename, &Device::Cpu).unwrap();
+// ANCHOR_END: book_hub_1
+        assert_eq!(weights.len(), 206);
+    }
+
+    #[rustfmt::skip]
+    #[test]
+    fn book_hub_2() {
+        {
+// ANCHOR: book_hub_2
+use candle::Device;
+use hf_hub::api::sync::Api;
+use memmap2::Mmap;
+use std::fs;
+
+let api = Api::new().unwrap();
+let repo = api.model("bert-base-uncased".to_string());
+let weights_filename = repo.get("model.safetensors").unwrap();
+
+let file = fs::File::open(weights_filename).unwrap();
+let mmap = unsafe { Mmap::map(&file).unwrap() };
+let weights = candle::safetensors::load_buffer(&mmap[..], &Device::Cpu).unwrap();
+// ANCHOR_END: book_hub_2
+        assert_eq!(weights.len(), 206);
+    }
+
+    // #[rustfmt::skip]
+    // #[test]
+    // fn book_hub_3() {
+    {
+// ANCHOR: book_hub_3
+use candle::{DType, Device, Tensor};
+use hf_hub::api::sync::Api;
+use memmap2::Mmap;
+use safetensors::slice::IndexOp;
+use safetensors::SafeTensors;
+use std::fs;
+
+let api = Api::new().unwrap();
+let repo = api.model("bert-base-uncased".to_string());
+let weights_filename = repo.get("model.safetensors").unwrap();
+
+let file = fs::File::open(weights_filename).unwrap();
+let mmap = unsafe { Mmap::map(&file).unwrap() };
+
+// Use safetensors directly
+let tensors = SafeTensors::deserialize(&mmap[..]).unwrap();
+let view = tensors
+    .tensor("bert.encoder.layer.0.attention.self.query.weight")
+    .unwrap();
+
+// We're going to load shard with rank 1, within a world_size of 4
+// We're going to split along dimension 0 doing VIEW[start..stop, :]
+let rank = 1;
+let world_size = 4;
+let dim = 0;
+let dtype = view.dtype();
+let mut tp_shape = view.shape().to_vec();
+let size = tp_shape[0];
+
+if size % world_size != 0 {
+    panic!("The dimension is not divisible by `world_size`");
+}
+let block_size = size / world_size;
+let start = rank * block_size;
+let stop = (rank + 1) * block_size;
+
+// Everything is expressed in tensor dimension
+// bytes offsets is handled automatically for safetensors.
+
+let iterator = view.slice(start..stop).unwrap();
+
+tp_shape[dim] = block_size;
+
+// Convert safetensors Dtype to candle DType
+let dtype: DType = dtype.try_into().unwrap();
+
+// TODO: Implement from_buffer_iterator so we can skip the extra CPU alloc.
+let raw: Vec<u8> = iterator.into_iter().flatten().cloned().collect();
+let tp_tensor = Tensor::from_raw_buffer(&raw, dtype, &tp_shape, &Device::Cpu).unwrap();
+// ANCHOR_END: book_hub_3
+        assert_eq!(view.shape(), &[768, 768]);
+        assert_eq!(tp_tensor.dims(), &[192, 768]);
+    }
+}
+
+    #[allow(unused)]
+    #[rustfmt::skip]
+    fn book_training_1() -> Result<()>{
+// ANCHOR: book_training_1
+use hf_hub::{api::sync::Api, Repo, RepoType};
+
+let dataset_id = "mnist".to_string();
+
+let api = Api::new()?;
+let repo = Repo::with_revision(
+    dataset_id,
+    RepoType::Dataset,
+    "refs/convert/parquet".to_string(),
+);
+let repo = api.repo(repo);
+let test_parquet_filename = repo.get("mnist/test/0000.parquet")?;
+let train_parquet_filename = repo.get("mnist/train/0000.parquet")?;
+let test_parquet = SerializedFileReader::new(std::fs::File::open(test_parquet_filename)?)?;
+let train_parquet = SerializedFileReader::new(std::fs::File::open(train_parquet_filename)?)?;
+// ANCHOR_END: book_training_1
+// Ignore unused
+let _train = train_parquet;
+// ANCHOR: book_training_2
+for row in test_parquet {
+    for (idx, (name, field)) in row?.get_column_iter().enumerate() {
+        println!("Column id {idx}, name {name}, value {field}");
+    }
+}
+// ANCHOR_END: book_training_2
+let test_parquet_filename = repo.get("mnist/test/0000.parquet")?;
+let train_parquet_filename = repo.get("mnist/train/0000.parquet")?;
+let test_parquet = SerializedFileReader::new(std::fs::File::open(test_parquet_filename)?)?;
+let train_parquet = SerializedFileReader::new(std::fs::File::open(train_parquet_filename)?)?;
+// ANCHOR: book_training_3
+
+let test_samples = 10_000;
+let mut test_buffer_images: Vec<u8> = Vec::with_capacity(test_samples * 784);
+let mut test_buffer_labels: Vec<u8> = Vec::with_capacity(test_samples);
+for row in test_parquet{
+    for (_name, field) in row?.get_column_iter() {
+        if let parquet::record::Field::Group(subrow) = field {
+            for (_name, field) in subrow.get_column_iter() {
+                if let parquet::record::Field::Bytes(value) = field {
+                    let image = image::load_from_memory(value.data()).unwrap();
+                    test_buffer_images.extend(image.to_luma8().as_raw());
+                }
+            }
+        }else if let parquet::record::Field::Long(label) = field {
+            test_buffer_labels.push(*label as u8);
+        }
+    }
+}
+let test_images = (Tensor::from_vec(test_buffer_images, (test_samples, 784), &Device::Cpu)?.to_dtype(DType::F32)? / 255.)?;
+let test_labels = Tensor::from_vec(test_buffer_labels, (test_samples, ), &Device::Cpu)?;
+
+let train_samples = 60_000;
+let mut train_buffer_images: Vec<u8> = Vec::with_capacity(train_samples * 784);
+let mut train_buffer_labels: Vec<u8> = Vec::with_capacity(train_samples);
+for row in train_parquet{
+    for (_name, field) in row?.get_column_iter() {
+        if let parquet::record::Field::Group(subrow) = field {
+            for (_name, field) in subrow.get_column_iter() {
+                if let parquet::record::Field::Bytes(value) = field {
+                    let image = image::load_from_memory(value.data()).unwrap();
+                    train_buffer_images.extend(image.to_luma8().as_raw());
+                }
+            }
+        }else if let parquet::record::Field::Long(label) = field {
+            train_buffer_labels.push(*label as u8);
+        }
+    }
+}
+let train_images = (Tensor::from_vec(train_buffer_images, (train_samples, 784), &Device::Cpu)?.to_dtype(DType::F32)? / 255.)?;
+let train_labels = Tensor::from_vec(train_buffer_labels, (train_samples, ), &Device::Cpu)?;
+
+let mnist = candle_datasets::vision::Dataset {
+    train_images,
+    train_labels,
+    test_images,
+    test_labels,
+    labels: 10,
+};
+
+// ANCHOR_END: book_training_3
+assert_eq!(mnist.test_images.dims(), &[10_000, 784]);
+assert_eq!(mnist.test_labels.dims(), &[10_000]);
+assert_eq!(mnist.train_images.dims(), &[60_000, 784]);
+assert_eq!(mnist.train_labels.dims(), &[60_000]);
+Ok(())
+    }
+}

candle-book/src/simplified.rs

@@ -0,0 +1,196 @@
+//! #A simplified example in Rust of training a neural network and then using it based on the Candle Framework by Hugging Face.
+//! Author: Evgeny Igumnov 2023 igumnovnsk@gmail.com
+//! This program implements a neural network to predict the winner of the second round of elections based on the results of the first round.
+//!
+//! ##Basic moments:
+//!
+//! A multilayer perceptron with two hidden layers is used. The first hidden layer has 4 neurons, the second has 2 neurons.
+//! The input is a vector of 2 numbers - the percentage of votes for the first and second candidates in the first stage.
+//! The output is the number 0 or 1, where 1 means that the first candidate will win in the second stage, 0 means that he will lose.
+//! For training, samples with real data on the results of the first and second stages of different elections are used.
+//! The model is trained by backpropagation using gradient descent and the cross-entropy loss function.
+//! Model parameters (weights of neurons) are initialized randomly, then optimized during training.
+//! After training, the model is tested on a deferred sample to evaluate the accuracy.
+//! If the accuracy on the test set is below 100%, the model is considered underfit and the learning process is repeated.
+//! Thus, this neural network learns to find hidden relationships between the results of the first and second rounds of voting in order to make predictions for new data.
+
+#[rustfmt::skip]
+mod tests {
+
+use candle::{DType, Result, Tensor, D, Device};
+use candle_nn::{loss, ops, Linear, Module, VarBuilder, VarMap, Optimizer};
+
+// ANCHOR: book_training_simplified1
+const VOTE_DIM: usize = 2;
+const RESULTS: usize = 1;
+const EPOCHS: usize = 10;
+const LAYER1_OUT_SIZE: usize = 4;
+const LAYER2_OUT_SIZE: usize = 2;
+const LEARNING_RATE: f64 = 0.05;
+
+#[derive(Clone)]
+pub struct Dataset {
+    pub train_votes: Tensor,
+    pub train_results: Tensor,
+    pub test_votes: Tensor,
+    pub test_results: Tensor,
+}
+
+struct MultiLevelPerceptron {
+    ln1: Linear,
+    ln2: Linear,
+    ln3: Linear,
+}
+
+impl MultiLevelPerceptron {
+    fn new(vs: VarBuilder) -> Result<Self> {
+        let ln1 = candle_nn::linear(VOTE_DIM, LAYER1_OUT_SIZE, vs.pp("ln1"))?;
+        let ln2 = candle_nn::linear(LAYER1_OUT_SIZE, LAYER2_OUT_SIZE, vs.pp("ln2"))?;
+        let ln3 = candle_nn::linear(LAYER2_OUT_SIZE, RESULTS + 1, vs.pp("ln3"))?;
+        Ok(Self { ln1, ln2, ln3 })
+    }
+
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let xs = self.ln1.forward(xs)?;
+        let xs = xs.relu()?;
+        let xs = self.ln2.forward(&xs)?;
+        let xs = xs.relu()?;
+        self.ln3.forward(&xs)
+    }
+}
+
+// ANCHOR_END: book_training_simplified1
+
+
+
+// ANCHOR: book_training_simplified3
+#[tokio::test]
+async fn simplified() -> anyhow::Result<()> {
+
+    let dev = Device::cuda_if_available(0)?;
+
+    let train_votes_vec: Vec<u32> = vec![
+        15, 10,
+        10, 15,
+        5, 12,
+        30, 20,
+        16, 12,
+        13, 25,
+        6, 14,
+        31, 21,
+    ];
+    let train_votes_tensor = Tensor::from_vec(train_votes_vec.clone(), (train_votes_vec.len() / VOTE_DIM, VOTE_DIM), &dev)?.to_dtype(DType::F32)?;
+
+    let train_results_vec: Vec<u32> = vec![
+        1,
+        0,
+        0,
+        1,
+        1,
+        0,
+        0,
+        1,
+    ];
+    let train_results_tensor = Tensor::from_vec(train_results_vec, train_votes_vec.len() / VOTE_DIM, &dev)?;
+
+    let test_votes_vec: Vec<u32> = vec![
+        13, 9,
+        8, 14,
+        3, 10,
+    ];
+    let test_votes_tensor = Tensor::from_vec(test_votes_vec.clone(), (test_votes_vec.len() / VOTE_DIM, VOTE_DIM), &dev)?.to_dtype(DType::F32)?;
+
+    let test_results_vec: Vec<u32> = vec![
+        1,
+        0,
+        0,
+    ];
+    let test_results_tensor = Tensor::from_vec(test_results_vec.clone(), test_results_vec.len(), &dev)?;
+
+    let m = Dataset {
+        train_votes: train_votes_tensor,
+        train_results: train_results_tensor,
+        test_votes: test_votes_tensor,
+        test_results: test_results_tensor,
+    };
+
+    let trained_model: MultiLevelPerceptron;
+    loop {
+        println!("Trying to train neural network.");
+        match train(m.clone(), &dev) {
+            Ok(model) => {
+                trained_model = model;
+                break;
+            },
+            Err(e) => {
+                println!("Error: {}", e);
+                continue;
+            }
+        }
+
+    }
+
+    let real_world_votes: Vec<u32> = vec![
+        13, 22,
+    ];
+
+    let tensor_test_votes = Tensor::from_vec(real_world_votes.clone(), (1, VOTE_DIM), &dev)?.to_dtype(DType::F32)?;
+
+    let final_result = trained_model.forward(&tensor_test_votes)?;
+
+    let result = final_result
+        .argmax(D::Minus1)?
+        .to_dtype(DType::F32)?
+        .get(0).map(|x| x.to_scalar::<f32>())??;
+    println!("real_life_votes: {:?}", real_world_votes);
+    println!("neural_network_prediction_result: {:?}", result);
+
+    Ok(())
+
+}
+// ANCHOR_END: book_training_simplified3
+
+// ANCHOR: book_training_simplified2
+fn train(m: Dataset, dev: &Device) -> anyhow::Result<MultiLevelPerceptron> {
+    let train_results = m.train_results.to_device(dev)?;
+    let train_votes = m.train_votes.to_device(dev)?;
+    let varmap = VarMap::new();
+    let vs = VarBuilder::from_varmap(&varmap, DType::F32, dev);
+    let model = MultiLevelPerceptron::new(vs.clone())?;
+    let mut sgd = candle_nn::SGD::new(varmap.all_vars(), LEARNING_RATE)?;
+    let test_votes = m.test_votes.to_device(dev)?;
+    let test_results = m.test_results.to_device(dev)?;
+    let mut final_accuracy: f32 = 0.0;
+    for epoch in 1..EPOCHS + 1 {
+        let logits = model.forward(&train_votes)?;
+        let log_sm = ops::log_softmax(&logits, D::Minus1)?;
+        let loss = loss::nll(&log_sm, &train_results)?;
+        sgd.backward_step(&loss)?;
+
+        let test_logits = model.forward(&test_votes)?;
+        let sum_ok = test_logits
+            .argmax(D::Minus1)?
+            .eq(&test_results)?
+            .to_dtype(DType::F32)?
+            .sum_all()?
+            .to_scalar::<f32>()?;
+        let test_accuracy = sum_ok / test_results.dims1()? as f32;
+        final_accuracy = 100. * test_accuracy;
+        println!("Epoch: {epoch:3} Train loss: {:8.5} Test accuracy: {:5.2}%",
+                 loss.to_scalar::<f32>()?,
+                 final_accuracy
+        );
+        if final_accuracy == 100.0 {
+            break;
+        }
+    }
+    if final_accuracy < 100.0 {
+        Err(anyhow::Error::msg("The model is not trained well enough."))
+    } else {
+        Ok(model)
+    }
+}
+// ANCHOR_END: book_training_simplified2
+
+
+}

candle-book/src/training/README.md

@@ -1 +0,0 @@
-# Training

candle-book/src/training/mnist.md

@@ -1 +1,10 @@
 # MNIST
+
+So we now have downloaded the MNIST parquet files, let's put them in a simple struct.
+
+```rust,ignore
+{{#include ../lib.rs:book_training_3}}
+```
+
+The parsing of the file and putting it into single tensors requires the dataset to fit the entire memory.
+It is quite rudimentary, but simple enough for a small dataset like MNIST.

candle-book/src/training/simplified.md

@@ -0,0 +1,45 @@
+# Simplified
+
+## How its works
+
+This program implements a neural network to predict the winner of the second round of elections based on the results of the first round.
+
+Basic moments:
+
+1. A multilayer perceptron with two hidden layers is used. The first hidden layer has 4 neurons, the second has 2 neurons.
+2. The input is a vector of 2 numbers - the percentage of votes for the first and second candidates in the first stage.
+3. The output is the number 0 or 1, where 1 means that the first candidate will win in the second stage, 0 means that he will lose.
+4. For training, samples with real data on the results of the first and second stages of different elections are used.
+5. The model is trained by backpropagation using gradient descent and the cross-entropy loss function.
+6. Model parameters (weights of neurons) are initialized randomly, then optimized during training.
+7. After training, the model is tested on a deferred sample to evaluate the accuracy.
+8. If the accuracy on the test set is below 100%, the model is considered underfit and the learning process is repeated.
+
+Thus, this neural network learns to find hidden relationships between the results of the first and second rounds of voting in order to make predictions for new data.
+
+
+```rust,ignore
+{{#include ../simplified.rs:book_training_simplified1}}
+```
+
+```rust,ignore
+{{#include ../simplified.rs:book_training_simplified2}}
+```
+
+```rust,ignore
+{{#include ../simplified.rs:book_training_simplified3}}
+```
+
+
+## Example output
+
+```bash
+Trying to train neural network.
+Epoch:   1 Train loss:  4.42555 Test accuracy:  0.00%
+Epoch:   2 Train loss:  0.84677 Test accuracy: 33.33%
+Epoch:   3 Train loss:  2.54335 Test accuracy: 33.33%
+Epoch:   4 Train loss:  0.37806 Test accuracy: 33.33%
+Epoch:   5 Train loss:  0.36647 Test accuracy: 100.00%
+real_life_votes: [13, 22]
+neural_network_prediction_result: 0.0
+```

candle-book/src/training/training.md

@@ -0,0 +1,39 @@
+# Training
+
+
+Training starts with data. We're going to use the huggingface hub and 
+start with the Hello world dataset of machine learning, MNIST.
+
+Let's start with downloading `MNIST` from [huggingface](https://huggingface.co/datasets/mnist).
+
+This requires [`hf-hub`](https://github.com/huggingface/hf-hub).
+```bash
+cargo add hf-hub
+```
+
+This is going to be very hands-on for now.
+
+```rust,ignore
+{{#include ../../../candle-examples/src/lib.rs:book_training_1}}
+```
+
+This uses the standardized `parquet` files from the `refs/convert/parquet` branch on every dataset.
+Our handles are now [`parquet::file::serialized_reader::SerializedFileReader`].
+
+We can inspect the content of the files with:
+
+```rust,ignore
+{{#include ../../../candle-examples/src/lib.rs:book_training_2}}
+```
+
+You should see something like:
+
+```bash
+Column id 1, name label, value 6
+Column id 0, name image, value {bytes: [137, ....]
+Column id 1, name label, value 8
+Column id 0, name image, value {bytes: [137, ....]
+```
+
+So each row contains 2 columns (image, label) with image being saved as bytes.
+Let's put them into a useful struct.

candle-core/Cargo.toml

@@ -12,7 +12,9 @@ readme = "README.md"
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
 byteorder = { workspace = true }
-candle-kernels = { path = "../candle-kernels", version = "0.1.1", optional = true }
+candle-kernels = { workspace = true, optional = true }
+candle-metal-kernels = { workspace = true, optional = true }
+metal = { workspace = true, optional = true}
 cudarc = { workspace = true, optional = true }
 gemm = { workspace = true }
 half = { workspace = true }
@@ -26,11 +28,14 @@ rand_distr = { workspace = true }
 rayon = { workspace = true }
 safetensors = { workspace = true }
 thiserror = { workspace = true }
+yoke = { workspace = true }
 zip = { workspace = true }
 
 [dev-dependencies]
 anyhow = { workspace = true }
 clap = { workspace = true }
+criterion = { workspace = true }
+
 
 [features]
 default = []
@@ -38,3 +43,12 @@ cuda = ["cudarc", "dep:candle-kernels"]
 cudnn = ["cuda", "cudarc/cudnn"]
 mkl = ["dep:libc", "dep:intel-mkl-src"]
 accelerate = ["dep:libc", "dep:accelerate-src"]
+metal = ["dep:metal", "dep:candle-metal-kernels"]
+
+[[bench]]
+name = "bench_main"
+harness = false
+
+[[example]]
+name = "metal_basics"
+required-features = ["metal"]

candle-core/benches/bench_main.rs

@@ -0,0 +1,12 @@
+mod benchmarks;
+
+use criterion::criterion_main;
+criterion_main!(
+    benchmarks::affine::benches,
+    benchmarks::matmul::benches,
+    benchmarks::random::benches,
+    benchmarks::where_cond::benches,
+    benchmarks::conv_transpose2d::benches,
+    benchmarks::qmatmul::benches,
+    benchmarks::unary::benches
+);

candle-core/benches/benchmarks/affine.rs

@@ -0,0 +1,43 @@
+use crate::benchmarks::{BenchDevice, BenchDeviceHandler};
+use candle_core::{DType, Device, Tensor};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use std::time::Instant;
+
+fn run(a: &Tensor) {
+    a.affine(12.34, 56.78).unwrap();
+}
+
+fn run_affine_benchmark(c: &mut Criterion, device: &Device, dtype: DType, name: &str) {
+    let b = 1;
+    let m = 1024;
+    let k = 1024;
+
+    let tensor = Tensor::zeros((b, m, k), dtype, device).unwrap();
+
+    let flops = b * m * k * dtype.size_in_bytes();
+
+    let mut group = c.benchmark_group(device.bench_name(name));
+    group.throughput(Throughput::Bytes(flops as u64));
+    group.bench_function("iter", move |b| {
+        b.iter_custom(|iters| {
+            let start = Instant::now();
+            for _i in 0..iters {
+                run(black_box(&tensor));
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    let handler = BenchDeviceHandler::new().unwrap();
+    for device in handler.devices {
+        run_affine_benchmark(c, &device, DType::F32, "affine_f32");
+        run_affine_benchmark(c, &device, DType::F16, "affine_f16");
+        run_affine_benchmark(c, &device, DType::BF16, "affine_bf16");
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/benches/benchmarks/conv_transpose2d.rs

@@ -0,0 +1,59 @@
+use crate::benchmarks::{BenchDevice, BenchDeviceHandler};
+use candle_core::{DType, Device, Tensor};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use std::time::Instant;
+
+fn run(
+    x: &Tensor,
+    k: &Tensor,
+    padding: usize,
+    output_padding: usize,
+    stride: usize,
+    dilation: usize,
+) {
+    x.conv_transpose2d(k, padding, output_padding, stride, dilation)
+        .unwrap();
+}
+
+fn run_benchmark(c: &mut Criterion, device: &Device, dtype: DType, name: &str) {
+    let t = Tensor::arange(0.0f32, 10000.0, device)
+        .unwrap()
+        .reshape((1, 4, 50, 50))
+        .unwrap()
+        .to_dtype(dtype)
+        .unwrap();
+
+    let kernel = Tensor::arange(0.0f32, 100.0, device)
+        .unwrap()
+        .reshape((4, 1, 5, 5))
+        .unwrap()
+        .to_dtype(dtype)
+        .unwrap();
+
+    let flops = t.dims().iter().product::<usize>() * dtype.size_in_bytes();
+
+    let mut group = c.benchmark_group(device.bench_name(name));
+    group.throughput(Throughput::Bytes(flops as u64));
+    group.bench_function("iter", move |b| {
+        b.iter_custom(|iters| {
+            let start = Instant::now();
+            for _i in 0..iters {
+                run(black_box(&t), black_box(&kernel), 1, 0, 1, 2);
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    let handler = BenchDeviceHandler::new().unwrap();
+    for device in handler.devices {
+        run_benchmark(c, &device, DType::F32, "conv_transpose2d_f32");
+        run_benchmark(c, &device, DType::F16, "conv_transpose2d_f16");
+        run_benchmark(c, &device, DType::BF16, "conv_transpose2d_bf16");
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/benches/benchmarks/matmul.rs

@@ -0,0 +1,44 @@
+use crate::benchmarks::{BenchDevice, BenchDeviceHandler};
+use candle_core::{DType, Device, Tensor};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use std::time::Instant;
+
+fn run(a: &Tensor, b: &Tensor) {
+    a.matmul(&b.t().unwrap()).unwrap();
+}
+
+fn run_bench(c: &mut Criterion, device: &Device) {
+    let b = 1;
+    let m = 1;
+    let n = 2048;
+    let k = 2048;
+
+    let dtype = DType::F32;
+    let lhs = Tensor::zeros((b, m, k), dtype, device).unwrap();
+    let rhs = Tensor::zeros((b, n, k), dtype, device).unwrap();
+
+    let flops = b * m * n * k;
+
+    let mut group = c.benchmark_group(device.bench_name("matmul"));
+    group.throughput(Throughput::Bytes(flops as u64));
+    group.bench_function("iter", move |b| {
+        b.iter_custom(|iters| {
+            let start = Instant::now();
+            for _i in 0..iters {
+                run(black_box(&lhs), black_box(&rhs));
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    let handler = BenchDeviceHandler::new().unwrap();
+    for device in handler.devices {
+        run_bench(c, &device);
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/benches/benchmarks/mod.rs

@@ -0,0 +1,69 @@
+pub(crate) mod affine;
+pub(crate) mod conv_transpose2d;
+pub(crate) mod matmul;
+pub(crate) mod qmatmul;
+pub(crate) mod random;
+pub(crate) mod unary;
+pub(crate) mod where_cond;
+
+use candle_core::{Device, Result};
+
+pub(crate) trait BenchDevice {
+    fn sync(&self) -> Result<()>;
+
+    fn bench_name<S: Into<String>>(&self, name: S) -> String;
+}
+
+impl BenchDevice for Device {
+    fn sync(&self) -> Result<()> {
+        match self {
+            Device::Cpu => Ok(()),
+            Device::Cuda(device) => {
+                #[cfg(feature = "cuda")]
+                return Ok(device.synchronize()?);
+                #[cfg(not(feature = "cuda"))]
+                panic!("Cuda device without cuda feature enabled: {:?}", device)
+            }
+            Device::Metal(device) => {
+                #[cfg(feature = "metal")]
+                return Ok(device.wait_until_completed()?);
+                #[cfg(not(feature = "metal"))]
+                panic!("Metal device without metal feature enabled: {:?}", device)
+            }
+        }
+    }
+
+    fn bench_name<S: Into<String>>(&self, name: S) -> String {
+        match self {
+            Device::Cpu => {
+                let cpu_type = if cfg!(feature = "accelerate") {
+                    "accelerate"
+                } else if cfg!(feature = "mkl") {
+                    "mkl"
+                } else {
+                    "cpu"
+                };
+                format!("{}_{}", cpu_type, name.into())
+            }
+            Device::Cuda(_) => format!("cuda_{}", name.into()),
+            Device::Metal(_) => format!("metal_{}", name.into()),
+        }
+    }
+}
+
+struct BenchDeviceHandler {
+    devices: Vec<Device>,
+}
+
+impl BenchDeviceHandler {
+    pub fn new() -> Result<Self> {
+        let mut devices = Vec::new();
+        if cfg!(feature = "metal") {
+            devices.push(Device::new_metal(0)?);
+        } else if cfg!(feature = "cuda") {
+            devices.push(Device::new_cuda(0)?);
+        }
+        devices.push(Device::Cpu);
+        Ok(Self { devices })
+    }
+}

candle-core/benches/benchmarks/qmatmul.rs

@@ -0,0 +1,72 @@
+use crate::benchmarks::{BenchDevice, BenchDeviceHandler};
+use candle_core::{
+    quantized::{self, GgmlDType, QMatMul},
+    Device, Module, Tensor,
+};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use std::time::Instant;
+
+fn run(matmul: &QMatMul, x: &Tensor) {
+    matmul.forward(x).unwrap();
+}
+
+fn run_bench(c: &mut Criterion, device: &Device, dtype: GgmlDType) {
+    let b = 1;
+    let m = 1;
+    let n = 1024;
+    let k = 1024;
+
+    let lhs = (0..(m * k))
+        .map(|v| v as f32 / (m * k) as f32)
+        .collect::<Vec<_>>();
+    let rhs = (0..(k * n))
+        .map(|v| v as f32 / (n * k) as f32)
+        .collect::<Vec<_>>();
+
+    let lhs = Tensor::from_slice(&lhs, (m, k), device).unwrap();
+    let rhs = Tensor::from_slice(&rhs, (k, n), device).unwrap();
+
+    let qtensor = quantized::QTensor::quantize(&rhs.t().unwrap(), dtype).unwrap();
+    let matmul = quantized::QMatMul::from_qtensor(qtensor).unwrap();
+
+    let flops = b * m * n * k;
+
+    let mut group = c.benchmark_group(device.bench_name(format!("qmatmul_{:?}", dtype)));
+    group.sample_size(200);
+    group.throughput(Throughput::Bytes(flops as u64));
+    group.bench_function("iter", move |b| {
+        b.iter_custom(|iters| {
+            let start = Instant::now();
+            for _i in 0..iters {
+                run(black_box(&matmul), black_box(&lhs));
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    let handler = BenchDeviceHandler::new().unwrap();
+    for device in handler.devices {
+        for dtype in [
+            GgmlDType::F32,
+            GgmlDType::F16,
+            GgmlDType::Q4_0,
+            GgmlDType::Q4_1,
+            GgmlDType::Q5_0,
+            GgmlDType::Q5_1,
+            GgmlDType::Q8_0,
+            GgmlDType::Q2K,
+            GgmlDType::Q3K,
+            GgmlDType::Q4K,
+            GgmlDType::Q5K,
+            GgmlDType::Q6K,
+        ] {
+            run_bench(c, &device, dtype);
+        }
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/benches/benchmarks/random.rs

@@ -0,0 +1,63 @@
+use crate::benchmarks::{BenchDevice, BenchDeviceHandler};
+use candle_core::{DType, Device, Tensor};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use std::time::Instant;
+
+fn rand_uniform(a: &Tensor) {
+    a.rand_like(-1.0, 123.0).unwrap();
+}
+
+fn rand_normal(a: &Tensor) {
+    a.randn_like(100.0, 15.0).unwrap();
+}
+
+fn run_random_bench(c: &mut Criterion, device: &Device) {
+    let b = 1;
+
+    let rows = 2048;
+    let cols = 2048;
+
+    let dtype = DType::F32;
+    let tensor = Tensor::zeros((b, rows, cols), dtype, device).unwrap();
+
+    let flops = b * rows * cols * dtype.size_in_bytes();
+
+    let mut group = c.benchmark_group(device.bench_name("random_uniform"));
+    group.throughput(Throughput::Bytes(flops as u64));
+    group.bench_function("iter", move |benches| {
+        benches.iter_custom(|iters| {
+            let start = Instant::now();
+            for _i in 0..iters {
+                rand_uniform(black_box(&tensor));
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+
+    let tensor = Tensor::zeros((b, rows, cols), dtype, device).unwrap();
+
+    let mut group = c.benchmark_group(device.bench_name("random_normal"));
+    group.throughput(Throughput::Bytes(flops as u64));
+    group.bench_function("iter", move |benches| {
+        benches.iter_custom(|iters| {
+            let start = Instant::now();
+            for _i in 0..iters {
+                rand_normal(black_box(&tensor));
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    let handler = BenchDeviceHandler::new().unwrap();
+    for device in handler.devices {
+        run_random_bench(c, &device);
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/benches/benchmarks/unary.rs

@@ -0,0 +1,49 @@
+use crate::benchmarks::{BenchDevice, BenchDeviceHandler};
+use candle_core::{DType, Device, Tensor};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use std::time::Instant;
+
+fn run(a: &Tensor) {
+    a.sqrt().unwrap();
+}
+
+fn run_unary_benchmark(c: &mut Criterion, device: &Device, dtype: DType, name: &str) {
+    let b = 1;
+    let m = 1024;
+    let k = 1024;
+
+    let tensor = Tensor::arange(0.0f32, (b * m * k) as f32, device)
+        .unwrap()
+        .to_dtype(dtype)
+        .unwrap()
+        .reshape((b, m, k))
+        .unwrap();
+
+    let flops = b * m * k * dtype.size_in_bytes();
+
+    let mut group = c.benchmark_group(device.bench_name(name));
+    group.throughput(Throughput::Bytes(flops as u64));
+    group.bench_function("iter", move |b| {
+        b.iter_custom(|iters| {
+            let start = Instant::now();
+            for _i in 0..iters {
+                run(black_box(&tensor));
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    let handler = BenchDeviceHandler::new().unwrap();
+    for device in handler.devices {
+        for dtype in [DType::F32, DType::BF16, DType::F16] {
+            let name = format!("sqrt_{:?}", dtype);
+            run_unary_benchmark(c, &device, dtype, &name);
+        }
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/benches/benchmarks/where_cond.rs

@@ -0,0 +1,64 @@
+use crate::benchmarks::{BenchDevice, BenchDeviceHandler};
+use candle_core::{DType, Device, Tensor};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use std::time::Instant;
+
+fn run(a: &Tensor, b: &Tensor, c: &Tensor) {
+    a.where_cond(b, c).unwrap();
+}
+
+const fn create_cond_arr<const N: usize>() -> [u8; N] {
+    let mut arr = [0u8; N];
+    let mut i = 0;
+    while i < N {
+        arr[i] = (i % 2) as u8;
+        i += 1;
+    }
+    arr
+}
+
+const B: usize = 1;
+const M: usize = 1024;
+const K: usize = 1024;
+const SIZE: usize = B * M * K;
+
+const DATA: [u8; SIZE] = create_cond_arr::<SIZE>();
+
+fn run_where_cond_benchmark(c: &mut Criterion, device: &Device, dtype: DType, name: &str) {
+    let tensor = Tensor::from_slice(DATA.as_slice(), (B, M, K), device).unwrap();
+    let on_true = Tensor::ones((B, M, K), dtype, device).unwrap();
+    let on_false = Tensor::zeros((B, M, K), dtype, device).unwrap();
+
+    let elements = B * M * K;
+    // E.g. 2 f32 tensors + 1 u8 tensor
+    let flops = (2 * elements * dtype.size_in_bytes()) + elements;
+
+    let mut group = c.benchmark_group(device.bench_name(name));
+    group.throughput(Throughput::Bytes(flops as u64));
+    group.bench_function("iter", move |b| {
+        b.iter_custom(|iters| {
+            let start = Instant::now();
+            for _i in 0..iters {
+                run(
+                    black_box(&tensor),
+                    black_box(&on_true),
+                    black_box(&on_false),
+                );
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    let device = BenchDeviceHandler::new().unwrap();
+    for d in device.devices {
+        run_where_cond_benchmark(c, &d, DType::F32, "where_cond_f32");
+        run_where_cond_benchmark(c, &d, DType::BF16, "where_cond_bf16");
+        run_where_cond_benchmark(c, &d, DType::F16, "where_cond_f16");
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/examples/basics.rs

@@ -8,11 +8,10 @@ use anyhow::Result;
 use candle_core::{Device, Tensor};
 
 fn main() -> Result<()> {
-    let inp = Tensor::randn(0f32, 1., (2, 320, 96, 96), &Device::Cpu)?;
-    let w = Tensor::randn(0f32, 1., (320, 320, 3, 3), &Device::Cpu)?;
-    let start = std::time::Instant::now();
-    let res = inp.conv2d(&w, 0, 1);
-    println!("{:?}", start.elapsed());
-    println!("{res:?}");
+    let a = Tensor::new(&[[0.0f32, 1.0, 2.0], [3.0, 4.0, 5.0]], &Device::Cpu)?;
+    let b = Tensor::new(&[[88.0f32, 99.0]], &Device::Cpu)?;
+    let new_a = a.slice_scatter(&b, 1, 2)?;
+    assert_eq!(a.to_vec2::<f32>()?, [[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]]);
+    assert_eq!(new_a.to_vec2::<f32>()?, [[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]]);
     Ok(())
 }

candle-core/examples/cpu_benchmarks.rs

@@ -1,142 +0,0 @@
-/// This example contains some simple benchmarks so that it's easy to run them in perf etc.
-#[cfg(feature = "mkl")]
-extern crate intel_mkl_src;
-
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-
-use candle_core::{Device, Result, Tensor, D};
-use clap::{Parser, Subcommand};
-
-fn softmax<D: candle_core::shape::Dim>(xs: &Tensor, dim: D) -> Result<Tensor> {
-    let dim = dim.to_index(xs.shape(), "softmax")?;
-    let max = xs.max_keepdim(dim)?;
-    let diff = xs.broadcast_sub(&max)?;
-    let num = diff.exp()?;
-    let den = num.sum_keepdim(dim)?;
-    num.broadcast_div(&den)
-}
-
-trait Benchmark {
-    type PreProcessData;
-    type RunResult;
-
-    fn preprocess() -> Result<Self::PreProcessData>;
-    fn run_one(_: &Self::PreProcessData) -> Result<Self::RunResult>;
-
-    const ITERS: usize;
-}
-
-// Conv1d example as used in whisper.
-struct Conv1d;
-impl Benchmark for Conv1d {
-    type PreProcessData = (Tensor, Tensor);
-    type RunResult = Tensor;
-    fn preprocess() -> Result<Self::PreProcessData> {
-        let inp = Tensor::randn(0f32, 1., (1, 384, 3000), &Device::Cpu)?;
-        let w = Tensor::randn(0f32, 1., (384, 384, 3), &Device::Cpu)?;
-        Ok((inp, w))
-    }
-
-    fn run_one(d: &Self::PreProcessData) -> Result<Self::RunResult> {
-        d.0.conv1d(&d.1, 0, 1)
-    }
-
-    const ITERS: usize = 5;
-}
-
-// Conv2d example as used in stable-diffusion.
-struct Conv2d;
-impl Benchmark for Conv2d {
-    type PreProcessData = (Tensor, Tensor);
-    type RunResult = Tensor;
-
-    fn preprocess() -> Result<Self::PreProcessData> {
-        let inp = Tensor::randn(0f32, 1., (2, 320, 96, 96), &Device::Cpu)?;
-        let w = Tensor::randn(0f32, 1., (320, 320, 3, 3), &Device::Cpu)?;
-        Ok((inp, w))
-    }
-
-    fn run_one(d: &Self::PreProcessData) -> Result<Self::RunResult> {
-        d.0.conv2d(&d.1, 0, 1)
-    }
-
-    const ITERS: usize = 1;
-}
-
-struct Matmul;
-impl Benchmark for Matmul {
-    type PreProcessData = (Tensor, Tensor);
-    type RunResult = Tensor;
-    fn preprocess() -> Result<Self::PreProcessData> {
-        let lhs = Tensor::randn(0f32, 1., (1024, 1024), &Device::Cpu)?;
-        let rhs = Tensor::randn(0f32, 1., (1024, 1024), &Device::Cpu)?;
-        Ok((lhs, rhs))
-    }
-
-    fn run_one(d: &Self::PreProcessData) -> Result<Self::RunResult> {
-        d.0.matmul(&d.1)
-    }
-
-    const ITERS: usize = 100;
-}
-
-struct Softmax;
-impl Benchmark for Softmax {
-    type PreProcessData = Tensor;
-    type RunResult = Tensor;
-    fn preprocess() -> Result<Self::PreProcessData> {
-        // Typical whisper tiny size.
-        let x = Tensor::randn(0f32, 1., (1, 6, 200, 1500), &Device::Cpu)?;
-        Ok(x)
-    }
-
-    fn run_one(d: &Self::PreProcessData) -> Result<Self::RunResult> {
-        softmax(d, D::Minus1)
-    }
-
-    const ITERS: usize = 100;
-}
-
-fn run<B: Benchmark>(iters: Option<usize>) -> Result<()> {
-    use std::hint::black_box;
-
-    let iters = iters.unwrap_or(B::ITERS);
-    let d = B::preprocess()?;
-    let start = std::time::Instant::now();
-    for _iter in 0..iters {
-        let _res = black_box(B::run_one(black_box(&d))?);
-    }
-    println!("{:?}", start.elapsed() / iters as u32);
-    Ok(())
-}
-
-#[derive(Subcommand, Debug, Clone)]
-enum Task {
-    Conv1d,
-    Conv2d,
-    Matmul,
-    Softmax,
-}
-
-#[derive(Parser, Debug)]
-#[command(author, version, about, long_about = None)]
-pub struct Args {
-    /// The benchmark to be run.
-    #[command(subcommand)]
-    task: Task,
-
-    #[arg(long)]
-    iters: Option<usize>,
-}
-
-fn main() -> Result<()> {
-    let args = Args::parse();
-    match args.task {
-        Task::Conv1d => run::<Conv1d>(args.iters)?,
-        Task::Conv2d => run::<Conv2d>(args.iters)?,
-        Task::Matmul => run::<Matmul>(args.iters)?,
-        Task::Softmax => run::<Softmax>(args.iters)?,
-    }
-    Ok(())
-}

candle-core/examples/cuda_basics.rs

@@ -9,9 +9,25 @@ use candle_core::{Device, Tensor};
 
 fn main() -> Result<()> {
     let device = Device::new_cuda(0)?;
-    let t = Tensor::randn(0f32, 1f32, (2, 4, 96, 96), &device)?;
-    let w = Tensor::randn(0f32, 1f32, (320, 4, 3, 3), &device)?;
-    let res = t.conv2d(&w, 1, 1)?;
-    println!("{res:?}");
+    let x = Tensor::randn(0f32, 1.0, (8 * 4096, 8 * 4096), &device)?
+        .to_dtype(candle_core::DType::BF16)?;
+    candle_core::cuda::set_gemm_reduced_precision_f32(false);
+    candle_core::cuda::set_gemm_reduced_precision_bf16(false);
+    let _x1 = x.matmul(&x)?;
+    drop(_x1);
+    let start_time = std::time::Instant::now();
+    let _x1 = x.matmul(&x)?;
+    device.synchronize()?;
+    println!("fp32: {:?}", start_time.elapsed());
+    drop(_x1);
+    candle_core::cuda::set_gemm_reduced_precision_f32(true);
+    candle_core::cuda::set_gemm_reduced_precision_bf16(true);
+    let _x1 = x.matmul(&x)?;
+    drop(_x1);
+    let start_time = std::time::Instant::now();
+    let _x1 = x.matmul(&x)?;
+    device.synchronize()?;
+    println!("tf32: {:?}", start_time.elapsed());
+    drop(_x1);
     Ok(())
 }

candle-core/examples/metal_basics.rs

@@ -0,0 +1,28 @@
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+use anyhow::Result;
+use candle_core::{Device, Tensor};
+
+fn main() -> Result<()> {
+    // This requires the code to be run with MTL_CAPTURE_ENABLED=1
+    let device = Device::new_metal(0)?;
+    let metal_device = match &device {
+        Device::Metal(m) => m,
+        _ => anyhow::bail!("unexpected device"),
+    };
+    metal_device.capture("/tmp/candle.gputrace")?;
+    // This first synchronize ensures that a new command buffer gets created after setting up the
+    // capture scope.
+    device.synchronize()?;
+    let x = Tensor::randn(0f32, 1.0, (128, 128), &device)?;
+    let x1 = x.add(&x)?;
+    println!("{x1:?}");
+    // This second synchronize ensures that the command buffer gets commited before the end of the
+    // capture scope.
+    device.synchronize()?;
+    Ok(())
+}

candle-core/src/accelerate.rs

@@ -50,6 +50,8 @@ mod ffi {
         pub fn vvcos(dst: *mut c_double, src: *const c_double, len: *const c_int);
         pub fn vvlogf(dst: *mut c_float, src: *const c_float, len: *const c_int);
         pub fn vvlog(dst: *mut c_double, src: *const c_double, len: *const c_int);
+        pub fn vvtanhf(dst: *mut c_float, src: *const c_float, len: *const c_int);
+        pub fn vvtanh(dst: *mut c_double, src: *const c_double, len: *const c_int);
 
         pub fn vDSP_vaddD(
             _: *const c_double,
@@ -123,6 +125,42 @@ mod ffi {
             _: c_long,
             _: c_ulong,
         );
+        pub fn vDSP_vminD(
+            _: *const c_double,
+            _: c_long,
+            _: *const c_double,
+            _: c_long,
+            _: *mut c_double,
+            _: c_long,
+            _: c_ulong,
+        );
+        pub fn vDSP_vmin(
+            _: *const c_float,
+            _: c_long,
+            _: *const c_float,
+            _: c_long,
+            _: *mut c_float,
+            _: c_long,
+            _: c_ulong,
+        );
+        pub fn vDSP_vmaxD(
+            _: *const c_double,
+            _: c_long,
+            _: *const c_double,
+            _: c_long,
+            _: *mut c_double,
+            _: c_long,
+            _: c_ulong,
+        );
+        pub fn vDSP_vmax(
+            _: *const c_float,
+            _: c_long,
+            _: *const c_float,
+            _: c_long,
+            _: *mut c_float,
+            _: c_long,
+            _: c_ulong,
+        );
     }
 }
 
@@ -272,6 +310,26 @@ pub fn vd_cos(a: &[f64], y: &mut [f64]) {
     }
     unsafe { ffi::vvcos(y.as_mut_ptr(), a.as_ptr(), &(a_len as i32)) }
 }
+#[inline]
+pub fn vs_tanh(a: &[f32], y: &mut [f32]) {
+    let a_len = a.len();
+    let y_len = y.len();
+    if a_len != y_len {
+        panic!("a and y have different lengths {a_len} <> {y_len}")
+    }
+    unsafe { ffi::vvtanhf(y.as_mut_ptr(), a.as_ptr(), &(a_len as i32)) }
+}
+
+#[inline]
+pub fn vd_tanh(a: &[f64], y: &mut [f64]) {
+    let a_len = a.len();
+    let y_len = y.len();
+    if a_len != y_len {
+        panic!("a and y have different lengths {a_len} <> {y_len}")
+    }
+    unsafe { ffi::vvtanh(y.as_mut_ptr(), a.as_ptr(), &(a_len as i32)) }
+}
+
 #[inline]
 pub fn vs_ln(a: &[f32], y: &mut [f32]) {
     let a_len = a.len();
@@ -312,6 +370,70 @@ pub fn vd_sqr(a: &[f64], y: &mut [f64]) {
     y.iter_mut().zip(a.iter()).for_each(|(y, a)| *y = *a * *a)
 }
 
+#[inline]
+pub fn vs_tanh_inplace(y: &mut [f32]) {
+    unsafe { ffi::vvtanhf(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
+}
+
+#[inline]
+pub fn vd_tanh_inplace(y: &mut [f64]) {
+    unsafe { ffi::vvtanh(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
+}
+
+#[inline]
+pub fn vs_exp_inplace(y: &mut [f32]) {
+    unsafe { ffi::vvexpf(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
+}
+
+#[inline]
+pub fn vd_exp_inplace(y: &mut [f64]) {
+    unsafe { ffi::vvexp(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
+}
+
+#[inline]
+pub fn vs_gelu(vs: &[f32], ys: &mut [f32]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = (2.0f32 / std::f32::consts::PI).sqrt() * v * (1.0 + 0.044715 * v * v)
+    }
+    vs_tanh_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = 0.5 * v * (1.0 + *y)
+    }
+}
+
+#[inline]
+pub fn vd_gelu(vs: &[f64], ys: &mut [f64]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = (2.0f64 / std::f64::consts::PI).sqrt() * v * (1.0 + 0.044715 * v * v)
+    }
+    vd_tanh_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = 0.5 * v * (1.0 + *y)
+    }
+}
+
+#[inline]
+pub fn vs_silu(vs: &[f32], ys: &mut [f32]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vs_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
+#[inline]
+pub fn vd_silu(vs: &[f64], ys: &mut [f64]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vd_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
 macro_rules! binary_op {
     ($fn_name:ident, $ty:ty, $accelerate_name:ident) => {
         #[inline]
@@ -348,3 +470,7 @@ binary_op!(vs_mul, f32, vDSP_vmul);
 binary_op!(vd_mul, f64, vDSP_vmulD);
 binary_op!(vs_div, f32, vDSP_vdiv);
 binary_op!(vd_div, f64, vDSP_vdivD);
+binary_op!(vs_max, f32, vDSP_vmax);
+binary_op!(vd_max, f64, vDSP_vmaxD);
+binary_op!(vs_min, f32, vDSP_vmin);
+binary_op!(vd_min, f64, vDSP_vminD);

candle-core/src/backend.rs

@@ -15,6 +15,8 @@ pub trait BackendStorage: Sized {
 
     fn affine(&self, _: &Layout, _: f64, _: f64) -> Result<Self>;
 
+    fn powf(&self, _: &Layout, _: f64) -> Result<Self>;
+
     fn elu(&self, _: &Layout, _: f64) -> Result<Self>;
 
     fn reduce_op(&self, _: ReduceOp, _: &Layout, _: &[usize]) -> Result<Self>;
@@ -37,6 +39,14 @@ pub trait BackendStorage: Sized {
         _params: &crate::conv::ParamsConv1D,
     ) -> Result<Self>;
 
+    fn conv_transpose1d(
+        &self,
+        _l: &Layout,
+        _kernel: &Self,
+        _kernel_l: &Layout,
+        _params: &crate::conv::ParamsConvTranspose1D,
+    ) -> Result<Self>;
+
     fn conv2d(
         &self,
         _l: &Layout,
@@ -45,8 +55,17 @@ pub trait BackendStorage: Sized {
         _params: &crate::conv::ParamsConv2D,
     ) -> Result<Self>;
 
+    fn conv_transpose2d(
+        &self,
+        _l: &Layout,
+        _kernel: &Self,
+        _kernel_l: &Layout,
+        _params: &crate::conv::ParamsConvTranspose2D,
+    ) -> Result<Self>;
+
     fn avg_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self>;
     fn max_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self>;
+    fn upsample_nearest1d(&self, _: &Layout, _: usize) -> Result<Self>;
     fn upsample_nearest2d(&self, _: &Layout, _: usize, _: usize) -> Result<Self>;
 
     fn gather(&self, _: &Layout, _: &Self, _: &Layout, _: usize) -> Result<Self>;
@@ -79,6 +98,19 @@ pub trait BackendStorage: Sized {
     ) -> Result<Self>;
 
     fn copy_strided_src(&self, _: &mut Self, _: usize, _: &Layout) -> Result<()>;
+
+    #[allow(clippy::too_many_arguments)]
+    // Similar to cudaMemcpy2D, though values are in elements and not in bytes.
+    fn copy2d(
+        &self,
+        _: &mut Self,
+        _d1: usize,
+        _d2: usize,
+        _src_stride1: usize,
+        _dst_stride1: usize,
+        _src_offset: usize,
+        _dst_offset: usize,
+    ) -> Result<()>;
 }
 
 pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
@@ -95,9 +127,24 @@ pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
 
     fn ones_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;
 
+    /// # Safety
+    /// This function is unsafe as it doesn't initialize the underlying data store.
+    /// The caller should ensure that the data is properly initialized as early as possible
+    /// after this call.
+    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;
+
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage>;
+
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage>;
 
+    fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage>;
+
     fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;
 
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;
+
+    fn set_seed(&self, _: u64) -> Result<()>;
+
+    /// Synchronize should block until all the operations on the device are completed.
+    fn synchronize(&self) -> Result<()>;
 }

candle-core/src/backprop.rs

@@ -1,3 +1,4 @@
+/// Methods for backpropagation of gradients.
 use crate::op::{BinaryOp, Op, ReduceOp, UnaryOp};
 use crate::{Error, Result, Tensor, TensorId};
 use std::collections::HashMap;
@@ -15,6 +16,17 @@ fn broadcast_back(arg: &Tensor, node: &Tensor, reduced_dims: &[usize]) -> Result
     }
 }
 
+thread_local! {
+    static CANDLE_GRAD_DO_NOT_DETACH: bool = {
+        match std::env::var("CANDLE_GRAD_DO_NOT_DETACH") {
+            Ok(s) => {
+                !s.is_empty() && s != "0"
+            },
+            Err(_) => false,
+        }
+    }
+}
+
 impl Tensor {
     /// Return all the nodes that lead to this value in a topologically sorted vec, the first
     /// elements having dependencies on the latter ones, e.g. the first element if any is the
@@ -36,6 +48,8 @@ impl Tensor {
                 // Do not call recursively on the "leaf" nodes.
                 track_grad = true;
                 nodes
+            } else if node.dtype().is_int() {
+                nodes
             } else if let Some(op) = node.op() {
                 match op {
                     Op::IndexAdd(t1, t2, t3, _)
@@ -55,16 +69,27 @@ impl Tensor {
                         kernel: rhs,
                         ..
                     }
+                    | Op::ConvTranspose1D {
+                        arg: lhs,
+                        kernel: rhs,
+                        ..
+                    }
                     | Op::Conv2D {
                         arg: lhs,
                         kernel: rhs,
                         ..
                     }
+                    | Op::ConvTranspose2D {
+                        arg: lhs,
+                        kernel: rhs,
+                        ..
+                    }
                     | Op::CustomOp2(lhs, rhs, _)
                     | Op::Binary(lhs, rhs, _)
                     | Op::Gather(lhs, rhs, _)
                     | Op::IndexSelect(lhs, rhs, _)
-                    | Op::Matmul(lhs, rhs) => {
+                    | Op::Matmul(lhs, rhs)
+                    | Op::SliceScatter0(lhs, rhs, _) => {
                         let (tg, nodes) = walk(lhs, nodes, already_seen);
                         track_grad |= tg;
                         let (tg, nodes) = walk(rhs, nodes, already_seen);
@@ -85,25 +110,41 @@ impl Tensor {
                             nodes
                         }
                     }
+                    Op::Unary(_node, UnaryOp::Ceil)
+                    | Op::Unary(_node, UnaryOp::Floor)
+                    | Op::Unary(_node, UnaryOp::Round)
+                    | Op::Unary(_node, UnaryOp::Sign) => nodes,
                     Op::Reshape(node)
-                    | Op::UpsampleNearest2D(node)
+                    | Op::UpsampleNearest1D { arg: node, .. }
+                    | Op::UpsampleNearest2D { arg: node, .. }
                     | Op::AvgPool2D { arg: node, .. }
                     | Op::MaxPool2D { arg: node, .. }
                     | Op::Copy(node)
                     | Op::Broadcast(node)
                     | Op::Cmp(node, _)
-                    | Op::Reduce(node, _, _)
-                    | Op::ToDType(node)
+                    | Op::Reduce(node, ReduceOp::Min | ReduceOp::Sum | ReduceOp::Max, _)
                     | Op::ToDevice(node)
                     | Op::Transpose(node, _, _)
+                    | Op::Permute(node, _)
                     | Op::Narrow(node, _, _, _)
                     | Op::Unary(node, _)
                     | Op::Elu(node, _)
+                    | Op::Powf(node, _)
                     | Op::CustomOp1(node, _) => {
                         let (tg, nodes) = walk(node, nodes, already_seen);
                         track_grad |= tg;
                         nodes
                     }
+                    Op::ToDType(node) => {
+                        if node.dtype().is_float() {
+                            let (tg, nodes) = walk(node, nodes, already_seen);
+                            track_grad |= tg;
+                            nodes
+                        } else {
+                            nodes
+                        }
+                    }
+                    Op::Reduce(_, ReduceOp::ArgMin | ReduceOp::ArgMax, _) => nodes,
                 }
             } else {
                 nodes
@@ -127,10 +168,16 @@ impl Tensor {
             if node.is_variable() {
                 continue;
             }
-            let grad = grads.remove(node).unwrap();
-            // TODO: We should perform all these operations in place (or at least not track the
-            // whole graph). The only drawback would be if we wanted to support grad of grad but
-            // this is out of scope.
+            let grad = grads
+                .remove(node)
+                .expect("candle internal error - grad not populated");
+            // https://github.com/huggingface/candle/issues/1241
+            // Ideally, we would make these operations in place where possible to ensure that we
+            // do not have to allocate too often. Here we just call `.detach` to avoid computing
+            // the backprop graph of the backprop itself. This would be an issue for second order
+            // derivatives but these are out of scope at the moment.
+            let do_not_detach = CANDLE_GRAD_DO_NOT_DETACH.with(|b| *b);
+            let grad = if do_not_detach { grad } else { grad.detach() };
             if let Some(op) = node.op() {
                 match op {
                     Op::Binary(lhs, rhs, BinaryOp::Add) => {
@@ -161,6 +208,21 @@ impl Tensor {
                         let rhs_sum_grad = grads.or_insert(rhs)?;
                         *rhs_sum_grad = rhs_sum_grad.sub(&rhs_grad)?;
                     }
+                    Op::Binary(lhs, rhs, BinaryOp::Minimum)
+                    | Op::Binary(lhs, rhs, BinaryOp::Maximum) => {
+                        let mask_lhs = node.eq(lhs)?.to_dtype(grad.dtype())?;
+                        let mask_rhs = node.eq(rhs)?.to_dtype(grad.dtype())?;
+
+                        // If both masks are 1 one the same point, we want to scale the
+                        // gradient by 0.5 rather than 1.
+                        let lhs_grad = mask_lhs.mul(&grad)?.div(&(&mask_rhs + 1.)?)?;
+                        let lhs_sum_grad = grads.or_insert(lhs)?;
+                        *lhs_sum_grad = lhs_sum_grad.add(&lhs_grad)?;
+
+                        let rhs_grad = mask_rhs.mul(&grad)?.div(&(&mask_lhs + 1.)?)?;
+                        let rhs_sum_grad = grads.or_insert(rhs)?;
+                        *rhs_sum_grad = rhs_sum_grad.add(&rhs_grad)?;
+                    }
                     Op::WhereCond(pred, t, f) => {
                         let zeros = grad.zeros_like()?;
                         let t_sum_grad = grads.or_insert(t)?;
@@ -170,13 +232,189 @@ impl Tensor {
                         let f_grad = pred.where_cond(&zeros, &grad)?;
                         *f_sum_grad = f_sum_grad.add(&f_grad)?;
                     }
-                    Op::Conv1D { .. } => Err(Error::BackwardNotSupported { op: "conv1d" })?,
-                    Op::Conv2D { .. } => Err(Error::BackwardNotSupported { op: "conv2d" })?,
-                    Op::AvgPool2D { .. } => Err(Error::BackwardNotSupported { op: "avg-pool2d" })?,
-                    Op::MaxPool2D { .. } => Err(Error::BackwardNotSupported { op: "max-pool2d" })?,
-                    Op::UpsampleNearest2D { .. } => Err(Error::BackwardNotSupported {
-                        op: "upsample-nearest2d",
+                    Op::Conv1D {
+                        arg,
+                        kernel,
+                        padding,
+                        stride,
+                        dilation,
+                    } => {
+                        // The output height for conv_transpose1d is:
+                        // (l_in - 1) * stride - 2 * padding + dilation * (k_size - 1) + out_padding + 1
+                        let grad_l_in = grad.dim(2)?;
+                        let k_size = kernel.dim(2)?;
+                        let out_size =
+                            (grad_l_in - 1) * stride + dilation * (k_size - 1) + 1 - 2 * padding;
+                        let out_padding = arg.dim(2)? - out_size;
+                        let grad_arg = grad.conv_transpose1d(
+                            kernel,
+                            *padding,
+                            out_padding,
+                            *stride,
+                            *dilation,
+                            /* groups */ 1,
+                        )?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = sum_grad.add(&grad_arg)?;
+
+                        let grad_kernel = arg
+                            .transpose(0, 1)?
+                            .conv1d(&grad.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
+                            .transpose(0, 1)?;
+                        let sum_grad = grads.or_insert(kernel)?;
+                        let (_, _, k0) = kernel.dims3()?;
+                        let (_, _, g_k0) = grad_kernel.dims3()?;
+                        let grad_kernel = if g_k0 != k0 {
+                            grad_kernel.narrow(2, 0, k0)?
+                        } else {
+                            grad_kernel
+                        };
+                        *sum_grad = sum_grad.add(&grad_kernel)?;
+                    }
+                    Op::Conv2D {
+                        arg,
+                        kernel,
+                        padding,
+                        stride,
+                        dilation,
+                    } => {
+                        // The output height for conv_transpose2d is:
+                        // (i_h - 1) * stride - 2 * padding + dilation * (k_h - 1) + out_padding + 1
+                        let grad_h = grad.dim(2)?;
+                        let k_h = kernel.dim(2)?;
+                        let out_size =
+                            (grad_h - 1) * stride + dilation * (k_h - 1) + 1 - 2 * padding;
+                        let out_padding = arg.dim(2)? - out_size;
+                        let grad_arg = grad.conv_transpose2d(
+                            kernel,
+                            *padding,
+                            out_padding,
+                            *stride,
+                            *dilation,
+                        )?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = sum_grad.add(&grad_arg)?;
+
+                        let grad_kernel = arg
+                            .transpose(0, 1)?
+                            .conv2d(&grad.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
+                            .transpose(0, 1)?;
+                        let sum_grad = grads.or_insert(kernel)?;
+                        let (_, _, k0, k1) = kernel.dims4()?;
+                        let (_, _, g_k0, g_k1) = grad_kernel.dims4()?;
+                        let grad_kernel = if g_k0 != k0 || g_k1 != k1 {
+                            grad_kernel.narrow(2, 0, k0)?.narrow(3, 0, k1)?
+                        } else {
+                            grad_kernel
+                        };
+                        *sum_grad = sum_grad.add(&grad_kernel)?;
+                    }
+                    Op::ConvTranspose1D { .. } => Err(Error::BackwardNotSupported {
+                        op: "conv-transpose1d",
                     })?,
+                    Op::ConvTranspose2D {
+                        arg,
+                        kernel,
+                        padding,
+                        stride,
+                        dilation,
+                        output_padding: _output_padding,
+                    } => {
+                        let grad_arg = grad.conv2d(kernel, *padding, *stride, *dilation, 1)?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = sum_grad.add(&grad_arg)?;
+
+                        let grad_kernel = grad
+                            .transpose(0, 1)?
+                            .conv2d(&arg.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
+                            .transpose(0, 1)?;
+                        let sum_grad = grads.or_insert(kernel)?;
+                        let (_, _, k0, k1) = kernel.dims4()?;
+                        let (_, _, g_k0, g_k1) = grad_kernel.dims4()?;
+                        let grad_kernel = if g_k0 != k0 || g_k1 != k1 {
+                            grad_kernel.narrow(2, 0, k0)?.narrow(3, 0, k1)?
+                        } else {
+                            grad_kernel
+                        };
+                        *sum_grad = sum_grad.add(&grad_kernel)?;
+                    }
+                    Op::AvgPool2D {
+                        arg,
+                        kernel_size,
+                        stride,
+                    } => {
+                        if kernel_size != stride {
+                            crate::bail!("backward not supported for avgpool2d if ksize {kernel_size:?} != stride {stride:?}")
+                        }
+                        let (_n, _c, h, w) = arg.dims4()?;
+                        let grad_arg = grad.upsample_nearest2d(h, w)?;
+                        let grad_arg =
+                            (grad_arg * (1f64 / (kernel_size.0 * kernel_size.1) as f64))?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = sum_grad.add(&grad_arg)?;
+                    }
+                    Op::MaxPool2D {
+                        arg,
+                        kernel_size,
+                        stride,
+                    } => {
+                        if kernel_size != stride {
+                            crate::bail!("backward not supported for maxpool2d if ksize {kernel_size:?} != stride {stride:?}")
+                        }
+                        let (_n, _c, h, w) = arg.dims4()?;
+                        // For computing the max-pool gradient, we compute a mask where a 1 means
+                        // that the element is the maximum, then we apply this mask to the
+                        // upsampled gradient (taking into account that multiple max may exist so
+                        // we scale the gradient for this case).
+                        let node_upsampled = node.upsample_nearest2d(h, w)?;
+                        let mask = arg.eq(&node_upsampled)?.to_dtype(arg.dtype())?;
+                        let avg = mask.avg_pool2d_with_stride(*kernel_size, *stride)?;
+                        let grad_arg = ((grad * avg)?.upsample_nearest2d(h, w)? * mask)?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = sum_grad.add(&grad_arg)?;
+                    }
+                    Op::UpsampleNearest1D { arg, target_size } => {
+                        let (_n, c, size) = arg.dims3()?;
+                        if target_size % size != 0 {
+                            crate::bail!("backward not supported for non integer upscaling factors")
+                        }
+                        let scale = target_size / size;
+
+                        let kernel = Tensor::ones((c, 1, scale), arg.dtype(), arg.device())?;
+                        let conv_sum = grad.conv1d(&kernel, 0, scale, 1, c)?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = conv_sum;
+                    }
+                    Op::UpsampleNearest2D {
+                        arg,
+                        target_h,
+                        target_w,
+                    } => {
+                        let (_n, c, h, w) = arg.dims4()?;
+                        if target_h % h != 0 || target_w % w != 0 {
+                            crate::bail!("backward not supported for non integer upscaling factors")
+                        }
+                        let scale_h = target_h / h;
+                        let scale_w = target_w / w;
+
+                        if scale_h != scale_w {
+                            crate::bail!("backward not supported for non uniform upscaling factors")
+                        };
+                        let kernel =
+                            Tensor::ones((c, 1, scale_h, scale_w), arg.dtype(), arg.device())?;
+                        let conv_sum = grad.conv2d(&kernel, 0, scale_h, 1, c)?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = conv_sum;
+                    }
+                    Op::SliceScatter0(lhs, rhs, start_rhs) => {
+                        let rhs_sum_grad = grads.or_insert(rhs)?;
+                        let rhs_grad = grad.narrow(0, *start_rhs, rhs.dim(0)?)?;
+                        *rhs_sum_grad = rhs_sum_grad.add(&rhs_grad)?;
+
+                        let lhs_sum_grad = grads.or_insert(lhs)?;
+                        let lhs_grad = grad.slice_scatter0(&rhs.zeros_like()?, *start_rhs)?;
+                        *lhs_sum_grad = lhs_sum_grad.add(&lhs_grad)?
+                    }
                     Op::Gather(arg, indexes, dim) => {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.scatter_add(indexes, &grad, *dim)?;
@@ -251,7 +489,6 @@ impl Tensor {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&grad)?;
                     }
-                    Op::Cmp(_args, _) => {}
                     Op::Reduce(arg, ReduceOp::Max, reduced_dims) => {
                         let node = broadcast_back(arg, node, reduced_dims)?;
                         let grad = broadcast_back(arg, &grad, reduced_dims)?;
@@ -268,7 +505,7 @@ impl Tensor {
                     }
                     Op::ToDType(arg) => {
                         let sum_grad = grads.or_insert(arg)?;
-                        *sum_grad = sum_grad.add(&grad.to_dtype(node.dtype())?)?
+                        *sum_grad = sum_grad.add(&grad.to_dtype(arg.dtype())?)?
                     }
                     Op::Copy(arg) => {
                         let sum_grad = grads.or_insert(arg)?;
@@ -291,6 +528,11 @@ impl Tensor {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.sub(&(&grad * arg.sin())?)?
                     }
+                    Op::Unary(arg, UnaryOp::Tanh) => {
+                        let sum_grad = grads.or_insert(arg)?;
+                        let minus_dtanh = (node.sqr()? - 1.)?;
+                        *sum_grad = sum_grad.sub(&(&grad * &minus_dtanh)?)?
+                    }
                     Op::Unary(arg, UnaryOp::Abs) => {
                         let sum_grad = grads.or_insert(arg)?;
                         let ones = arg.ones_like()?;
@@ -336,20 +578,72 @@ impl Tensor {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&arg_grad)?
                     }
-                    Op::Reduce(_, ReduceOp::ArgMin, _) => {}
-                    Op::Reduce(_, ReduceOp::ArgMax, _) => {}
+                    Op::Unary(_, UnaryOp::Floor)
+                    | Op::Unary(_, UnaryOp::Round)
+                    | Op::Reduce(_, ReduceOp::ArgMin, _)
+                    | Op::Reduce(_, ReduceOp::ArgMax, _)
+                    | Op::Unary(_, UnaryOp::Sign)
+                    | Op::Cmp(_, _) => {}
                     Op::Reshape(arg) => {
                         let arg_grad = grad.reshape(arg.dims())?;
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&arg_grad)?
                     }
-                    Op::Unary(_, UnaryOp::Gelu) => Err(Error::BackwardNotSupported { op: "gelu" })?,
+                    Op::Unary(_, UnaryOp::Ceil) => Err(Error::BackwardNotSupported { op: "ceil" })?,
+                    Op::Unary(arg, UnaryOp::Gelu) => {
+                        let sum_grad = grads.or_insert(arg)?;
+                        let cube = arg.powf(3.)?;
+                        let tanh = (0.0356774 * &cube + (0.797885 * arg)?)?.tanh()?;
+                        let gelu_grad = (((0.5 * &tanh)?
+                            + (0.0535161 * cube + (0.398942 * arg)?)? * (1. - tanh.powf(2.)?))?
+                            + 0.5)?;
+                        *sum_grad = sum_grad.add(&(&grad * gelu_grad)?)?
+                    }
+                    Op::Unary(arg, UnaryOp::Erf) => {
+                        let sum_grad = grads.or_insert(arg)?;
+                        // d/dx erf(x) = 2/sqrt(pi) * e^(-x^2)
+                        let erf_grad =
+                            (2. / std::f64::consts::PI.sqrt()) * (arg.sqr()?.neg()?).exp()?;
+                        *sum_grad = sum_grad.add(&(&grad * erf_grad)?)?
+                    }
+                    Op::Unary(arg, UnaryOp::GeluErf) => {
+                        let sum_grad = grads.or_insert(arg)?;
+                        // d/dx gelu_erf(x) = 0.5 + 0.398942 e^(-x^2/2) x + 0.5 erf(x/sqrt(2))
+                        let neg_half_square = (arg.sqr()?.neg()? / 2.)?;
+                        let scaled_exp_arg = (0.398942 * neg_half_square.exp()? * arg)?;
+                        let arg_scaled_sqrt = (arg / 2f64.sqrt())?;
+                        let erf_scaled_sqrt = (0.5 * arg_scaled_sqrt.erf()?)?;
+                        let gelu_erf_grad = (0.5 + scaled_exp_arg + erf_scaled_sqrt)?;
+                        *sum_grad = sum_grad.add(&(&grad * gelu_erf_grad)?)?;
+                    }
                     Op::Unary(arg, UnaryOp::Relu) => {
                         let sum_grad = grads.or_insert(arg)?;
                         let relu_grad = arg.ge(&arg.zeros_like()?)?.to_dtype(arg.dtype())?;
                         *sum_grad = sum_grad.add(&(&grad * relu_grad)?)?
                     }
-                    Op::Elu(..) => Err(Error::BackwardNotSupported { op: "elu" })?,
+                    Op::Unary(arg, UnaryOp::Silu) => {
+                        let sum_grad = grads.or_insert(arg)?;
+                        // d/dx silu = sigmoid(x) * (1 + x * (1 - sigmoid(x))) = sigmoid(x) * (1 - node) + node
+                        let sigmoid_arg = (arg.neg()?.exp()? + 1.)?.recip()?;
+                        let silu_grad = &sigmoid_arg * (1. - *node) + *node;
+                        *sum_grad = sum_grad.add(&(&grad * silu_grad)?)?
+                    }
+                    Op::Elu(arg, alpha) => {
+                        // d/dx elu(x) = 1 for x > 0, alpha * e^x for x <= 0
+                        let sum_grad = grads.or_insert(arg)?;
+                        let zeros = arg.zeros_like()?;
+                        let positive_mask = arg.gt(&zeros)?.to_dtype(arg.dtype())?;
+                        let negative_mask = arg.le(&zeros)?.to_dtype(arg.dtype())?;
+                        // node == alpha * (e^x - 1) for x <= 0, reuse it
+                        let negative_exp_mask = (negative_mask * (*node + *alpha))?;
+                        let combined_mask = (positive_mask + negative_exp_mask)?;
+                        *sum_grad = sum_grad.add(&(grad * combined_mask)?)?
+                    }
+                    Op::Powf(arg, e) => {
+                        let arg_grad = (&(grad * arg.powf(e - 1.)?)? * *e)?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = sum_grad.add(&arg_grad)?
+                    }
                     Op::CustomOp1(arg, c) => {
                         if let Some(arg_grad) = c.bwd(arg, node, &grad)? {
                             let sum_grad = grads.or_insert(arg)?;
@@ -403,6 +697,15 @@ impl Tensor {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&arg_grad)?
                     }
+                    Op::Permute(arg, dims) => {
+                        let mut inv_dims = vec![0; dims.len()];
+                        for (i, &dim_idx) in dims.iter().enumerate() {
+                            inv_dims[dim_idx] = i
+                        }
+                        let arg_grad = grad.permute(inv_dims)?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = sum_grad.add(&arg_grad)?
+                    }
                 };
             }
         }
@@ -410,29 +713,38 @@ impl Tensor {
     }
 }
 
+/// A store for gradients, associating a tensor id to the corresponding gradient tensor, used for back propagation.
+#[derive(Debug)]
 pub struct GradStore(HashMap<TensorId, Tensor>);
 
 impl GradStore {
+    /// Create a new gradient store
     fn new() -> Self {
         GradStore(HashMap::new())
     }
 
+    /// Get the gradient tensor corresponding to the given tensor id
     pub fn get_id(&self, id: TensorId) -> Option<&Tensor> {
         self.0.get(&id)
     }
 
+    /// Get the gradient tensor associated with the given tensor
     pub fn get(&self, tensor: &Tensor) -> Option<&Tensor> {
         self.0.get(&tensor.id())
     }
 
+    /// Remove the gradient tensor associated with the given tensor, returning it if it exists
     pub fn remove(&mut self, tensor: &Tensor) -> Option<Tensor> {
         self.0.remove(&tensor.id())
     }
 
+    /// Insert a gradient tensor associated with the given tensor, returning the previous gradient tensor if it existed
     pub fn insert(&mut self, tensor: &Tensor, grad: Tensor) -> Option<Tensor> {
         self.0.insert(tensor.id(), grad)
     }
 
+    /// Get the gradient tensor associated with the given tensor, or, if it does not exist,
+    /// insert a tensor of zeroes, with the same shape and type as the given tensors and return it
     fn or_insert(&mut self, tensor: &Tensor) -> Result<&mut Tensor> {
         use std::collections::hash_map::Entry;
         let grad = match self.0.entry(tensor.id()) {
@@ -444,4 +756,9 @@ impl GradStore {
         };
         Ok(grad)
     }
+
+    /// Get the tensor ids of the stored gradient tensors
+    pub fn get_ids(&self) -> impl Iterator<Item = &TensorId> {
+        self.0.keys()
+    }
 }

candle-core/src/conv.rs

@@ -1,3 +1,5 @@
+use crate::{op::BackpropOp, op::Op, Error, Result, Tensor};
+
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ParamsConv1D {
     pub(crate) b_size: usize,
@@ -9,12 +11,12 @@ pub struct ParamsConv1D {
     pub(crate) k_size: usize,
     pub(crate) padding: usize,
     pub(crate) stride: usize,
+    pub(crate) dilation: usize,
 }
 
 impl ParamsConv1D {
     pub(crate) fn l_out(&self) -> usize {
-        let dilation = 1;
-        (self.l_in + 2 * self.padding - dilation * (self.k_size - 1) - 1) / self.stride + 1
+        (self.l_in + 2 * self.padding - self.dilation * (self.k_size - 1) - 1) / self.stride + 1
     }
 
     pub(crate) fn out_dims(&self) -> Vec<usize> {
@@ -23,6 +25,46 @@ impl ParamsConv1D {
     }
 }
 
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ParamsConvTranspose1D {
+    pub(crate) b_size: usize,
+    pub(crate) l_in: usize,
+    pub(crate) c_out: usize,
+    pub(crate) c_in: usize,
+    pub(crate) k_size: usize,
+    pub(crate) padding: usize,
+    pub(crate) output_padding: usize,
+    pub(crate) stride: usize,
+    pub(crate) dilation: usize,
+}
+
+impl ParamsConvTranspose1D {
+    pub(crate) fn l_out(&self) -> usize {
+        (self.l_in - 1) * self.stride - 2 * self.padding
+            + self.dilation * (self.k_size - 1)
+            + self.output_padding
+            + 1
+    }
+
+    pub(crate) fn out_dims(&self) -> Vec<usize> {
+        let l_out = self.l_out();
+        vec![self.b_size, self.c_out, l_out]
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+pub enum CudnnFwdAlgo {
+    ImplicitGemm,
+    ImplicitPrecompGemm,
+    Gemm,
+    Direct,
+    Fft,
+    FftTiling,
+    Winograd,
+    WinogradNonFused,
+    Count,
+}
+
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ParamsConv2D {
     pub(crate) b_size: usize,
@@ -34,20 +76,283 @@ pub struct ParamsConv2D {
     pub(crate) c_in: usize,
     pub(crate) padding: usize,
     pub(crate) stride: usize,
+    pub(crate) dilation: usize,
+    pub cudnn_fwd_algo: Option<CudnnFwdAlgo>,
 }
 
 impl ParamsConv2D {
     pub(crate) fn out_h(&self) -> usize {
-        let dilation = 1;
-        (self.i_h + 2 * self.padding - dilation * (self.k_h - 1) - 1) / self.stride + 1
+        (self.i_h + 2 * self.padding - self.dilation * (self.k_h - 1) - 1) / self.stride + 1
     }
 
     pub(crate) fn out_w(&self) -> usize {
-        let dilation = 1;
-        (self.i_w + 2 * self.padding - dilation * (self.k_w - 1) - 1) / self.stride + 1
+        (self.i_w + 2 * self.padding - self.dilation * (self.k_w - 1) - 1) / self.stride + 1
     }
 
     pub(crate) fn out_dims(&self) -> Vec<usize> {
         vec![self.b_size, self.c_out, self.out_h(), self.out_w()]
     }
 }
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ParamsConvTranspose2D {
+    pub(crate) b_size: usize,
+    pub(crate) i_h: usize,
+    pub(crate) i_w: usize,
+    pub(crate) k_h: usize,
+    pub(crate) k_w: usize,
+    pub(crate) c_out: usize,
+    pub(crate) c_in: usize,
+    pub(crate) padding: usize,
+    pub(crate) output_padding: usize,
+    pub(crate) stride: usize,
+    pub(crate) dilation: usize,
+}
+
+impl ParamsConvTranspose2D {
+    pub(crate) fn out_h(&self) -> usize {
+        (self.i_h - 1) * self.stride + self.dilation * (self.k_h - 1) + self.output_padding + 1
+            - 2 * self.padding
+    }
+
+    pub(crate) fn out_w(&self) -> usize {
+        (self.i_w - 1) * self.stride + self.dilation * (self.k_w - 1) + self.output_padding + 1
+            - 2 * self.padding
+    }
+
+    pub(crate) fn out_dims(&self) -> Vec<usize> {
+        vec![self.b_size, self.c_out, self.out_h(), self.out_w()]
+    }
+}
+
+impl Tensor {
+    fn conv1d_single_group(&self, kernel: &Self, params: &ParamsConv1D) -> Result<Self> {
+        let storage =
+            self.storage()
+                .conv1d(self.layout(), &kernel.storage(), kernel.layout(), params)?;
+        let op = BackpropOp::new2(self, kernel, |arg, kernel| Op::Conv1D {
+            arg,
+            kernel,
+            padding: params.padding,
+            stride: params.stride,
+            dilation: params.dilation,
+        });
+        let out_dims = params.out_dims();
+        Ok(crate::tensor::from_storage(storage, out_dims, op, false))
+    }
+
+    /// Applies a 1D convolution over the input tensor.
+    pub fn conv1d(
+        &self,
+        kernel: &Self,
+        padding: usize,
+        stride: usize,
+        dilation: usize,
+        groups: usize,
+    ) -> Result<Self> {
+        let (c_out, c_in_k, k_size) = kernel.dims3()?;
+        let (b_size, c_in, l_in) = self.dims3()?;
+        if c_in != c_in_k * groups {
+            Err(Error::Conv1dInvalidArgs {
+                inp_shape: self.shape().clone(),
+                k_shape: kernel.shape().clone(),
+                padding,
+                stride,
+                msg: "the number of in-channels on the input doesn't match the kernel size",
+            }
+            .bt())?
+        }
+
+        let params = ParamsConv1D {
+            b_size,
+            l_in,
+            c_out: c_out / groups,
+            c_in: c_in / groups,
+            k_size,
+            padding,
+            stride,
+            dilation,
+        };
+        if groups == 1 {
+            self.conv1d_single_group(kernel, &params)
+        } else {
+            let blocks = self.chunk(groups, 1)?;
+            let kernel = kernel.chunk(groups, 0)?;
+            let blocks = blocks
+                .iter()
+                .zip(&kernel)
+                .map(|(block, kernel)| block.conv1d_single_group(kernel, &params))
+                .collect::<Result<Vec<_>>>()?;
+            Tensor::cat(&blocks, 1)
+        }
+    }
+
+    fn conv_transpose1d_single_group(
+        &self,
+        kernel: &Self,
+        params: &ParamsConvTranspose1D,
+    ) -> Result<Self> {
+        let storage = self.storage().conv_transpose1d(
+            self.layout(),
+            &kernel.storage(),
+            kernel.layout(),
+            params,
+        )?;
+        let op = BackpropOp::new2(self, kernel, |arg, kernel| Op::ConvTranspose1D {
+            arg,
+            kernel,
+            padding: params.padding,
+            output_padding: params.output_padding,
+            stride: params.stride,
+            dilation: params.dilation,
+        });
+        let out_dims = params.out_dims();
+        Ok(crate::tensor::from_storage(storage, out_dims, op, false))
+    }
+
+    /// Applies a 1D transposed convolution over the input tensor.
+    pub fn conv_transpose1d(
+        &self,
+        kernel: &Self,
+        padding: usize,
+        output_padding: usize,
+        stride: usize,
+        dilation: usize,
+        groups: usize,
+    ) -> Result<Self> {
+        let (c_in_k, c_out, k_size) = kernel.dims3()?;
+        let (b_size, c_in, l_in) = self.dims3()?;
+        if c_in != c_in_k {
+            crate::bail!("in_channel mismatch between input ({c_in}) and kernel ({c_in_k})")
+        }
+        if c_in % groups != 0 {
+            crate::bail!("in_channel {c_in} is not divisible by the number of groups")
+        }
+        let params = ParamsConvTranspose1D {
+            b_size,
+            l_in,
+            k_size,
+            c_out,
+            c_in: c_in / groups,
+            padding,
+            output_padding,
+            stride,
+            dilation,
+        };
+        if groups == 1 {
+            self.conv_transpose1d_single_group(kernel, &params)
+        } else {
+            let blocks = self.chunk(groups, 1)?;
+            let kernel = kernel.chunk(groups, 0)?;
+            let blocks = blocks
+                .iter()
+                .zip(&kernel)
+                .map(|(block, kernel)| block.conv_transpose1d_single_group(kernel, &params))
+                .collect::<Result<Vec<_>>>()?;
+            Tensor::cat(&blocks, 1)
+        }
+    }
+
+    fn conv2d_single_group(&self, kernel: &Self, params: &ParamsConv2D) -> Result<Self> {
+        let storage =
+            self.storage()
+                .conv2d(self.layout(), &kernel.storage(), kernel.layout(), params)?;
+        let op = BackpropOp::new2(self, kernel, |arg, kernel| Op::Conv2D {
+            arg,
+            kernel,
+            padding: params.padding,
+            stride: params.stride,
+            dilation: params.dilation,
+        });
+        let out_dims = params.out_dims();
+        Ok(crate::tensor::from_storage(storage, out_dims, op, false))
+    }
+
+    /// Applies a 2D convolution over the input tensor.
+    pub fn conv2d(
+        &self,
+        kernel: &Self,
+        padding: usize,
+        stride: usize,
+        dilation: usize,
+        groups: usize,
+    ) -> Result<Self> {
+        let (b_size, c_in, i_h, i_w) = self.dims4()?;
+        let (c_out, c_in_k, k_h, k_w) = kernel.dims4()?;
+        if c_in != c_in_k * groups {
+            crate::bail!(
+                "in_channel mismatch between input ({c_in}, groups {groups}) and kernel ({c_in_k})"
+            )
+        }
+        let params = ParamsConv2D {
+            b_size,
+            i_h,
+            i_w,
+            k_h,
+            k_w,
+            c_out: c_out / groups,
+            c_in: c_in / groups,
+            padding,
+            stride,
+            dilation,
+            cudnn_fwd_algo: None,
+        };
+        if groups == 1 {
+            self.conv2d_single_group(kernel, &params)
+        } else {
+            let blocks = self.chunk(groups, 1)?;
+            let kernel = kernel.chunk(groups, 0)?;
+            let blocks = blocks
+                .iter()
+                .zip(&kernel)
+                .map(|(block, kernel)| block.conv2d_single_group(kernel, &params))
+                .collect::<Result<Vec<_>>>()?;
+            Tensor::cat(&blocks, 1)
+        }
+    }
+
+    /// Applies a 2D transposed convolution over the input tensor.
+    pub fn conv_transpose2d(
+        &self,
+        kernel: &Self,
+        padding: usize,
+        output_padding: usize,
+        stride: usize,
+        dilation: usize,
+    ) -> Result<Self> {
+        let (b_size, c_in, i_h, i_w) = self.dims4()?;
+        let (c_in_k, c_out, k_h, k_w) = kernel.dims4()?;
+        if c_in != c_in_k {
+            crate::bail!("in_channel mismatch between input ({c_in}) and kernel ({c_in_k})")
+        }
+        let params = ParamsConvTranspose2D {
+            b_size,
+            i_h,
+            i_w,
+            k_h,
+            k_w,
+            c_out,
+            c_in,
+            padding,
+            output_padding,
+            stride,
+            dilation,
+        };
+        let storage = self.storage().conv_transpose2d(
+            self.layout(),
+            &kernel.storage(),
+            kernel.layout(),
+            &params,
+        )?;
+        let op = BackpropOp::new2(self, kernel, |arg, kernel| Op::ConvTranspose2D {
+            arg,
+            kernel,
+            padding: params.padding,
+            output_padding: params.output_padding,
+            stride: params.stride,
+            dilation: params.dilation,
+        });
+        let out_dims = params.out_dims();
+        Ok(crate::tensor::from_storage(storage, out_dims, op, false))
+    }
+}

candle-core/src/convert.rs

@@ -92,6 +92,7 @@ from_tensor!(f64);
 from_tensor!(f32);
 from_tensor!(f16);
 from_tensor!(bf16);
+from_tensor!(i64);
 from_tensor!(u32);
 from_tensor!(u8);
 
@@ -129,6 +130,11 @@ impl Tensor {
                     f.write_u32::<LittleEndian>(v)?
                 }
             }
+            DType::I64 => {
+                for v in vs.to_vec1::<i64>()? {
+                    f.write_i64::<LittleEndian>(v)?
+                }
+            }
             DType::U8 => {
                 let vs = vs.to_vec1::<u8>()?;
                 f.write_all(&vs)?;

candle-core/src/cpu/avx.rs

@@ -103,7 +103,7 @@ impl CpuF16<ARR> for CurrentCpuF16 {
         for i in 0..8 {
             tmp[i] = (*mem_addr.add(i)).to_f32();
         }
-        _mm_loadu_ps(tmp.as_ptr())
+        _mm256_loadu_ps(tmp.as_ptr())
     }
 
     unsafe fn vec_add(a: Self::Unit, b: Self::Unit) -> Self::Unit {

candle-core/src/cpu/erf.rs

@@ -0,0 +1,763 @@
+#![allow(clippy::excessive_precision)]
+// Code taken from https://github.com/statrs-dev/statrs
+//! Provides the [error](https://en.wikipedia.org/wiki/Error_function) and
+//! related functions
+
+mod evaluate {
+    //! Provides functions that don't have a numerical solution and must
+    //! be solved computationally (e.g. evaluation of a polynomial)
+
+    /// evaluates a polynomial at `z` where `coeff` are the coeffecients
+    /// to a polynomial of order `k` where `k` is the length of `coeff` and the
+    /// coeffecient
+    /// to the `k`th power is the `k`th element in coeff. E.g. [3,-1,2] equates to
+    /// `2z^2 - z + 3`
+    ///
+    /// # Remarks
+    ///
+    /// Returns 0 for a 0 length coefficient slice
+    pub fn polynomial(z: f64, coeff: &[f64]) -> f64 {
+        let n = coeff.len();
+        if n == 0 {
+            return 0.0;
+        }
+
+        let mut sum = *coeff.last().unwrap();
+        for c in coeff[0..n - 1].iter().rev() {
+            sum = *c + z * sum;
+        }
+        sum
+    }
+}
+use std::f64;
+
+/// `erf` calculates the error function at `x`.
+pub fn erf(x: f64) -> f64 {
+    if x.is_nan() {
+        f64::NAN
+    } else if x >= 0.0 && x.is_infinite() {
+        1.0
+    } else if x <= 0.0 && x.is_infinite() {
+        -1.0
+    } else if x == 0. {
+        0.0
+    } else {
+        erf_impl(x, false)
+    }
+}
+
+/// `erf_inv` calculates the inverse error function
+/// at `x`.
+pub fn erf_inv(x: f64) -> f64 {
+    if x == 0.0 {
+        0.0
+    } else if x >= 1.0 {
+        f64::INFINITY
+    } else if x <= -1.0 {
+        f64::NEG_INFINITY
+    } else if x < 0.0 {
+        erf_inv_impl(-x, 1.0 + x, -1.0)
+    } else {
+        erf_inv_impl(x, 1.0 - x, 1.0)
+    }
+}
+
+/// `erfc` calculates the complementary error function
+/// at `x`.
+pub fn erfc(x: f64) -> f64 {
+    if x.is_nan() {
+        f64::NAN
+    } else if x == f64::INFINITY {
+        0.0
+    } else if x == f64::NEG_INFINITY {
+        2.0
+    } else {
+        erf_impl(x, true)
+    }
+}
+
+/// `erfc_inv` calculates the complementary inverse
+/// error function at `x`.
+pub fn erfc_inv(x: f64) -> f64 {
+    if x <= 0.0 {
+        f64::INFINITY
+    } else if x >= 2.0 {
+        f64::NEG_INFINITY
+    } else if x > 1.0 {
+        erf_inv_impl(-1.0 + x, 2.0 - x, -1.0)
+    } else {
+        erf_inv_impl(1.0 - x, x, 1.0)
+    }
+}
+
+// **********************************************************
+// ********** Coefficients for erf_impl polynomial **********
+// **********************************************************
+
+/// Polynomial coefficients for a numerator of `erf_impl`
+/// in the interval [1e-10, 0.5].
+const ERF_IMPL_AN: &[f64] = &[
+    0.00337916709551257388990745,
+    -0.00073695653048167948530905,
+    -0.374732337392919607868241,
+    0.0817442448733587196071743,
+    -0.0421089319936548595203468,
+    0.0070165709512095756344528,
+    -0.00495091255982435110337458,
+    0.000871646599037922480317225,
+];
+
+/// Polynomial coefficients for a denominator of `erf_impl`
+/// in the interval [1e-10, 0.5]
+const ERF_IMPL_AD: &[f64] = &[
+    1.0,
+    -0.218088218087924645390535,
+    0.412542972725442099083918,
+    -0.0841891147873106755410271,
+    0.0655338856400241519690695,
+    -0.0120019604454941768171266,
+    0.00408165558926174048329689,
+    -0.000615900721557769691924509,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [0.5, 0.75].
+const ERF_IMPL_BN: &[f64] = &[
+    -0.0361790390718262471360258,
+    0.292251883444882683221149,
+    0.281447041797604512774415,
+    0.125610208862766947294894,
+    0.0274135028268930549240776,
+    0.00250839672168065762786937,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [0.5, 0.75].
+const ERF_IMPL_BD: &[f64] = &[
+    1.0,
+    1.8545005897903486499845,
+    1.43575803037831418074962,
+    0.582827658753036572454135,
+    0.124810476932949746447682,
+    0.0113724176546353285778481,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [0.75, 1.25].
+const ERF_IMPL_CN: &[f64] = &[
+    -0.0397876892611136856954425,
+    0.153165212467878293257683,
+    0.191260295600936245503129,
+    0.10276327061989304213645,
+    0.029637090615738836726027,
+    0.0046093486780275489468812,
+    0.000307607820348680180548455,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [0.75, 1.25].
+const ERF_IMPL_CD: &[f64] = &[
+    1.0,
+    1.95520072987627704987886,
+    1.64762317199384860109595,
+    0.768238607022126250082483,
+    0.209793185936509782784315,
+    0.0319569316899913392596356,
+    0.00213363160895785378615014,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [1.25, 2.25].
+const ERF_IMPL_DN: &[f64] = &[
+    -0.0300838560557949717328341,
+    0.0538578829844454508530552,
+    0.0726211541651914182692959,
+    0.0367628469888049348429018,
+    0.00964629015572527529605267,
+    0.00133453480075291076745275,
+    0.778087599782504251917881e-4,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [1.25, 2.25].
+const ERF_IMPL_DD: &[f64] = &[
+    1.0,
+    1.75967098147167528287343,
+    1.32883571437961120556307,
+    0.552528596508757581287907,
+    0.133793056941332861912279,
+    0.0179509645176280768640766,
+    0.00104712440019937356634038,
+    -0.106640381820357337177643e-7,
+];
+
+///  Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [2.25, 3.5].
+const ERF_IMPL_EN: &[f64] = &[
+    -0.0117907570137227847827732,
+    0.014262132090538809896674,
+    0.0202234435902960820020765,
+    0.00930668299990432009042239,
+    0.00213357802422065994322516,
+    0.00025022987386460102395382,
+    0.120534912219588189822126e-4,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [2.25, 3.5].
+const ERF_IMPL_ED: &[f64] = &[
+    1.0,
+    1.50376225203620482047419,
+    0.965397786204462896346934,
+    0.339265230476796681555511,
+    0.0689740649541569716897427,
+    0.00771060262491768307365526,
+    0.000371421101531069302990367,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [3.5, 5.25].
+const ERF_IMPL_FN: &[f64] = &[
+    -0.00546954795538729307482955,
+    0.00404190278731707110245394,
+    0.0054963369553161170521356,
+    0.00212616472603945399437862,
+    0.000394984014495083900689956,
+    0.365565477064442377259271e-4,
+    0.135485897109932323253786e-5,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [3.5, 5.25].
+const ERF_IMPL_FD: &[f64] = &[
+    1.0,
+    1.21019697773630784832251,
+    0.620914668221143886601045,
+    0.173038430661142762569515,
+    0.0276550813773432047594539,
+    0.00240625974424309709745382,
+    0.891811817251336577241006e-4,
+    -0.465528836283382684461025e-11,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [5.25, 8].
+const ERF_IMPL_GN: &[f64] = &[
+    -0.00270722535905778347999196,
+    0.0013187563425029400461378,
+    0.00119925933261002333923989,
+    0.00027849619811344664248235,
+    0.267822988218331849989363e-4,
+    0.923043672315028197865066e-6,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [5.25, 8].
+const ERF_IMPL_GD: &[f64] = &[
+    1.0,
+    0.814632808543141591118279,
+    0.268901665856299542168425,
+    0.0449877216103041118694989,
+    0.00381759663320248459168994,
+    0.000131571897888596914350697,
+    0.404815359675764138445257e-11,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [8, 11.5].
+const ERF_IMPL_HN: &[f64] = &[
+    -0.00109946720691742196814323,
+    0.000406425442750422675169153,
+    0.000274499489416900707787024,
+    0.465293770646659383436343e-4,
+    0.320955425395767463401993e-5,
+    0.778286018145020892261936e-7,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [8, 11.5].
+const ERF_IMPL_HD: &[f64] = &[
+    1.0,
+    0.588173710611846046373373,
+    0.139363331289409746077541,
+    0.0166329340417083678763028,
+    0.00100023921310234908642639,
+    0.24254837521587225125068e-4,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [11.5, 17].
+const ERF_IMPL_IN: &[f64] = &[
+    -0.00056907993601094962855594,
+    0.000169498540373762264416984,
+    0.518472354581100890120501e-4,
+    0.382819312231928859704678e-5,
+    0.824989931281894431781794e-7,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [11.5, 17].
+const ERF_IMPL_ID: &[f64] = &[
+    1.0,
+    0.339637250051139347430323,
+    0.043472647870310663055044,
+    0.00248549335224637114641629,
+    0.535633305337152900549536e-4,
+    -0.117490944405459578783846e-12,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [17, 24].
+const ERF_IMPL_JN: &[f64] = &[
+    -0.000241313599483991337479091,
+    0.574224975202501512365975e-4,
+    0.115998962927383778460557e-4,
+    0.581762134402593739370875e-6,
+    0.853971555085673614607418e-8,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [17, 24].
+const ERF_IMPL_JD: &[f64] = &[
+    1.0,
+    0.233044138299687841018015,
+    0.0204186940546440312625597,
+    0.000797185647564398289151125,
+    0.117019281670172327758019e-4,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [24, 38].
+const ERF_IMPL_KN: &[f64] = &[
+    -0.000146674699277760365803642,
+    0.162666552112280519955647e-4,
+    0.269116248509165239294897e-5,
+    0.979584479468091935086972e-7,
+    0.101994647625723465722285e-8,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [24, 38].
+const ERF_IMPL_KD: &[f64] = &[
+    1.0,
+    0.165907812944847226546036,
+    0.0103361716191505884359634,
+    0.000286593026373868366935721,
+    0.298401570840900340874568e-5,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [38, 60].
+const ERF_IMPL_LN: &[f64] = &[
+    -0.583905797629771786720406e-4,
+    0.412510325105496173512992e-5,
+    0.431790922420250949096906e-6,
+    0.993365155590013193345569e-8,
+    0.653480510020104699270084e-10,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [38, 60].
+const ERF_IMPL_LD: &[f64] = &[
+    1.0,
+    0.105077086072039915406159,
+    0.00414278428675475620830226,
+    0.726338754644523769144108e-4,
+    0.477818471047398785369849e-6,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [60, 85].
+const ERF_IMPL_MN: &[f64] = &[
+    -0.196457797609229579459841e-4,
+    0.157243887666800692441195e-5,
+    0.543902511192700878690335e-7,
+    0.317472492369117710852685e-9,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [60, 85].
+const ERF_IMPL_MD: &[f64] = &[
+    1.0,
+    0.052803989240957632204885,
+    0.000926876069151753290378112,
+    0.541011723226630257077328e-5,
+    0.535093845803642394908747e-15,
+];
+
+/// Polynomial coefficients for a numerator in `erf_impl`
+/// in the interval [85, 110].
+const ERF_IMPL_NN: &[f64] = &[
+    -0.789224703978722689089794e-5,
+    0.622088451660986955124162e-6,
+    0.145728445676882396797184e-7,
+    0.603715505542715364529243e-10,
+];
+
+/// Polynomial coefficients for a denominator in `erf_impl`
+/// in the interval [85, 110].
+const ERF_IMPL_ND: &[f64] = &[
+    1.0,
+    0.0375328846356293715248719,
+    0.000467919535974625308126054,
+    0.193847039275845656900547e-5,
+];
+
+// **********************************************************
+// ********** Coefficients for erf_inv_impl polynomial ******
+// **********************************************************
+
+/// Polynomial coefficients for a numerator of `erf_inv_impl`
+/// in the interval [0, 0.5].
+const ERF_INV_IMPL_AN: &[f64] = &[
+    -0.000508781949658280665617,
+    -0.00836874819741736770379,
+    0.0334806625409744615033,
+    -0.0126926147662974029034,
+    -0.0365637971411762664006,
+    0.0219878681111168899165,
+    0.00822687874676915743155,
+    -0.00538772965071242932965,
+];
+
+/// Polynomial coefficients for a denominator of `erf_inv_impl`
+/// in the interval [0, 0.5].
+const ERF_INV_IMPL_AD: &[f64] = &[
+    1.0,
+    -0.970005043303290640362,
+    -1.56574558234175846809,
+    1.56221558398423026363,
+    0.662328840472002992063,
+    -0.71228902341542847553,
+    -0.0527396382340099713954,
+    0.0795283687341571680018,
+    -0.00233393759374190016776,
+    0.000886216390456424707504,
+];
+
+/// Polynomial coefficients for a numerator of `erf_inv_impl`
+/// in the interval [0.5, 0.75].
+const ERF_INV_IMPL_BN: &[f64] = &[
+    -0.202433508355938759655,
+    0.105264680699391713268,
+    8.37050328343119927838,
+    17.6447298408374015486,
+    -18.8510648058714251895,
+    -44.6382324441786960818,
+    17.445385985570866523,
+    21.1294655448340526258,
+    -3.67192254707729348546,
+];
+
+/// Polynomial coefficients for a denominator of `erf_inv_impl`
+/// in the interval [0.5, 0.75].
+const ERF_INV_IMPL_BD: &[f64] = &[
+    1.0,
+    6.24264124854247537712,
+    3.9713437953343869095,
+    -28.6608180499800029974,
+    -20.1432634680485188801,
+    48.5609213108739935468,
+    10.8268667355460159008,
+    -22.6436933413139721736,
+    1.72114765761200282724,
+];
+
+/// Polynomial coefficients for a numerator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x less than 3.
+const ERF_INV_IMPL_CN: &[f64] = &[
+    -0.131102781679951906451,
+    -0.163794047193317060787,
+    0.117030156341995252019,
+    0.387079738972604337464,
+    0.337785538912035898924,
+    0.142869534408157156766,
+    0.0290157910005329060432,
+    0.00214558995388805277169,
+    -0.679465575181126350155e-6,
+    0.285225331782217055858e-7,
+    -0.681149956853776992068e-9,
+];
+
+/// Polynomial coefficients for a denominator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x less than 3.
+const ERF_INV_IMPL_CD: &[f64] = &[
+    1.0,
+    3.46625407242567245975,
+    5.38168345707006855425,
+    4.77846592945843778382,
+    2.59301921623620271374,
+    0.848854343457902036425,
+    0.152264338295331783612,
+    0.01105924229346489121,
+];
+
+/// Polynomial coefficients for a numerator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x between 3 and 6.
+const ERF_INV_IMPL_DN: &[f64] = &[
+    -0.0350353787183177984712,
+    -0.00222426529213447927281,
+    0.0185573306514231072324,
+    0.00950804701325919603619,
+    0.00187123492819559223345,
+    0.000157544617424960554631,
+    0.460469890584317994083e-5,
+    -0.230404776911882601748e-9,
+    0.266339227425782031962e-11,
+];
+
+/// Polynomial coefficients for a denominator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x between 3 and 6.
+const ERF_INV_IMPL_DD: &[f64] = &[
+    1.0,
+    1.3653349817554063097,
+    0.762059164553623404043,
+    0.220091105764131249824,
+    0.0341589143670947727934,
+    0.00263861676657015992959,
+    0.764675292302794483503e-4,
+];
+
+/// Polynomial coefficients for a numerator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x between 6 and 18.
+const ERF_INV_IMPL_EN: &[f64] = &[
+    -0.0167431005076633737133,
+    -0.00112951438745580278863,
+    0.00105628862152492910091,
+    0.000209386317487588078668,
+    0.149624783758342370182e-4,
+    0.449696789927706453732e-6,
+    0.462596163522878599135e-8,
+    -0.281128735628831791805e-13,
+    0.99055709973310326855e-16,
+];
+
+/// Polynomial coefficients for a denominator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x between 6 and 18.
+const ERF_INV_IMPL_ED: &[f64] = &[
+    1.0,
+    0.591429344886417493481,
+    0.138151865749083321638,
+    0.0160746087093676504695,
+    0.000964011807005165528527,
+    0.275335474764726041141e-4,
+    0.282243172016108031869e-6,
+];
+
+/// Polynomial coefficients for a numerator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x between 18 and 44.
+const ERF_INV_IMPL_FN: &[f64] = &[
+    -0.0024978212791898131227,
+    -0.779190719229053954292e-5,
+    0.254723037413027451751e-4,
+    0.162397777342510920873e-5,
+    0.396341011304801168516e-7,
+    0.411632831190944208473e-9,
+    0.145596286718675035587e-11,
+    -0.116765012397184275695e-17,
+];
+
+/// Polynomial coefficients for a denominator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x between 18 and 44.
+const ERF_INV_IMPL_FD: &[f64] = &[
+    1.0,
+    0.207123112214422517181,
+    0.0169410838120975906478,
+    0.000690538265622684595676,
+    0.145007359818232637924e-4,
+    0.144437756628144157666e-6,
+    0.509761276599778486139e-9,
+];
+
+/// Polynomial coefficients for a numerator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x greater than 44.
+const ERF_INV_IMPL_GN: &[f64] = &[
+    -0.000539042911019078575891,
+    -0.28398759004727721098e-6,
+    0.899465114892291446442e-6,
+    0.229345859265920864296e-7,
+    0.225561444863500149219e-9,
+    0.947846627503022684216e-12,
+    0.135880130108924861008e-14,
+    -0.348890393399948882918e-21,
+];
+
+/// Polynomial coefficients for a denominator of `erf_inv_impl`
+/// in the interval [0.75, 1] with x greater than 44.
+const ERF_INV_IMPL_GD: &[f64] = &[
+    1.0,
+    0.0845746234001899436914,
+    0.00282092984726264681981,
+    0.468292921940894236786e-4,
+    0.399968812193862100054e-6,
+    0.161809290887904476097e-8,
+    0.231558608310259605225e-11,
+];
+
+/// `erf_impl` computes the error function at `z`.
+/// If `inv` is true, `1 - erf` is calculated as opposed to `erf`
+fn erf_impl(z: f64, inv: bool) -> f64 {
+    if z < 0.0 {
+        if !inv {
+            return -erf_impl(-z, false);
+        }
+        if z < -0.5 {
+            return 2.0 - erf_impl(-z, true);
+        }
+        return 1.0 + erf_impl(-z, false);
+    }
+
+    let result = if z < 0.5 {
+        if z < 1e-10 {
+            z * 1.125 + z * 0.003379167095512573896158903121545171688
+        } else {
+            z * 1.125
+                + z * evaluate::polynomial(z, ERF_IMPL_AN) / evaluate::polynomial(z, ERF_IMPL_AD)
+        }
+    } else if z < 110.0 {
+        let (r, b) = if z < 0.75 {
+            (
+                evaluate::polynomial(z - 0.5, ERF_IMPL_BN)
+                    / evaluate::polynomial(z - 0.5, ERF_IMPL_BD),
+                0.3440242112,
+            )
+        } else if z < 1.25 {
+            (
+                evaluate::polynomial(z - 0.75, ERF_IMPL_CN)
+                    / evaluate::polynomial(z - 0.75, ERF_IMPL_CD),
+                0.419990927,
+            )
+        } else if z < 2.25 {
+            (
+                evaluate::polynomial(z - 1.25, ERF_IMPL_DN)
+                    / evaluate::polynomial(z - 1.25, ERF_IMPL_DD),
+                0.4898625016,
+            )
+        } else if z < 3.5 {
+            (
+                evaluate::polynomial(z - 2.25, ERF_IMPL_EN)
+                    / evaluate::polynomial(z - 2.25, ERF_IMPL_ED),
+                0.5317370892,
+            )
+        } else if z < 5.25 {
+            (
+                evaluate::polynomial(z - 3.5, ERF_IMPL_FN)
+                    / evaluate::polynomial(z - 3.5, ERF_IMPL_FD),
+                0.5489973426,
+            )
+        } else if z < 8.0 {
+            (
+                evaluate::polynomial(z - 5.25, ERF_IMPL_GN)
+                    / evaluate::polynomial(z - 5.25, ERF_IMPL_GD),
+                0.5571740866,
+            )
+        } else if z < 11.5 {
+            (
+                evaluate::polynomial(z - 8.0, ERF_IMPL_HN)
+                    / evaluate::polynomial(z - 8.0, ERF_IMPL_HD),
+                0.5609807968,
+            )
+        } else if z < 17.0 {
+            (
+                evaluate::polynomial(z - 11.5, ERF_IMPL_IN)
+                    / evaluate::polynomial(z - 11.5, ERF_IMPL_ID),
+                0.5626493692,
+            )
+        } else if z < 24.0 {
+            (
+                evaluate::polynomial(z - 17.0, ERF_IMPL_JN)
+                    / evaluate::polynomial(z - 17.0, ERF_IMPL_JD),
+                0.5634598136,
+            )
+        } else if z < 38.0 {
+            (
+                evaluate::polynomial(z - 24.0, ERF_IMPL_KN)
+                    / evaluate::polynomial(z - 24.0, ERF_IMPL_KD),
+                0.5638477802,
+            )
+        } else if z < 60.0 {
+            (
+                evaluate::polynomial(z - 38.0, ERF_IMPL_LN)
+                    / evaluate::polynomial(z - 38.0, ERF_IMPL_LD),
+                0.5640528202,
+            )
+        } else if z < 85.0 {
+            (
+                evaluate::polynomial(z - 60.0, ERF_IMPL_MN)
+                    / evaluate::polynomial(z - 60.0, ERF_IMPL_MD),
+                0.5641309023,
+            )
+        } else {
+            (
+                evaluate::polynomial(z - 85.0, ERF_IMPL_NN)
+                    / evaluate::polynomial(z - 85.0, ERF_IMPL_ND),
+                0.5641584396,
+            )
+        };
+        let g = (-z * z).exp() / z;
+        g * b + g * r
+    } else {
+        0.0
+    };
+
+    if inv && z >= 0.5 {
+        result
+    } else if z >= 0.5 || inv {
+        1.0 - result
+    } else {
+        result
+    }
+}
+
+// `erf_inv_impl` computes the inverse error function where
+// `p`,`q`, and `s` are the first, second, and third intermediate
+// parameters respectively
+fn erf_inv_impl(p: f64, q: f64, s: f64) -> f64 {
+    let result = if p <= 0.5 {
+        let y = 0.0891314744949340820313;
+        let g = p * (p + 10.0);
+        let r = evaluate::polynomial(p, ERF_INV_IMPL_AN) / evaluate::polynomial(p, ERF_INV_IMPL_AD);
+        g * y + g * r
+    } else if q >= 0.25 {
+        let y = 2.249481201171875;
+        let g = (-2.0 * q.ln()).sqrt();
+        let xs = q - 0.25;
+        let r =
+            evaluate::polynomial(xs, ERF_INV_IMPL_BN) / evaluate::polynomial(xs, ERF_INV_IMPL_BD);
+        g / (y + r)
+    } else {
+        let x = (-q.ln()).sqrt();
+        if x < 3.0 {
+            let y = 0.807220458984375;
+            let xs = x - 1.125;
+            let r = evaluate::polynomial(xs, ERF_INV_IMPL_CN)
+                / evaluate::polynomial(xs, ERF_INV_IMPL_CD);
+            y * x + r * x
+        } else if x < 6.0 {
+            let y = 0.93995571136474609375;
+            let xs = x - 3.0;
+            let r = evaluate::polynomial(xs, ERF_INV_IMPL_DN)
+                / evaluate::polynomial(xs, ERF_INV_IMPL_DD);
+            y * x + r * x
+        } else if x < 18.0 {
+            let y = 0.98362827301025390625;
+            let xs = x - 6.0;
+            let r = evaluate::polynomial(xs, ERF_INV_IMPL_EN)
+                / evaluate::polynomial(xs, ERF_INV_IMPL_ED);
+            y * x + r * x
+        } else if x < 44.0 {
+            let y = 0.99714565277099609375;
+            let xs = x - 18.0;
+            let r = evaluate::polynomial(xs, ERF_INV_IMPL_FN)
+                / evaluate::polynomial(xs, ERF_INV_IMPL_FD);
+            y * x + r * x
+        } else {
+            let y = 0.99941349029541015625;
+            let xs = x - 44.0;
+            let r = evaluate::polynomial(xs, ERF_INV_IMPL_GN)
+                / evaluate::polynomial(xs, ERF_INV_IMPL_GD);
+            y * x + r * x
+        }
+    };
+    s * result
+}

candle-core/src/cpu/kernels.rs

@@ -1,4 +1,7 @@
 pub trait VecOps: num_traits::NumAssign + Copy {
+    fn min(self, rhs: Self) -> Self;
+    fn max(self, rhs: Self) -> Self;
+
     /// Dot-product of two vectors.
     ///
     /// # Safety
@@ -26,9 +29,47 @@ pub trait VecOps: num_traits::NumAssign + Copy {
             *res += *xs.add(i)
         }
     }
+
+    /// Maximum element in a non-empty vector.
+    ///
+    /// # Safety
+    ///
+    /// The length of `xs` must be at least `len` and positive. `res` has to point to a valid
+    /// element.
+    #[inline(always)]
+    unsafe fn vec_reduce_max(xs: *const Self, res: *mut Self, len: usize) {
+        *res = *xs;
+        for i in 1..len {
+            *res = (*res).max(*xs.add(i))
+        }
+    }
+
+    /// Minimum element in a non-empty vector.
+    ///
+    /// # Safety
+    ///
+    /// The length of `xs` must be at least `len` and positive. `res` has to point to a valid
+    /// element.
+    #[inline(always)]
+    unsafe fn vec_reduce_min(xs: *const Self, res: *mut Self, len: usize) {
+        *res = *xs;
+        for i in 1..len {
+            *res = (*res).min(*xs.add(i))
+        }
+    }
 }
 
 impl VecOps for f32 {
+    #[inline(always)]
+    fn min(self, other: Self) -> Self {
+        Self::min(self, other)
+    }
+
+    #[inline(always)]
+    fn max(self, other: Self) -> Self {
+        Self::max(self, other)
+    }
+
     #[inline(always)]
     unsafe fn vec_dot(lhs: *const Self, rhs: *const Self, res: *mut Self, len: usize) {
         super::vec_dot_f32(lhs, rhs, res, len)
@@ -41,6 +82,16 @@ impl VecOps for f32 {
 }
 
 impl VecOps for half::f16 {
+    #[inline(always)]
+    fn min(self, other: Self) -> Self {
+        Self::min(self, other)
+    }
+
+    #[inline(always)]
+    fn max(self, other: Self) -> Self {
+        Self::max(self, other)
+    }
+
     #[inline(always)]
     unsafe fn vec_dot(lhs: *const Self, rhs: *const Self, res: *mut Self, len: usize) {
         let mut res_f32 = 0f32;
@@ -49,10 +100,61 @@ impl VecOps for half::f16 {
     }
 }
 
-impl VecOps for f64 {}
-impl VecOps for half::bf16 {}
-impl VecOps for u8 {}
-impl VecOps for u32 {}
+impl VecOps for f64 {
+    #[inline(always)]
+    fn min(self, other: Self) -> Self {
+        Self::min(self, other)
+    }
+
+    #[inline(always)]
+    fn max(self, other: Self) -> Self {
+        Self::max(self, other)
+    }
+}
+impl VecOps for half::bf16 {
+    #[inline(always)]
+    fn min(self, other: Self) -> Self {
+        Self::min(self, other)
+    }
+
+    #[inline(always)]
+    fn max(self, other: Self) -> Self {
+        Self::max(self, other)
+    }
+}
+impl VecOps for u8 {
+    #[inline(always)]
+    fn min(self, other: Self) -> Self {
+        <Self as Ord>::min(self, other)
+    }
+
+    #[inline(always)]
+    fn max(self, other: Self) -> Self {
+        <Self as Ord>::max(self, other)
+    }
+}
+impl VecOps for u32 {
+    #[inline(always)]
+    fn min(self, other: Self) -> Self {
+        <Self as Ord>::min(self, other)
+    }
+
+    #[inline(always)]
+    fn max(self, other: Self) -> Self {
+        <Self as Ord>::max(self, other)
+    }
+}
+impl VecOps for i64 {
+    #[inline(always)]
+    fn min(self, other: Self) -> Self {
+        <Self as Ord>::min(self, other)
+    }
+
+    #[inline(always)]
+    fn max(self, other: Self) -> Self {
+        <Self as Ord>::max(self, other)
+    }
+}
 
 #[inline(always)]
 pub fn par_for_each(n_threads: usize, func: impl Fn(usize) + Send + Sync) {

candle-core/src/cpu/mod.rs

@@ -1,5 +1,7 @@
+pub mod erf;
 pub mod kernels;
 
+#[allow(unused)]
 trait Cpu<const ARR: usize> {
     type Unit;
     type Array;
@@ -17,6 +19,7 @@ trait Cpu<const ARR: usize> {
     unsafe fn vec_store(mem_addr: *mut f32, a: Self::Unit);
 }
 
+#[allow(unused)]
 trait CpuF16<const ARR: usize> {
     type Unit;
     type Array;

candle-core/src/cpu_backend/utils.rs

@@ -0,0 +1,360 @@
+/// Helper functions to write CPU kernels.
+use crate::backend::BackendStorage;
+use crate::{Error, Layout, Result, WithDType};
+
+type C = super::CpuStorage;
+pub trait Map1 {
+    fn f<T: WithDType>(&self, vs: &[T], layout: &Layout) -> Result<Vec<T>>;
+
+    fn map(&self, vs: &C, layout: &Layout) -> Result<C> {
+        match vs {
+            C::U8(vs) => Ok(C::U8(self.f(vs, layout)?)),
+            C::U32(vs) => Ok(C::U32(self.f(vs, layout)?)),
+            C::I64(vs) => Ok(C::I64(self.f(vs, layout)?)),
+            C::BF16(vs) => Ok(C::BF16(self.f(vs, layout)?)),
+            C::F16(vs) => Ok(C::F16(self.f(vs, layout)?)),
+            C::F32(vs) => Ok(C::F32(self.f(vs, layout)?)),
+            C::F64(vs) => Ok(C::F64(self.f(vs, layout)?)),
+        }
+    }
+}
+
+pub trait Map1Any {
+    fn f<T: WithDType, W: Fn(Vec<T>) -> C>(&self, vs: &[T], layout: &Layout, wrap: W) -> Result<C>;
+
+    fn map(&self, vs: &C, layout: &Layout) -> Result<C> {
+        match vs {
+            C::U8(vs) => Ok(self.f(vs, layout, C::U8)?),
+            C::U32(vs) => Ok(self.f(vs, layout, C::U32)?),
+            C::I64(vs) => Ok(self.f(vs, layout, C::I64)?),
+            C::BF16(vs) => Ok(self.f(vs, layout, C::BF16)?),
+            C::F16(vs) => Ok(self.f(vs, layout, C::F16)?),
+            C::F32(vs) => Ok(self.f(vs, layout, C::F32)?),
+            C::F64(vs) => Ok(self.f(vs, layout, C::F64)?),
+        }
+    }
+}
+
+pub trait Map2 {
+    const OP: &'static str;
+    fn f<T: WithDType>(&self, v1: &[T], l1: &Layout, v2: &[T], l2: &Layout) -> Result<Vec<T>>;
+
+    fn map(&self, v1: &C, l1: &Layout, v2: &C, l2: &Layout) -> Result<C> {
+        match (v1, v2) {
+            (C::U8(v1), C::U8(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::U32(v1), C::U32(v2)) => Ok(C::U32(self.f(v1, l1, v2, l2)?)),
+            (C::I64(v1), C::I64(v2)) => Ok(C::I64(self.f(v1, l1, v2, l2)?)),
+            (C::BF16(v1), C::BF16(v2)) => Ok(C::BF16(self.f(v1, l1, v2, l2)?)),
+            (C::F16(v1), C::F16(v2)) => Ok(C::F16(self.f(v1, l1, v2, l2)?)),
+            (C::F32(v1), C::F32(v2)) => Ok(C::F32(self.f(v1, l1, v2, l2)?)),
+            (C::F64(v1), C::F64(v2)) => Ok(C::F64(self.f(v1, l1, v2, l2)?)),
+            _ => Err(Error::DTypeMismatchBinaryOp {
+                lhs: v1.dtype(),
+                rhs: v2.dtype(),
+                op: Self::OP,
+            }
+            .bt()),
+        }
+    }
+}
+
+pub trait Map2U8 {
+    const OP: &'static str;
+    fn f<T: WithDType>(&self, v1: &[T], l1: &Layout, v2: &[T], l2: &Layout) -> Result<Vec<u8>>;
+
+    fn map(&self, v1: &C, l1: &Layout, v2: &C, l2: &Layout) -> Result<C> {
+        match (v1, v2) {
+            (C::U8(v1), C::U8(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::U32(v1), C::U32(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::I64(v1), C::I64(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::BF16(v1), C::BF16(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::F16(v1), C::F16(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::F32(v1), C::F32(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::F64(v1), C::F64(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            _ => Err(Error::DTypeMismatchBinaryOp {
+                lhs: v1.dtype(),
+                rhs: v2.dtype(),
+                op: Self::OP,
+            }
+            .bt()),
+        }
+    }
+}
+
+pub fn binary_map<T: Copy, U: Copy, F: FnMut(T, T) -> U>(
+    lhs_l: &Layout,
+    rhs_l: &Layout,
+    lhs: &[T],
+    rhs: &[T],
+    mut f: F,
+) -> Vec<U> {
+    match (lhs_l.contiguous_offsets(), rhs_l.contiguous_offsets()) {
+        (Some((o_l1, o_l2)), Some((o_r1, o_r2))) => lhs[o_l1..o_l2]
+            .iter()
+            .zip(rhs[o_r1..o_r2].iter())
+            .map(|(&l, &r)| f(l, r))
+            .collect(),
+        (Some((o_l1, o_l2)), None) => {
+            // TODO: Maybe we want to avoid going through the layout twice.
+            match rhs_l.offsets_b() {
+                Some(ob) => {
+                    let mut i_in_block = 0;
+                    let mut i_right_broadcast = 0;
+                    lhs[o_l1..o_l2]
+                        .iter()
+                        .map(|&l| {
+                            let r = unsafe { rhs.get_unchecked(i_in_block + ob.start) };
+                            i_right_broadcast += 1;
+                            if i_right_broadcast >= ob.right_broadcast {
+                                i_in_block += 1;
+                                i_right_broadcast = 0;
+                            }
+                            if i_in_block >= ob.len {
+                                i_in_block = 0
+                            }
+                            f(l, *r)
+                        })
+                        .collect()
+                }
+                None => lhs_l
+                    .strided_index()
+                    .zip(rhs_l.strided_index())
+                    .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+                    .collect(),
+            }
+        }
+        (None, Some((o_r1, o_r2))) => {
+            // TODO: Maybe we want to avoid going through the layout twice.
+            match lhs_l.offsets_b() {
+                Some(ob) => {
+                    let mut i_in_block = 0;
+                    let mut i_right_broadcast = 0;
+                    rhs[o_r1..o_r2]
+                        .iter()
+                        .map(|&r| {
+                            let l = unsafe { lhs.get_unchecked(i_in_block + ob.start) };
+                            i_right_broadcast += 1;
+                            if i_right_broadcast >= ob.right_broadcast {
+                                i_in_block += 1;
+                                i_right_broadcast = 0;
+                            }
+                            if i_in_block >= ob.len {
+                                i_in_block = 0
+                            }
+                            f(*l, r)
+                        })
+                        .collect()
+                }
+                None => lhs_l
+                    .strided_index()
+                    .zip(rhs_l.strided_index())
+                    .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+                    .collect(),
+            }
+        }
+        _ => lhs_l
+            .strided_index()
+            .zip(rhs_l.strided_index())
+            .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+            .collect(),
+    }
+}
+
+// Similar to binary_map but with vectorized variants.
+pub fn binary_map_vec<T: Copy, F: FnMut(T, T) -> T, FV: FnMut(&[T], &[T], &mut [T])>(
+    lhs_l: &Layout,
+    rhs_l: &Layout,
+    lhs: &[T],
+    rhs: &[T],
+    mut f: F,
+    mut f_vec: FV,
+) -> Vec<T> {
+    let el_count = lhs_l.shape().elem_count();
+    match (lhs_l.contiguous_offsets(), rhs_l.contiguous_offsets()) {
+        (Some((o_l1, o_l2)), Some((o_r1, o_r2))) => {
+            let mut ys: Vec<T> = Vec::with_capacity(el_count);
+            let ys_to_set = ys.spare_capacity_mut();
+            let ys_to_set = unsafe {
+                std::mem::transmute::<&mut [std::mem::MaybeUninit<T>], &mut [T]>(ys_to_set)
+            };
+            f_vec(&lhs[o_l1..o_l2], &rhs[o_r1..o_r2], ys_to_set);
+            // SAFETY: values are all set by f_vec.
+            unsafe { ys.set_len(el_count) };
+            ys
+        }
+        (Some((o_l1, o_l2)), None) => match rhs_l.offsets_b() {
+            Some(ob) if ob.right_broadcast == 1 => {
+                let rhs = &rhs[ob.start..ob.start + ob.len];
+                let mut ys: Vec<T> = Vec::with_capacity(el_count);
+                let ys_to_set = ys.spare_capacity_mut();
+                let ys_to_set = unsafe {
+                    std::mem::transmute::<&mut [std::mem::MaybeUninit<T>], &mut [T]>(ys_to_set)
+                };
+                let mut dst_i = 0;
+                for src_i in (o_l1..o_l2).step_by(ob.len) {
+                    f_vec(
+                        &lhs[src_i..src_i + ob.len],
+                        rhs,
+                        &mut ys_to_set[dst_i..dst_i + ob.len],
+                    );
+                    dst_i += ob.len;
+                }
+                // SAFETY: values are all set by f_vec.
+                unsafe { ys.set_len(el_count) };
+                ys
+            }
+            Some(ob) => {
+                let rhs = &rhs[ob.start..ob.start + ob.len];
+                let mut ys = lhs[o_l1..o_l2].to_vec();
+                for idx_l in 0..ob.left_broadcast {
+                    let start = idx_l * ob.len * ob.right_broadcast;
+                    for (i, &r) in rhs.iter().enumerate() {
+                        let start = start + i * ob.right_broadcast;
+                        for v in ys[start..start + ob.right_broadcast].iter_mut() {
+                            *v = f(*v, r)
+                        }
+                    }
+                }
+                ys
+            }
+            None => lhs_l
+                .strided_index()
+                .zip(rhs_l.strided_index())
+                .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+                .collect(),
+        },
+        (None, Some((o_r1, o_r2))) => match lhs_l.offsets_b() {
+            Some(ob) if ob.right_broadcast == 1 => {
+                let lhs = &lhs[ob.start..ob.start + ob.len];
+                let mut ys: Vec<T> = Vec::with_capacity(el_count);
+                let ys_to_set = ys.spare_capacity_mut();
+                let ys_to_set = unsafe {
+                    std::mem::transmute::<&mut [std::mem::MaybeUninit<T>], &mut [T]>(ys_to_set)
+                };
+                let mut dst_i = 0;
+                for src_i in (o_r1..o_r2).step_by(ob.len) {
+                    f_vec(
+                        lhs,
+                        &rhs[src_i..src_i + ob.len],
+                        &mut ys_to_set[dst_i..dst_i + ob.len],
+                    );
+                    dst_i += ob.len;
+                }
+                // SAFETY: values are all set by f_vec.
+                unsafe { ys.set_len(el_count) };
+                ys
+            }
+            Some(ob) => {
+                let lhs = &lhs[ob.start..ob.start + ob.len];
+                let mut ys = rhs[o_r1..o_r2].to_vec();
+                for idx_l in 0..ob.left_broadcast {
+                    let start = idx_l * ob.len * ob.right_broadcast;
+                    for (i, &l) in lhs.iter().enumerate() {
+                        let start = start + i * ob.right_broadcast;
+                        for v in ys[start..start + ob.right_broadcast].iter_mut() {
+                            *v = f(l, *v)
+                        }
+                    }
+                }
+                ys
+            }
+            None => lhs_l
+                .strided_index()
+                .zip(rhs_l.strided_index())
+                .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+                .collect(),
+        },
+        _ => lhs_l
+            .strided_index()
+            .zip(rhs_l.strided_index())
+            .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+            .collect(),
+    }
+}
+
+pub fn unary_map<T: Copy, U: Copy, F: FnMut(T) -> U>(
+    vs: &[T],
+    layout: &Layout,
+    mut f: F,
+) -> Vec<U> {
+    match layout.strided_blocks() {
+        crate::StridedBlocks::SingleBlock { start_offset, len } => vs
+            [start_offset..start_offset + len]
+            .iter()
+            .map(|&v| f(v))
+            .collect(),
+        crate::StridedBlocks::MultipleBlocks {
+            block_start_index,
+            block_len,
+        } => {
+            let mut result = Vec::with_capacity(layout.shape().elem_count());
+            // Specialize the case where block_len is one to avoid the second loop.
+            if block_len == 1 {
+                for index in block_start_index {
+                    let v = unsafe { vs.get_unchecked(index) };
+                    result.push(f(*v))
+                }
+            } else {
+                for index in block_start_index {
+                    for offset in 0..block_len {
+                        let v = unsafe { vs.get_unchecked(index + offset) };
+                        result.push(f(*v))
+                    }
+                }
+            }
+            result
+        }
+    }
+}
+
+pub fn unary_map_vec<T: Copy, U: Copy, F: FnMut(T) -> U, FV: FnMut(&[T], &mut [U])>(
+    vs: &[T],
+    layout: &Layout,
+    mut f: F,
+    mut f_vec: FV,
+) -> Vec<U> {
+    match layout.strided_blocks() {
+        crate::StridedBlocks::SingleBlock { start_offset, len } => {
+            let mut ys: Vec<U> = Vec::with_capacity(len);
+            let ys_to_set = ys.spare_capacity_mut();
+            let ys_to_set = unsafe {
+                std::mem::transmute::<&mut [std::mem::MaybeUninit<U>], &mut [U]>(ys_to_set)
+            };
+            f_vec(&vs[start_offset..start_offset + len], ys_to_set);
+            // SAFETY: values are all set by f_vec.
+            unsafe { ys.set_len(len) };
+            ys
+        }
+        crate::StridedBlocks::MultipleBlocks {
+            block_start_index,
+            block_len,
+        } => {
+            let el_count = layout.shape().elem_count();
+            // Specialize the case where block_len is one to avoid the second loop.
+            if block_len == 1 {
+                let mut result = Vec::with_capacity(el_count);
+                for index in block_start_index {
+                    let v = unsafe { vs.get_unchecked(index) };
+                    result.push(f(*v))
+                }
+                result
+            } else {
+                let mut ys: Vec<U> = Vec::with_capacity(el_count);
+                let ys_to_set = ys.spare_capacity_mut();
+                let ys_to_set = unsafe {
+                    std::mem::transmute::<&mut [std::mem::MaybeUninit<U>], &mut [U]>(ys_to_set)
+                };
+                let mut dst_index = 0;
+                for src_index in block_start_index {
+                    let vs = &vs[src_index..src_index + block_len];
+                    let ys = &mut ys_to_set[dst_index..dst_index + block_len];
+                    f_vec(vs, ys);
+                    dst_index += block_len;
+                }
+                // SAFETY: values are all set by f_vec.
+                unsafe { ys.set_len(el_count) };
+                ys
+            }
+        }
+    }
+}

candle-core/src/cuda_backend/cudnn.rs

@@ -1,6 +1,6 @@
 use crate::WithDType;
 use cudarc;
-use cudarc::cudnn::safe::{Conv2dForward, Cudnn};
+use cudarc::cudnn::safe::{ConvForward, Cudnn};
 use cudarc::driver::{CudaSlice, CudaView, DeviceRepr, ValidAsZeroBits};
 use std::cell::RefCell;
 use std::collections::HashMap;
@@ -34,6 +34,9 @@ pub(crate) fn launch_conv2d<
     params: &crate::conv::ParamsConv2D,
     dev: &crate::cuda_backend::CudaDevice,
 ) -> crate::Result<()> {
+    use crate::conv::CudnnFwdAlgo as CandleAlgo;
+    use cudarc::cudnn::sys::cudnnConvolutionFwdAlgo_t as A;
+
     let device_id = dev.id();
     let cudnn = CUDNN.with(|cudnn| {
         if let Some(cudnn) = cudnn.borrow().get(&device_id) {
@@ -48,14 +51,14 @@ pub(crate) fn launch_conv2d<
     let conv = cudnn.create_conv2d::<T>(
         /* pad */ [params.padding as i32, params.padding as i32],
         /* stride */ [params.stride as i32, params.stride as i32],
-        /* dilation */ [1, 1],
+        /* dilation */ [params.dilation as i32, params.dilation as i32],
         cudarc::cudnn::sys::cudnnConvolutionMode_t::CUDNN_CROSS_CORRELATION,
     )?;
     let x_shape = [
         params.b_size as i32,
         params.c_in as i32,
-        params.i_w as i32,
         params.i_h as i32,
+        params.i_w as i32,
     ];
     // Note that `src` already starts at the proper offset.
     let x = if src_l.is_contiguous() {
@@ -75,22 +78,35 @@ pub(crate) fn launch_conv2d<
         [
             params.c_out as i32,
             params.c_in as i32,
-            params.k_w as i32,
             params.k_h as i32,
+            params.k_w as i32,
         ],
     )?;
     let (w_out, h_out) = (params.out_w() as i32, params.out_h() as i32);
     let y = cudnn.create_4d_tensor(
         cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
-        [params.b_size as i32, params.c_out as i32, w_out, h_out],
+        [params.b_size as i32, params.c_out as i32, h_out, w_out],
     )?;
-    let conv2d = Conv2dForward {
+    let conv2d = ConvForward {
         conv: &conv,
         x: &x,
         w: &w,
         y: &y,
     };
-    let alg = conv2d.pick_algorithm()?;
+    let alg = match params.cudnn_fwd_algo {
+        None => conv2d.pick_algorithm()?,
+        Some(CandleAlgo::ImplicitGemm) => A::CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM,
+        Some(CandleAlgo::ImplicitPrecompGemm) => {
+            A::CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM
+        }
+        Some(CandleAlgo::Gemm) => A::CUDNN_CONVOLUTION_FWD_ALGO_GEMM,
+        Some(CandleAlgo::Direct) => A::CUDNN_CONVOLUTION_FWD_ALGO_DIRECT,
+        Some(CandleAlgo::Fft) => A::CUDNN_CONVOLUTION_FWD_ALGO_FFT,
+        Some(CandleAlgo::FftTiling) => A::CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING,
+        Some(CandleAlgo::Winograd) => A::CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD,
+        Some(CandleAlgo::WinogradNonFused) => A::CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED,
+        Some(CandleAlgo::Count) => A::CUDNN_CONVOLUTION_FWD_ALGO_COUNT,
+    };
     let workspace_size = conv2d.get_workspace_size(alg)?;
     let mut workspace = dev.cuda_device().alloc_zeros::<u8>(workspace_size)?;
     unsafe {

candle-core/src/cuda_backend/device.rs

@@ -0,0 +1,452 @@
+use crate::backend::BackendDevice;
+use crate::{CpuStorage, CpuStorageRef, DType, Layout, Result, Shape};
+pub use candle_kernels as kernels;
+pub use cudarc;
+use cudarc::driver::{CudaFunction, LaunchAsync, LaunchConfig};
+use half::{bf16, f16};
+use std::sync::{Arc, Mutex};
+
+use super::{CudaError, CudaStorage, CudaStorageSlice, WrapErr};
+
+/// Unique identifier for cuda devices.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
+pub struct DeviceId(usize);
+
+impl DeviceId {
+    fn new() -> Self {
+        // https://users.rust-lang.org/t/idiomatic-rust-way-to-generate-unique-id/33805
+        use std::sync::atomic;
+        static COUNTER: atomic::AtomicUsize = atomic::AtomicUsize::new(1);
+        Self(COUNTER.fetch_add(1, atomic::Ordering::Relaxed))
+    }
+}
+
+struct CudaRng(cudarc::curand::CudaRng);
+unsafe impl Send for CudaRng {}
+
+#[derive(Clone)]
+pub struct CudaDevice {
+    id: DeviceId,
+    device: Arc<cudarc::driver::CudaDevice>,
+    pub(crate) blas: Arc<cudarc::cublas::CudaBlas>,
+    curand: Arc<Mutex<CudaRng>>,
+}
+
+impl std::fmt::Debug for CudaDevice {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "CudaDevice({:?})", self.id)
+    }
+}
+
+impl std::ops::Deref for CudaDevice {
+    type Target = Arc<cudarc::driver::CudaDevice>;
+
+    fn deref(&self) -> &Self::Target {
+        &self.device
+    }
+}
+
+impl CudaDevice {
+    pub fn cuda_device(&self) -> Arc<cudarc::driver::CudaDevice> {
+        self.device.clone()
+    }
+
+    pub fn id(&self) -> DeviceId {
+        self.id
+    }
+
+    fn const_impl(&self, v: f64, shape: &Shape, dtype: DType) -> Result<CudaStorage> {
+        let elem_count = shape.elem_count();
+        let cfg = LaunchConfig::for_num_elems(elem_count as u32);
+        let slice = match dtype {
+            DType::U8 => {
+                // SAFETY: Set later by running the fill kernel.
+                let data = unsafe { self.alloc::<u8>(elem_count) }.w()?;
+                let func = self.get_or_load_func("fill_u8", kernels::FILL)?;
+                let params = (&data, v as u8, elem_count);
+                unsafe { func.launch(cfg, params) }.w()?;
+                CudaStorageSlice::U8(data)
+            }
+            DType::U32 => {
+                // SAFETY: Set later by running the fill kernel.
+                let data = unsafe { self.alloc::<u32>(elem_count) }.w()?;
+                let func = self.get_or_load_func("fill_u32", kernels::FILL)?;
+                let params = (&data, v as u32, elem_count);
+                unsafe { func.launch(cfg, params) }.w()?;
+                CudaStorageSlice::U32(data)
+            }
+            DType::I64 => {
+                // SAFETY: Set later by running the fill kernel.
+                let data = unsafe { self.alloc::<i64>(elem_count) }.w()?;
+                let func = self.get_or_load_func("fill_i64", kernels::FILL)?;
+                let params = (&data, v as i64, elem_count);
+                unsafe { func.launch(cfg, params) }.w()?;
+                CudaStorageSlice::I64(data)
+            }
+            DType::BF16 => {
+                // SAFETY: Set later by running the fill kernel.
+                let data = unsafe { self.alloc::<bf16>(elem_count) }.w()?;
+                let func = self.get_or_load_func("fill_bf16", kernels::FILL)?;
+                let params = (&data, bf16::from_f64(v), elem_count);
+                unsafe { func.launch(cfg, params) }.w()?;
+                CudaStorageSlice::BF16(data)
+            }
+            DType::F16 => {
+                // SAFETY: Set later by running the fill kernel.
+                let data = unsafe { self.alloc::<f16>(elem_count) }.w()?;
+                let func = self.get_or_load_func("fill_f16", kernels::FILL)?;
+                let params = (&data, f16::from_f64(v), elem_count);
+                unsafe { func.launch(cfg, params) }.w()?;
+                CudaStorageSlice::F16(data)
+            }
+            DType::F32 => {
+                // SAFETY: Set later by running the fill kernel.
+                let data = unsafe { self.alloc::<f32>(elem_count) }.w()?;
+                let func = self.get_or_load_func("fill_f32", kernels::FILL)?;
+                let params = (&data, v as f32, elem_count);
+                unsafe { func.launch(cfg, params) }.w()?;
+                CudaStorageSlice::F32(data)
+            }
+            DType::F64 => {
+                // SAFETY: Set later by running the fill kernel.
+                let data = unsafe { self.alloc::<f64>(elem_count) }.w()?;
+                let func = self.get_or_load_func("fill_f64", kernels::FILL)?;
+                let params = (&data, v, elem_count);
+                unsafe { func.launch(cfg, params) }.w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
+    pub fn get_or_load_func(&self, module_name: &str, ptx: &'static str) -> Result<CudaFunction> {
+        if !self.has_func(module_name, module_name) {
+            // Leaking the string here is a bit sad but we need a &'static str and this is only
+            // done once per kernel name.
+            let static_module_name = Box::leak(module_name.to_string().into_boxed_str());
+            self.load_ptx(ptx.into(), module_name, &[static_module_name])
+                .map_err(|cuda| CudaError::Load {
+                    cuda,
+                    module_name: module_name.to_string(),
+                })
+                .w()?;
+        }
+        self.get_func(module_name, module_name)
+            // Clippy recommends this `ok_or` rather than `ok_or_else` so hopefully the compiler is
+            // able to only build the error value if needed.
+            .ok_or(CudaError::MissingKernel {
+                module_name: module_name.to_string(),
+            })
+            .w()
+    }
+}
+
+impl BackendDevice for CudaDevice {
+    type Storage = CudaStorage;
+
+    fn new(ordinal: usize) -> Result<Self> {
+        let device = cudarc::driver::CudaDevice::new(ordinal).w()?;
+        let blas = cudarc::cublas::CudaBlas::new(device.clone()).w()?;
+        let curand = cudarc::curand::CudaRng::new(299792458, device.clone()).w()?;
+        Ok(Self {
+            id: DeviceId::new(),
+            device,
+            blas: Arc::new(blas),
+            curand: Arc::new(Mutex::new(CudaRng(curand))),
+        })
+    }
+
+    fn set_seed(&self, seed: u64) -> Result<()> {
+        // We do not call set_seed but instead create a new curand object. This ensures that the
+        // state will be identical and the same random numbers will be generated.
+        let mut curand = self.curand.lock().unwrap();
+        curand.0 = cudarc::curand::CudaRng::new(seed, self.device.clone()).w()?;
+        Ok(())
+    }
+
+    fn location(&self) -> crate::DeviceLocation {
+        crate::DeviceLocation::Cuda {
+            gpu_id: self.device.ordinal(),
+        }
+    }
+
+    fn same_device(&self, rhs: &Self) -> bool {
+        self.id == rhs.id
+    }
+
+    fn zeros_impl(&self, shape: &Shape, dtype: DType) -> Result<CudaStorage> {
+        let elem_count = shape.elem_count();
+        let slice = match dtype {
+            DType::U8 => {
+                let data = self.alloc_zeros::<u8>(elem_count).w()?;
+                CudaStorageSlice::U8(data)
+            }
+            DType::U32 => {
+                let data = self.alloc_zeros::<u32>(elem_count).w()?;
+                CudaStorageSlice::U32(data)
+            }
+            DType::I64 => {
+                let data = self.alloc_zeros::<i64>(elem_count).w()?;
+                CudaStorageSlice::I64(data)
+            }
+            DType::BF16 => {
+                let data = self.alloc_zeros::<bf16>(elem_count).w()?;
+                CudaStorageSlice::BF16(data)
+            }
+            DType::F16 => {
+                let data = self.alloc_zeros::<f16>(elem_count).w()?;
+                CudaStorageSlice::F16(data)
+            }
+            DType::F32 => {
+                let data = self.alloc_zeros::<f32>(elem_count).w()?;
+                CudaStorageSlice::F32(data)
+            }
+            DType::F64 => {
+                let data = self.alloc_zeros::<f64>(elem_count).w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
+    fn rand_uniform(&self, shape: &Shape, dtype: DType, lo: f64, up: f64) -> Result<CudaStorage> {
+        let elem_count = shape.elem_count();
+        let curand = self.curand.lock().unwrap();
+        let slice = match dtype {
+            // TODO: Add support for F16 and BF16 though this is likely to require some upstream
+            // cudarc changes.
+            DType::U8 | DType::U32 | DType::I64 | DType::F16 | DType::BF16 => {
+                Err(CudaError::UnsupportedDtype {
+                    dtype,
+                    op: "rand_uniform",
+                })
+                .w()?
+            }
+            DType::F32 => {
+                let mut data = unsafe { self.alloc::<f32>(elem_count) }.w()?;
+                curand.0.fill_with_uniform(&mut data).w()?;
+                CudaStorageSlice::F32(data)
+            }
+            DType::F64 => {
+                let mut data = unsafe { self.alloc::<f64>(elem_count) }.w()?;
+                curand.0.fill_with_uniform(&mut data).w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        let slice = if lo == 0. && up == 1.0 {
+            slice
+        } else {
+            use super::utils::Map1;
+            let layout = Layout::contiguous(shape);
+            super::Affine(up - lo, lo).map(&slice, self, &layout)?
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
+    fn rand_normal(&self, shape: &Shape, dtype: DType, mean: f64, std: f64) -> Result<CudaStorage> {
+        // TODO: Add support for F16 and BF16 though this is likely to require some upstream
+        // cudarc changes.
+        let elem_count = shape.elem_count();
+        let curand = self.curand.lock().unwrap();
+        // curand can only generate an odd number of values.
+        // https://github.com/huggingface/candle/issues/734
+        let elem_count_round = if elem_count % 2 == 1 {
+            elem_count + 1
+        } else {
+            elem_count
+        };
+        let slice = match dtype {
+            DType::U8 | DType::U32 | DType::I64 | DType::F16 | DType::BF16 => {
+                Err(CudaError::UnsupportedDtype {
+                    dtype,
+                    op: "rand_normal",
+                })
+                .w()?
+            }
+            DType::F32 => {
+                let mut data = unsafe { self.alloc::<f32>(elem_count_round) }.w()?;
+                curand
+                    .0
+                    .fill_with_normal(&mut data, mean as f32, std as f32)
+                    .w()?;
+                CudaStorageSlice::F32(data)
+            }
+            DType::F64 => {
+                let mut data = unsafe { self.alloc::<f64>(elem_count_round) }.w()?;
+                curand.0.fill_with_normal(&mut data, mean, std).w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
+    fn ones_impl(&self, shape: &Shape, dtype: DType) -> Result<CudaStorage> {
+        self.const_impl(1., shape, dtype)
+    }
+
+    unsafe fn alloc_uninit(&self, shape: &Shape, dtype: DType) -> Result<Self::Storage> {
+        let elem_count = shape.elem_count();
+        let slice = match dtype {
+            DType::U8 => {
+                let data = self.alloc::<u8>(elem_count).w()?;
+                CudaStorageSlice::U8(data)
+            }
+            DType::U32 => {
+                let data = self.alloc::<u32>(elem_count).w()?;
+                CudaStorageSlice::U32(data)
+            }
+            DType::I64 => {
+                let data = self.alloc::<i64>(elem_count).w()?;
+                CudaStorageSlice::I64(data)
+            }
+            DType::BF16 => {
+                let data = self.alloc::<bf16>(elem_count).w()?;
+                CudaStorageSlice::BF16(data)
+            }
+            DType::F16 => {
+                let data = self.alloc::<f16>(elem_count).w()?;
+                CudaStorageSlice::F16(data)
+            }
+            DType::F32 => {
+                let data = self.alloc::<f32>(elem_count).w()?;
+                CudaStorageSlice::F32(data)
+            }
+            DType::F64 => {
+                let data = self.alloc::<f64>(elem_count).w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
+    fn storage_from_slice<T: crate::WithDType>(&self, s: &[T]) -> Result<Self::Storage> {
+        let slice = match T::cpu_storage_ref(s) {
+            CpuStorageRef::U8(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::U8(data)
+            }
+            CpuStorageRef::U32(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::U32(data)
+            }
+            CpuStorageRef::I64(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::I64(data)
+            }
+            CpuStorageRef::BF16(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::BF16(data)
+            }
+            CpuStorageRef::F16(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::F16(data)
+            }
+            CpuStorageRef::F32(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::F32(data)
+            }
+            CpuStorageRef::F64(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
+    fn storage_from_cpu_storage(&self, storage: &CpuStorage) -> Result<CudaStorage> {
+        let slice = match storage {
+            CpuStorage::U8(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::U8(data)
+            }
+            CpuStorage::U32(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::U32(data)
+            }
+            CpuStorage::I64(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::I64(data)
+            }
+            CpuStorage::BF16(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::BF16(data)
+            }
+            CpuStorage::F16(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::F16(data)
+            }
+            CpuStorage::F32(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::F32(data)
+            }
+            CpuStorage::F64(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
+    fn storage_from_cpu_storage_owned(&self, storage: CpuStorage) -> Result<CudaStorage> {
+        let slice = match storage {
+            CpuStorage::U8(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::U8(data)
+            }
+            CpuStorage::U32(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::U32(data)
+            }
+            CpuStorage::I64(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::I64(data)
+            }
+            CpuStorage::BF16(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::BF16(data)
+            }
+            CpuStorage::F16(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::F16(data)
+            }
+            CpuStorage::F32(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::F32(data)
+            }
+            CpuStorage::F64(storage) => {
+                let data = self.htod_copy(storage).w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
+    fn synchronize(&self) -> Result<()> {
+        self.device.synchronize().map_err(crate::Error::wrap)?;
+        Ok(())
+    }
+}

candle-core/src/cuda_backend/error.rs

@@ -0,0 +1,62 @@
+use crate::{DType, Layout};
+
+/// cudarc related errors
+#[derive(thiserror::Error, Debug)]
+pub enum CudaError {
+    #[error(transparent)]
+    Cuda(#[from] cudarc::driver::DriverError),
+
+    #[error(transparent)]
+    Compiler(#[from] cudarc::nvrtc::CompileError),
+
+    #[error(transparent)]
+    Cublas(#[from] cudarc::cublas::result::CublasError),
+
+    #[error(transparent)]
+    Curand(#[from] cudarc::curand::result::CurandError),
+
+    #[error("missing kernel '{module_name}'")]
+    MissingKernel { module_name: String },
+
+    #[error("unsupported dtype {dtype:?} for {op}")]
+    UnsupportedDtype { dtype: DType, op: &'static str },
+
+    #[error("internal error '{0}'")]
+    InternalError(&'static str),
+
+    #[error("matmul is only supported for contiguous tensors lstride: {lhs_stride:?} rstride: {rhs_stride:?} mnk: {mnk:?}")]
+    MatMulNonContiguous {
+        lhs_stride: Layout,
+        rhs_stride: Layout,
+        mnk: (usize, usize, usize),
+    },
+
+    #[error("{msg}, expected: {expected:?}, got: {got:?}")]
+    UnexpectedDType {
+        msg: &'static str,
+        expected: DType,
+        got: DType,
+    },
+
+    #[error("{cuda} when loading {module_name}")]
+    Load {
+        cuda: cudarc::driver::DriverError,
+        module_name: String,
+    },
+}
+
+impl From<CudaError> for crate::Error {
+    fn from(val: CudaError) -> Self {
+        crate::Error::Cuda(Box::new(val)).bt()
+    }
+}
+
+pub trait WrapErr<O> {
+    fn w(self) -> std::result::Result<O, crate::Error>;
+}
+
+impl<O, E: Into<CudaError>> WrapErr<O> for std::result::Result<O, E> {
+    fn w(self) -> std::result::Result<O, crate::Error> {
+        self.map_err(|e| crate::Error::Cuda(Box::new(e.into())).bt())
+    }
+}

candle-core/src/cuda_backend/utils.rs

@@ -0,0 +1,172 @@
+/// Helper functions to plug cuda kernels in candle.
+use crate::{Layout, Result, Shape, WithDType};
+pub use cudarc;
+use cudarc::driver::{CudaSlice, DeviceRepr, ValidAsZeroBits};
+
+use super::{CudaDevice, CudaError, WrapErr};
+
+pub type S = super::CudaStorageSlice;
+
+pub trait Map1 {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        src: &CudaSlice<T>,
+        dev: &CudaDevice,
+        layout: &Layout,
+    ) -> Result<CudaSlice<T>>;
+
+    fn map(&self, s: &S, d: &CudaDevice, l: &Layout) -> Result<S> {
+        let out = match s {
+            S::U8(s) => S::U8(self.f(s, d, l)?),
+            S::U32(s) => S::U32(self.f(s, d, l)?),
+            S::I64(s) => S::I64(self.f(s, d, l)?),
+            S::BF16(s) => S::BF16(self.f(s, d, l)?),
+            S::F16(s) => S::F16(self.f(s, d, l)?),
+            S::F32(s) => S::F32(self.f(s, d, l)?),
+            S::F64(s) => S::F64(self.f(s, d, l)?),
+        };
+        Ok(out)
+    }
+}
+
+pub trait Map2 {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        src1: &CudaSlice<T>,
+        layout1: &Layout,
+        src2: &CudaSlice<T>,
+        layout2: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<CudaSlice<T>>;
+
+    fn map(&self, s1: &S, l1: &Layout, s2: &S, l2: &Layout, d: &CudaDevice) -> Result<S> {
+        let out = match (s1, s2) {
+            (S::U8(s1), S::U8(s2)) => S::U8(self.f(s1, l1, s2, l2, d)?),
+            (S::U32(s1), S::U32(s2)) => S::U32(self.f(s1, l1, s2, l2, d)?),
+            (S::I64(s1), S::I64(s2)) => S::I64(self.f(s1, l1, s2, l2, d)?),
+            (S::BF16(s1), S::BF16(s2)) => S::BF16(self.f(s1, l1, s2, l2, d)?),
+            (S::F16(s1), S::F16(s2)) => S::F16(self.f(s1, l1, s2, l2, d)?),
+            (S::F32(s1), S::F32(s2)) => S::F32(self.f(s1, l1, s2, l2, d)?),
+            (S::F64(s1), S::F64(s2)) => S::F64(self.f(s1, l1, s2, l2, d)?),
+            _ => Err(CudaError::InternalError("dtype mismatch in binary op"))?,
+        };
+        Ok(out)
+    }
+}
+
+pub trait Map3 {
+    #[allow(clippy::too_many_arguments)]
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        src1: &CudaSlice<T>,
+        layout1: &Layout,
+        src2: &CudaSlice<T>,
+        layout2: &Layout,
+        src3: &CudaSlice<T>,
+        layout3: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<CudaSlice<T>>;
+
+    #[allow(clippy::too_many_arguments)]
+    fn map(
+        &self,
+        s1: &S,
+        l1: &Layout,
+        s2: &S,
+        l2: &Layout,
+        s3: &S,
+        l3: &Layout,
+        d: &CudaDevice,
+    ) -> Result<S> {
+        let out = match (s1, s2, s3) {
+            (S::U8(s1), S::U8(s2), S::U8(s3)) => S::U8(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::U32(s1), S::U32(s2), S::U32(s3)) => S::U32(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::I64(s1), S::I64(s2), S::I64(s3)) => S::I64(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::BF16(s1), S::BF16(s2), S::BF16(s3)) => S::BF16(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::F16(s1), S::F16(s2), S::F16(s3)) => S::F16(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::F32(s1), S::F32(s2), S::F32(s3)) => S::F32(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::F64(s1), S::F64(s2), S::F64(s3)) => S::F64(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            _ => Err(CudaError::InternalError("dtype mismatch in ternary op"))?,
+        };
+        Ok(out)
+    }
+}
+
+pub trait Map2InPlace {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        dst: &mut CudaSlice<T>,
+        dst_shape: &Shape,
+        src: &CudaSlice<T>,
+        src_l: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<()>;
+
+    fn map(
+        &self,
+        dst: &mut S,
+        dst_s: &Shape,
+        src: &S,
+        src_l: &Layout,
+        d: &CudaDevice,
+    ) -> Result<()> {
+        match (dst, src) {
+            (S::U8(dst), S::U8(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::U32(dst), S::U32(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::I64(dst), S::I64(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::BF16(dst), S::BF16(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::F16(dst), S::F16(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::F32(dst), S::F32(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::F64(dst), S::F64(src)) => self.f(dst, dst_s, src, src_l, d),
+            _ => Err(CudaError::InternalError("dtype mismatch in binary op"))?,
+        }
+    }
+}
+
+pub trait Map1Any {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits, W: Fn(CudaSlice<T>) -> S>(
+        &self,
+        src: &CudaSlice<T>,
+        dev: &CudaDevice,
+        layout: &Layout,
+        wrap: W,
+    ) -> Result<S>;
+
+    fn map(&self, s: &S, d: &CudaDevice, l: &Layout) -> Result<S> {
+        let out = match s {
+            S::U8(s) => self.f(s, d, l, S::U8)?,
+            S::U32(s) => self.f(s, d, l, S::U32)?,
+            S::I64(s) => self.f(s, d, l, S::I64)?,
+            S::BF16(s) => self.f(s, d, l, S::BF16)?,
+            S::F16(s) => self.f(s, d, l, S::F16)?,
+            S::F32(s) => self.f(s, d, l, S::F32)?,
+            S::F64(s) => self.f(s, d, l, S::F64)?,
+        };
+        Ok(out)
+    }
+}
+
+pub trait Map2Any {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        src1: &CudaSlice<T>,
+        layout1: &Layout,
+        src2: &CudaSlice<T>,
+        layout2: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<S>;
+
+    fn map(&self, s1: &S, l1: &Layout, s2: &S, l2: &Layout, d: &CudaDevice) -> Result<S> {
+        let out = match (s1, s2) {
+            (S::U8(s1), S::U8(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::U32(s1), S::U32(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::I64(s1), S::I64(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::BF16(s1), S::BF16(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::F16(s1), S::F16(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::F32(s1), S::F32(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::F64(s1), S::F64(s2)) => self.f(s1, l1, s2, l2, d)?,
+            _ => Err(CudaError::InternalError("dtype mismatch in binary op")).w()?,
+        };
+        Ok(out)
+    }
+}

candle-core/src/custom_op.rs

@@ -0,0 +1,377 @@
+use crate::op::{BackpropOp, Op};
+use crate::tensor::from_storage;
+use crate::{CpuStorage, CudaStorage, Layout, MetalStorage, Result, Shape, Tensor};
+use std::sync::Arc;
+
+/// Unary ops that can be defined in user-land.
+pub trait CustomOp1 {
+    // Box<dyn> does not support const yet, so use a function to get the name.
+    fn name(&self) -> &'static str;
+
+    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cpu_fwd(&self, storage: &CpuStorage, layout: &Layout) -> Result<(CpuStorage, Shape)>;
+
+    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cuda_fwd(&self, _storage: &CudaStorage, _layout: &Layout) -> Result<(CudaStorage, Shape)> {
+        Err(crate::Error::Cuda(
+            format!("no cuda implementation for {}", self.name()).into(),
+        ))
+    }
+
+    /// The forward pass, as run on a metal gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn metal_fwd(
+        &self,
+        _storage: &MetalStorage,
+        _layout: &Layout,
+    ) -> Result<(MetalStorage, Shape)> {
+        Err(crate::Error::Metal(
+            format!("no metal implementation for {}", self.name()).into(),
+        ))
+    }
+
+    /// This function takes as argument the argument `arg` used in the forward pass, the result
+    /// produced by the forward operation `res` and the gradient of the result `grad_res`.
+    /// The function should return the gradient of the argument.
+    fn bwd(&self, _arg: &Tensor, _res: &Tensor, _grad_res: &Tensor) -> Result<Option<Tensor>> {
+        Err(crate::Error::BackwardNotSupported { op: self.name() })
+    }
+}
+
+pub trait CustomOp2 {
+    fn name(&self) -> &'static str;
+
+    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cpu_fwd(
+        &self,
+        s1: &CpuStorage,
+        l1: &Layout,
+        s2: &CpuStorage,
+        l2: &Layout,
+    ) -> Result<(CpuStorage, Shape)>;
+
+    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cuda_fwd(
+        &self,
+        _: &CudaStorage,
+        _: &Layout,
+        _: &CudaStorage,
+        _: &Layout,
+    ) -> Result<(CudaStorage, Shape)> {
+        Err(crate::Error::Cuda(
+            format!("no cuda implementation for {}", self.name()).into(),
+        ))
+    }
+
+    /// The forward pass, as run on a metal gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn metal_fwd(
+        &self,
+        _: &MetalStorage,
+        _: &Layout,
+        _: &MetalStorage,
+        _: &Layout,
+    ) -> Result<(MetalStorage, Shape)> {
+        Err(crate::Error::Metal(
+            format!("no metal implementation for {}", self.name()).into(),
+        ))
+    }
+
+    fn bwd(
+        &self,
+        _arg1: &Tensor,
+        _arg2: &Tensor,
+        _res: &Tensor,
+        _grad_res: &Tensor,
+    ) -> Result<(Option<Tensor>, Option<Tensor>)> {
+        Err(crate::Error::BackwardNotSupported { op: self.name() })
+    }
+}
+
+pub trait CustomOp3 {
+    fn name(&self) -> &'static str;
+
+    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cpu_fwd(
+        &self,
+        s1: &CpuStorage,
+        l1: &Layout,
+        s2: &CpuStorage,
+        l2: &Layout,
+        s3: &CpuStorage,
+        l3: &Layout,
+    ) -> Result<(CpuStorage, Shape)>;
+
+    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cuda_fwd(
+        &self,
+        _: &CudaStorage,
+        _: &Layout,
+        _: &CudaStorage,
+        _: &Layout,
+        _: &CudaStorage,
+        _: &Layout,
+    ) -> Result<(CudaStorage, Shape)> {
+        Err(crate::Error::Cuda(
+            format!("no cuda implementation for {}", self.name()).into(),
+        ))
+    }
+
+    /// The forward pass, as run on a metal gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn metal_fwd(
+        &self,
+        _: &MetalStorage,
+        _: &Layout,
+        _: &MetalStorage,
+        _: &Layout,
+        _: &MetalStorage,
+        _: &Layout,
+    ) -> Result<(MetalStorage, Shape)> {
+        Err(crate::Error::Metal(
+            format!("no metal implementation for {}", self.name()).into(),
+        ))
+    }
+
+    fn bwd(
+        &self,
+        _arg1: &Tensor,
+        _arg2: &Tensor,
+        _arg3: &Tensor,
+        _res: &Tensor,
+        _grad_res: &Tensor,
+    ) -> Result<(Option<Tensor>, Option<Tensor>, Option<Tensor>)> {
+        Err(crate::Error::BackwardNotSupported { op: self.name() })
+    }
+}
+
+impl Tensor {
+    /// Applies a unary custom op without backward support
+    pub fn apply_op1_no_bwd<C: CustomOp1>(&self, c: &C) -> Result<Self> {
+        let (storage, shape) = self.storage().apply_op1(self.layout(), c)?;
+        Ok(from_storage(storage, shape, BackpropOp::none(), false))
+    }
+
+    /// Applies a binary custom op without backward support
+    pub fn apply_op2_no_bwd<C: CustomOp2>(&self, rhs: &Self, c: &C) -> Result<Self> {
+        let (storage, shape) =
+            self.storage()
+                .apply_op2(self.layout(), &rhs.storage(), rhs.layout(), c)?;
+        Ok(from_storage(storage, shape, BackpropOp::none(), false))
+    }
+
+    /// Applies a ternary custom op without backward support
+    pub fn apply_op3_no_bwd<C: CustomOp3>(&self, t2: &Self, t3: &Self, c: &C) -> Result<Self> {
+        let (storage, shape) = self.storage().apply_op3(
+            self.layout(),
+            &t2.storage(),
+            t2.layout(),
+            &t3.storage(),
+            t3.layout(),
+            c,
+        )?;
+        Ok(from_storage(storage, shape, BackpropOp::none(), false))
+    }
+
+    /// Applies a unary custom op.
+    pub fn apply_op1_arc(&self, c: Arc<Box<dyn CustomOp1 + Send + Sync>>) -> Result<Self> {
+        let (storage, shape) = self
+            .storage()
+            .apply_op1(self.layout(), c.as_ref().as_ref())?;
+        let op = BackpropOp::new1(self, |s| Op::CustomOp1(s, c.clone()));
+        Ok(from_storage(storage, shape, op, false))
+    }
+
+    pub fn apply_op1<C: 'static + CustomOp1 + Send + Sync>(&self, c: C) -> Result<Self> {
+        self.apply_op1_arc(Arc::new(Box::new(c)))
+    }
+
+    /// Applies a binary custom op.
+    pub fn apply_op2_arc(
+        &self,
+        rhs: &Self,
+        c: Arc<Box<dyn CustomOp2 + Send + Sync>>,
+    ) -> Result<Self> {
+        let (storage, shape) = self.storage().apply_op2(
+            self.layout(),
+            &rhs.storage(),
+            rhs.layout(),
+            c.as_ref().as_ref(),
+        )?;
+        let op = BackpropOp::new2(self, rhs, |t1, t2| Op::CustomOp2(t1, t2, c.clone()));
+        Ok(from_storage(storage, shape, op, false))
+    }
+
+    pub fn apply_op2<C: 'static + CustomOp2 + Send + Sync>(&self, r: &Self, c: C) -> Result<Self> {
+        self.apply_op2_arc(r, Arc::new(Box::new(c)))
+    }
+
+    /// Applies a ternary custom op.
+    pub fn apply_op3_arc(
+        &self,
+        t2: &Self,
+        t3: &Self,
+        c: Arc<Box<dyn CustomOp3 + Send + Sync>>,
+    ) -> Result<Self> {
+        let (storage, shape) = self.storage().apply_op3(
+            self.layout(),
+            &t2.storage(),
+            t2.layout(),
+            &t3.storage(),
+            t3.layout(),
+            c.as_ref().as_ref(),
+        )?;
+        let op = BackpropOp::new3(self, t2, t3, |t1, t2, t3| {
+            Op::CustomOp3(t1, t2, t3, c.clone())
+        });
+        Ok(from_storage(storage, shape, op, false))
+    }
+
+    pub fn apply_op3<C: 'static + CustomOp3 + Send + Sync>(
+        &self,
+        t2: &Self,
+        t3: &Self,
+        c: C,
+    ) -> Result<Self> {
+        self.apply_op3_arc(t2, t3, Arc::new(Box::new(c)))
+    }
+}
+
+// In place ops.
+
+/// Unary ops that can be defined in user-land.
+/// These ops work in place and as such back-prop is unsupported.
+pub trait InplaceOp1 {
+    // Box<dyn> does not support const yet, so use a function to get the name.
+    fn name(&self) -> &'static str;
+
+    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cpu_fwd(&self, storage: &mut CpuStorage, layout: &Layout) -> Result<()>;
+
+    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cuda_fwd(&self, _storage: &mut CudaStorage, _layout: &Layout) -> Result<()> {
+        Err(crate::Error::Cuda(
+            format!("no cuda implementation for {}", self.name()).into(),
+        ))
+    }
+
+    /// The forward pass, as run on a metal gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn metal_fwd(&self, _storage: &mut MetalStorage, _layout: &Layout) -> Result<()> {
+        Err(crate::Error::Metal(
+            format!("no metal implementation for {}", self.name()).into(),
+        ))
+    }
+}
+
+pub trait InplaceOp2 {
+    fn name(&self) -> &'static str;
+
+    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cpu_fwd(&self, s1: &mut CpuStorage, l1: &Layout, s2: &CpuStorage, l2: &Layout)
+        -> Result<()>;
+
+    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cuda_fwd(&self, _: &mut CudaStorage, _: &Layout, _: &CudaStorage, _: &Layout) -> Result<()> {
+        Err(crate::Error::Cuda(
+            format!("no cuda implementation for {}", self.name()).into(),
+        ))
+    }
+
+    /// The forward pass, as run on a metal gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn metal_fwd(
+        &self,
+        _: &mut MetalStorage,
+        _: &Layout,
+        _: &MetalStorage,
+        _: &Layout,
+    ) -> Result<()> {
+        Err(crate::Error::Metal(
+            format!("no metal implementation for {}", self.name()).into(),
+        ))
+    }
+}
+
+pub trait InplaceOp3 {
+    fn name(&self) -> &'static str;
+
+    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cpu_fwd(
+        &self,
+        s1: &mut CpuStorage,
+        l1: &Layout,
+        s2: &CpuStorage,
+        l2: &Layout,
+        s3: &CpuStorage,
+        l3: &Layout,
+    ) -> Result<()>;
+
+    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn cuda_fwd(
+        &self,
+        _: &mut CudaStorage,
+        _: &Layout,
+        _: &CudaStorage,
+        _: &Layout,
+        _: &CudaStorage,
+        _: &Layout,
+    ) -> Result<()> {
+        Err(crate::Error::Cuda(
+            format!("no cuda implementation for {}", self.name()).into(),
+        ))
+    }
+
+    /// The forward pass, as run on a metal gpu device. Note that the storage can use arbitrary strides,
+    /// offsets etc so the associated layout should be used to access it.
+    fn metal_fwd(
+        &self,
+        _: &mut MetalStorage,
+        _: &Layout,
+        _: &MetalStorage,
+        _: &Layout,
+        _: &MetalStorage,
+        _: &Layout,
+    ) -> Result<()> {
+        Err(crate::Error::Metal(
+            format!("no metal implementation for {}", self.name()).into(),
+        ))
+    }
+}
+
+impl Tensor {
+    /// Applies a unary custom op in place.
+    pub fn inplace_op1<C: InplaceOp1>(&self, c: &C) -> Result<()> {
+        self.storage_mut().inplace_op1(self.layout(), c)
+    }
+
+    /// Applies a unary custom op in place (for the first tensor).
+    pub fn inplace_op2<C: InplaceOp2>(&self, rhs: &Self, c: &C) -> Result<()> {
+        self.storage_mut()
+            .inplace_op2(self.layout(), &rhs.storage(), rhs.layout(), c)
+    }
+
+    /// Applies a ternary custom op in place (for the first tensor).
+    pub fn inplace_op3<C: InplaceOp3>(&self, t2: &Self, t3: &Self, c: &C) -> Result<()> {
+        self.storage_mut().inplace_op3(
+            self.layout(),
+            &t2.storage(),
+            t2.layout(),
+            &t3.storage(),
+            t3.layout(),
+            c,
+        )
+    }
+}

candle-core/src/device.rs

@@ -8,15 +8,16 @@ use crate::{CpuStorage, DType, Result, Shape, Storage, WithDType};
 pub enum DeviceLocation {
     Cpu,
     Cuda { gpu_id: usize },
+    Metal { gpu_id: usize },
 }
 
 #[derive(Debug, Clone)]
 pub enum Device {
     Cpu,
     Cuda(crate::CudaDevice),
+    Metal(crate::MetalDevice),
 }
 
-// TODO: Should we back the cpu implementation using the NdArray crate or similar?
 pub trait NdArray {
     fn shape(&self) -> Result<Shape>;
 
@@ -81,15 +82,71 @@ impl<S: WithDType, const N1: usize, const N2: usize, const N3: usize> NdArray
     }
 }
 
+impl<S: WithDType, const N1: usize, const N2: usize, const N3: usize, const N4: usize> NdArray
+    for &[[[[S; N4]; N3]; N2]; N1]
+{
+    fn shape(&self) -> Result<Shape> {
+        Ok(Shape::from((N1, N2, N3, N4)))
+    }
+
+    fn to_cpu_storage(&self) -> CpuStorage {
+        let mut vec = Vec::with_capacity(N1 * N2 * N3 * N4);
+        for i1 in 0..N1 {
+            for i2 in 0..N2 {
+                for i3 in 0..N3 {
+                    vec.extend(self[i1][i2][i3])
+                }
+            }
+        }
+        S::to_cpu_storage_owned(vec)
+    }
+}
+
+impl<S: NdArray> NdArray for Vec<S> {
+    fn shape(&self) -> Result<Shape> {
+        if self.is_empty() {
+            crate::bail!("empty array")
+        }
+        let shape0 = self[0].shape()?;
+        let n = self.len();
+        for v in self.iter() {
+            let shape = v.shape()?;
+            if shape != shape0 {
+                crate::bail!("two elements have different shapes {shape:?} {shape0:?}")
+            }
+        }
+        Ok(Shape::from([[n].as_slice(), shape0.dims()].concat()))
+    }
+
+    fn to_cpu_storage(&self) -> CpuStorage {
+        // This allocates intermediary memory and shouldn't be necessary.
+        let storages = self.iter().map(|v| v.to_cpu_storage()).collect::<Vec<_>>();
+        CpuStorage::concat(storages.as_slice()).unwrap()
+    }
+}
+
 impl Device {
     pub fn new_cuda(ordinal: usize) -> Result<Self> {
         Ok(Self::Cuda(crate::CudaDevice::new(ordinal)?))
     }
 
+    pub fn new_metal(ordinal: usize) -> Result<Self> {
+        Ok(Self::Metal(crate::MetalDevice::new(ordinal)?))
+    }
+
+    pub fn set_seed(&self, seed: u64) -> Result<()> {
+        match self {
+            Self::Cpu => CpuDevice.set_seed(seed),
+            Self::Cuda(c) => c.set_seed(seed),
+            Self::Metal(m) => m.set_seed(seed),
+        }
+    }
+
     pub fn same_device(&self, rhs: &Self) -> bool {
         match (self, rhs) {
             (Self::Cpu, Self::Cpu) => true,
             (Self::Cuda(lhs), Self::Cuda(rhs)) => lhs.same_device(rhs),
+            (Self::Metal(lhs), Self::Metal(rhs)) => lhs.same_device(rhs),
             _ => false,
         }
     }
@@ -98,20 +155,35 @@ impl Device {
         match self {
             Self::Cpu => DeviceLocation::Cpu,
             Self::Cuda(device) => device.location(),
+            Device::Metal(device) => device.location(),
         }
     }
 
     pub fn is_cpu(&self) -> bool {
-        match self {
-            Self::Cpu => true,
-            Self::Cuda(_) => false,
-        }
+        matches!(self, Self::Cpu)
     }
 
     pub fn is_cuda(&self) -> bool {
+        matches!(self, Self::Cuda(_))
+    }
+
+    pub fn is_metal(&self) -> bool {
+        matches!(self, Self::Metal(_))
+    }
+
+    pub fn supports_bf16(&self) -> bool {
         match self {
+            Self::Cuda(_) | Self::Metal(_) => true,
             Self::Cpu => false,
-            Self::Cuda(_) => true,
+        }
+    }
+
+    /// Return `BF16` for devices that support it, otherwise default to `F32`.
+    pub fn bf16_default_to_f32(&self) -> DType {
+        if self.supports_bf16() {
+            DType::BF16
+        } else {
+            DType::F32
         }
     }
 
@@ -136,8 +208,18 @@ impl Device {
                 Ok(Storage::Cpu(storage))
             }
             Device::Cuda(device) => {
+                // TODO: Remove the special case if we start supporting generating f16/bf16 directly.
+                if dtype == DType::F16 || dtype == DType::BF16 {
+                    let storage = device.rand_uniform(shape, DType::F32, lo, up)?;
+                    Storage::Cuda(storage).to_dtype(&crate::Layout::contiguous(shape), dtype)
+                } else {
+                    let storage = device.rand_uniform(shape, dtype, lo, up)?;
+                    Ok(Storage::Cuda(storage))
+                }
+            }
+            Device::Metal(device) => {
                 let storage = device.rand_uniform(shape, dtype, lo, up)?;
-                Ok(Storage::Cuda(storage))
+                Ok(Storage::Metal(storage))
             }
         }
     }
@@ -164,8 +246,18 @@ impl Device {
                 Ok(Storage::Cpu(storage))
             }
             Device::Cuda(device) => {
+                // TODO: Remove the special case if we start supporting generating f16/bf16 directly.
+                if dtype == DType::F16 || dtype == DType::BF16 {
+                    let storage = device.rand_normal(shape, DType::F32, mean, std)?;
+                    Storage::Cuda(storage).to_dtype(&crate::Layout::contiguous(shape), dtype)
+                } else {
+                    let storage = device.rand_normal(shape, dtype, mean, std)?;
+                    Ok(Storage::Cuda(storage))
+                }
+            }
+            Device::Metal(device) => {
                 let storage = device.rand_normal(shape, dtype, mean, std)?;
-                Ok(Storage::Cuda(storage))
+                Ok(Storage::Metal(storage))
             }
         }
     }
@@ -189,6 +281,10 @@ impl Device {
                 let storage = device.ones_impl(shape, dtype)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = device.ones_impl(shape, dtype)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 
@@ -202,6 +298,41 @@ impl Device {
                 let storage = device.zeros_impl(shape, dtype)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = device.zeros_impl(shape, dtype)?;
+                Ok(Storage::Metal(storage))
+            }
+        }
+    }
+
+    pub(crate) unsafe fn alloc_uninit(&self, shape: &Shape, dtype: DType) -> Result<Storage> {
+        match self {
+            Device::Cpu => {
+                let storage = CpuDevice.alloc_uninit(shape, dtype)?;
+                Ok(Storage::Cpu(storage))
+            }
+            Device::Cuda(device) => {
+                let storage = device.alloc_uninit(shape, dtype)?;
+                Ok(Storage::Cuda(storage))
+            }
+            Device::Metal(device) => {
+                let storage = device.alloc_uninit(shape, dtype)?;
+                Ok(Storage::Metal(storage))
+            }
+        }
+    }
+
+    pub(crate) fn storage_from_slice<D: WithDType>(&self, data: &[D]) -> Result<Storage> {
+        match self {
+            Device::Cpu => Ok(Storage::Cpu(data.to_cpu_storage())),
+            Device::Cuda(device) => {
+                let storage = device.storage_from_slice(data)?;
+                Ok(Storage::Cuda(storage))
+            }
+            Device::Metal(device) => {
+                let storage = device.storage_from_slice(data)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 
@@ -210,9 +341,14 @@ impl Device {
             Device::Cpu => Ok(Storage::Cpu(array.to_cpu_storage())),
             Device::Cuda(device) => {
                 let storage = array.to_cpu_storage();
-                let storage = device.storage_from_cpu_storage(&storage)?;
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = array.to_cpu_storage();
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 
@@ -221,9 +357,22 @@ impl Device {
             Device::Cpu => Ok(Storage::Cpu(S::to_cpu_storage_owned(data))),
             Device::Cuda(device) => {
                 let storage = S::to_cpu_storage_owned(data);
-                let storage = device.storage_from_cpu_storage(&storage)?;
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = S::to_cpu_storage_owned(data);
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
+                Ok(Storage::Metal(storage))
+            }
+        }
+    }
+
+    pub fn synchronize(&self) -> Result<()> {
+        match self {
+            Self::Cpu => Ok(()),
+            Self::Cuda(d) => d.synchronize(),
+            Self::Metal(d) => d.synchronize(),
         }
     }
 }

candle-core/src/display.rs

@@ -9,11 +9,17 @@ impl Tensor {
         &self,
         f: &mut std::fmt::Formatter,
     ) -> std::fmt::Result {
-        let prefix = match self.device() {
-            crate::Device::Cpu => "Cpu",
-            crate::Device::Cuda(_) => "Cuda",
+        let device_str = match self.device().location() {
+            crate::DeviceLocation::Cpu => "".to_owned(),
+            crate::DeviceLocation::Cuda { gpu_id } => {
+                format!(", cuda:{}", gpu_id)
+            }
+            crate::DeviceLocation::Metal { gpu_id } => {
+                format!(", metal:{}", gpu_id)
+            }
         };
-        write!(f, "{prefix}Tensor[")?;
+
+        write!(f, "Tensor[")?;
         match self.dims() {
             [] => {
                 if let Ok(v) = self.to_scalar::<T>() {
@@ -40,7 +46,7 @@ impl Tensor {
                 }
             }
         }
-        write!(f, "; {}]", self.dtype().as_str())
+        write!(f, "; {}{}]", self.dtype().as_str(), device_str)
     }
 }
 
@@ -49,6 +55,7 @@ impl std::fmt::Debug for Tensor {
         match self.dtype() {
             DType::U8 => self.fmt_dt::<u8>(f),
             DType::U32 => self.fmt_dt::<u32>(f),
+            DType::I64 => self.fmt_dt::<i64>(f),
             DType::BF16 => self.fmt_dt::<bf16>(f),
             DType::F16 => self.fmt_dt::<f16>(f),
             DType::F32 => self.fmt_dt::<f32>(f),
@@ -58,12 +65,13 @@ impl std::fmt::Debug for Tensor {
 }
 
 /// Options for Tensor pretty printing
+#[derive(Debug, Clone)]
 pub struct PrinterOptions {
-    precision: usize,
-    threshold: usize,
-    edge_items: usize,
-    line_width: usize,
-    sci_mode: Option<bool>,
+    pub precision: usize,
+    pub threshold: usize,
+    pub edge_items: usize,
+    pub line_width: usize,
+    pub sci_mode: Option<bool>,
 }
 
 static PRINT_OPTS: std::sync::Mutex<PrinterOptions> =
@@ -82,6 +90,10 @@ impl PrinterOptions {
     }
 }
 
+pub fn print_options() -> &'static std::sync::Mutex<PrinterOptions> {
+    &PRINT_OPTS
+}
+
 pub fn set_print_options(options: PrinterOptions) {
     *PRINT_OPTS.lock().unwrap() = options
 }
@@ -110,6 +122,26 @@ pub fn set_print_options_full() {
     }
 }
 
+pub fn set_line_width(line_width: usize) {
+    PRINT_OPTS.lock().unwrap().line_width = line_width
+}
+
+pub fn set_precision(precision: usize) {
+    PRINT_OPTS.lock().unwrap().precision = precision
+}
+
+pub fn set_edge_items(edge_items: usize) {
+    PRINT_OPTS.lock().unwrap().edge_items = edge_items
+}
+
+pub fn set_threshold(threshold: usize) {
+    PRINT_OPTS.lock().unwrap().threshold = threshold
+}
+
+pub fn set_sci_mode(sci_mode: Option<bool>) {
+    PRINT_OPTS.lock().unwrap().sci_mode = sci_mode
+}
+
 struct FmtSize {
     current_size: usize,
 }
@@ -431,6 +463,12 @@ impl std::fmt::Display for Tensor {
                 tf.fmt_tensor(self, 1, max_w, summarize, &po, f)?;
                 writeln!(f)?;
             }
+            DType::I64 => {
+                let tf: IntFormatter<i64> = IntFormatter::new();
+                let max_w = tf.max_width(&to_display);
+                tf.fmt_tensor(self, 1, max_w, summarize, &po, f)?;
+                writeln!(f)?;
+            }
             DType::BF16 => {
                 if let Ok(tf) = FloatFormatter::<bf16>::new(&to_display, &po) {
                     let max_w = tf.max_width(&to_display);
@@ -460,6 +498,23 @@ impl std::fmt::Display for Tensor {
                 }
             }
         };
-        write!(f, "Tensor[{:?}, {}]", self.dims(), self.dtype().as_str())
+
+        let device_str = match self.device().location() {
+            crate::DeviceLocation::Cpu => "".to_owned(),
+            crate::DeviceLocation::Cuda { gpu_id } => {
+                format!(", cuda:{}", gpu_id)
+            }
+            crate::DeviceLocation::Metal { gpu_id } => {
+                format!(", metal:{}", gpu_id)
+            }
+        };
+
+        write!(
+            f,
+            "Tensor[{:?}, {}{}]",
+            self.dims(),
+            self.dtype().as_str(),
+            device_str
+        )
     }
 }

candle-core/src/dtype.rs

@@ -1,18 +1,37 @@
+//! Types for elements that can be stored and manipulated using tensors.
+#![allow(clippy::redundant_closure_call)]
 use crate::backend::BackendStorage;
-use crate::{CpuStorage, Error, Result};
+use crate::{CpuStorage, CpuStorageRef, Error, Result};
 
+/// The different types of elements allowed in tensors.
 #[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
 pub enum DType {
+    // Unsigned 8 bits integer.
     U8,
+    // Unsigned 32 bits integer.
     U32,
+    // Signed 64 bits integer.
+    I64,
+    // Brain floating-point using half precision (16 bits).
     BF16,
+    // Floating-point using half precision (16 bits).
     F16,
+    // Floating-point using single precision (32 bits).
     F32,
+    // Floating-point using double precision (64 bits).
     F64,
 }
 
 #[derive(Debug, PartialEq, Eq)]
-pub struct DTypeParseError;
+pub struct DTypeParseError(String);
+
+impl std::fmt::Display for DTypeParseError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "cannot parse '{}' as a dtype", self.0)
+    }
+}
+
+impl std::error::Error for DTypeParseError {}
 
 impl std::str::FromStr for DType {
     type Err = DTypeParseError;
@@ -20,20 +39,23 @@ impl std::str::FromStr for DType {
         match s {
             "u8" => Ok(Self::U8),
             "u32" => Ok(Self::U32),
+            "i64" => Ok(Self::I64),
             "bf16" => Ok(Self::BF16),
             "f16" => Ok(Self::F16),
             "f32" => Ok(Self::F32),
             "f64" => Ok(Self::F64),
-            _ => Err(DTypeParseError),
+            _ => Err(DTypeParseError(s.to_string())),
         }
     }
 }
 
 impl DType {
+    /// String representation for dtypes.
     pub fn as_str(&self) -> &'static str {
         match self {
             Self::U8 => "u8",
             Self::U32 => "u32",
+            Self::I64 => "i64",
             Self::BF16 => "bf16",
             Self::F16 => "f16",
             Self::F32 => "f32",
@@ -41,16 +63,32 @@ impl DType {
         }
     }
 
+    /// The size used by each element in bytes, i.e. 1 for `U8`, 4 for `F32`.
     pub fn size_in_bytes(&self) -> usize {
         match self {
             Self::U8 => 1,
             Self::U32 => 4,
+            Self::I64 => 8,
             Self::BF16 => 2,
             Self::F16 => 2,
             Self::F32 => 4,
             Self::F64 => 8,
         }
     }
+
+    pub fn is_int(&self) -> bool {
+        match self {
+            Self::U8 | Self::U32 | Self::I64 => true,
+            Self::BF16 | Self::F16 | Self::F32 | Self::F64 => false,
+        }
+    }
+
+    pub fn is_float(&self) -> bool {
+        match self {
+            Self::U8 | Self::U32 | Self::I64 => false,
+            Self::BF16 | Self::F16 | Self::F32 | Self::F64 => true,
+        }
+    }
 }
 
 pub trait WithDType:
@@ -62,12 +100,14 @@ pub trait WithDType:
     + 'static
     + Send
     + Sync
+    + std::any::Any
     + crate::cpu::kernels::VecOps
 {
     const DTYPE: DType;
 
     fn from_f64(v: f64) -> Self;
     fn to_f64(self) -> f64;
+    fn cpu_storage_ref(data: &[Self]) -> CpuStorageRef<'_>;
     fn to_cpu_storage_owned(data: Vec<Self>) -> CpuStorage;
 
     fn to_cpu_storage(data: &[Self]) -> CpuStorage {
@@ -91,6 +131,10 @@ macro_rules! with_dtype {
                 $to_f64(self)
             }
 
+            fn cpu_storage_ref(data: &[Self]) -> CpuStorageRef<'_> {
+                CpuStorageRef::$dtype(data)
+            }
+
             fn to_cpu_storage_owned(data: Vec<Self>) -> CpuStorage {
                 CpuStorage::$dtype(data)
             }
@@ -125,6 +169,7 @@ use half::{bf16, f16};
 
 with_dtype!(u8, U8, |v: f64| v as u8, |v: u8| v as f64);
 with_dtype!(u32, U32, |v: f64| v as u32, |v: u32| v as f64);
+with_dtype!(i64, I64, |v: f64| v as i64, |v: i64| v as f64);
 with_dtype!(f16, F16, f16::from_f64, f16::to_f64);
 with_dtype!(bf16, BF16, bf16::from_f64, bf16::to_f64);
 with_dtype!(f32, F32, |v: f64| v as f32, |v: f32| v as f64);
@@ -135,6 +180,15 @@ pub trait IntDType: WithDType {
     fn as_usize(&self) -> usize;
 }
 
+impl IntDType for i64 {
+    fn is_true(&self) -> bool {
+        *self != 0
+    }
+    fn as_usize(&self) -> usize {
+        *self as usize
+    }
+}
+
 impl IntDType for u32 {
     fn is_true(&self) -> bool {
         *self != 0

candle-core/src/dummy_cuda_backend.rs

@@ -37,6 +37,10 @@ impl crate::backend::BackendStorage for CudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn powf(&self, _: &Layout, _: f64) -> Result<Self> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn elu(&self, _: &Layout, _: f64) -> Result<Self> {
         Err(Error::NotCompiledWithCudaSupport)
     }
@@ -75,6 +79,16 @@ impl crate::backend::BackendStorage for CudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn conv_transpose1d(
+        &self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: &crate::conv::ParamsConvTranspose1D,
+    ) -> Result<Self> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn conv2d(
         &self,
         _: &Layout,
@@ -85,6 +99,16 @@ impl crate::backend::BackendStorage for CudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn conv_transpose2d(
+        &self,
+        _l: &Layout,
+        _kernel: &Self,
+        _kernel_l: &Layout,
+        _params: &crate::conv::ParamsConvTranspose2D,
+    ) -> Result<Self> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn index_select(&self, _: &Self, _: &Layout, _: &Layout, _: usize) -> Result<Self> {
         Err(Error::NotCompiledWithCudaSupport)
     }
@@ -130,6 +154,19 @@ impl crate::backend::BackendStorage for CudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn copy2d(
+        &self,
+        _: &mut Self,
+        _: usize,
+        _: usize,
+        _: usize,
+        _: usize,
+        _: usize,
+        _: usize,
+    ) -> Result<()> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn avg_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self> {
         Err(Error::NotCompiledWithCudaSupport)
     }
@@ -138,6 +175,10 @@ impl crate::backend::BackendStorage for CudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn upsample_nearest1d(&self, _: &Layout, _: usize) -> Result<Self> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn upsample_nearest2d(&self, _: &Layout, _: usize, _: usize) -> Result<Self> {
         Err(Error::NotCompiledWithCudaSupport)
     }
@@ -149,6 +190,10 @@ impl crate::backend::BackendDevice for CudaDevice {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn set_seed(&self, _: u64) -> Result<()> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn location(&self) -> crate::DeviceLocation {
         fail!()
     }
@@ -165,10 +210,22 @@ impl crate::backend::BackendDevice for CudaDevice {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
@@ -176,4 +233,38 @@ impl crate::backend::BackendDevice for CudaDevice {
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
 }
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn gemm_reduced_precision_f16() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn set_gemm_reduced_precision_f16(_: bool) {}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn gemm_reduced_precision_bf16() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn set_gemm_reduced_precision_bf16(_: bool) {}
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn gemm_reduced_precision_f32() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn set_gemm_reduced_precision_f32(_b: bool) {}

candle-core/src/dummy_metal_backend.rs

@@ -0,0 +1,252 @@
+#![allow(dead_code)]
+use crate::op::{BinaryOpT, CmpOp, ReduceOp, UnaryOpT};
+use crate::{CpuStorage, DType, Error, Layout, Result, Shape};
+
+#[derive(Debug, Clone)]
+pub struct MetalDevice;
+
+#[derive(Debug)]
+pub struct MetalStorage;
+
+#[derive(thiserror::Error, Debug)]
+pub enum MetalError {
+    #[error("{0}")]
+    Message(String),
+}
+
+impl From<String> for MetalError {
+    fn from(e: String) -> Self {
+        MetalError::Message(e)
+    }
+}
+
+macro_rules! fail {
+    () => {
+        unimplemented!("metal support has not been enabled, add `metal` feature to enable.")
+    };
+}
+
+impl crate::backend::BackendStorage for MetalStorage {
+    type Device = MetalDevice;
+
+    fn try_clone(&self, _: &Layout) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn dtype(&self) -> DType {
+        fail!()
+    }
+
+    fn device(&self) -> &Self::Device {
+        fail!()
+    }
+
+    fn to_cpu_storage(&self) -> Result<CpuStorage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn affine(&self, _: &Layout, _: f64, _: f64) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn powf(&self, _: &Layout, _: f64) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn elu(&self, _: &Layout, _: f64) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn reduce_op(&self, _: ReduceOp, _: &Layout, _: &[usize]) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn cmp(&self, _: CmpOp, _: &Self, _: &Layout, _: &Layout) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn to_dtype(&self, _: &Layout, _: DType) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn unary_impl<B: UnaryOpT>(&self, _: &Layout) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn binary_impl<B: BinaryOpT>(&self, _: &Self, _: &Layout, _: &Layout) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn where_cond(&self, _: &Layout, _: &Self, _: &Layout, _: &Self, _: &Layout) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn conv1d(
+        &self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: &crate::conv::ParamsConv1D,
+    ) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn conv_transpose1d(
+        &self,
+        _l: &Layout,
+        _kernel: &Self,
+        _kernel_l: &Layout,
+        _params: &crate::conv::ParamsConvTranspose1D,
+    ) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn conv2d(
+        &self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: &crate::conv::ParamsConv2D,
+    ) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn conv_transpose2d(
+        &self,
+        _l: &Layout,
+        _kernel: &Self,
+        _kernel_l: &Layout,
+        _params: &crate::conv::ParamsConvTranspose2D,
+    ) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn index_select(&self, _: &Self, _: &Layout, _: &Layout, _: usize) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+    fn gather(&self, _: &Layout, _: &Self, _: &Layout, _: usize) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn scatter_add(
+        &self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: usize,
+    ) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn index_add(
+        &self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: usize,
+    ) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn matmul(
+        &self,
+        _: &Self,
+        _: (usize, usize, usize, usize),
+        _: &Layout,
+        _: &Layout,
+    ) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn copy_strided_src(&self, _: &mut Self, _: usize, _: &Layout) -> Result<()> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn copy2d(
+        &self,
+        _: &mut Self,
+        _: usize,
+        _: usize,
+        _: usize,
+        _: usize,
+        _: usize,
+        _: usize,
+    ) -> Result<()> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn avg_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn max_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn upsample_nearest1d(&self, _: &Layout, _: usize) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn upsample_nearest2d(&self, _: &Layout, _: usize, _: usize) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+}
+
+impl crate::backend::BackendDevice for MetalDevice {
+    type Storage = MetalStorage;
+    fn new(_: usize) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn set_seed(&self, _: u64) -> Result<()> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn location(&self) -> crate::DeviceLocation {
+        fail!()
+    }
+
+    fn same_device(&self, _: &Self) -> bool {
+        fail!()
+    }
+
+    fn zeros_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn ones_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
+}

candle-core/src/error.rs

@@ -1,4 +1,4 @@
-use crate::{DType, DeviceLocation, Layout, Shape};
+use crate::{DType, DeviceLocation, Layout, MetalError, Shape};
 
 #[derive(Debug, Clone)]
 pub struct MatMulUnexpectedStriding {
@@ -30,7 +30,7 @@ pub enum Error {
     UnsupportedDTypeForOp(DType, &'static str),
 
     // === Dimension Index Errors ===
-    #[error("{op}: dimension index {dim} out of range for {shape:?}")]
+    #[error("{op}: dimension index {dim} out of range for shape {shape:?}")]
     DimOutOfRange {
         shape: Shape,
         dim: i32,
@@ -142,6 +142,9 @@ pub enum Error {
     #[error("{op} expects at least one tensor")]
     OpRequiresAtLeastOneTensor { op: &'static str },
 
+    #[error("{op} expects at least two tensors")]
+    OpRequiresAtLeastTwoTensors { op: &'static str },
+
     #[error("backward is not supported for {op}")]
     BackwardNotSupported { op: &'static str },
 
@@ -149,6 +152,9 @@ pub enum Error {
     #[error("the candle crate has not been built with cuda support")]
     NotCompiledWithCudaSupport,
 
+    #[error("the candle crate has not been built with metal support")]
+    NotCompiledWithMetalSupport,
+
     #[error("cannot find tensor {path}")]
     CannotFindTensor { path: String },
 
@@ -156,6 +162,9 @@ pub enum Error {
     #[error(transparent)]
     Cuda(Box<dyn std::error::Error + Send + Sync>),
 
+    #[error("Metal error {0}")]
+    Metal(#[from] MetalError),
+
     #[error(transparent)]
     TryFromIntError(#[from] core::num::TryFromIntError),
 
@@ -207,7 +216,15 @@ pub type Result<T> = std::result::Result<T, Error>;
 
 impl Error {
     pub fn wrap(err: impl std::error::Error + Send + Sync + 'static) -> Self {
-        Self::Wrapped(Box::new(err))
+        Self::Wrapped(Box::new(err)).bt()
+    }
+
+    pub fn msg(err: impl std::error::Error) -> Self {
+        Self::Msg(err.to_string()).bt()
+    }
+
+    pub fn debug(err: impl std::fmt::Debug) -> Self {
+        Self::Msg(format!("{err:?}")).bt()
     }
 
     pub fn bt(self) -> Self {

candle-core/src/indexer.rs

@@ -46,19 +46,31 @@ impl Tensor {
                     current_dim += 1;
                     out
                 }
+                TensorIndexer::IndexSelect(indexes) => {
+                    if indexes.rank() != 1 {
+                        crate::bail!("multi-dimensional tensor indexing is not supported")
+                    }
+                    let out = x.index_select(&indexes.to_device(x.device())?, current_dim)?;
+                    current_dim += 1;
+                    out
+                }
+                TensorIndexer::Err(e) => crate::bail!("indexing error {e:?}"),
             };
         }
         Ok(x)
     }
 }
 
-#[derive(Debug, Clone)]
+#[derive(Debug)]
 /// Generic structure used to index a slice of the tensor
 pub enum TensorIndexer {
-    /// This selects the elemnts for which an index has some specific value.
+    /// This selects the elements for which an index has some specific value.
     Select(usize),
     /// This is a regular slice, purely indexing a chunk of the tensor
     Narrow(Bound<usize>, Bound<usize>),
+    /// Indexing via a 1d tensor
+    IndexSelect(Tensor),
+    Err(Error),
 }
 
 impl From<usize> for TensorIndexer {
@@ -67,36 +79,55 @@ impl From<usize> for TensorIndexer {
     }
 }
 
-macro_rules! impl_from_range {
-    ($range_type:ty) => {
-        impl From<$range_type> for TensorIndexer {
-            fn from(range: $range_type) -> Self {
-                use std::ops::Bound::*;
-
-                let start = match range.start_bound() {
-                    Included(idx) => Included(*idx),
-                    Excluded(idx) => Excluded(*idx),
-                    Unbounded => Unbounded,
-                };
-
-                let end = match range.end_bound() {
-                    Included(idx) => Included(*idx),
-                    Excluded(idx) => Excluded(*idx),
-                    Unbounded => Unbounded,
-                };
-
-                TensorIndexer::Narrow(start, end)
-            }
+impl From<&[u32]> for TensorIndexer {
+    fn from(index: &[u32]) -> Self {
+        match Tensor::new(index, &crate::Device::Cpu) {
+            Ok(tensor) => TensorIndexer::IndexSelect(tensor),
+            Err(e) => TensorIndexer::Err(e),
         }
-    };
+    }
 }
 
-impl_from_range!(Range<usize>);
-impl_from_range!(RangeFrom<usize>);
-impl_from_range!(RangeFull);
-impl_from_range!(RangeInclusive<usize>);
-impl_from_range!(RangeTo<usize>);
-impl_from_range!(RangeToInclusive<usize>);
+impl From<Vec<u32>> for TensorIndexer {
+    fn from(index: Vec<u32>) -> Self {
+        let len = index.len();
+        match Tensor::from_vec(index, len, &crate::Device::Cpu) {
+            Ok(tensor) => TensorIndexer::IndexSelect(tensor),
+            Err(e) => TensorIndexer::Err(e),
+        }
+    }
+}
+
+impl From<&Tensor> for TensorIndexer {
+    fn from(tensor: &Tensor) -> Self {
+        TensorIndexer::IndexSelect(tensor.clone())
+    }
+}
+
+trait RB: RangeBounds<usize> {}
+impl RB for Range<usize> {}
+impl RB for RangeFrom<usize> {}
+impl RB for RangeFull {}
+impl RB for RangeInclusive<usize> {}
+impl RB for RangeTo<usize> {}
+impl RB for RangeToInclusive<usize> {}
+
+impl<T: RB> From<T> for TensorIndexer {
+    fn from(range: T) -> Self {
+        use std::ops::Bound::*;
+        let start = match range.start_bound() {
+            Included(idx) => Included(*idx),
+            Excluded(idx) => Excluded(*idx),
+            Unbounded => Unbounded,
+        };
+        let end = match range.end_bound() {
+            Included(idx) => Included(*idx),
+            Excluded(idx) => Excluded(*idx),
+            Unbounded => Unbounded,
+        };
+        TensorIndexer::Narrow(start, end)
+    }
+}
 
 /// Trait used to implement multiple signatures for ease of use of the slicing
 /// of a tensor
@@ -110,28 +141,117 @@ impl<T> IndexOp<T> for Tensor
 where
     T: Into<TensorIndexer>,
 {
+    ///```rust
+    /// use candle_core::{Tensor, DType, Device, IndexOp};
+    /// let a = Tensor::new(&[
+    ///     [0., 1.],
+    ///     [2., 3.],
+    ///     [4., 5.]
+    /// ], &Device::Cpu)?;
+    ///
+    /// let b = a.i(0)?;
+    /// assert_eq!(b.shape().dims(), &[2]);
+    /// assert_eq!(b.to_vec1::<f64>()?, &[0., 1.]);
+    ///
+    /// let c = a.i(..2)?;
+    /// assert_eq!(c.shape().dims(), &[2, 2]);
+    /// assert_eq!(c.to_vec2::<f64>()?, &[
+    ///     [0., 1.],
+    ///     [2., 3.]
+    /// ]);
+    ///
+    /// let d = a.i(1..)?;
+    /// assert_eq!(d.shape().dims(), &[2, 2]);
+    /// assert_eq!(d.to_vec2::<f64>()?, &[
+    ///     [2., 3.],
+    ///     [4., 5.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     fn i(&self, index: T) -> Result<Tensor, Error> {
         self.index(&[index.into()])
     }
 }
 
+impl<A> IndexOp<(A,)> for Tensor
+where
+    A: Into<TensorIndexer>,
+{
+    ///```rust
+    /// use candle_core::{Tensor, DType, Device, IndexOp};
+    /// let a = Tensor::new(&[
+    ///     [0f32, 1.],
+    ///     [2.  , 3.],
+    ///     [4.  , 5.]
+    /// ], &Device::Cpu)?;
+    ///
+    /// let b = a.i((0,))?;
+    /// assert_eq!(b.shape().dims(), &[2]);
+    /// assert_eq!(b.to_vec1::<f32>()?, &[0., 1.]);
+    ///
+    /// let c = a.i((..2,))?;
+    /// assert_eq!(c.shape().dims(), &[2, 2]);
+    /// assert_eq!(c.to_vec2::<f32>()?, &[
+    ///     [0., 1.],
+    ///     [2., 3.]
+    /// ]);
+    ///
+    /// let d = a.i((1..,))?;
+    /// assert_eq!(d.shape().dims(), &[2, 2]);
+    /// assert_eq!(d.to_vec2::<f32>()?, &[
+    ///     [2., 3.],
+    ///     [4., 5.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    fn i(&self, (a,): (A,)) -> Result<Tensor, Error> {
+        self.index(&[a.into()])
+    }
+}
+#[allow(non_snake_case)]
+impl<A, B> IndexOp<(A, B)> for Tensor
+where
+    A: Into<TensorIndexer>,
+    B: Into<TensorIndexer>,
+{
+    ///```rust
+    /// use candle_core::{Tensor, DType, Device, IndexOp};
+    /// let a = Tensor::new(&[[0f32, 1., 2.], [3., 4., 5.], [6., 7., 8.]], &Device::Cpu)?;
+    ///
+    /// let b = a.i((1, 0))?;
+    /// assert_eq!(b.to_vec0::<f32>()?, 3.);
+    ///
+    /// let c = a.i((..2, 1))?;
+    /// assert_eq!(c.shape().dims(), &[2]);
+    /// assert_eq!(c.to_vec1::<f32>()?, &[1., 4.]);
+    ///
+    /// let d = a.i((2.., ..))?;
+    /// assert_eq!(c.shape().dims(), &[2]);
+    /// assert_eq!(c.to_vec1::<f32>()?, &[1., 4.]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    fn i(&self, (a, b): (A, B)) -> Result<Tensor, Error> {
+        self.index(&[a.into(), b.into()])
+    }
+}
+
 macro_rules! index_op_tuple {
-    ($($t:ident),+) => {
+    ($doc:tt, $($t:ident),+) => {
         #[allow(non_snake_case)]
         impl<$($t),*> IndexOp<($($t,)*)> for Tensor
         where
             $($t: Into<TensorIndexer>,)*
         {
+            #[doc=$doc]
             fn i(&self, ($($t,)*): ($($t,)*)) -> Result<Tensor, Error> {
                 self.index(&[$($t.into(),)*])
             }
         }
     };
 }
-index_op_tuple!(A);
-index_op_tuple!(A, B);
-index_op_tuple!(A, B, C);
-index_op_tuple!(A, B, C, D);
-index_op_tuple!(A, B, C, D, E);
-index_op_tuple!(A, B, C, D, E, F);
-index_op_tuple!(A, B, C, D, E, F, G);
+
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C);
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C, D);
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C, D, E);
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C, D, E, F);
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C, D, E, F, G);

candle-core/src/layout.rs

@@ -9,6 +9,14 @@ pub struct Layout {
 }
 
 impl Layout {
+    pub fn new(shape: Shape, stride: Vec<usize>, start_offset: usize) -> Self {
+        Self {
+            shape,
+            stride,
+            start_offset,
+        }
+    }
+
     pub fn contiguous_with_offset<S: Into<Shape>>(shape: S, start_offset: usize) -> Self {
         let shape = shape.into();
         let stride = shape.stride_contiguous();
@@ -62,7 +70,7 @@ impl Layout {
         self.shape.is_fortran_contiguous(&self.stride)
     }
 
-    pub(crate) fn narrow(&self, dim: usize, start: usize, len: usize) -> Result<Self> {
+    pub fn narrow(&self, dim: usize, start: usize, len: usize) -> Result<Self> {
         let dims = self.shape().dims();
         if dim >= dims.len() {
             Err(Error::DimOutOfRange {
@@ -91,7 +99,7 @@ impl Layout {
         })
     }
 
-    pub(crate) fn transpose(&self, dim1: usize, dim2: usize) -> Result<Self> {
+    pub fn transpose(&self, dim1: usize, dim2: usize) -> Result<Self> {
         let rank = self.shape.rank();
         if rank <= dim1 || rank <= dim2 {
             Err(Error::UnexpectedNumberOfDims {
@@ -112,6 +120,31 @@ impl Layout {
         })
     }
 
+    pub fn permute(&self, idxs: &[usize]) -> Result<Self> {
+        let is_permutation =
+            idxs.len() == self.shape.rank() && (0..idxs.len()).all(|i| idxs.contains(&i));
+        if !is_permutation {
+            crate::bail!(
+                "dimension mismatch in permute, tensor {:?}, dims: {:?}",
+                self.dims(),
+                idxs
+            )
+        }
+        let stride = self.stride();
+        let dims = self.shape().dims();
+        let mut perm_stride = stride.to_vec();
+        let mut perm_dims = dims.to_vec();
+        for (i, &idx) in idxs.iter().enumerate() {
+            perm_stride[i] = stride[idx];
+            perm_dims[i] = dims[idx];
+        }
+        Ok(Self {
+            shape: Shape::from(perm_dims),
+            stride: perm_stride,
+            start_offset: self.start_offset,
+        })
+    }
+
     pub fn broadcast_as<S: Into<Shape>>(&self, shape: S) -> Result<Self> {
         let shape = shape.into();
         if shape.rank() < self.shape().rank() {

candle-core/src/lib.rs

@@ -14,7 +14,7 @@
 //!
 //! ## Features
 //!
-//! - Simple syntax (looks and like PyTorch)
+//! - Simple syntax (looks and feels like PyTorch)
 //! - CPU and Cuda backends (and M1 support)
 //! - Enable serverless (CPU) small and fast deployments
 //! - Model training
@@ -37,52 +37,123 @@
 mod accelerate;
 pub mod backend;
 pub mod backprop;
-mod conv;
+pub mod conv;
 mod convert;
 pub mod cpu;
 pub mod cpu_backend;
 #[cfg(feature = "cuda")]
 pub mod cuda_backend;
-#[cfg(feature = "cudnn")]
-pub mod cudnn;
+mod custom_op;
 mod device;
 pub mod display;
 mod dtype;
-mod dummy_cuda_backend;
+pub mod dummy_cuda_backend;
+mod dummy_metal_backend;
 pub mod error;
 mod indexer;
 pub mod layout;
+#[cfg(feature = "metal")]
+pub mod metal_backend;
 #[cfg(feature = "mkl")]
 mod mkl;
 pub mod npy;
-mod op;
+pub mod op;
+pub mod pickle;
 pub mod quantized;
 pub mod safetensors;
+pub mod scalar;
 pub mod shape;
+mod sort;
 mod storage;
+pub mod streaming;
 mod strided_index;
 mod tensor;
+mod tensor_cat;
+pub mod test_utils;
 pub mod utils;
 mod variable;
 
-pub use cpu_backend::CpuStorage;
-pub use device::{Device, DeviceLocation};
-pub use dtype::{DType, FloatDType, IntDType, WithDType};
+#[cfg(feature = "cudnn")]
+pub use cuda_backend::cudnn;
+
+pub use cpu_backend::{CpuStorage, CpuStorageRef};
+pub use custom_op::{CustomOp1, CustomOp2, CustomOp3, InplaceOp1, InplaceOp2, InplaceOp3};
+pub use device::{Device, DeviceLocation, NdArray};
+pub use dtype::{DType, DTypeParseError, FloatDType, IntDType, WithDType};
 pub use error::{Error, Result};
-pub use indexer::IndexOp;
+pub use indexer::{IndexOp, TensorIndexer};
 pub use layout::Layout;
-pub use op::{CustomOp1, CustomOp2, CustomOp3};
 pub use shape::{Shape, D};
 pub use storage::Storage;
+pub use streaming::{StreamTensor, StreamingBinOp, StreamingModule};
 pub use strided_index::{StridedBlocks, StridedIndex};
 pub use tensor::{Tensor, TensorId};
 pub use variable::Var;
 
 #[cfg(feature = "cuda")]
-pub use cuda_backend::{CudaDevice, CudaStorage};
+pub use cuda_backend as cuda;
 
 #[cfg(not(feature = "cuda"))]
-pub use dummy_cuda_backend::{CudaDevice, CudaStorage};
+pub use dummy_cuda_backend as cuda;
+
+pub use cuda::{CudaDevice, CudaStorage};
+
+#[cfg(feature = "metal")]
+pub use metal_backend::{MetalDevice, MetalError, MetalStorage};
+
+#[cfg(not(feature = "metal"))]
+pub use dummy_metal_backend::{MetalDevice, MetalError, MetalStorage};
 
 #[cfg(feature = "mkl")]
 extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+pub trait ToUsize2 {
+    fn to_usize2(self) -> (usize, usize);
+}
+
+impl ToUsize2 for usize {
+    fn to_usize2(self) -> (usize, usize) {
+        (self, self)
+    }
+}
+
+impl ToUsize2 for (usize, usize) {
+    fn to_usize2(self) -> (usize, usize) {
+        self
+    }
+}
+
+// A simple trait defining a module with forward method using a single argument.
+pub trait Module {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor>;
+}
+
+impl<T: Fn(&Tensor) -> Result<Tensor>> Module for T {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        self(xs)
+    }
+}
+
+impl<M: Module> Module for Option<&M> {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        match self {
+            None => Ok(xs.clone()),
+            Some(m) => m.forward(xs),
+        }
+    }
+}
+
+// A trait defining a module with forward method using a single tensor argument and a flag to
+// separate the training and evaluation behaviors.
+pub trait ModuleT {
+    fn forward_t(&self, xs: &Tensor, train: bool) -> Result<Tensor>;
+}
+
+impl<M: Module> ModuleT for M {
+    fn forward_t(&self, xs: &Tensor, _train: bool) -> Result<Tensor> {
+        self.forward(xs)
+    }
+}

candle-core/src/metal_backend/device.rs

@@ -0,0 +1,324 @@
+use crate::{DType, Result};
+use candle_metal_kernels::Kernels;
+use metal::{Buffer, CommandBuffer, CommandQueue, MTLResourceOptions, NSUInteger};
+use std::collections::HashMap;
+use std::ffi::c_void;
+use std::path::Path;
+use std::sync::{Arc, Mutex, RwLock};
+
+use super::MetalError;
+
+/// Unique identifier for cuda devices.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
+pub struct DeviceId(usize);
+
+impl DeviceId {
+    pub(crate) fn new() -> Self {
+        // https://users.rust-lang.org/t/idiomatic-rust-way-to-generate-unique-id/33805
+        use std::sync::atomic;
+        static COUNTER: atomic::AtomicUsize = atomic::AtomicUsize::new(1);
+        Self(COUNTER.fetch_add(1, atomic::Ordering::Relaxed))
+    }
+}
+
+type BufferMap = HashMap<(NSUInteger, MTLResourceOptions), Vec<Arc<Buffer>>>;
+pub(crate) struct Commands {
+    /// Single command queue for the entire device.
+    command_queue: CommandQueue,
+    /// One command buffer at a time.
+    /// The scheduler works by allowing multiple
+    /// [ComputeCommandEncoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc)
+    /// on a single command buffer. Using a single command buffer would be fastest on the GPU but
+    /// prevents overlapping of CPU and GPU commands (because command buffer needs to be committed
+    /// to start to work).
+    /// Despite what the documentation says, command buffers are NOT ordered. They are ordered
+    /// for their START time, but there's no guarantee that command buffer1 will finish before
+    /// command buffer2 starts (or there are metal bugs there)
+    command_buffer: CommandBuffer,
+    /// Keeps track of the current amount of compute command encoders on the current
+    /// command buffer
+    /// Arc, RwLock because of the interior mutability.
+    command_buffer_index: usize,
+    /// The maximum amount of [compute command encoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc) per [command buffer](https://developer.apple.com/documentation/metal/mtlcommandbuffer?language=objc)
+    compute_per_buffer: usize,
+}
+
+impl Commands {
+    pub(crate) fn new(command_queue: CommandQueue) -> Result<Self> {
+        let command_buffer = command_queue.new_command_buffer().to_owned();
+        command_buffer.enqueue();
+        let compute_per_buffer = match std::env::var("CANDLE_METAL_COMPUTE_PER_BUFFER") {
+            Ok(val) => val.parse()?,
+            _ => 50,
+        };
+        Ok(Self {
+            command_queue,
+            command_buffer,
+            command_buffer_index: 0,
+            compute_per_buffer,
+        })
+    }
+
+    pub fn command_buffer(&mut self) -> Result<(bool, CommandBuffer)> {
+        let mut command_buffer = self.command_buffer.to_owned();
+        let mut flushed = false;
+        if self.command_buffer_index > self.compute_per_buffer {
+            self.command_buffer.commit();
+            command_buffer = self.command_queue.new_command_buffer().to_owned();
+            self.command_buffer = command_buffer.clone();
+            self.command_buffer_index = 0;
+            flushed = true;
+        }
+        self.command_buffer_index += 1;
+        Ok((flushed, command_buffer))
+    }
+
+    pub fn wait_until_completed(&mut self) -> Result<()> {
+        match self.command_buffer.status() {
+            metal::MTLCommandBufferStatus::Committed
+            | metal::MTLCommandBufferStatus::Scheduled
+            | metal::MTLCommandBufferStatus::Completed => {
+                panic!("Already committed");
+            }
+            _ => {}
+        }
+        self.command_buffer.commit();
+        self.command_buffer.wait_until_completed();
+        self.command_buffer = self.command_queue.new_command_buffer().to_owned();
+
+        Ok(())
+    }
+}
+
+#[derive(Clone)]
+pub struct MetalDevice {
+    /// Unique identifier, the registryID is not sufficient as it identifies the GPU rather than
+    /// the device itself.
+    pub(crate) id: DeviceId,
+
+    /// Raw metal device: <https://developer.apple.com/documentation/metal/mtldevice?language=objc>
+    pub(crate) device: metal::Device,
+
+    pub(crate) commands: Arc<RwLock<Commands>>,
+
+    /// Simple allocator struct.
+    /// The buffers are stored in size buckets since ML tends to use similar shapes over and over.
+    /// We store the buffers in [`Arc`] because it's much faster than Obj-c internal ref counting
+    /// (could be linked to FFI communication overhead).
+    ///
+    /// Whenever a buffer has a strong_count==1, we can reuse it, it means it was dropped in the
+    /// graph calculation, and only we the allocator kept a reference to it, therefore it's free
+    /// to be reused. However, in order for this to work, we need to guarantee the order of
+    /// operation, so that this buffer is not being used by another kernel at the same time.
+    /// Arc is the CPU reference count, it doesn't mean anything on the GPU side of things.
+    ///
+    /// Whenever we actually allocate a new buffer, we make a full sweep to clean up unused buffers
+    /// (strong_count = 1).
+    pub(crate) buffers: Arc<RwLock<BufferMap>>,
+
+    /// Simple keeper struct to keep track of the already compiled kernels so we can reuse them.
+    /// Heavily used by [`candle_metal_kernels`]
+    pub(crate) kernels: Arc<Kernels>,
+    /// Seed for random number generation.
+    pub(crate) seed: Arc<Mutex<Buffer>>,
+    /// Whether to use the MLX matmul kernels instead of the MFA ones.
+    pub(crate) use_mlx_mm: bool,
+}
+
+impl std::fmt::Debug for MetalDevice {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "MetalDevice({:?})", self.id)
+    }
+}
+
+impl std::ops::Deref for MetalDevice {
+    type Target = metal::DeviceRef;
+
+    fn deref(&self) -> &Self::Target {
+        &self.device
+    }
+}
+
+impl MetalDevice {
+    pub fn set_use_mlx_mm(&mut self, use_mlx_mm: bool) {
+        self.use_mlx_mm = use_mlx_mm
+    }
+
+    pub fn id(&self) -> DeviceId {
+        self.id
+    }
+
+    pub fn metal_device(&self) -> &metal::Device {
+        &self.device
+    }
+
+    fn drop_unused_buffers(&self) -> Result<()> {
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+        for subbuffers in buffers.values_mut() {
+            let newbuffers = subbuffers
+                .iter()
+                .filter(|s| Arc::strong_count(*s) > 1)
+                .map(Arc::clone)
+                .collect();
+            *subbuffers = newbuffers;
+        }
+        Ok(())
+    }
+
+    pub fn command_buffer(&self) -> Result<CommandBuffer> {
+        let mut commands = self.commands.write().map_err(MetalError::from)?;
+        let (flushed, command_buffer) = commands.command_buffer()?;
+        if flushed {
+            self.drop_unused_buffers()?
+        }
+        Ok(command_buffer)
+    }
+
+    pub fn wait_until_completed(&self) -> Result<()> {
+        let mut commands = self.commands.write().map_err(MetalError::from)?;
+        commands.wait_until_completed()
+    }
+
+    pub fn kernels(&self) -> &Kernels {
+        &self.kernels
+    }
+
+    pub fn device(&self) -> &metal::Device {
+        &self.device
+    }
+
+    /// Creates a new buffer (not necessarily zeroed).
+    /// The buffer is [MTLPrivate](https://developer.apple.com/documentation/metal/mtlstoragemode)
+    /// This means the buffer data cannot be read on the CPU directly.
+    ///
+    /// [`name`] is only used to keep track of the resource origin in case of bugs
+    pub fn new_buffer(
+        &self,
+        element_count: usize,
+        dtype: DType,
+        name: &str,
+    ) -> Result<Arc<Buffer>> {
+        let size = (element_count * dtype.size_in_bytes()) as NSUInteger;
+        self.allocate_buffer(size, MTLResourceOptions::StorageModePrivate, name)
+    }
+
+    /// Creates a new buffer (not necessarily zeroed).
+    /// The buffer is [MTLManaged](https://developer.apple.com/documentation/metal/mtlstoragemode)
+    /// This means the buffer can be read on the CPU but will require manual
+    /// synchronization when the CPU memory is modified
+    /// Used as a bridge to gather data back from the GPU
+    pub fn new_buffer_managed(&self, size: NSUInteger) -> Result<Arc<Buffer>> {
+        self.allocate_buffer(size, MTLResourceOptions::StorageModeManaged, "managed")
+    }
+
+    /// Creates a new buffer from data.
+    /// The buffer is [MTLManaged](https://developer.apple.com/documentation/metal/mtlstoragemode)
+    ///
+    /// Does not require synchronization, as [newBufferWithBytes](https://developer.apple.com/documentation/metal/mtldevice/1433429-newbufferwithbytes)
+    /// allocates the buffer and copies over the existing data before returning the MTLBuffer.
+    pub fn new_buffer_with_data<T>(&self, data: &[T]) -> Result<Arc<Buffer>> {
+        let size = core::mem::size_of_val(data) as NSUInteger;
+        let new_buffer = self.device.new_buffer_with_data(
+            data.as_ptr() as *const c_void,
+            size,
+            MTLResourceOptions::StorageModeManaged,
+        );
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+
+        let subbuffers = buffers
+            .entry((size, MTLResourceOptions::StorageModeManaged))
+            .or_insert(vec![]);
+
+        let new_buffer = Arc::new(new_buffer);
+        subbuffers.push(new_buffer.clone());
+        Ok(new_buffer)
+    }
+
+    pub fn allocate_zeros(&self, size_in_bytes: usize) -> Result<Arc<Buffer>> {
+        let buffer = self.allocate_buffer(
+            size_in_bytes as NSUInteger,
+            MTLResourceOptions::StorageModePrivate,
+            "allocate_zeros",
+        )?;
+        let command_buffer = self.command_buffer()?;
+        command_buffer.set_label("zeros");
+        let blit = command_buffer.new_blit_command_encoder();
+        blit.fill_buffer(
+            &buffer,
+            metal::NSRange {
+                location: 0,
+                length: buffer.length(),
+            },
+            0,
+        );
+        blit.end_encoding();
+        Ok(buffer)
+    }
+
+    /// The critical allocator algorithm
+    fn allocate_buffer(
+        &self,
+        size: NSUInteger,
+        option: MTLResourceOptions,
+        _name: &str,
+    ) -> Result<Arc<Buffer>> {
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+        if let Some(b) = find_available_buffer(size, option, &buffers) {
+            // Cloning also ensures we increment the strong count
+            return Ok(b.clone());
+        }
+
+        let size = buf_size(size);
+        let subbuffers = buffers.entry((size, option)).or_insert(vec![]);
+
+        let new_buffer = self.device.new_buffer(size as NSUInteger, option);
+        let new_buffer = Arc::new(new_buffer);
+        subbuffers.push(new_buffer.clone());
+
+        Ok(new_buffer)
+    }
+
+    /// Create a metal GPU capture trace on [`path`].
+    pub fn capture<P: AsRef<Path>>(&self, path: P) -> Result<()> {
+        let capture = metal::CaptureManager::shared();
+        let descriptor = metal::CaptureDescriptor::new();
+        descriptor.set_destination(metal::MTLCaptureDestination::GpuTraceDocument);
+        descriptor.set_capture_device(self);
+        // The [set_output_url] call requires an absolute path so we convert it if needed.
+        if path.as_ref().is_absolute() {
+            descriptor.set_output_url(path);
+        } else {
+            let path = std::env::current_dir()?.join(path);
+            descriptor.set_output_url(path);
+        }
+
+        capture
+            .start_capture(&descriptor)
+            .map_err(MetalError::from)?;
+        Ok(())
+    }
+}
+
+fn buf_size(size: NSUInteger) -> NSUInteger {
+    size.saturating_sub(1).next_power_of_two() as NSUInteger
+}
+
+fn find_available_buffer(
+    size: NSUInteger,
+    option: MTLResourceOptions,
+    buffers: &BufferMap,
+) -> Option<Arc<Buffer>> {
+    let mut best_buffer: Option<&Arc<Buffer>> = None;
+    let mut best_buffer_size: NSUInteger = NSUInteger::MAX;
+    for ((buffer_size, buffer_option), subbuffers) in buffers.iter() {
+        if buffer_size >= &size && buffer_size < &best_buffer_size && buffer_option == &option {
+            for sub in subbuffers {
+                if Arc::strong_count(sub) == 1 {
+                    best_buffer = Some(sub);
+                    best_buffer_size = *buffer_size;
+                }
+            }
+        }
+    }
+    best_buffer.cloned()
+}

candle-core/src/mkl.rs

@@ -25,6 +25,10 @@ mod ffi {
         pub fn vdMul(n: c_int, a: *const c_double, b: *const c_double, y: *mut c_double);
         pub fn vsDiv(n: c_int, a: *const c_float, b: *const c_float, y: *mut c_float);
         pub fn vdDiv(n: c_int, a: *const c_double, b: *const c_double, y: *mut c_double);
+        pub fn vsFmax(n: c_int, a: *const c_float, b: *const c_float, y: *mut c_float);
+        pub fn vdFmax(n: c_int, a: *const c_double, b: *const c_double, y: *mut c_double);
+        pub fn vsFmin(n: c_int, a: *const c_float, b: *const c_float, y: *mut c_float);
+        pub fn vdFmin(n: c_int, a: *const c_double, b: *const c_double, y: *mut c_double);
 
         pub fn sgemm_(
             transa: *const c_char,
@@ -297,7 +301,7 @@ pub fn vd_sqr(a: &[f64], y: &mut [f64]) {
 }
 
 #[inline]
-fn vs_tanh(a: &[f32], y: &mut [f32]) {
+pub fn vs_tanh(a: &[f32], y: &mut [f32]) {
     let a_len = a.len();
     let y_len = y.len();
     if a_len != y_len {
@@ -307,7 +311,7 @@ fn vs_tanh(a: &[f32], y: &mut [f32]) {
 }
 
 #[inline]
-fn vd_tanh(a: &[f64], y: &mut [f64]) {
+pub fn vd_tanh(a: &[f64], y: &mut [f64]) {
     let a_len = a.len();
     let y_len = y.len();
     if a_len != y_len {
@@ -329,6 +333,16 @@ pub fn vd_tanh_inplace(y: &mut [f64]) {
     unsafe { ffi::vdTanh(y.len() as i32, y.as_ptr(), y.as_mut_ptr()) }
 }
 
+#[inline]
+pub fn vs_exp_inplace(y: &mut [f32]) {
+    unsafe { ffi::vsExp(y.len() as i32, y.as_ptr(), y.as_mut_ptr()) }
+}
+
+#[inline]
+pub fn vd_exp_inplace(y: &mut [f64]) {
+    unsafe { ffi::vdExp(y.len() as i32, y.as_ptr(), y.as_mut_ptr()) }
+}
+
 #[inline]
 pub fn vs_gelu(vs: &[f32], ys: &mut [f32]) {
     for (&v, y) in vs.iter().zip(ys.iter_mut()) {
@@ -351,6 +365,28 @@ pub fn vd_gelu(vs: &[f64], ys: &mut [f64]) {
     }
 }
 
+#[inline]
+pub fn vs_silu(vs: &[f32], ys: &mut [f32]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vs_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
+#[inline]
+pub fn vd_silu(vs: &[f64], ys: &mut [f64]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vd_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
 macro_rules! binary_op {
     ($fn_name:ident, $ty:ty, $mkl_name:ident) => {
         #[inline]
@@ -376,3 +412,7 @@ binary_op!(vs_mul, f32, vsMul);
 binary_op!(vd_mul, f64, vdMul);
 binary_op!(vs_div, f32, vsDiv);
 binary_op!(vd_div, f64, vdDiv);
+binary_op!(vs_max, f32, vsFmax);
+binary_op!(vd_max, f64, vdFmax);
+binary_op!(vs_min, f32, vsFmin);
+binary_op!(vd_min, f64, vdFmin);

candle-core/src/npy.rs

@@ -85,6 +85,7 @@ impl Header {
             DType::F16 => "f2",
             DType::F32 => "f4",
             DType::F64 => "f8",
+            DType::I64 => "i8",
             DType::U32 => "u4",
             DType::U8 => "u1",
         };
@@ -160,7 +161,7 @@ impl Header {
                     "f" | "f4" => DType::F32,
                     "d" | "f8" => DType::F64,
                     // "i" | "i4" => DType::S32,
-                    // "q" | "i8" => DType::S64,
+                    "q" | "i8" => DType::I64,
                     // "h" | "i2" => DType::S16,
                     // "b" | "i1" => DType::S8,
                     "B" | "u1" => DType::U8,
@@ -196,7 +197,11 @@ impl Header {
 
 impl Tensor {
     // TODO: Add the possibility to read directly to a device?
-    fn from_reader<R: std::io::Read>(shape: Shape, dtype: DType, reader: &mut R) -> Result<Self> {
+    pub(crate) fn from_reader<R: std::io::Read>(
+        shape: Shape,
+        dtype: DType,
+        reader: &mut R,
+    ) -> Result<Self> {
         let elem_count = shape.elem_count();
         match dtype {
             DType::BF16 => {
@@ -229,6 +234,11 @@ impl Tensor {
                 reader.read_u32_into::<LittleEndian>(&mut data_t)?;
                 Tensor::from_vec(data_t, shape, &Device::Cpu)
             }
+            DType::I64 => {
+                let mut data_t = vec![0i64; elem_count];
+                reader.read_i64_into::<LittleEndian>(&mut data_t)?;
+                Tensor::from_vec(data_t, shape, &Device::Cpu)
+            }
         }
     }
 
@@ -240,8 +250,6 @@ impl Tensor {
         if header.fortran_order {
             return Err(Error::Npy("fortran order not supported".to_string()));
         }
-        let mut data: Vec<u8> = vec![];
-        reader.read_to_end(&mut data)?;
         Self::from_reader(header.shape(), header.descr, &mut reader)
     }
 
@@ -322,7 +330,7 @@ impl Tensor {
         path: P,
     ) -> Result<()> {
         let mut zip = zip::ZipWriter::new(File::create(path.as_ref())?);
-        let options =
+        let options: zip::write::FileOptions<()> =
             zip::write::FileOptions::default().compression_method(zip::CompressionMethod::Stored);
 
         for (name, tensor) in ts.iter() {
@@ -361,6 +369,25 @@ impl NpzTensors {
         })
     }
 
+    pub fn names(&self) -> Vec<&String> {
+        self.index_per_name.keys().collect()
+    }
+
+    /// This only returns the shape and dtype for a named tensor. Compared to `get`, this avoids
+    /// reading the whole tensor data.
+    pub fn get_shape_and_dtype(&self, name: &str) -> Result<(Shape, DType)> {
+        let index = match self.index_per_name.get(name) {
+            None => crate::bail!("cannot find tensor {name}"),
+            Some(index) => *index,
+        };
+        let zip_reader = BufReader::new(File::open(&self.path)?);
+        let mut zip = zip::ZipArchive::new(zip_reader)?;
+        let mut reader = zip.by_index(index)?;
+        let header = read_header(&mut reader)?;
+        let header = Header::parse(&header)?;
+        Ok((header.shape(), header.descr))
+    }
+
     pub fn get(&self, name: &str) -> Result<Option<Tensor>> {
         let index = match self.index_per_name.get(name) {
             None => return Ok(None),

candle-core/src/op.rs

@@ -1,4 +1,5 @@
-use crate::{CpuStorage, CudaStorage, Layout, Result, Shape, Tensor};
+#![allow(clippy::redundant_closure_call)]
+use crate::Tensor;
 use half::{bf16, f16};
 use num_traits::float::Float;
 
@@ -40,6 +41,8 @@ pub enum BinaryOp {
     Mul,
     Sub,
     Div,
+    Maximum,
+    Minimum,
 }
 
 // Unary ops with no argument
@@ -55,7 +58,15 @@ pub enum UnaryOp {
     Sqr,
     Sqrt,
     Gelu,
+    GeluErf,
+    Erf,
     Relu,
+    Silu,
+    Tanh,
+    Floor,
+    Ceil,
+    Round,
+    Sign,
 }
 
 #[derive(Clone)]
@@ -78,6 +89,17 @@ pub enum Op {
         kernel: Tensor,
         padding: usize,
         stride: usize,
+        dilation: usize,
+    },
+
+    #[allow(dead_code)]
+    ConvTranspose1D {
+        arg: Tensor,
+        kernel: Tensor,
+        padding: usize,
+        output_padding: usize,
+        stride: usize,
+        dilation: usize,
     },
 
     #[allow(dead_code)]
@@ -86,6 +108,17 @@ pub enum Op {
         kernel: Tensor,
         padding: usize,
         stride: usize,
+        dilation: usize,
+    },
+
+    #[allow(dead_code)]
+    ConvTranspose2D {
+        arg: Tensor,
+        kernel: Tensor,
+        padding: usize,
+        output_padding: usize,
+        stride: usize,
+        dilation: usize,
     },
 
     AvgPool2D {
@@ -100,7 +133,15 @@ pub enum Op {
         stride: (usize, usize),
     },
 
-    UpsampleNearest2D(Tensor),
+    UpsampleNearest1D {
+        arg: Tensor,
+        target_size: usize,
+    },
+    UpsampleNearest2D {
+        arg: Tensor,
+        target_h: usize,
+        target_w: usize,
+    },
 
     Cat(Vec<Tensor>, usize),
 
@@ -114,119 +155,28 @@ pub enum Op {
     Copy(Tensor),
     Broadcast(Tensor),
     Narrow(Tensor, usize, usize, usize),
+    SliceScatter0(Tensor, Tensor, usize),
     Reshape(Tensor),
     ToDevice(Tensor),
     Transpose(Tensor, usize, usize),
+    Permute(Tensor, Vec<usize>),
     Elu(Tensor, f64),
-    CustomOp1(Tensor, std::sync::Arc<Box<dyn CustomOp1>>),
-    CustomOp2(Tensor, Tensor, std::sync::Arc<Box<dyn CustomOp2>>),
-    CustomOp3(Tensor, Tensor, Tensor, std::sync::Arc<Box<dyn CustomOp3>>),
-}
-
-/// Unary ops that can be defined in user-land.
-pub trait CustomOp1: Send + Sync {
-    // Box<dyn> does not support const yet, so use a function to get the name.
-    fn name(&self) -> &'static str;
-
-    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cpu_fwd(&self, storage: &CpuStorage, layout: &Layout) -> Result<(CpuStorage, Shape)>;
-
-    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cuda_fwd(&self, _storage: &CudaStorage, _layout: &Layout) -> Result<(CudaStorage, Shape)> {
-        Err(crate::Error::Cuda(
-            format!("no cuda implementation for {}", self.name()).into(),
-        ))
-    }
-
-    /// This function takes as argument the argument `arg` used in the forward pass, the result
-    /// produced by the forward operation `res` and the gradient of the result `grad_res`.
-    /// The function should return the gradient of the argument.
-    fn bwd(&self, _arg: &Tensor, _res: &Tensor, _grad_res: &Tensor) -> Result<Option<Tensor>> {
-        Err(crate::Error::BackwardNotSupported { op: self.name() })
-    }
-}
-
-pub trait CustomOp2: Send + Sync {
-    fn name(&self) -> &'static str;
-
-    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cpu_fwd(
-        &self,
-        s1: &CpuStorage,
-        l1: &Layout,
-        s2: &CpuStorage,
-        l2: &Layout,
-    ) -> Result<(CpuStorage, Shape)>;
-
-    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cuda_fwd(
-        &self,
-        _: &CudaStorage,
-        _: &Layout,
-        _: &CudaStorage,
-        _: &Layout,
-    ) -> Result<(CudaStorage, Shape)> {
-        Err(crate::Error::Cuda(
-            format!("no cuda implementation for {}", self.name()).into(),
-        ))
-    }
-
-    fn bwd(
-        &self,
-        _arg1: &Tensor,
-        _arg2: &Tensor,
-        _res: &Tensor,
-        _grad_res: &Tensor,
-    ) -> Result<(Option<Tensor>, Option<Tensor>)> {
-        Err(crate::Error::BackwardNotSupported { op: self.name() })
-    }
-}
-
-pub trait CustomOp3: Send + Sync {
-    fn name(&self) -> &'static str;
-
-    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cpu_fwd(
-        &self,
-        s1: &CpuStorage,
-        l1: &Layout,
-        s2: &CpuStorage,
-        l2: &Layout,
-        s3: &CpuStorage,
-        l3: &Layout,
-    ) -> Result<(CpuStorage, Shape)>;
-
-    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cuda_fwd(
-        &self,
-        _: &CudaStorage,
-        _: &Layout,
-        _: &CudaStorage,
-        _: &Layout,
-        _: &CudaStorage,
-        _: &Layout,
-    ) -> Result<(CudaStorage, Shape)> {
-        Err(crate::Error::Cuda(
-            format!("no cuda implementation for {}", self.name()).into(),
-        ))
-    }
-
-    fn bwd(
-        &self,
-        _arg1: &Tensor,
-        _arg2: &Tensor,
-        _arg3: &Tensor,
-        _res: &Tensor,
-        _grad_res: &Tensor,
-    ) -> Result<(Option<Tensor>, Option<Tensor>, Option<Tensor>)> {
-        Err(crate::Error::BackwardNotSupported { op: self.name() })
-    }
+    Powf(Tensor, f64),
+    CustomOp1(
+        Tensor,
+        std::sync::Arc<Box<dyn crate::CustomOp1 + Send + Sync>>,
+    ),
+    CustomOp2(
+        Tensor,
+        Tensor,
+        std::sync::Arc<Box<dyn crate::CustomOp2 + Send + Sync>>,
+    ),
+    CustomOp3(
+        Tensor,
+        Tensor,
+        Tensor,
+        std::sync::Arc<Box<dyn crate::CustomOp3 + Send + Sync>>,
+    ),
 }
 
 pub trait UnaryOpT {
@@ -239,6 +189,7 @@ pub trait UnaryOpT {
     fn f64(v1: f64) -> f64;
     fn u8(v1: u8) -> u8;
     fn u32(v1: u32) -> u32;
+    fn i64(v1: i64) -> i64;
 
     // There is no very good way to represent optional function in traits so we go for an explicit
     // boolean flag to mark the function as existing.
@@ -262,6 +213,7 @@ pub trait BinaryOpT {
     fn f64(v1: f64, v2: f64) -> f64;
     fn u8(v1: u8, v2: u8) -> u8;
     fn u32(v1: u32, v2: u32) -> u32;
+    fn i64(v1: i64, v2: i64) -> i64;
 
     const BF16_VEC: bool = false;
     fn bf16_vec(_xs1: &[bf16], _xs2: &[bf16], _ys: &mut [bf16]) {}
@@ -275,12 +227,16 @@ pub trait BinaryOpT {
     fn u8_vec(_xs1: &[u8], _xs2: &[u8], _ys: &mut [u8]) {}
     const U32_VEC: bool = false;
     fn u32_vec(_xs1: &[u32], _xs2: &[u32], _ys: &mut [u32]) {}
+    const I64_VEC: bool = false;
+    fn i64_vec(_xs1: &[i64], _xs2: &[i64], _ys: &mut [i64]) {}
 }
 
 pub(crate) struct Add;
 pub(crate) struct Div;
 pub(crate) struct Mul;
 pub(crate) struct Sub;
+pub(crate) struct Maximum;
+pub(crate) struct Minimum;
 pub(crate) struct Exp;
 pub(crate) struct Log;
 pub(crate) struct Sin;
@@ -291,7 +247,15 @@ pub(crate) struct Recip;
 pub(crate) struct Sqr;
 pub(crate) struct Sqrt;
 pub(crate) struct Gelu;
+pub(crate) struct GeluErf;
+pub(crate) struct Erf;
 pub(crate) struct Relu;
+pub(crate) struct Silu;
+pub(crate) struct Tanh;
+pub(crate) struct Floor;
+pub(crate) struct Ceil;
+pub(crate) struct Round;
+pub(crate) struct Sign;
 
 macro_rules! bin_op {
     ($op:ident, $name: literal, $e: expr, $f32_vec: ident, $f64_vec: ident) => {
@@ -323,6 +287,10 @@ macro_rules! bin_op {
             fn u32(v1: u32, v2: u32) -> u32 {
                 $e(v1, v2)
             }
+            #[inline(always)]
+            fn i64(v1: i64, v2: i64) -> i64 {
+                $e(v1, v2)
+            }
 
             #[cfg(feature = "mkl")]
             const F32_VEC: bool = true;
@@ -361,7 +329,22 @@ bin_op!(Add, "add", |v1, v2| v1 + v2, vs_add, vd_add);
 bin_op!(Sub, "sub", |v1, v2| v1 - v2, vs_sub, vd_sub);
 bin_op!(Mul, "mul", |v1, v2| v1 * v2, vs_mul, vd_mul);
 bin_op!(Div, "div", |v1, v2| v1 / v2, vs_div, vd_div);
+bin_op!(
+    Minimum,
+    "minimum",
+    |v1, v2| if v1 > v2 { v2 } else { v1 },
+    vs_min,
+    vd_min
+);
+bin_op!(
+    Maximum,
+    "maximum",
+    |v1, v2| if v1 < v2 { v2 } else { v1 },
+    vs_max,
+    vd_max
+);
 
+#[allow(clippy::redundant_closure_call)]
 macro_rules! unary_op {
     ($op: ident, $name: literal, $a: ident, $e: expr) => {
         impl UnaryOpT for $op {
@@ -392,6 +375,10 @@ macro_rules! unary_op {
             fn u32(_: u32) -> u32 {
                 todo!("no unary function for u32")
             }
+            #[inline(always)]
+            fn i64(_: i64) -> i64 {
+                todo!("no unary function for i64")
+            }
         }
     };
 
@@ -424,6 +411,10 @@ macro_rules! unary_op {
             fn u32(_: u32) -> u32 {
                 todo!("no unary function for u32")
             }
+            #[inline(always)]
+            fn i64(_: i64) -> i64 {
+                todo!("no unary function for i64")
+            }
 
             #[cfg(feature = "mkl")]
             const F32_VEC: bool = true;
@@ -462,13 +453,21 @@ unary_op!(Exp, "exp", v, v.exp(), vs_exp, vd_exp);
 unary_op!(Log, "log", v, v.ln(), vs_ln, vd_ln);
 unary_op!(Sin, "sin", v, v.sin(), vs_sin, vd_sin);
 unary_op!(Cos, "cos", v, v.cos(), vs_cos, vd_cos);
-unary_op!(Abs, "abs", v, v.abs());
+unary_op!(Tanh, "tanh", v, v.tanh(), vs_tanh, vd_tanh);
 unary_op!(Neg, "neg", v, -v);
 unary_op!(Recip, "recip", v, v.recip());
 unary_op!(Sqr, "sqr", v, v * v, vs_sqr, vd_sqr);
 unary_op!(Sqrt, "sqrt", v, v.sqrt(), vs_sqrt, vd_sqrt);
 
-/// `gelu` operation
+// Hardcode the value for sqrt(2/pi)
+// https://github.com/huggingface/candle/issues/1982
+#[allow(clippy::excessive_precision)]
+const SQRT_TWO_OVER_PI_F32: f32 = 0.79788456080286535587989211986876373;
+#[allow(clippy::excessive_precision)]
+const SQRT_TWO_OVER_PI_F64: f64 = 0.79788456080286535587989211986876373;
+
+/// Tanh based approximation of the `gelu` operation
+/// GeluErf is the more precise one.
 /// <https://en.wikipedia.org/wiki/Activation_function#Comparison_of_activation_functions>
 impl UnaryOpT for Gelu {
     const NAME: &'static str = "gelu";
@@ -479,7 +478,7 @@ impl UnaryOpT for Gelu {
             * v
             * (bf16::ONE
                 + bf16::tanh(
-                    (bf16::from_f32_const(2.0) / bf16::PI).sqrt()
+                    bf16::from_f32_const(SQRT_TWO_OVER_PI_F32)
                         * v
                         * (bf16::ONE + bf16::from_f32_const(0.044715) * v * v),
                 ))
@@ -490,22 +489,18 @@ impl UnaryOpT for Gelu {
             * v
             * (f16::ONE
                 + f16::tanh(
-                    (f16::from_f32_const(2.0) / f16::PI).sqrt()
+                    f16::from_f32_const(SQRT_TWO_OVER_PI_F32)
                         * v
                         * (f16::ONE + f16::from_f32_const(0.044715) * v * v),
                 ))
     }
     #[inline(always)]
     fn f32(v: f32) -> f32 {
-        0.5 * v
-            * (1.0
-                + f32::tanh((2.0f32 / std::f32::consts::PI).sqrt() * v * (1.0 + 0.044715 * v * v)))
+        0.5 * v * (1.0 + f32::tanh(SQRT_TWO_OVER_PI_F32 * v * (1.0 + 0.044715 * v * v)))
     }
     #[inline(always)]
     fn f64(v: f64) -> f64 {
-        0.5 * v
-            * (1.0
-                + f64::tanh((2.0f64 / std::f64::consts::PI).sqrt() * v * (1.0 + 0.044715 * v * v)))
+        0.5 * v * (1.0 + f64::tanh(SQRT_TWO_OVER_PI_F64 * v * (1.0 + 0.044715 * v * v)))
     }
     #[inline(always)]
     fn u8(_: u8) -> u8 {
@@ -515,6 +510,10 @@ impl UnaryOpT for Gelu {
     fn u32(_: u32) -> u32 {
         0
     }
+    #[inline(always)]
+    fn i64(_: i64) -> i64 {
+        0
+    }
     const KERNEL: &'static str = "ugelu";
 
     #[cfg(feature = "mkl")]
@@ -534,6 +533,301 @@ impl UnaryOpT for Gelu {
     fn f64_vec(xs: &[f64], ys: &mut [f64]) {
         crate::mkl::vd_gelu(xs, ys)
     }
+
+    #[cfg(feature = "accelerate")]
+    const F32_VEC: bool = true;
+
+    #[cfg(feature = "accelerate")]
+    #[inline(always)]
+    fn f32_vec(xs: &[f32], ys: &mut [f32]) {
+        crate::accelerate::vs_gelu(xs, ys)
+    }
+
+    #[cfg(feature = "accelerate")]
+    const F64_VEC: bool = true;
+
+    #[cfg(feature = "accelerate")]
+    #[inline(always)]
+    fn f64_vec(xs: &[f64], ys: &mut [f64]) {
+        crate::accelerate::vd_gelu(xs, ys)
+    }
+}
+
+/// `erf` operation
+/// <https://en.wikipedia.org/wiki/Error_function>
+impl UnaryOpT for Erf {
+    const NAME: &'static str = "erf";
+    const KERNEL: &'static str = "uerf";
+    const V: Self = Erf;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        bf16::from_f64(Self::f64(v.to_f64()))
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        f16::from_f64(Self::f64(v.to_f64()))
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        Self::f64(v as f64) as f32
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        crate::cpu::erf::erf(v)
+    }
+    #[inline(always)]
+    fn u8(_: u8) -> u8 {
+        0
+    }
+    #[inline(always)]
+    fn u32(_: u32) -> u32 {
+        0
+    }
+    #[inline(always)]
+    fn i64(_: i64) -> i64 {
+        0
+    }
+}
+
+/// Silu operation
+impl UnaryOpT for Silu {
+    const NAME: &'static str = "silu";
+    const V: Self = Silu;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        v / (bf16::ONE + (-v).exp())
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        v / (f16::ONE + (-v).exp())
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        v / (1.0 + (-v).exp())
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        v / (1.0 + (-v).exp())
+    }
+    #[inline(always)]
+    fn u8(_: u8) -> u8 {
+        0
+    }
+    #[inline(always)]
+    fn u32(_: u32) -> u32 {
+        0
+    }
+    #[inline(always)]
+    fn i64(_: i64) -> i64 {
+        0
+    }
+    const KERNEL: &'static str = "usilu";
+
+    #[cfg(feature = "mkl")]
+    const F32_VEC: bool = true;
+
+    #[cfg(feature = "mkl")]
+    #[inline(always)]
+    fn f32_vec(xs: &[f32], ys: &mut [f32]) {
+        crate::mkl::vs_silu(xs, ys)
+    }
+
+    #[cfg(feature = "mkl")]
+    const F64_VEC: bool = true;
+
+    #[cfg(feature = "mkl")]
+    #[inline(always)]
+    fn f64_vec(xs: &[f64], ys: &mut [f64]) {
+        crate::mkl::vd_silu(xs, ys)
+    }
+
+    #[cfg(feature = "accelerate")]
+    const F32_VEC: bool = true;
+
+    #[cfg(feature = "accelerate")]
+    #[inline(always)]
+    fn f32_vec(xs: &[f32], ys: &mut [f32]) {
+        crate::accelerate::vs_silu(xs, ys)
+    }
+
+    #[cfg(feature = "accelerate")]
+    const F64_VEC: bool = true;
+
+    #[cfg(feature = "accelerate")]
+    #[inline(always)]
+    fn f64_vec(xs: &[f64], ys: &mut [f64]) {
+        crate::accelerate::vd_silu(xs, ys)
+    }
+}
+
+impl UnaryOpT for Abs {
+    const NAME: &'static str = "abs";
+    const KERNEL: &'static str = "uabs";
+    const V: Self = Abs;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        v.abs()
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        v.abs()
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        v.abs()
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        v.abs()
+    }
+    #[inline(always)]
+    fn u8(v: u8) -> u8 {
+        v
+    }
+    #[inline(always)]
+    fn u32(v: u32) -> u32 {
+        v
+    }
+    #[inline(always)]
+    fn i64(v: i64) -> i64 {
+        v.abs()
+    }
+}
+
+impl UnaryOpT for Ceil {
+    const NAME: &'static str = "ceil";
+    const KERNEL: &'static str = "uceil";
+    const V: Self = Ceil;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        v.ceil()
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        v.ceil()
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        v.ceil()
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        v.ceil()
+    }
+    #[inline(always)]
+    fn u8(v: u8) -> u8 {
+        v
+    }
+    #[inline(always)]
+    fn u32(v: u32) -> u32 {
+        v
+    }
+    #[inline(always)]
+    fn i64(v: i64) -> i64 {
+        v
+    }
+}
+
+impl UnaryOpT for Floor {
+    const NAME: &'static str = "floor";
+    const KERNEL: &'static str = "ufloor";
+    const V: Self = Floor;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        v.floor()
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        v.floor()
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        v.floor()
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        v.floor()
+    }
+    #[inline(always)]
+    fn u8(v: u8) -> u8 {
+        v
+    }
+    #[inline(always)]
+    fn u32(v: u32) -> u32 {
+        v
+    }
+    #[inline(always)]
+    fn i64(v: i64) -> i64 {
+        v
+    }
+}
+
+impl UnaryOpT for Round {
+    const NAME: &'static str = "round";
+    const KERNEL: &'static str = "uround";
+    const V: Self = Round;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        v.round()
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        v.round()
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        v.round()
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        v.round()
+    }
+    #[inline(always)]
+    fn u8(v: u8) -> u8 {
+        v
+    }
+    #[inline(always)]
+    fn u32(v: u32) -> u32 {
+        v
+    }
+    #[inline(always)]
+    fn i64(v: i64) -> i64 {
+        v
+    }
+}
+
+impl UnaryOpT for GeluErf {
+    const NAME: &'static str = "gelu_erf";
+    const KERNEL: &'static str = "ugelu_erf";
+    const V: Self = GeluErf;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        bf16::from_f64(Self::f64(v.to_f64()))
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        f16::from_f64(Self::f64(v.to_f64()))
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        Self::f64(v as f64) as f32
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        (crate::cpu::erf::erf(v / 2f64.sqrt()) + 1.) * 0.5 * v
+    }
+    #[inline(always)]
+    fn u8(_: u8) -> u8 {
+        0
+    }
+    #[inline(always)]
+    fn u32(_: u32) -> u32 {
+        0
+    }
+    #[inline(always)]
+    fn i64(_: i64) -> i64 {
+        0
+    }
 }
 
 impl UnaryOpT for Relu {
@@ -564,6 +858,10 @@ impl UnaryOpT for Relu {
     fn u32(v: u32) -> u32 {
         v
     }
+    #[inline(always)]
+    fn i64(v: i64) -> i64 {
+        v
+    }
 }
 
 /// `BackpropOp` is a wrapper around `Option<Op>`. The main goal is to ensure that dependencies are
@@ -617,6 +915,10 @@ impl BackpropOp {
         };
         Self(op)
     }
+
+    pub(crate) fn is_none(&self) -> bool {
+        self.0.is_none()
+    }
 }
 
 impl std::ops::Deref for BackpropOp {
@@ -625,3 +927,37 @@ impl std::ops::Deref for BackpropOp {
         &self.0
     }
 }
+
+impl UnaryOpT for Sign {
+    const NAME: &'static str = "sign";
+    const KERNEL: &'static str = "usign";
+    const V: Self = Sign;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        bf16::from((v > bf16::ZERO) as i8) - bf16::from((v < bf16::ZERO) as i8)
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        f16::from((v > f16::ZERO) as i8) - f16::from((v < f16::ZERO) as i8)
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        f32::from(v > 0.) - f32::from(v < 0.)
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        f64::from(v > 0.) - f64::from(v < 0.)
+    }
+    #[inline(always)]
+    fn u8(v: u8) -> u8 {
+        u8::min(1, v)
+    }
+    #[inline(always)]
+    fn u32(v: u32) -> u32 {
+        u32::min(1, v)
+    }
+    #[inline(always)]
+    fn i64(v: i64) -> i64 {
+        (v > 0) as i64 - (v < 0) as i64
+    }
+}

candle-core/src/pickle.rs

@@ -0,0 +1,817 @@
+// Just enough pickle support to be able to read PyTorch checkpoints.
+// This hardcodes objects that are required for tensor reading, we may want to make this a bit more
+// composable/tensor agnostic at some point.
+use crate::{DType, Error as E, Layout, Result, Tensor};
+use byteorder::{LittleEndian, ReadBytesExt};
+use std::collections::HashMap;
+use std::io::BufRead;
+
+const VERBOSE: bool = false;
+
+// https://docs.juliahub.com/Pickle/LAUNc/0.1.0/opcode/
+#[repr(u8)]
+#[derive(Debug, Eq, PartialEq, Clone)]
+pub enum OpCode {
+    // https://github.com/python/cpython/blob/ed25f097160b5cbb0c9a1f9a746d2f1bbc96515a/Lib/pickletools.py#L2123
+    Proto = 0x80,
+    Global = b'c',
+    BinPut = b'q',
+    LongBinPut = b'r',
+    EmptyTuple = b')',
+    Reduce = b'R',
+    Mark = b'(',
+    BinUnicode = b'X',
+    BinInt = b'J',
+    Tuple = b't',
+    BinPersId = b'Q',
+    BinInt1 = b'K',
+    BinInt2 = b'M',
+    Tuple1 = 0x85,
+    Tuple2 = 0x86,
+    Tuple3 = 0x87,
+    NewTrue = 0x88,
+    NewFalse = 0x89,
+    None = b'N',
+    BinGet = b'h',
+    LongBinGet = b'j',
+    SetItem = b's',
+    SetItems = b'u',
+    EmptyDict = b'}',
+    Dict = b'd',
+    Build = b'b',
+    Stop = b'.',
+    NewObj = 0x81,
+    EmptyList = b']',
+    BinFloat = b'G',
+    Append = b'a',
+    Appends = b'e',
+}
+
+// Avoid using FromPrimitive so as not to drag another dependency.
+impl TryFrom<u8> for OpCode {
+    type Error = u8;
+    fn try_from(value: u8) -> std::result::Result<Self, Self::Error> {
+        match value {
+            0x80 => Ok(Self::Proto),
+            b'c' => Ok(Self::Global),
+            b'q' => Ok(Self::BinPut),
+            b'r' => Ok(Self::LongBinPut),
+            b')' => Ok(Self::EmptyTuple),
+            b'R' => Ok(Self::Reduce),
+            b'(' => Ok(Self::Mark),
+            b'X' => Ok(Self::BinUnicode),
+            b'J' => Ok(Self::BinInt),
+            b't' => Ok(Self::Tuple),
+            b'Q' => Ok(Self::BinPersId),
+            b'K' => Ok(Self::BinInt1),
+            b'M' => Ok(Self::BinInt2),
+            b'N' => Ok(Self::None),
+            0x85 => Ok(Self::Tuple1),
+            0x86 => Ok(Self::Tuple2),
+            0x87 => Ok(Self::Tuple3),
+            0x88 => Ok(Self::NewTrue),
+            0x89 => Ok(Self::NewFalse),
+            b'h' => Ok(Self::BinGet),
+            b'j' => Ok(Self::LongBinGet),
+            b's' => Ok(Self::SetItem),
+            b'u' => Ok(Self::SetItems),
+            b'}' => Ok(Self::EmptyDict),
+            b'd' => Ok(Self::EmptyDict),
+            b'b' => Ok(Self::Build),
+            b'.' => Ok(Self::Stop),
+            0x81 => Ok(Self::NewObj),
+            b']' => Ok(Self::EmptyList),
+            b'G' => Ok(Self::BinFloat),
+            b'a' => Ok(Self::Append),
+            b'e' => Ok(Self::Appends),
+            value => Err(value),
+        }
+    }
+}
+
+fn read_to_newline<R: BufRead>(r: &mut R) -> Result<Vec<u8>> {
+    let mut data: Vec<u8> = Vec::with_capacity(32);
+    r.read_until(b'\n', &mut data)?;
+    data.pop();
+    if data.last() == Some(&b'\r') {
+        data.pop();
+    }
+    Ok(data)
+}
+
+#[derive(Debug, Clone, PartialEq)]
+pub enum Object {
+    Class {
+        module_name: String,
+        class_name: String,
+    },
+    Int(i32),
+    Float(f64),
+    Unicode(String),
+    Bool(bool),
+    None,
+    Tuple(Vec<Object>),
+    List(Vec<Object>),
+    Mark,
+    Dict(Vec<(Object, Object)>),
+    Reduce {
+        callable: Box<Object>,
+        args: Box<Object>,
+    },
+    Build {
+        callable: Box<Object>,
+        args: Box<Object>,
+    },
+    PersistentLoad(Box<Object>),
+}
+
+type OResult<T> = std::result::Result<T, Object>;
+
+impl Object {
+    pub fn unicode(self) -> OResult<String> {
+        match self {
+            Self::Unicode(t) => Ok(t),
+            _ => Err(self),
+        }
+    }
+
+    pub fn reduce(self) -> OResult<(Self, Self)> {
+        match self {
+            Self::Reduce { callable, args } => Ok((*callable, *args)),
+            _ => Err(self),
+        }
+    }
+
+    pub fn none(self) -> OResult<()> {
+        match self {
+            Self::None => Ok(()),
+            _ => Err(self),
+        }
+    }
+
+    pub fn persistent_load(self) -> OResult<Self> {
+        match self {
+            Self::PersistentLoad(t) => Ok(*t),
+            _ => Err(self),
+        }
+    }
+
+    pub fn bool(self) -> OResult<bool> {
+        match self {
+            Self::Bool(t) => Ok(t),
+            _ => Err(self),
+        }
+    }
+
+    pub fn int(self) -> OResult<i32> {
+        match self {
+            Self::Int(t) => Ok(t),
+            _ => Err(self),
+        }
+    }
+
+    pub fn tuple(self) -> OResult<Vec<Self>> {
+        match self {
+            Self::Tuple(t) => Ok(t),
+            _ => Err(self),
+        }
+    }
+
+    pub fn dict(self) -> OResult<Vec<(Self, Self)>> {
+        match self {
+            Self::Dict(t) => Ok(t),
+            _ => Err(self),
+        }
+    }
+
+    pub fn class(self) -> OResult<(String, String)> {
+        match self {
+            Self::Class {
+                module_name,
+                class_name,
+            } => Ok((module_name, class_name)),
+            _ => Err(self),
+        }
+    }
+
+    pub fn into_tensor_info(
+        self,
+        name: Self,
+        dir_name: &std::path::Path,
+    ) -> Result<Option<TensorInfo>> {
+        let name = match name.unicode() {
+            Ok(name) => name,
+            Err(_) => return Ok(None),
+        };
+        let (callable, args) = match self.reduce() {
+            Ok(callable_args) => callable_args,
+            _ => return Ok(None),
+        };
+        let (callable, args) = match callable {
+            Object::Class {
+                module_name,
+                class_name,
+            } if module_name == "torch._tensor" && class_name == "_rebuild_from_type_v2" => {
+                let mut args = args.tuple()?;
+                let callable = args.remove(0);
+                let args = args.remove(1);
+                (callable, args)
+            }
+            Object::Class {
+                module_name,
+                class_name,
+            } if module_name == "torch._utils" && class_name == "_rebuild_parameter" => {
+                let mut args = args.tuple()?;
+                args.remove(0).reduce()?
+            }
+            _ => (callable, args),
+        };
+        match callable {
+            Object::Class {
+                module_name,
+                class_name,
+            } if module_name == "torch._utils" && class_name == "_rebuild_tensor_v2" => {}
+            _ => return Ok(None),
+        };
+        let (layout, dtype, file_path, storage_size) = rebuild_args(args)?;
+        Ok(Some(TensorInfo {
+            name,
+            dtype,
+            layout,
+            path: format!("{}/{}", dir_name.to_string_lossy(), file_path),
+            storage_size,
+        }))
+    }
+}
+
+impl TryFrom<Object> for String {
+    type Error = Object;
+    fn try_from(value: Object) -> std::result::Result<Self, Self::Error> {
+        match value {
+            Object::Unicode(s) => Ok(s),
+            other => Err(other),
+        }
+    }
+}
+
+impl TryFrom<Object> for usize {
+    type Error = Object;
+    fn try_from(value: Object) -> std::result::Result<Self, Self::Error> {
+        match value {
+            Object::Int(s) if s >= 0 => Ok(s as usize),
+            other => Err(other),
+        }
+    }
+}
+
+impl<T: TryFrom<Object, Error = Object>> TryFrom<Object> for Vec<T> {
+    type Error = Object;
+    fn try_from(value: Object) -> std::result::Result<Self, Self::Error> {
+        match value {
+            Object::Tuple(values) => {
+                // This does not return the appropriate value in the error case but instead return
+                // the object related to the first error.
+                values
+                    .into_iter()
+                    .map(|v| T::try_from(v))
+                    .collect::<std::result::Result<Vec<T>, Self::Error>>()
+            }
+            other => Err(other),
+        }
+    }
+}
+
+#[derive(Debug)]
+pub struct Stack {
+    stack: Vec<Object>,
+    memo: HashMap<u32, Object>,
+}
+
+impl Stack {
+    pub fn empty() -> Self {
+        Self {
+            stack: Vec::with_capacity(512),
+            memo: HashMap::new(),
+        }
+    }
+
+    pub fn stack(&self) -> &[Object] {
+        self.stack.as_slice()
+    }
+
+    pub fn read_loop<R: BufRead>(&mut self, r: &mut R) -> Result<()> {
+        loop {
+            if self.read(r)? {
+                break;
+            }
+        }
+        Ok(())
+    }
+
+    pub fn finalize(mut self) -> Result<Object> {
+        self.pop()
+    }
+
+    fn push(&mut self, obj: Object) {
+        self.stack.push(obj)
+    }
+
+    fn pop(&mut self) -> Result<Object> {
+        match self.stack.pop() {
+            None => crate::bail!("unexpected empty stack"),
+            Some(obj) => Ok(obj),
+        }
+    }
+
+    // https://docs.juliahub.com/Pickle/LAUNc/0.1.0/opcode/#Pickle.OpCodes.BUILD
+    fn build(&mut self) -> Result<()> {
+        let args = self.pop()?;
+        let obj = self.pop()?;
+        let obj = match (obj, args) {
+            (Object::Dict(mut obj), Object::Dict(mut args)) => {
+                obj.append(&mut args);
+                Object::Dict(obj)
+            }
+            (obj, args) => Object::Build {
+                callable: Box::new(obj),
+                args: Box::new(args),
+            },
+        };
+        self.push(obj);
+        Ok(())
+    }
+
+    fn reduce(&mut self) -> Result<()> {
+        let args = self.pop()?;
+        let callable = self.pop()?;
+        #[allow(clippy::single_match)]
+        let reduced = match &callable {
+            Object::Class {
+                module_name,
+                class_name,
+            } => {
+                if module_name == "collections"
+                    && (class_name == "OrderedDict" || class_name == "defaultdict")
+                {
+                    // TODO: have a separate ordered dict and a separate default dict.
+                    Some(Object::Dict(vec![]))
+                } else {
+                    None
+                }
+            }
+            _ => None,
+        };
+        let reduced = reduced.unwrap_or_else(|| Object::Reduce {
+            callable: Box::new(callable),
+            args: Box::new(args),
+        });
+        self.push(reduced);
+        Ok(())
+    }
+
+    fn last(&mut self) -> Result<&mut Object> {
+        match self.stack.last_mut() {
+            None => crate::bail!("unexpected empty stack"),
+            Some(obj) => Ok(obj),
+        }
+    }
+
+    fn memo_get(&self, id: u32) -> Result<Object> {
+        match self.memo.get(&id) {
+            None => crate::bail!("missing object in memo {id}"),
+            Some(obj) => {
+                // Maybe we should use refcounting rather than doing potential large clones here.
+                Ok(obj.clone())
+            }
+        }
+    }
+
+    fn memo_put(&mut self, id: u32) -> Result<()> {
+        let obj = self.last()?.clone();
+        self.memo.insert(id, obj);
+        Ok(())
+    }
+
+    fn persistent_load(&self, id: Object) -> Result<Object> {
+        Ok(Object::PersistentLoad(Box::new(id)))
+    }
+
+    fn new_obj(&self, class: Object, args: Object) -> Result<Object> {
+        Ok(Object::Reduce {
+            callable: Box::new(class),
+            args: Box::new(args),
+        })
+    }
+
+    fn pop_to_marker(&mut self) -> Result<Vec<Object>> {
+        let mut mark_idx = None;
+        for (idx, obj) in self.stack.iter().enumerate().rev() {
+            if obj == &Object::Mark {
+                mark_idx = Some(idx);
+                break;
+            }
+        }
+        match mark_idx {
+            Some(mark_idx) => {
+                let objs = self.stack.split_off(mark_idx + 1);
+                self.stack.pop();
+                Ok(objs)
+            }
+            None => {
+                crate::bail!("marker object not found")
+            }
+        }
+    }
+
+    pub fn read<R: BufRead>(&mut self, r: &mut R) -> Result<bool> {
+        let op_code = match OpCode::try_from(r.read_u8()?) {
+            Ok(op_code) => op_code,
+            Err(op_code) => {
+                crate::bail!("unknown op-code {op_code}")
+            }
+        };
+        // println!("op: {op_code:?}");
+        // println!("{:?}", self.stack);
+        match op_code {
+            OpCode::Proto => {
+                let version = r.read_u8()?;
+                if VERBOSE {
+                    println!("proto {version}");
+                }
+            }
+            OpCode::Global => {
+                let module_name = read_to_newline(r)?;
+                let class_name = read_to_newline(r)?;
+                let module_name = String::from_utf8_lossy(&module_name).to_string();
+                let class_name = String::from_utf8_lossy(&class_name).to_string();
+                self.push(Object::Class {
+                    module_name,
+                    class_name,
+                })
+            }
+            OpCode::BinInt1 => {
+                let arg = r.read_u8()?;
+                self.push(Object::Int(arg as i32))
+            }
+            OpCode::BinInt2 => {
+                let arg = r.read_u16::<LittleEndian>()?;
+                self.push(Object::Int(arg as i32))
+            }
+            OpCode::BinInt => {
+                let arg = r.read_i32::<LittleEndian>()?;
+                self.push(Object::Int(arg))
+            }
+            OpCode::BinFloat => {
+                // Somehow floats are encoded using BigEndian whereas int types use LittleEndian.
+                // https://github.com/python/cpython/blob/0c80da4c14d904a367968955544dd6ae58c8101c/Lib/pickletools.py#L855
+                // https://github.com/pytorch/pytorch/blob/372d078f361e726bb4ac0884ac334b04c58179ef/torch/_weights_only_unpickler.py#L243
+                let arg = r.read_f64::<byteorder::BigEndian>()?;
+                self.push(Object::Float(arg))
+            }
+            OpCode::BinUnicode => {
+                let len = r.read_u32::<LittleEndian>()?;
+                let mut data = vec![0u8; len as usize];
+                r.read_exact(&mut data)?;
+                let data = String::from_utf8(data).map_err(E::wrap)?;
+                self.push(Object::Unicode(data))
+            }
+            OpCode::BinPersId => {
+                let id = self.pop()?;
+                let obj = self.persistent_load(id)?;
+                self.push(obj)
+            }
+            OpCode::Tuple => {
+                let objs = self.pop_to_marker()?;
+                self.push(Object::Tuple(objs))
+            }
+            OpCode::Tuple1 => {
+                let obj = self.pop()?;
+                self.push(Object::Tuple(vec![obj]))
+            }
+            OpCode::Tuple2 => {
+                let obj2 = self.pop()?;
+                let obj1 = self.pop()?;
+                self.push(Object::Tuple(vec![obj1, obj2]))
+            }
+            OpCode::Tuple3 => {
+                let obj3 = self.pop()?;
+                let obj2 = self.pop()?;
+                let obj1 = self.pop()?;
+                self.push(Object::Tuple(vec![obj1, obj2, obj3]))
+            }
+            OpCode::NewTrue => self.push(Object::Bool(true)),
+            OpCode::NewFalse => self.push(Object::Bool(false)),
+            OpCode::Append => {
+                let value = self.pop()?;
+                let pylist = self.last()?;
+                if let Object::List(d) = pylist {
+                    d.push(value)
+                } else {
+                    crate::bail!("expected a list, got {pylist:?}")
+                }
+            }
+            OpCode::Appends => {
+                let objs = self.pop_to_marker()?;
+                let pylist = self.last()?;
+                if let Object::List(d) = pylist {
+                    d.extend(objs)
+                } else {
+                    crate::bail!("expected a list, got {pylist:?}")
+                }
+            }
+            OpCode::SetItem => {
+                let value = self.pop()?;
+                let key = self.pop()?;
+                let pydict = self.last()?;
+                if let Object::Dict(d) = pydict {
+                    d.push((key, value))
+                } else {
+                    crate::bail!("expected a dict, got {pydict:?}")
+                }
+            }
+            OpCode::SetItems => {
+                let mut objs = self.pop_to_marker()?;
+                let pydict = self.last()?;
+                if let Object::Dict(d) = pydict {
+                    if objs.len() % 2 != 0 {
+                        crate::bail!("setitems: not an even number of objects")
+                    }
+                    while let Some(value) = objs.pop() {
+                        let key = objs.pop().unwrap();
+                        d.push((key, value))
+                    }
+                } else {
+                    crate::bail!("expected a dict, got {pydict:?}")
+                }
+            }
+            OpCode::None => self.push(Object::None),
+            OpCode::Stop => {
+                return Ok(true);
+            }
+            OpCode::Build => self.build()?,
+            OpCode::EmptyDict => self.push(Object::Dict(vec![])),
+            OpCode::Dict => {
+                let mut objs = self.pop_to_marker()?;
+                let mut pydict = vec![];
+                if objs.len() % 2 != 0 {
+                    crate::bail!("setitems: not an even number of objects")
+                }
+                while let Some(value) = objs.pop() {
+                    let key = objs.pop().unwrap();
+                    pydict.push((key, value))
+                }
+                self.push(Object::Dict(pydict))
+            }
+            OpCode::Mark => self.push(Object::Mark),
+            OpCode::Reduce => self.reduce()?,
+            OpCode::EmptyTuple => self.push(Object::Tuple(vec![])),
+            OpCode::EmptyList => self.push(Object::List(vec![])),
+            OpCode::BinGet => {
+                let arg = r.read_u8()?;
+                let obj = self.memo_get(arg as u32)?;
+                self.push(obj)
+            }
+            OpCode::LongBinGet => {
+                let arg = r.read_u32::<LittleEndian>()?;
+                let obj = self.memo_get(arg)?;
+                self.push(obj)
+            }
+            OpCode::BinPut => {
+                let arg = r.read_u8()?;
+                self.memo_put(arg as u32)?
+            }
+            OpCode::LongBinPut => {
+                let arg = r.read_u32::<LittleEndian>()?;
+                self.memo_put(arg)?
+            }
+            OpCode::NewObj => {
+                let args = self.pop()?;
+                let class = self.pop()?;
+                let obj = self.new_obj(class, args)?;
+                self.push(obj)
+            }
+        }
+        Ok(false)
+    }
+}
+
+impl From<Object> for E {
+    fn from(value: Object) -> Self {
+        E::Msg(format!("conversion error on {value:?}"))
+    }
+}
+
+// https://github.com/pytorch/pytorch/blob/4eac43d046ded0f0a5a5fa8db03eb40f45bf656e/torch/_utils.py#L198
+// Arguments: storage, storage_offset, size, stride, requires_grad, backward_hooks
+fn rebuild_args(args: Object) -> Result<(Layout, DType, String, usize)> {
+    let mut args = args.tuple()?;
+    let stride = Vec::<usize>::try_from(args.remove(3))?;
+    let size = Vec::<usize>::try_from(args.remove(2))?;
+    let offset = args.remove(1).int()? as usize;
+    let storage = args.remove(0).persistent_load()?;
+    let mut storage = storage.tuple()?;
+    let storage_size = storage.remove(4).int()? as usize;
+    let path = storage.remove(2).unicode()?;
+    let (_module_name, class_name) = storage.remove(1).class()?;
+    let dtype = match class_name.as_str() {
+        "FloatStorage" => DType::F32,
+        "DoubleStorage" => DType::F64,
+        "HalfStorage" => DType::F16,
+        "BFloat16Storage" => DType::BF16,
+        "ByteStorage" => DType::U8,
+        "LongStorage" => DType::I64,
+        other => {
+            crate::bail!("unsupported storage type {other}")
+        }
+    };
+    let layout = Layout::new(crate::Shape::from(size), stride, offset);
+    Ok((layout, dtype, path, storage_size))
+}
+
+#[derive(Debug, Clone)]
+pub struct TensorInfo {
+    pub name: String,
+    pub dtype: DType,
+    pub layout: Layout,
+    pub path: String,
+    pub storage_size: usize,
+}
+
+/// Read the tensor info from a .pth file.
+///
+/// # Arguments
+/// * `file` - The path to the .pth file.
+/// * `verbose` - Whether to print debug information.
+/// * `key` - Optional key to retrieve `state_dict` from the pth file.
+pub fn read_pth_tensor_info<P: AsRef<std::path::Path>>(
+    file: P,
+    verbose: bool,
+    key: Option<&str>,
+) -> Result<Vec<TensorInfo>> {
+    let file = std::fs::File::open(file)?;
+    let zip_reader = std::io::BufReader::new(file);
+    let mut zip = zip::ZipArchive::new(zip_reader)?;
+    let zip_file_names = zip
+        .file_names()
+        .map(|f| f.to_string())
+        .collect::<Vec<String>>();
+
+    let mut tensor_infos = vec![];
+    for file_name in zip_file_names.iter() {
+        if !file_name.ends_with("data.pkl") {
+            continue;
+        }
+        let dir_name = std::path::PathBuf::from(file_name.strip_suffix(".pkl").unwrap());
+        let reader = zip.by_name(file_name)?;
+        let mut reader = std::io::BufReader::new(reader);
+        let mut stack = Stack::empty();
+        stack.read_loop(&mut reader)?;
+        let obj = stack.finalize()?;
+        if VERBOSE || verbose {
+            println!("{obj:#?}");
+        }
+
+        let obj = match obj {
+            Object::Build { callable, args } => match *callable {
+                Object::Reduce { callable, args: _ } => match *callable {
+                    Object::Class {
+                        module_name,
+                        class_name,
+                    } if module_name == "__torch__" && class_name == "Module" => *args,
+                    _ => continue,
+                },
+                _ => continue,
+            },
+            obj => obj,
+        };
+
+        // If key is provided, then we need to extract the state_dict from the object.
+        let obj = if let Some(key) = key {
+            if let Object::Dict(key_values) = obj {
+                key_values
+                    .into_iter()
+                    .find(|(k, _)| *k == Object::Unicode(key.to_owned()))
+                    .map(|(_, v)| v)
+                    .ok_or_else(|| E::Msg(format!("key {key} not found")))?
+            } else {
+                obj
+            }
+        } else {
+            obj
+        };
+
+        // If the object is a dict, then we can extract the tensor info from it.
+        // NOTE: We are assuming that the `obj` is state_dict by this stage.
+        if let Object::Dict(key_values) = obj {
+            for (name, value) in key_values.into_iter() {
+                match value.into_tensor_info(name, &dir_name) {
+                    Ok(Some(tensor_info)) => tensor_infos.push(tensor_info),
+                    Ok(None) => {}
+                    Err(err) => eprintln!("skipping: {err:?}"),
+                }
+            }
+        }
+    }
+    Ok(tensor_infos)
+}
+
+/// Lazy tensor loader.
+pub struct PthTensors {
+    tensor_infos: HashMap<String, TensorInfo>,
+    path: std::path::PathBuf,
+    // We do not store a zip reader as it needs mutable access to extract data. Instead we
+    // re-create a zip reader for each tensor.
+}
+
+impl PthTensors {
+    pub fn new<P: AsRef<std::path::Path>>(path: P, key: Option<&str>) -> Result<Self> {
+        let tensor_infos = read_pth_tensor_info(path.as_ref(), false, key)?;
+        let tensor_infos = tensor_infos
+            .into_iter()
+            .map(|ti| (ti.name.to_string(), ti))
+            .collect();
+        let path = path.as_ref().to_owned();
+        Ok(Self { tensor_infos, path })
+    }
+
+    pub fn tensor_infos(&self) -> &HashMap<String, TensorInfo> {
+        &self.tensor_infos
+    }
+
+    pub fn get(&self, name: &str) -> Result<Option<Tensor>> {
+        use std::io::Read;
+        let tensor_info = match self.tensor_infos.get(name) {
+            None => return Ok(None),
+            Some(tensor_info) => tensor_info,
+        };
+        // We hope that the file has not changed since first reading it.
+        let zip_reader = std::io::BufReader::new(std::fs::File::open(&self.path)?);
+        let mut zip = zip::ZipArchive::new(zip_reader)?;
+        let mut reader = zip.by_name(&tensor_info.path)?;
+        let is_fortran_contiguous = tensor_info.layout.is_fortran_contiguous();
+        let rank = tensor_info.layout.shape().rank();
+
+        // Reading the data is a bit tricky as it can be strided, for now only support the basic
+        // case and when the tensor is fortran contiguous.
+        if !tensor_info.layout.is_contiguous() && !is_fortran_contiguous {
+            crate::bail!(
+                "cannot retrieve non-contiguous tensors {:?}",
+                tensor_info.layout
+            )
+        }
+        let start_offset = tensor_info.layout.start_offset();
+        if start_offset > 0 {
+            std::io::copy(
+                &mut reader.by_ref().take(start_offset as u64),
+                &mut std::io::sink(),
+            )?;
+        }
+        let tensor = Tensor::from_reader(
+            tensor_info.layout.shape().clone(),
+            tensor_info.dtype,
+            &mut reader,
+        )?;
+
+        if rank > 1 && is_fortran_contiguous {
+            // Reverse the shape, e.g. Shape(2, 3, 4) -> Shape(4, 3, 2)
+            let shape_reversed: Vec<_> = tensor_info.layout.dims().iter().rev().cloned().collect();
+            let tensor = tensor.reshape(shape_reversed)?;
+
+            // Permute (transpose) the dimensions, e.g. Shape(4, 3, 2) -> Shape(2, 3, 4)
+            let dim_indeces_reversed: Vec<_> = (0..rank).rev().collect();
+            let tensor = tensor.permute(dim_indeces_reversed)?;
+            Ok(Some(tensor))
+        } else {
+            Ok(Some(tensor))
+        }
+    }
+}
+
+/// Read all the tensors from a PyTorch pth file with a given key.
+///
+/// # Arguments
+/// * `path` - Path to the pth file.
+/// * `key` - Optional key to retrieve `state_dict` from the pth file. Sometimes the pth file
+///           contains multiple objects and the state_dict is the one we are interested in.
+pub fn read_all_with_key<P: AsRef<std::path::Path>>(
+    path: P,
+    key: Option<&str>,
+) -> Result<Vec<(String, Tensor)>> {
+    let pth = PthTensors::new(path, key)?;
+    let tensor_names = pth.tensor_infos.keys();
+    let mut tensors = Vec::with_capacity(tensor_names.len());
+    for name in tensor_names {
+        if let Some(tensor) = pth.get(name)? {
+            tensors.push((name.to_string(), tensor))
+        }
+    }
+    Ok(tensors)
+}
+
+/// Read all the tensors from a PyTorch pth file.
+///
+/// # Arguments
+/// * `path` - Path to the pth file.
+pub fn read_all<P: AsRef<std::path::Path>>(path: P) -> Result<Vec<(String, Tensor)>> {
+    read_all_with_key(path, None)
+}

candle-core/src/quantized/avx.rs

@@ -0,0 +1,667 @@
+use super::k_quants::{
+    BlockQ2K, BlockQ3K, BlockQ4K, BlockQ4_0, BlockQ5K, BlockQ6K, BlockQ8K, BlockQ8_0, QK8_0, QK_K,
+};
+use crate::Result;
+use byteorder::{ByteOrder, LittleEndian};
+use half::f16;
+
+#[cfg(target_arch = "x86")]
+use core::arch::x86::*;
+#[cfg(target_arch = "x86_64")]
+use core::arch::x86_64::*;
+
+#[inline(always)]
+pub(crate) unsafe fn sum_i16_pairs_float(x: __m256i) -> __m256 {
+    let ones = _mm256_set1_epi16(1);
+    let summed_pairs = _mm256_madd_epi16(ones, x);
+    _mm256_cvtepi32_ps(summed_pairs)
+}
+
+#[inline(always)]
+pub(crate) unsafe fn mul_sum_us8_pairs_float(ax: __m256i, sy: __m256i) -> __m256 {
+    let dot = _mm256_maddubs_epi16(ax, sy);
+    sum_i16_pairs_float(dot)
+}
+
+#[inline(always)]
+pub(crate) unsafe fn hsum_float_8(x: __m256) -> f32 {
+    let res = _mm256_extractf128_ps(x, 1);
+    let res = _mm_add_ps(res, _mm256_castps256_ps128(x));
+    let res = _mm_add_ps(res, _mm_movehl_ps(res, res));
+    let res = _mm_add_ss(res, _mm_movehdup_ps(res));
+    _mm_cvtss_f32(res)
+}
+
+#[inline(always)]
+pub(crate) unsafe fn bytes_from_nibbles_32(rsi: *const u8) -> __m256i {
+    let tmp = _mm_loadu_si128(rsi as *const __m128i);
+    let bytes = _mm256_insertf128_si256::<1>(_mm256_castsi128_si256(tmp), _mm_srli_epi16(tmp, 4));
+    let low_mask = _mm256_set1_epi8(0xF);
+    _mm256_and_si256(low_mask, bytes)
+}
+
+#[inline(always)]
+pub(crate) unsafe fn mul_sum_i8_pairs_float(x: __m256i, y: __m256i) -> __m256 {
+    let ax = _mm256_sign_epi8(x, x);
+    let sy = _mm256_sign_epi8(y, x);
+    mul_sum_us8_pairs_float(ax, sy)
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q4_0_q8_0(n: usize, xs: &[BlockQ4_0], ys: &[BlockQ8_0]) -> Result<f32> {
+    let qk = QK8_0;
+    if n % QK8_0 != 0 {
+        crate::bail!("vec_dot_q4_0_q8_0: {n} is not divisible by {qk}")
+    }
+    unsafe {
+        let mut acc = _mm256_setzero_ps();
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = _mm256_set1_ps(f16::to_f32(x.d) * f16::to_f32(y.d));
+            let bx = bytes_from_nibbles_32(x.qs.as_ptr());
+            let off = _mm256_set1_epi8(8);
+            let bx = _mm256_sub_epi8(bx, off);
+            let by = _mm256_loadu_si256(y.qs.as_ptr() as *const __m256i);
+            let q = mul_sum_i8_pairs_float(bx, by);
+            acc = _mm256_fmadd_ps(d, q, acc);
+        }
+        Ok(hsum_float_8(acc))
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q8_0_q8_0(n: usize, xs: &[BlockQ8_0], ys: &[BlockQ8_0]) -> Result<f32> {
+    let qk = QK8_0;
+    if n % QK8_0 != 0 {
+        crate::bail!("vec_dot_q8_0_q8_0: {n} is not divisible by {qk}")
+    }
+    unsafe {
+        let mut acc = _mm256_setzero_ps();
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = _mm256_set1_ps(f16::to_f32(x.d) * f16::to_f32(y.d));
+            let bx = _mm256_loadu_si256(x.qs.as_ptr() as *const __m256i);
+            let by = _mm256_loadu_si256(y.qs.as_ptr() as *const __m256i);
+            let q = mul_sum_i8_pairs_float(bx, by);
+            acc = _mm256_fmadd_ps(d, q, acc);
+        }
+        Ok(hsum_float_8(acc))
+    }
+}
+
+#[inline(always)]
+unsafe fn get_scale_shuffle(i: usize) -> __m128i {
+    const K_SHUFFLE: [u8; 128] = [
+        0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3,
+        3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7,
+        7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10, 10, 10, 10, 10,
+        11, 11, 11, 11, 11, 11, 11, 11, 12, 12, 12, 12, 12, 12, 12, 12, 13, 13, 13, 13, 13, 13, 13,
+        13, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15,
+    ];
+    _mm_loadu_si128((K_SHUFFLE.as_ptr() as *const __m128i).add(i))
+}
+
+#[inline(always)]
+unsafe fn get_scale_shuffle_k4(i: usize) -> __m256i {
+    const K_SHUFFLE: [u8; 256] = [
+        0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
+        0, 1, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+        2, 3, 2, 3, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
+        4, 5, 4, 5, 4, 5, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+        6, 7, 6, 7, 6, 7, 6, 7, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
+        8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 10, 11, 10, 11, 10, 11, 10, 11, 10, 11, 10, 11, 10, 11, 10,
+        11, 10, 11, 10, 11, 10, 11, 10, 11, 10, 11, 10, 11, 10, 11, 10, 11, 12, 13, 12, 13, 12, 13,
+        12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12,
+        13, 12, 13, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15,
+        14, 15, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15,
+    ];
+    _mm256_loadu_si256((K_SHUFFLE.as_ptr() as *const __m256i).add(i))
+}
+
+#[inline(always)]
+unsafe fn get_scale_shuffle_q3k(i: usize) -> __m256i {
+    const K_SHUFFLE: [u8; 128] = [
+        0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+        2, 3, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+        6, 7, 6, 7, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 10, 11, 10, 11, 10, 11, 10, 11,
+        10, 11, 10, 11, 10, 11, 10, 11, 12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12, 13, 12,
+        13, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15, 14, 15,
+    ];
+    _mm256_loadu_si256((K_SHUFFLE.as_ptr() as *const __m256i).add(i))
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q6k_q8k(n: usize, xs: &[BlockQ6K], ys: &[BlockQ8K]) -> Result<f32> {
+    let qk = QK_K;
+    if n % qk != 0 {
+        crate::bail!("vec_dot_q6k_8k: {n} is not divisible by {qk}")
+    }
+
+    unsafe {
+        let m4 = _mm256_set1_epi8(0xF);
+        let m2 = _mm256_set1_epi8(3);
+        let m32s = _mm256_set1_epi8(32);
+        let mut acc = _mm256_setzero_ps();
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = y.d * x.d.to_f32();
+            let mut q4 = x.ql.as_ptr();
+            let mut qh = x.qh.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+
+            let scales = _mm_loadu_si128(x.scales.as_ptr() as *const __m128i);
+            let mut sumi = _mm256_setzero_si256();
+
+            for j in 0..QK_K / 128 {
+                let is = j * 4;
+                let scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is));
+                let scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+                let scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+                let scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+
+                let q4bits1 = _mm256_loadu_si256(q4 as *const __m256i);
+                q4 = q4.add(32);
+                let q4bits2 = _mm256_loadu_si256(q4 as *const __m256i);
+                q4 = q4.add(32);
+                let q4bits_h = _mm256_loadu_si256(qh as *const __m256i);
+                qh = qh.add(32);
+
+                let q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bits_h, m2), 4);
+                let q4h_1 =
+                    _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bits_h, 2), m2), 4);
+                let q4h_2 =
+                    _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bits_h, 4), m2), 4);
+                let q4h_3 =
+                    _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bits_h, 6), m2), 4);
+
+                let q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
+                let q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
+                let q4_2 =
+                    _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
+                let q4_3 =
+                    _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
+
+                let q8_0 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_1 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_2 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_3 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+
+                let q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
+                let q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
+                let q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
+                let q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
+
+                let p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
+                let p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
+                let p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
+                let p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
+
+                let p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+                let p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+                let p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
+                let p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+
+                let p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
+                let p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
+                let p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
+                let p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
+
+                sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+                sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
+            }
+            acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+        }
+        Ok(hsum_float_8(acc))
+    }
+}
+
+#[inline(always)]
+unsafe fn mm256_set_m128i(a: __m128i, b: __m128i) -> __m256i {
+    _mm256_insertf128_si256(_mm256_castsi128_si256(b), a, 1)
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q2k_q8k(n: usize, xs: &[BlockQ2K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q2k_q8k: {n} is not divisible by {QK_K}")
+    }
+
+    unsafe {
+        let m3 = _mm256_set1_epi8(3);
+        let m4 = _mm_set1_epi8(0xF);
+
+        let mut acc = _mm256_setzero_ps();
+
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = y.d * x.d.to_f32();
+            let dmin = -y.d * x.dmin.to_f32();
+
+            let mut q2 = x.qs.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+
+            let mins_and_scales = _mm_loadu_si128(x.scales.as_ptr() as *const __m128i);
+            let scales8 = _mm_and_si128(mins_and_scales, m4);
+            let mins8 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+            let mins = _mm256_cvtepi8_epi16(mins8);
+            let prod =
+                _mm256_madd_epi16(mins, _mm256_loadu_si256(y.bsums.as_ptr() as *const __m256i));
+
+            acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(prod), acc);
+
+            let all_scales = _mm256_cvtepi8_epi16(scales8);
+            let l_scales = _mm256_extracti128_si256(all_scales, 0);
+            let h_scales = _mm256_extracti128_si256(all_scales, 1);
+            let scales = [
+                mm256_set_m128i(l_scales, l_scales),
+                mm256_set_m128i(h_scales, h_scales),
+            ];
+
+            let mut sumi = _mm256_setzero_si256();
+
+            for scale in scales {
+                let q2bits = _mm256_loadu_si256(q2 as *const __m256i);
+                q2 = q2.add(32);
+
+                let q8_0 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_1 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_2 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_3 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+
+                let q2_0 = _mm256_and_si256(q2bits, m3);
+                let q2_1 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 2), m3);
+                let q2_2 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 4), m3);
+                let q2_3 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 6), m3);
+
+                let p0 = _mm256_maddubs_epi16(q2_0, q8_0);
+                let p1 = _mm256_maddubs_epi16(q2_1, q8_1);
+                let p2 = _mm256_maddubs_epi16(q2_2, q8_2);
+                let p3 = _mm256_maddubs_epi16(q2_3, q8_3);
+
+                let p0 =
+                    _mm256_madd_epi16(_mm256_shuffle_epi8(scale, get_scale_shuffle_q3k(0)), p0);
+                let p1 =
+                    _mm256_madd_epi16(_mm256_shuffle_epi8(scale, get_scale_shuffle_q3k(1)), p1);
+                let p2 =
+                    _mm256_madd_epi16(_mm256_shuffle_epi8(scale, get_scale_shuffle_q3k(2)), p2);
+                let p3 =
+                    _mm256_madd_epi16(_mm256_shuffle_epi8(scale, get_scale_shuffle_q3k(3)), p3);
+
+                let p0 = _mm256_add_epi32(p0, p1);
+                let p2 = _mm256_add_epi32(p2, p3);
+
+                sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p0, p2));
+            }
+            acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+        }
+
+        Ok(hsum_float_8(acc))
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q3k_q8k: {n} is not divisible by {QK_K}")
+    }
+
+    const KMASK1: u32 = 0x03030303;
+    const KMASK2: u32 = 0x0f0f0f0f;
+
+    let mut aux = [0u32; 3];
+
+    unsafe {
+        let m3 = _mm256_set1_epi8(3);
+        let mone = _mm256_set1_epi8(1);
+        let m32 = _mm_set1_epi8(32);
+
+        let mut acc = _mm256_setzero_ps();
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = y.d * x.d.to_f32();
+
+            let mut q3 = x.qs.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+
+            LittleEndian::read_u32_into(&x.scales, &mut aux);
+            let scales128 = _mm_set_epi32(
+                (((aux[1] >> 4) & KMASK2) | (((aux[2] >> 6) & KMASK1) << 4)) as i32,
+                (((aux[0] >> 4) & KMASK2) | (((aux[2] >> 4) & KMASK1) << 4)) as i32,
+                ((aux[1] & KMASK2) | (((aux[2] >> 2) & KMASK1) << 4)) as i32,
+                ((aux[0] & KMASK2) | (((aux[2]) & KMASK1) << 4)) as i32,
+            );
+            let scales128 = _mm_sub_epi8(scales128, m32);
+            let all_scales = _mm256_cvtepi8_epi16(scales128);
+            let l_scales = _mm256_extracti128_si256(all_scales, 0);
+            let h_scales = _mm256_extracti128_si256(all_scales, 1);
+            let scales = [
+                mm256_set_m128i(l_scales, l_scales),
+                mm256_set_m128i(h_scales, h_scales),
+            ];
+
+            // high bit
+            let hbits = _mm256_loadu_si256(x.hmask.as_ptr() as *const __m256i);
+
+            let mut sumi = _mm256_setzero_si256();
+
+            for (j, scale) in scales.iter().enumerate() {
+                // load low 2 bits
+                let q3bits = _mm256_loadu_si256(q3 as *const __m256i);
+                q3 = q3.add(32);
+
+                // Prepare low and high bits
+                // We hardcode the shifts here to avoid loading them into a separate register
+                let q3l_0 = _mm256_and_si256(q3bits, m3);
+                let q3h_0 = if j == 0 {
+                    _mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, 0)), 0)
+                } else {
+                    _mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, 4)), 4)
+                };
+                let q3h_0 = _mm256_slli_epi16(q3h_0, 2);
+
+                let q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 2), m3);
+                let q3h_1 = if j == 0 {
+                    _mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, 1)), 1)
+                } else {
+                    _mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, 5)), 5)
+                };
+                let q3h_1 = _mm256_slli_epi16(q3h_1, 2);
+
+                let q3l_2 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 4), m3);
+                let q3h_2 = if j == 0 {
+                    _mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, 2)), 2)
+                } else {
+                    _mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, 6)), 6)
+                };
+                let q3h_2 = _mm256_slli_epi16(q3h_2, 2);
+
+                let q3l_3 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 6), m3);
+                let q3h_3 = if j == 0 {
+                    _mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, 3)), 3)
+                } else {
+                    _mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, 7)), 7)
+                };
+                let q3h_3 = _mm256_slli_epi16(q3h_3, 2);
+
+                // load Q8 quants
+                let q8_0 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_1 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_2 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_3 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+
+                // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we
+                // can use _mm256_maddubs_epi16, and then subtract. The high bit part has the 2
+                // already subtracted (and so, it is zero if the high bit was not set, and 2 if the
+                // high bit was set)
+                let q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
+                let q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
+                let q8s_2 = _mm256_maddubs_epi16(q3h_2, q8_2);
+                let q8s_3 = _mm256_maddubs_epi16(q3h_3, q8_3);
+
+                let p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
+                let p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
+                let p16_2 = _mm256_maddubs_epi16(q3l_2, q8_2);
+                let p16_3 = _mm256_maddubs_epi16(q3l_3, q8_3);
+
+                let p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+                let p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+                let p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
+                let p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+
+                // multiply with scales
+                let p16_0 =
+                    _mm256_madd_epi16(_mm256_shuffle_epi8(*scale, get_scale_shuffle_q3k(0)), p16_0);
+                let p16_1 =
+                    _mm256_madd_epi16(_mm256_shuffle_epi8(*scale, get_scale_shuffle_q3k(1)), p16_1);
+                let p16_2 =
+                    _mm256_madd_epi16(_mm256_shuffle_epi8(*scale, get_scale_shuffle_q3k(2)), p16_2);
+                let p16_3 =
+                    _mm256_madd_epi16(_mm256_shuffle_epi8(*scale, get_scale_shuffle_q3k(3)), p16_3);
+
+                // accumulate
+                let p16_0 = _mm256_add_epi32(p16_0, p16_1);
+                let p16_2 = _mm256_add_epi32(p16_2, p16_3);
+                sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_2));
+            }
+
+            // multiply with block scale and accumulate
+            acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+        }
+        Ok(hsum_float_8(acc))
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q4k_q8k(n: usize, xs: &[BlockQ4K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q4k_q8k: {n} is not divisible by {QK_K}")
+    }
+    let mut utmp = [0u32; 4];
+    const KMASK1: u32 = 0x3f3f3f3f;
+    const KMASK2: u32 = 0x0f0f0f0f;
+    const KMASK3: u32 = 0x03030303;
+
+    unsafe {
+        let m4 = _mm256_set1_epi8(0xF);
+
+        let mut acc = _mm256_setzero_ps();
+        let mut acc_m = _mm_setzero_ps();
+
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = y.d * x.d.to_f32();
+            let dmin = -y.d * x.dmin.to_f32();
+
+            LittleEndian::read_u32_into(&x.scales, &mut utmp[0..3]);
+
+            utmp[3] = ((utmp[2] >> 4) & KMASK2) | (((utmp[1] >> 6) & KMASK3) << 4);
+            let uaux = utmp[1] & KMASK1;
+            utmp[1] = (utmp[2] & KMASK2) | (((utmp[0] >> 6) & KMASK3) << 4);
+            utmp[2] = uaux;
+            utmp[0] &= KMASK1;
+
+            let mut q4 = x.qs.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+
+            let mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(
+                utmp[3] as i32,
+                utmp[2] as i32,
+                utmp[1] as i32,
+                utmp[0] as i32,
+            ));
+
+            let q8sums = _mm256_loadu_si256(y.bsums.as_ptr() as *const __m256i);
+            let q8s = _mm_hadd_epi16(
+                _mm256_extracti128_si256(q8sums, 0),
+                _mm256_extracti128_si256(q8sums, 1),
+            );
+            let prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+            acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
+
+            let sc128 = _mm256_extracti128_si256(mins_and_scales, 0);
+            let scales = mm256_set_m128i(sc128, sc128);
+
+            let mut sumi = _mm256_setzero_si256();
+
+            for j in 0..QK_K / 64 {
+                let scale_l = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2 * j));
+                let scale_h = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2 * j + 1));
+
+                let q4bits = _mm256_loadu_si256(q4 as *const __m256i);
+                q4 = q4.add(32);
+                let q4l = _mm256_and_si256(q4bits, m4);
+                let q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
+
+                let q8l = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let p16l = _mm256_maddubs_epi16(q4l, q8l);
+                let p16l = _mm256_madd_epi16(scale_l, p16l);
+                sumi = _mm256_add_epi32(sumi, p16l);
+
+                let q8h = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let p16h = _mm256_maddubs_epi16(q4h, q8h);
+                let p16h = _mm256_madd_epi16(scale_h, p16h);
+                sumi = _mm256_add_epi32(sumi, p16h);
+            }
+
+            let vd = _mm256_set1_ps(d);
+            acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+        }
+
+        let acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+        let acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+        Ok(hsum_float_8(acc) + _mm_cvtss_f32(acc_m))
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q5k_q8k(n: usize, xs: &[BlockQ5K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q5k_q8k: {n} is not divisible by {QK_K}")
+    }
+    let mut utmp = [0u32; 4];
+    const KMASK1: u32 = 0x3f3f3f3f;
+    const KMASK2: u32 = 0x0f0f0f0f;
+    const KMASK3: u32 = 0x03030303;
+
+    unsafe {
+        let m4 = _mm256_set1_epi8(0xF);
+        let mzero = _mm_setzero_si128();
+        let mone = _mm256_set1_epi8(1);
+
+        let mut acc = _mm256_setzero_ps();
+        let mut summs = 0.0;
+
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = y.d * x.d.to_f32();
+            let dmin = -y.d * x.dmin.to_f32();
+
+            LittleEndian::read_u32_into(&x.scales, &mut utmp[0..3]);
+
+            utmp[3] = ((utmp[2] >> 4) & KMASK2) | (((utmp[1] >> 6) & KMASK3) << 4);
+            let uaux = utmp[1] & KMASK1;
+            utmp[1] = (utmp[2] & KMASK2) | (((utmp[0] >> 6) & KMASK3) << 4);
+            utmp[2] = uaux;
+            utmp[0] &= KMASK1;
+
+            let mut q5 = x.qs.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+
+            let mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(
+                utmp[3] as i32,
+                utmp[2] as i32,
+                utmp[1] as i32,
+                utmp[0] as i32,
+            ));
+
+            let q8sums = _mm256_loadu_si256(y.bsums.as_ptr() as *const __m256i);
+            let q8s = _mm_hadd_epi16(
+                _mm256_extracti128_si256(q8sums, 0),
+                _mm256_extracti128_si256(q8sums, 1),
+            );
+            let prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+            let hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+            summs += dmin * _mm_extract_epi32(hsum, 0) as f32;
+
+            let sc128 = _mm256_extracti128_si256(mins_and_scales, 0);
+            let scales = mm256_set_m128i(sc128, sc128);
+
+            let hbits = _mm256_loadu_si256(x.qh.as_ptr() as *const __m256i);
+            let mut hmask = mone;
+
+            let mut sumi = _mm256_setzero_si256();
+
+            for j in 0..QK_K / 64 {
+                let scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2 * j));
+                let scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2 * j + 1));
+
+                let q5bits = _mm256_loadu_si256(q5 as *const __m256i);
+                q5 = q5.add(32);
+
+                //Similar to q3k we hardcode the shifts here to avoid loading them into a separate register
+                let q5l_0 = _mm256_and_si256(q5bits, m4);
+                let q5l_0_shift_input = _mm256_and_si256(hbits, hmask);
+                let q5l_0_right_shift = match j {
+                    0 => _mm256_srli_epi16(q5l_0_shift_input, 0),
+                    1 => _mm256_srli_epi16(q5l_0_shift_input, 2),
+                    2 => _mm256_srli_epi16(q5l_0_shift_input, 4),
+                    3 => _mm256_srli_epi16(q5l_0_shift_input, 6),
+                    _ => unreachable!(),
+                };
+                let q5h_0 = _mm256_slli_epi16(q5l_0_right_shift, 4);
+                let q5_0 = _mm256_add_epi8(q5l_0, q5h_0);
+                hmask = _mm256_slli_epi16(hmask, 1);
+
+                let q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
+                let q5l_1_shift_input = _mm256_and_si256(hbits, hmask);
+                let q5l_1_right_shift = match j {
+                    0 => _mm256_srli_epi16(q5l_1_shift_input, 1),
+                    1 => _mm256_srli_epi16(q5l_1_shift_input, 3),
+                    2 => _mm256_srli_epi16(q5l_1_shift_input, 5),
+                    3 => _mm256_srli_epi16(q5l_1_shift_input, 7),
+                    _ => unreachable!(),
+                };
+
+                let q5h_1 = _mm256_slli_epi16(q5l_1_right_shift, 4);
+                let q5_1 = _mm256_add_epi8(q5l_1, q5h_1);
+                hmask = _mm256_slli_epi16(hmask, 1);
+
+                let q8_0 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+                let q8_1 = _mm256_loadu_si256(q8 as *const __m256i);
+                q8 = q8.add(32);
+
+                let p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
+                let p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
+
+                let p16_0 = _mm256_madd_epi16(scale_0, p16_0);
+                let p16_1 = _mm256_madd_epi16(scale_1, p16_1);
+
+                sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+            }
+            let vd = _mm256_set1_ps(d);
+            acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+        }
+        Ok(hsum_float_8(acc) + summs)
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q8k_q8k(n: usize, xs: &[BlockQ8K], ys: &[BlockQ8K]) -> Result<f32> {
+    let qk = QK_K;
+    if n % qk != 0 {
+        crate::bail!("vec_dot_q8k_8k: {n} is not divisible by {qk}")
+    }
+
+    unsafe {
+        let mut acc = _mm256_setzero_ps();
+        for (xs, ys) in xs.iter().zip(ys.iter()) {
+            let mut sumi = _mm256_setzero_si256();
+            let x_qs = xs.qs.as_ptr();
+            let y_qs = ys.qs.as_ptr();
+            for j in (0..QK_K).step_by(32) {
+                let xs = _mm256_loadu_si256(x_qs.add(j) as *const __m256i);
+                let ys = _mm256_loadu_si256(y_qs.add(j) as *const __m256i);
+
+                let xs0 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(xs, 0));
+                let ys0 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(ys, 0));
+                sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(xs0, ys0));
+
+                let xs1 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(xs, 1));
+                let ys1 = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(ys, 1));
+                sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(xs1, ys1));
+            }
+            let d = _mm256_set1_ps(xs.d * ys.d);
+            acc = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi), acc);
+        }
+        Ok(hsum_float_8(acc))
+    }
+}

candle-core/src/quantized/cuda.rs

@@ -0,0 +1,680 @@
+use super::{GgmlDType, QStorage};
+use crate::quantized::k_quants::GgmlType;
+use crate::{backend::BackendDevice, cuda_backend::WrapErr};
+use crate::{CudaDevice, CudaStorage, Result};
+use half::f16;
+
+use cudarc::driver::{CudaSlice, CudaView, DeviceSlice};
+
+#[derive(Clone, Debug)]
+pub struct QCudaStorage {
+    data: CudaSlice<u8>,
+    dtype: GgmlDType,
+    device: CudaDevice,
+}
+
+static FORCE_DMMV: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);
+
+pub fn set_force_dmmv(f: bool) {
+    FORCE_DMMV.store(f, std::sync::atomic::Ordering::Relaxed)
+}
+
+pub const WARP_SIZE: usize = 32;
+pub const MMQ_X_Q4_0_AMPERE: usize = 4;
+pub const MMQ_Y_Q4_0_AMPERE: usize = 32;
+pub const NWARPS_Q4_0_AMPERE: usize = 4;
+pub const GGML_CUDA_MMV_X: usize = 32;
+pub const GGML_CUDA_MMV_Y: usize = 1;
+pub const CUDA_QUANTIZE_BLOCK_SIZE: usize = 256;
+pub const CUDA_DEQUANTIZE_BLOCK_SIZE: usize = 256;
+pub const MATRIX_ROW_PADDING: usize = 512;
+
+fn ceil_div(p: usize, q: usize) -> usize {
+    (p + q - 1) / q
+}
+
+fn pad(p: usize, q: usize) -> usize {
+    ceil_div(p, q) * q
+}
+
+fn quantize_q8_1(
+    src: &CudaView<f32>,
+    dst: &mut CudaSlice<u8>,
+    elem_count: usize,
+    ky: usize,
+    dev: &CudaDevice,
+) -> Result<()> {
+    use cudarc::driver::LaunchAsync;
+
+    let kx = elem_count;
+    let kx_padded = pad(kx, MATRIX_ROW_PADDING);
+    let num_blocks = ceil_div(kx_padded, CUDA_QUANTIZE_BLOCK_SIZE);
+    let func = dev.get_or_load_func("quantize_q8_1", candle_kernels::QUANTIZED)?;
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (num_blocks as u32, ky as u32, 1),
+        block_dim: (CUDA_QUANTIZE_BLOCK_SIZE as u32, 1, 1),
+        shared_mem_bytes: 0,
+    };
+    let params = (src, dst, kx as i32, kx_padded as i32);
+    unsafe { func.launch(cfg, params) }.w()?;
+    Ok(())
+}
+
+fn dequantize_f32(
+    data: &CudaSlice<u8>,
+    dtype: GgmlDType,
+    elem_count: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    use cudarc::driver::LaunchAsync;
+
+    let nb = (elem_count + 255) / 256;
+    let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
+        GgmlDType::Q4_0 => ("dequantize_block_q4_0_f32", false, 32, nb),
+        GgmlDType::Q4_1 => ("dequantize_block_q4_1_f32", false, 32, nb),
+        GgmlDType::Q5_0 => (
+            "dequantize_block_q5_0_f32",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q5_1 => (
+            "dequantize_block_q5_1_f32",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q8_0 => ("dequantize_block_q8_0_f32", false, 32, nb),
+        GgmlDType::Q2K => ("dequantize_block_q2_K_f32", true, 64, nb),
+        GgmlDType::Q3K => ("dequantize_block_q3_K_f32", true, 64, nb),
+        GgmlDType::Q4K => ("dequantize_block_q4_K_f32", true, 32, nb),
+        GgmlDType::Q5K => ("dequantize_block_q5_K_f32", true, 64, nb),
+        GgmlDType::Q6K => ("dequantize_block_q6_K_f32", true, 64, nb),
+        GgmlDType::Q8K => ("dequantize_block_q8_K_f32", true, 32, nb),
+        _ => crate::bail!("unsupported dtype for dequantize {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(elem_count).w()? };
+    // See e.g.
+    // https://github.com/ggerganov/llama.cpp/blob/cbbd1efa06f8c09f9dff58ff9d9af509cc4c152b/ggml-cuda.cu#L7270
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (num_blocks as u32, 1, 1),
+        block_dim: (block_dim as u32, 1, 1),
+        shared_mem_bytes: 0,
+    };
+
+    if is_k {
+        let params = (data, &dst);
+        unsafe { func.launch(cfg, params) }.w()?;
+    } else {
+        let nb32 = match dtype {
+            GgmlDType::Q5_0 | GgmlDType::Q5_1 => elem_count,
+            _ => elem_count / 32,
+        };
+        let params = (data, &dst, nb32 as i32);
+        unsafe { func.launch(cfg, params) }.w()?;
+    }
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
+fn dequantize_f16(
+    data: &CudaSlice<u8>,
+    dtype: GgmlDType,
+    elem_count: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    use cudarc::driver::LaunchAsync;
+
+    let nb = (elem_count + 255) / 256;
+    let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
+        GgmlDType::Q4_0 => ("dequantize_block_q4_0_f16", false, 32, nb),
+        GgmlDType::Q4_1 => ("dequantize_block_q4_1_f16", false, 32, nb),
+        GgmlDType::Q5_0 => (
+            "dequantize_block_q5_0_f16",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q5_1 => (
+            "dequantize_block_q5_1_f16",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q8_0 => ("dequantize_block_q8_0_f16", false, 32, nb),
+        GgmlDType::Q2K => ("dequantize_block_q2_K_f16", true, 64, nb),
+        GgmlDType::Q3K => ("dequantize_block_q3_K_f16", true, 64, nb),
+        GgmlDType::Q4K => ("dequantize_block_q4_K_f16", true, 32, nb),
+        GgmlDType::Q5K => ("dequantize_block_q5_K_f16", true, 64, nb),
+        GgmlDType::Q6K => ("dequantize_block_q6_K_f16", true, 64, nb),
+        GgmlDType::Q8K => ("dequantize_block_q8_K_f16", true, 32, nb),
+        _ => crate::bail!("unsupported dtype for dequantize {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f16>(elem_count).w()? };
+    // See e.g.
+    // https://github.com/ggerganov/llama.cpp/blob/cbbd1efa06f8c09f9dff58ff9d9af509cc4c152b/ggml-cuda.cu#L7270
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (num_blocks as u32, 1, 1),
+        block_dim: (block_dim as u32, 1, 1),
+        shared_mem_bytes: 0,
+    };
+
+    if is_k {
+        let params = (data, &dst);
+        unsafe { func.launch(cfg, params) }.w()?;
+    } else {
+        let nb32 = match dtype {
+            GgmlDType::Q5_0 | GgmlDType::Q5_1 => elem_count,
+            _ => elem_count / 32,
+        };
+        let params = (data, &dst, nb32 as i32);
+        unsafe { func.launch(cfg, params) }.w()?;
+    }
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
+fn dequantize_mul_mat_vec(
+    data: &CudaSlice<u8>,
+    y: &CudaView<f32>,
+    dtype: GgmlDType,
+    ncols: usize,
+    nrows: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    use cudarc::driver::LaunchAsync;
+
+    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    if data_elems < ncols * nrows {
+        crate::bail!("unexpected data size {}, ncols {ncols} {nrows}", data_elems)
+    }
+    if y.len() != ncols {
+        crate::bail!("unexpected y size {}, ncols {ncols} {nrows}", y.len())
+    }
+    let kernel_name = match dtype {
+        GgmlDType::Q4_0 => "dequantize_mul_mat_vec_q4_0_cuda",
+        GgmlDType::Q4_1 => "dequantize_mul_mat_vec_q4_1_cuda",
+        GgmlDType::Q5_0 => "dequantize_mul_mat_vec_q5_0_cuda",
+        GgmlDType::Q5_1 => "dequantize_mul_mat_vec_q5_1_cuda",
+        GgmlDType::Q8_0 => "dequantize_mul_mat_vec_q8_0_cuda",
+        GgmlDType::Q2K => "dequantize_mul_mat_vec_q2_k",
+        GgmlDType::Q3K => "dequantize_mul_mat_vec_q3_k",
+        GgmlDType::Q4K => "dequantize_mul_mat_vec_q4_k",
+        GgmlDType::Q5K => "dequantize_mul_mat_vec_q5_k",
+        GgmlDType::Q6K => "dequantize_mul_mat_vec_q6_k",
+        _ => crate::bail!("unsupported dtype for quantized matmul {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(nrows).w()? };
+    let block_num_y = ceil_div(nrows, GGML_CUDA_MMV_Y);
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (block_num_y as u32, 1, 1),
+        block_dim: (WARP_SIZE as u32, GGML_CUDA_MMV_Y as u32, 1),
+        shared_mem_bytes: 0,
+    };
+
+    let params = (data, y, &dst, ncols as i32, nrows as i32);
+    unsafe { func.launch(cfg, params) }.w()?;
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
+fn mul_mat_vec_via_q8_1(
+    data: &CudaSlice<u8>,
+    y: &CudaView<f32>,
+    dtype: GgmlDType,
+    ncols: usize,
+    nrows: usize,
+    b_size: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    use cudarc::driver::LaunchAsync;
+
+    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    if data_elems < ncols * nrows {
+        crate::bail!("unexpected data size {}, ncols {ncols} {nrows}", data_elems)
+    }
+    if y.len() != ncols * b_size {
+        crate::bail!("unexpected y size {}, ncols {ncols} {nrows}", y.len())
+    }
+    if b_size == 0 || b_size > 8 {
+        crate::bail!("only bsize between 1 and 8 are supported, got {b_size}")
+    }
+    // Start by quantizing y
+    let ncols_padded = pad(ncols, MATRIX_ROW_PADDING);
+    let y_size_in_bytes =
+        b_size * ncols_padded * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
+    let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes).w()? };
+    quantize_q8_1(y, &mut y_q8_1, ncols, b_size, dev)?;
+
+    let kernel_name = match dtype {
+        GgmlDType::Q4_0 => "mul_mat_vec_q4_0_q8_1_cuda",
+        GgmlDType::Q4_1 => "mul_mat_vec_q4_1_q8_1_cuda",
+        GgmlDType::Q5_0 => "mul_mat_vec_q5_0_q8_1_cuda",
+        GgmlDType::Q5_1 => "mul_mat_vec_q5_1_q8_1_cuda",
+        GgmlDType::Q8_0 => "mul_mat_vec_q8_0_q8_1_cuda",
+        GgmlDType::Q2K => "mul_mat_vec_q2_K_q8_1_cuda",
+        GgmlDType::Q3K => "mul_mat_vec_q3_K_q8_1_cuda",
+        GgmlDType::Q4K => "mul_mat_vec_q4_K_q8_1_cuda",
+        GgmlDType::Q5K => "mul_mat_vec_q5_K_q8_1_cuda",
+        GgmlDType::Q6K => "mul_mat_vec_q6_K_q8_1_cuda",
+        _ => crate::bail!("unsupported dtype for quantized matmul {dtype:?}"),
+    };
+    let kernel_name = format!("{kernel_name}{b_size}");
+    let func = dev.get_or_load_func(&kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(nrows * b_size).w()? };
+    // https://github.com/ggerganov/llama.cpp/blob/facb8b56f8fd3bb10a693bf0943ae9d69d0828ef/ggml-cuda/mmvq.cu#L98
+    let (nblocks, nwarps) = match b_size {
+        1 => (nrows as u32, 4),
+        2..=4 => ((nrows as u32 + 1) / 2, 4),
+        5..=8 => ((nrows as u32 + 1) / 2, 2),
+        _ => crate::bail!("unexpected bsize {b_size}"),
+    };
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (nblocks, 1, 1),
+        block_dim: (WARP_SIZE as u32, nwarps, 1),
+        shared_mem_bytes: 0,
+    };
+
+    let params = (
+        data,
+        &y_q8_1,
+        &dst,
+        /* ncols_x */ ncols as i32,
+        /* nrows_x */ nrows as i32,
+        /* nrows_y */ ncols_padded as i32,
+        /* nrows_dst */ nrows as i32,
+    );
+    unsafe { func.launch(cfg, params) }.w()?;
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
+#[allow(clippy::too_many_arguments)]
+fn mul_mat_via_q8_1(
+    data: &CudaSlice<u8>,
+    y: &CudaView<f32>,
+    dtype: GgmlDType,
+    x_rows: usize,
+    x_cols: usize,
+    y_rows: usize,
+    y_cols: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    use cudarc::driver::LaunchAsync;
+
+    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    if data_elems < x_rows * x_cols {
+        crate::bail!("unexpected lhs size {}, {x_rows} {x_cols}", data_elems)
+    }
+    if y.len() != y_rows * y_cols {
+        crate::bail!("unexpected y size {}, {y_rows} {y_cols}", y.len())
+    }
+    if x_cols != y_rows {
+        crate::bail!("unexpected x/y size {x_rows} {x_cols} {y_rows} {y_cols}")
+    }
+    let k = x_cols;
+    // Start by quantizing y
+    let k_padded = pad(k, MATRIX_ROW_PADDING);
+    let y_size_in_bytes =
+        k_padded * y_rows * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
+    let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes).w()? };
+    quantize_q8_1(y, &mut y_q8_1, k, y_cols, dev)?;
+
+    let (kernel_name, mmq_x, mmq_y) = match dtype {
+        GgmlDType::Q4_0 => ("mul_mat_q4_0", 64, 128),
+        GgmlDType::Q4_1 => ("mul_mat_q4_1", 64, 128),
+        GgmlDType::Q5_0 => ("mul_mat_q5_0", 128, 64),
+        GgmlDType::Q5_1 => ("mul_mat_q5_1", 128, 64),
+        GgmlDType::Q8_0 => ("mul_mat_q8_0", 128, 64),
+        GgmlDType::Q2K => ("mul_mat_q2_K", 64, 128),
+        GgmlDType::Q3K => ("mul_mat_q3_K", 128, 128),
+        GgmlDType::Q4K => ("mul_mat_q4_K", 64, 128),
+        GgmlDType::Q5K => ("mul_mat_q5_K", 64, 128),
+        GgmlDType::Q6K => ("mul_mat_q6_K", 64, 64),
+        _ => crate::bail!("unsupported dtype for quantized matmul {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(x_rows * y_cols).w()? };
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (
+            ceil_div(x_rows, mmq_y) as u32,
+            ceil_div(y_cols, mmq_x) as u32,
+            1,
+        ),
+        block_dim: (WARP_SIZE as u32, 4, 1),
+        shared_mem_bytes: 0,
+    };
+
+    let params = (
+        /* vx */ data,
+        /* vy */ &y_q8_1,
+        /* dst */ &dst,
+        /* ncols_x */ x_cols as i32,
+        /* nrows_x */ x_rows as i32,
+        /* ncols_y */ y_cols as i32,
+        /* nrows_y */ k_padded as i32,
+        /* nrows_dst */ x_rows as i32,
+    );
+    unsafe { func.launch(cfg, params) }.w()?;
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
+impl QCudaStorage {
+    pub fn zeros(device: &CudaDevice, el_count: usize, dtype: GgmlDType) -> Result<Self> {
+        let size_in_bytes = ceil_div(el_count, dtype.block_size()) * dtype.type_size();
+        let data = device.alloc_zeros::<u8>(size_in_bytes).w()?;
+        Ok(QCudaStorage {
+            data,
+            device: device.clone(),
+            dtype,
+        })
+    }
+
+    pub fn dtype(&self) -> GgmlDType {
+        self.dtype
+    }
+
+    pub fn device(&self) -> &CudaDevice {
+        &self.device
+    }
+
+    pub fn dequantize(&self, elem_count: usize) -> Result<CudaStorage> {
+        fn deq<T: GgmlType>(buffer: &[u8], n: usize, dst: &mut [f32]) -> Result<()> {
+            let slice = unsafe { std::slice::from_raw_parts(buffer.as_ptr() as *const T, n) };
+            let vec = slice.to_vec();
+            T::to_float(&vec, dst)
+        }
+
+        let fast_kernel = matches!(
+            self.dtype,
+            GgmlDType::Q4_0
+                | GgmlDType::Q4_1
+                | GgmlDType::Q5_0
+                | GgmlDType::Q5_1
+                | GgmlDType::Q8_0
+                | GgmlDType::Q2K
+                | GgmlDType::Q3K
+                | GgmlDType::Q4K
+                | GgmlDType::Q5K
+                | GgmlDType::Q6K
+                | GgmlDType::Q8K
+        );
+        if fast_kernel {
+            return dequantize_f32(&self.data, self.dtype, elem_count, self.device());
+        }
+        // Run the dequantization on cpu.
+
+        let buffer = self.device.dtoh_sync_copy(&self.data).w()?;
+        let mut out = vec![0.0; elem_count];
+        let block_len = elem_count / self.dtype.block_size();
+        match self.dtype {
+            GgmlDType::F32 => deq::<f32>(&buffer, block_len, &mut out)?,
+            GgmlDType::F16 => deq::<half::f16>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q4_0 => deq::<crate::quantized::BlockQ4_0>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q4_1 => deq::<crate::quantized::BlockQ4_1>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q5_0 => deq::<crate::quantized::BlockQ5_0>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q5_1 => deq::<crate::quantized::BlockQ5_1>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q8_0 => deq::<crate::quantized::BlockQ8_0>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q8_1 => deq::<crate::quantized::BlockQ8_1>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q2K => deq::<crate::quantized::BlockQ2K>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q3K => deq::<crate::quantized::BlockQ3K>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q4K => deq::<crate::quantized::BlockQ4K>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q5K => deq::<crate::quantized::BlockQ5K>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q6K => deq::<crate::quantized::BlockQ6K>(&buffer, block_len, &mut out)?,
+            GgmlDType::Q8K => deq::<crate::quantized::BlockQ8K>(&buffer, block_len, &mut out)?,
+        }
+
+        self.device
+            .storage_from_cpu_storage(&crate::CpuStorage::F32(out))
+    }
+
+    pub fn dequantize_f16(&self, elem_count: usize) -> Result<CudaStorage> {
+        dequantize_f16(&self.data, self.dtype, elem_count, self.device())
+    }
+
+    pub fn quantize(&mut self, src: &CudaStorage) -> Result<()> {
+        // Run the quantization on cpu.
+        let src = match &src.slice {
+            crate::cuda_backend::CudaStorageSlice::F32(data) => {
+                self.device.dtoh_sync_copy(data).w()?
+            }
+            _ => crate::bail!("only f32 can be quantized"),
+        };
+        let src_len = src.len();
+        let src = crate::Storage::Cpu(crate::CpuStorage::F32(src));
+        let mut qcpu_storage = crate::Device::Cpu.qzeros(src_len, self.dtype)?;
+        qcpu_storage.quantize(&src)?;
+        let data = qcpu_storage.data()?;
+        let data = self.device.htod_sync_copy(data.as_ref()).w()?;
+        self.data = data;
+        Ok(())
+    }
+
+    pub fn storage_size_in_bytes(&self) -> usize {
+        self.data.len()
+    }
+
+    pub fn fwd(
+        &self,
+        self_shape: &crate::Shape,
+        storage: &CudaStorage,
+        layout: &crate::Layout,
+    ) -> Result<(CudaStorage, crate::Shape)> {
+        let max_bm = if FORCE_DMMV.load(std::sync::atomic::Ordering::Relaxed) {
+            1
+        } else {
+            8
+        };
+        let use_vec_kernel = match layout.shape().dims() {
+            [b, m, _k] => b * m <= max_bm,
+            [b, _k] => *b <= max_bm,
+            _ => false,
+        };
+        if use_vec_kernel {
+            self.dequantize_matmul_vec(self_shape, storage, layout)
+        } else {
+            self.dequantize_matmul(self_shape, storage, layout)
+        }
+    }
+}
+
+impl QCudaStorage {
+    fn dequantize_matmul_vec(
+        &self,
+        self_shape: &crate::Shape,
+        rhs: &CudaStorage,
+        rhs_l: &crate::Layout,
+    ) -> Result<(CudaStorage, crate::Shape)> {
+        let (nrows, ncols) = self_shape.dims2()?;
+        let rhs = rhs.as_cuda_slice::<f32>()?;
+        let rhs = match rhs_l.contiguous_offsets() {
+            Some((o1, o2)) => rhs.slice(o1..o2),
+            None => Err(crate::Error::RequiresContiguous { op: "dmmv" }.bt())?,
+        };
+        let (b_size, k) = match rhs_l.shape().dims() {
+            [b, m, k] => (b * m, *k),
+            [b, k] => (*b, *k),
+            _ => crate::bail!("unexpected rhs shape in dmmv {:?}", rhs_l.shape()),
+        };
+        if ncols != k {
+            crate::bail!("mismatch on matmul dim {self_shape:?} {:?}", rhs_l.shape())
+        }
+
+        let out = if FORCE_DMMV.load(std::sync::atomic::Ordering::Relaxed) {
+            dequantize_mul_mat_vec(&self.data, &rhs, self.dtype, ncols, nrows, self.device())?
+        } else {
+            mul_mat_vec_via_q8_1(
+                &self.data,
+                &rhs,
+                self.dtype,
+                ncols,
+                nrows,
+                b_size,
+                self.device(),
+            )?
+        };
+        let mut out_shape = rhs_l.shape().dims().to_vec();
+        out_shape.pop();
+        out_shape.push(nrows);
+        Ok((out, out_shape.into()))
+    }
+
+    fn dequantize_matmul(
+        &self,
+        self_shape: &crate::Shape,
+        storage: &CudaStorage,
+        layout: &crate::Layout,
+    ) -> Result<(CudaStorage, crate::Shape)> {
+        use crate::backend::BackendStorage;
+        let (n, k) = self_shape.dims2()?;
+        let (b, m, k2) = match layout.shape().dims() {
+            &[b, m, k2] => (b, m, k2),
+            &[m, k2] => (1, m, k2),
+            s => crate::bail!("unexpected shape for input {s:?}"),
+        };
+        if k2 != k {
+            crate::bail!("mismatch on matmul dim {self_shape:?} {:?}", layout.shape())
+        }
+
+        let out = if FORCE_DMMV.load(std::sync::atomic::Ordering::Relaxed) {
+            let data_f32 = self.dequantize(n * k)?;
+            let rhs_l = crate::Layout::new((k, n).into(), vec![1, k], 0).broadcast_as((b, k, n))?;
+            storage.matmul(&data_f32, (b, m, n, k), layout, &rhs_l)?
+        } else {
+            let storage = storage.as_cuda_slice::<f32>()?;
+            let storage = match layout.contiguous_offsets() {
+                Some((o1, o2)) => storage.slice(o1..o2),
+                None => Err(crate::Error::RequiresContiguous {
+                    op: "quantized-matmul",
+                }
+                .bt())?,
+            };
+            mul_mat_via_q8_1(
+                &self.data,
+                &storage,
+                self.dtype,
+                /* x_rows */ n,
+                /* x_cols */ k,
+                /* y_rows */ k,
+                /* y_cols */ b * m,
+                self.device(),
+            )?
+        };
+        let mut out_shape = layout.shape().dims().to_vec();
+        out_shape.pop();
+        out_shape.push(n);
+        Ok((out, out_shape.into()))
+    }
+}
+
+pub fn load_quantized<T: super::GgmlType + Send + Sync + 'static>(
+    device: &CudaDevice,
+    data: &[T],
+) -> Result<super::QStorage> {
+    let data = unsafe {
+        std::slice::from_raw_parts(data.as_ptr() as *const u8, core::mem::size_of_val(data))
+    };
+    let data = device.htod_sync_copy(data).w()?;
+    Ok(QStorage::Cuda(QCudaStorage {
+        data,
+        device: device.clone(),
+        dtype: T::DTYPE,
+    }))
+}
+
+#[cfg(test)]
+mod test {
+    use super::*;
+
+    #[test]
+    fn cuda_quantize_q8_1() -> Result<()> {
+        let dev = CudaDevice::new(0)?;
+        let el = 256;
+        let el_padded = pad(el, MATRIX_ROW_PADDING);
+        let y_size_in_bytes =
+            el_padded * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
+        let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes).w()? };
+        let vs: Vec<f32> = (0..el).map(|v| v as f32).collect();
+        let y = dev.htod_sync_copy(&vs).w()?;
+        quantize_q8_1(&y.slice(..), &mut y_q8_1, el, 1, &dev)?;
+        Ok(())
+    }
+
+    #[test]
+    fn cuda_mmv_q8_1() -> Result<()> {
+        let dev = CudaDevice::new(0)?;
+        let ncols = 256;
+        let vs: Vec<f32> = (0..ncols).map(|v| v as f32).collect();
+        let y = dev.htod_sync_copy(&vs).w()?;
+        let mut xs = QCudaStorage::zeros(&dev, ncols, GgmlDType::Q4_0)?;
+        xs.quantize(&CudaStorage::wrap_cuda_slice(y.clone(), dev.clone()))?;
+        let cuda_storage = mul_mat_vec_via_q8_1(
+            &xs.data,
+            &y.slice(..),
+            /* dtype */ GgmlDType::Q4_0,
+            /* ncols */ ncols,
+            /* nrows */ 1,
+            /* b_size */ 1,
+            &dev,
+        )?;
+        let vs = cuda_storage.as_cuda_slice::<f32>()?;
+        let vs = dev.dtoh_sync_copy(&vs.slice(..)).unwrap();
+        assert_eq!(vs.len(), 1);
+        // for n = 255, n.(n+1).(2n+1) / 6 = 5559680
+        // Q8 means 1/256 precision.
+        assert_eq!(vs[0], 5561664.5);
+
+        let cuda_storage = dequantize_mul_mat_vec(
+            &xs.data,
+            &y.slice(..),
+            /* dtype */ GgmlDType::Q4_0,
+            /* ncols */ ncols,
+            /* nrows */ 1,
+            &dev,
+        )?;
+        let vs = cuda_storage.as_cuda_slice::<f32>()?;
+        let vs = dev.dtoh_sync_copy(&vs.slice(..)).unwrap();
+        assert_eq!(vs.len(), 1);
+        assert_eq!(vs[0], 5561851.0);
+        Ok(())
+    }
+
+    #[test]
+    fn cuda_mm_q8_1() -> Result<()> {
+        let dev = CudaDevice::new(0)?;
+        let ncols = 256;
+        let vs: Vec<f32> = (0..ncols * 4).map(|v| v as f32 / 4.).collect();
+        let y = dev.htod_sync_copy(&vs).w()?;
+        let mut xs = QCudaStorage::zeros(&dev, ncols * 4, GgmlDType::Q4_0)?;
+        xs.quantize(&CudaStorage::wrap_cuda_slice(y.clone(), dev.clone()))?;
+        let cuda_storage = mul_mat_via_q8_1(
+            &xs.data,
+            &y.slice(..),
+            /* dtype */ GgmlDType::Q4_0,
+            /* x_rows */ 4,
+            /* x_cols */ ncols,
+            /* y_rows */ ncols,
+            /* y_cols */ 4,
+            &dev,
+        )?;
+        let vs = cuda_storage.as_cuda_slice::<f32>()?;
+        let vs = dev.dtoh_sync_copy(&vs.slice(..)).unwrap();
+
+        /*
+           x = torch.tensor([float(v) for v in range(1024)]).reshape(4, 256)
+           x @ x.t() / 16
+        tensor([[  347480.0000,   869720.0000,  1391960.0000,  1914200.0000],
+                [  869720.0000,  2440536.0000,  4011352.0000,  5582166.5000],
+                [ 1391960.0000,  4011352.0000,  6630742.0000,  9250132.0000],
+                [ 1914200.0000,  5582166.5000,  9250132.0000, 12918099.0000]])
+                */
+        assert_eq!(vs.len(), 16);
+        assert_eq!(vs[0], 347604.0);
+        assert_eq!(vs[1], 888153.06);
+        assert_eq!(vs[4], 869780.7);
+        assert_eq!(vs[5], 2483145.0);
+        assert_eq!(vs[11], 9407368.0);
+        assert_eq!(vs[14], 9470856.0);
+        assert_eq!(vs[15], 13138824.0);
+        Ok(())
+    }
+}

candle-core/src/quantized/dummy_cuda.rs

@@ -0,0 +1,54 @@
+#![allow(unused)]
+use super::GgmlDType;
+use crate::{CudaDevice, CudaStorage, Error, Result};
+
+pub struct QCudaStorage {
+    dtype: GgmlDType,
+    device: CudaDevice,
+}
+
+impl QCudaStorage {
+    pub fn zeros(_: &CudaDevice, _: usize, _: GgmlDType) -> Result<Self> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    pub fn dtype(&self) -> GgmlDType {
+        self.dtype
+    }
+
+    pub fn device(&self) -> &CudaDevice {
+        &self.device
+    }
+
+    pub fn dequantize(&self, _elem_count: usize) -> Result<CudaStorage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    pub fn dequantize_f16(&self, _elem_count: usize) -> Result<CudaStorage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    pub fn quantize(&mut self, _src: &CudaStorage) -> Result<()> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    pub fn storage_size_in_bytes(&self) -> usize {
+        0
+    }
+
+    pub fn fwd(
+        &self,
+        _self_shape: &crate::Shape,
+        _storage: &CudaStorage,
+        _layout: &crate::Layout,
+    ) -> Result<(CudaStorage, crate::Shape)> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+}
+
+pub fn load_quantized<T: super::GgmlType + Send + Sync + 'static>(
+    _device: &CudaDevice,
+    _data: &[T],
+) -> Result<super::QStorage> {
+    Err(Error::NotCompiledWithCudaSupport)
+}

candle-core/src/quantized/dummy_metal.rs

@@ -0,0 +1,50 @@
+#![allow(unused)]
+use super::GgmlDType;
+use crate::{Error, MetalDevice, MetalStorage, Result};
+
+pub struct QMetalStorage {
+    dtype: GgmlDType,
+    device: MetalDevice,
+}
+
+impl QMetalStorage {
+    pub fn zeros(_: &MetalDevice, _: usize, _: GgmlDType) -> Result<Self> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    pub fn dtype(&self) -> GgmlDType {
+        self.dtype
+    }
+
+    pub fn device(&self) -> &MetalDevice {
+        &self.device
+    }
+
+    pub fn dequantize(&self, _elem_count: usize) -> Result<MetalStorage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    pub fn quantize(&mut self, _src: &MetalStorage) -> Result<()> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    pub fn storage_size_in_bytes(&self) -> usize {
+        0
+    }
+
+    pub fn fwd(
+        &self,
+        _self_shape: &crate::Shape,
+        _storage: &MetalStorage,
+        _layout: &crate::Layout,
+    ) -> Result<(MetalStorage, crate::Shape)> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+}
+
+pub fn load_quantized<T: super::GgmlType + Send + Sync + 'static>(
+    _device: &MetalDevice,
+    _data: &[T],
+) -> Result<super::QStorage> {
+    Err(Error::NotCompiledWithMetalSupport)
+}

candle-core/src/quantized/ggml_file.rs

@@ -1,8 +1,9 @@
 //! Support for the GGML file format.
 
-use super::{k_quants, GgmlDType};
-use crate::Result;
+use super::{k_quants, GgmlDType, QStorage};
+use crate::{Device, Result};
 use byteorder::{LittleEndian, ReadBytesExt};
+use std::collections::HashMap;
 
 // https://github.com/ggerganov/llama.cpp/blob/468ea24fb4633a0d681f7ac84089566c1c6190cb/llama.h#L37
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -120,11 +121,17 @@ fn from_raw_data<T: super::GgmlType + Send + Sync + 'static>(
     raw_data: &[u8],
     size_in_bytes: usize,
     dims: Vec<usize>,
+    device: &Device,
 ) -> Result<super::QTensor> {
     let raw_data_ptr = raw_data.as_ptr();
     let n_blocks = size_in_bytes / std::mem::size_of::<T>();
     let data = unsafe { std::slice::from_raw_parts(raw_data_ptr as *const T, n_blocks) };
-    Ok(super::QTensor::new(data.to_vec(), dims))
+    let data: QStorage = match device {
+        Device::Cpu => QStorage::Cpu(Box::new(data.to_vec())),
+        Device::Metal(metal) => super::metal::load_quantized(metal, data)?,
+        Device::Cuda(cuda) => super::cuda::load_quantized(cuda, data)?,
+    };
+    super::QTensor::new(data, dims)
 }
 
 /// Creates a [Tensor] from a raw GGML tensor.
@@ -132,23 +139,50 @@ pub fn qtensor_from_ggml(
     ggml_dtype: GgmlDType,
     raw_data: &[u8],
     dims: Vec<usize>,
+    device: &Device,
 ) -> Result<super::QTensor> {
     let tensor_elems = dims.iter().product::<usize>();
-    let size_in_bytes = tensor_elems * ggml_dtype.type_size() / ggml_dtype.blck_size();
+    let block_size = ggml_dtype.block_size();
+    if tensor_elems % block_size != 0 {
+        crate::bail!(
+            "the number of elements {tensor_elems} is not divisible by the block size {block_size}"
+        )
+    }
+    let size_in_bytes = tensor_elems / block_size * ggml_dtype.type_size();
 
     match ggml_dtype {
-        GgmlDType::F32 => from_raw_data::<f32>(raw_data, size_in_bytes, dims),
-        GgmlDType::F16 => from_raw_data::<half::f16>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q4_0 => from_raw_data::<k_quants::BlockQ4_0>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q4_1 => from_raw_data::<k_quants::BlockQ4_1>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q5_0 => from_raw_data::<k_quants::BlockQ5_0>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q5_1 => from_raw_data::<k_quants::BlockQ5_1>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q8_0 => from_raw_data::<k_quants::BlockQ8_0>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q2K => from_raw_data::<k_quants::BlockQ2K>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q3K => from_raw_data::<k_quants::BlockQ3K>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q4K => from_raw_data::<k_quants::BlockQ4K>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q5K => from_raw_data::<k_quants::BlockQ5K>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q6K => from_raw_data::<k_quants::BlockQ6K>(raw_data, size_in_bytes, dims),
+        GgmlDType::F32 => from_raw_data::<f32>(raw_data, size_in_bytes, dims, device),
+        GgmlDType::F16 => from_raw_data::<half::f16>(raw_data, size_in_bytes, dims, device),
+        GgmlDType::Q4_0 => {
+            from_raw_data::<k_quants::BlockQ4_0>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q4_1 => {
+            from_raw_data::<k_quants::BlockQ4_1>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q5_0 => {
+            from_raw_data::<k_quants::BlockQ5_0>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q5_1 => {
+            from_raw_data::<k_quants::BlockQ5_1>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q8_0 => {
+            from_raw_data::<k_quants::BlockQ8_0>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q2K => {
+            from_raw_data::<k_quants::BlockQ2K>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q3K => {
+            from_raw_data::<k_quants::BlockQ3K>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q4K => {
+            from_raw_data::<k_quants::BlockQ4K>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q5K => {
+            from_raw_data::<k_quants::BlockQ5K>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q6K => {
+            from_raw_data::<k_quants::BlockQ6K>(raw_data, size_in_bytes, dims, device)
+        }
         _ => crate::bail!("quantized type {ggml_dtype:?} is not supported yet"),
     }
 }
@@ -156,6 +190,7 @@ pub fn qtensor_from_ggml(
 fn read_one_tensor<R: std::io::Seek + std::io::Read>(
     reader: &mut R,
     magic: VersionedMagic,
+    device: &Device,
 ) -> Result<(String, super::QTensor)> {
     let n_dims = reader.read_u32::<LittleEndian>()?;
     let name_len = reader.read_u32::<LittleEndian>()?;
@@ -163,6 +198,9 @@ fn read_one_tensor<R: std::io::Seek + std::io::Read>(
     let ggml_dtype = GgmlDType::from_u32(ggml_dtype)?;
     let mut dims = vec![0u32; n_dims as usize];
     reader.read_u32_into::<LittleEndian>(&mut dims)?;
+    // The dimensions are stored in reverse order, see for example:
+    // https://github.com/ggerganov/llama.cpp/blob/b5ffb2849d23afe73647f68eec7b68187af09be6/convert.py#L969
+    dims.reverse();
     let mut name = vec![0u8; name_len as usize];
     reader.read_exact(&mut name)?;
     let name = String::from_utf8_lossy(&name).into_owned();
@@ -173,12 +211,11 @@ fn read_one_tensor<R: std::io::Seek + std::io::Read>(
     }
     let dims = dims.iter().map(|&u| u as usize).collect::<Vec<_>>();
     let tensor_elems = dims.iter().product::<usize>();
-    let size_in_bytes = tensor_elems * ggml_dtype.type_size() / ggml_dtype.blck_size();
-    println!("{name} {ggml_dtype:?} {dims:?}");
+    let size_in_bytes = tensor_elems * ggml_dtype.type_size() / ggml_dtype.block_size();
     // TODO: Mmap version to avoid copying the data around?
     let mut raw_data = vec![0u8; size_in_bytes];
     reader.read_exact(&mut raw_data)?;
-    match qtensor_from_ggml(ggml_dtype, &raw_data, dims) {
+    match qtensor_from_ggml(ggml_dtype, &raw_data, dims, device) {
         Ok(tensor) => Ok((name, tensor)),
         Err(e) => crate::bail!("Error creating tensor {name}: {e}"),
     }
@@ -188,28 +225,41 @@ pub struct Content {
     pub magic: VersionedMagic,
     pub hparams: HParams,
     pub vocab: Vocab,
-    pub tensors: Vec<(String, super::QTensor)>,
+    pub tensors: HashMap<String, super::QTensor>,
+    pub device: Device,
 }
 
 impl Content {
-    pub fn read<R: std::io::Seek + std::io::Read>(reader: &mut R) -> Result<Content> {
+    pub fn read<R: std::io::Seek + std::io::Read>(
+        reader: &mut R,
+        device: &Device,
+    ) -> Result<Content> {
         // https://github.com/ggerganov/llama.cpp/blob/468ea24fb4633a0d681f7ac84089566c1c6190cb/llama.cpp#L505
         let last_position = reader.seek(std::io::SeekFrom::End(0))?;
         reader.seek(std::io::SeekFrom::Start(0))?;
         let magic = VersionedMagic::read(reader)?;
         let hparams = HParams::read(reader)?;
         let vocab = Vocab::read(reader, hparams.n_vocab as usize)?;
-        let mut tensors = vec![];
+        let mut tensors = HashMap::new();
 
         while reader.stream_position()? != last_position {
-            let (name, tensor) = read_one_tensor(reader, magic)?;
-            tensors.push((name, tensor))
+            let (name, tensor) = read_one_tensor(reader, magic, device)?;
+            tensors.insert(name, tensor);
         }
+        let device = device.clone();
         Ok(Self {
             magic,
             hparams,
             vocab,
             tensors,
+            device,
         })
     }
+
+    pub fn remove(&mut self, name: &str) -> Result<super::QTensor> {
+        match self.tensors.remove(name) {
+            None => crate::bail!("cannot find tensor with name '{name}'"),
+            Some(tensor) => Ok(tensor),
+        }
+    }
 }

candle-core/src/quantized/gguf_file.rs

@@ -0,0 +1,539 @@
+//! Support for the GGUF file format.
+//!
+//! Spec: https://github.com/philpax/ggml/blob/gguf-spec/docs/gguf.md
+
+use super::{GgmlDType, QTensor};
+use crate::{Device, Result};
+use byteorder::{LittleEndian, ReadBytesExt, WriteBytesExt};
+use std::collections::HashMap;
+
+pub const DEFAULT_ALIGNMENT: u64 = 32;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum Magic {
+    Gguf,
+}
+
+impl TryFrom<u32> for Magic {
+    type Error = crate::Error;
+    fn try_from(value: u32) -> Result<Self> {
+        let magic = match value {
+            0x46554747 | 0x47475546 => Self::Gguf,
+            _ => crate::bail!("unknown magic 0x{value:08x}"),
+        };
+        Ok(magic)
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum VersionedMagic {
+    GgufV1,
+    GgufV2,
+    GgufV3,
+}
+
+impl VersionedMagic {
+    fn read<R: std::io::Read>(reader: &mut R) -> Result<Self> {
+        let magic = reader.read_u32::<LittleEndian>()?;
+        let magic = Magic::try_from(magic)?;
+        let version = reader.read_u32::<LittleEndian>()?;
+        let versioned_magic = match (magic, version) {
+            (Magic::Gguf, 1) => Self::GgufV1,
+            (Magic::Gguf, 2) => Self::GgufV2,
+            (Magic::Gguf, 3) => Self::GgufV3,
+            _ => crate::bail!("gguf: unsupported magic/version {magic:?}/{version}"),
+        };
+        Ok(versioned_magic)
+    }
+}
+
+#[derive(Debug)]
+pub struct TensorInfo {
+    pub ggml_dtype: GgmlDType,
+    pub shape: crate::Shape,
+    pub offset: u64,
+}
+
+impl TensorInfo {
+    pub fn read<R: std::io::Seek + std::io::Read>(
+        &self,
+        reader: &mut R,
+        tensor_data_offset: u64,
+        device: &Device,
+    ) -> Result<QTensor> {
+        let tensor_elems = self.shape.elem_count();
+        let block_size = self.ggml_dtype.block_size();
+        if tensor_elems % block_size != 0 {
+            crate::bail!(
+            "the number of elements {tensor_elems} is not divisible by the block size {block_size}"
+        )
+        }
+        let size_in_bytes = tensor_elems / block_size * self.ggml_dtype.type_size();
+        let mut raw_data = vec![0u8; size_in_bytes];
+        reader.seek(std::io::SeekFrom::Start(tensor_data_offset + self.offset))?;
+        reader.read_exact(&mut raw_data)?;
+        super::ggml_file::qtensor_from_ggml(
+            self.ggml_dtype,
+            &raw_data,
+            self.shape.dims().to_vec(),
+            device,
+        )
+    }
+}
+
+#[derive(Debug)]
+pub struct Content {
+    pub magic: VersionedMagic,
+    pub metadata: HashMap<String, Value>,
+    pub tensor_infos: HashMap<String, TensorInfo>,
+    pub tensor_data_offset: u64,
+}
+
+fn read_string<R: std::io::Read>(reader: &mut R, magic: &VersionedMagic) -> Result<String> {
+    let len = match magic {
+        VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
+        VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
+            reader.read_u64::<LittleEndian>()? as usize
+        }
+    };
+    let mut v = vec![0u8; len];
+    reader.read_exact(&mut v)?;
+    // GGUF strings are supposed to be non-null terminated but in practice this happens.
+    while let Some(0) = v.last() {
+        v.pop();
+    }
+    // GGUF strings are utf8 encoded but there are cases that don't seem to be valid.
+    Ok(String::from_utf8_lossy(&v).into_owned())
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum ValueType {
+    // The value is a 8-bit unsigned integer.
+    U8,
+    // The value is a 8-bit signed integer.
+    I8,
+    // The value is a 16-bit unsigned little-endian integer.
+    U16,
+    // The value is a 16-bit signed little-endian integer.
+    I16,
+    // The value is a 32-bit unsigned little-endian integer.
+    U32,
+    // The value is a 32-bit signed little-endian integer.
+    I32,
+    // The value is a 64-bit unsigned little-endian integer.
+    U64,
+    // The value is a 64-bit signed little-endian integer.
+    I64,
+    // The value is a 32-bit IEEE754 floating point number.
+    F32,
+    // The value is a 64-bit IEEE754 floating point number.
+    F64,
+    // The value is a boolean.
+    // 1-byte value where 0 is false and 1 is true.
+    // Anything else is invalid, and should be treated as either the model being invalid or the reader being buggy.
+    Bool,
+    // The value is a UTF-8 non-null-terminated string, with length prepended.
+    String,
+    // The value is an array of other values, with the length and type prepended.
+    // Arrays can be nested, and the length of the array is the number of elements in the array, not the number of bytes.
+    Array,
+}
+
+#[derive(Debug, Clone)]
+pub enum Value {
+    U8(u8),
+    I8(i8),
+    U16(u16),
+    I16(i16),
+    U32(u32),
+    I32(i32),
+    U64(u64),
+    I64(i64),
+    F32(f32),
+    F64(f64),
+    Bool(bool),
+    String(String),
+    Array(Vec<Value>),
+}
+
+impl Value {
+    pub fn value_type(&self) -> ValueType {
+        match self {
+            Self::U8(_) => ValueType::U8,
+            Self::I8(_) => ValueType::I8,
+            Self::U16(_) => ValueType::U16,
+            Self::I16(_) => ValueType::I16,
+            Self::U32(_) => ValueType::U32,
+            Self::I32(_) => ValueType::I32,
+            Self::U64(_) => ValueType::U64,
+            Self::I64(_) => ValueType::I64,
+            Self::F32(_) => ValueType::F32,
+            Self::F64(_) => ValueType::F64,
+            Self::Bool(_) => ValueType::Bool,
+            Self::String(_) => ValueType::String,
+            Self::Array(_) => ValueType::Array,
+        }
+    }
+
+    pub fn to_u8(&self) -> Result<u8> {
+        match self {
+            Self::U8(v) => Ok(*v),
+            v => crate::bail!("not a u8 {v:?}"),
+        }
+    }
+
+    pub fn to_i8(&self) -> Result<i8> {
+        match self {
+            Self::I8(v) => Ok(*v),
+            v => crate::bail!("not a i8 {v:?}"),
+        }
+    }
+
+    pub fn to_u16(&self) -> Result<u16> {
+        match self {
+            Self::U16(v) => Ok(*v),
+            v => crate::bail!("not a u16 {v:?}"),
+        }
+    }
+
+    pub fn to_i16(&self) -> Result<i16> {
+        match self {
+            Self::I16(v) => Ok(*v),
+            v => crate::bail!("not a i16 {v:?}"),
+        }
+    }
+
+    pub fn to_u32(&self) -> Result<u32> {
+        match self {
+            Self::U32(v) => Ok(*v),
+            v => crate::bail!("not a u32 {v:?}"),
+        }
+    }
+
+    pub fn to_i32(&self) -> Result<i32> {
+        match self {
+            Self::I32(v) => Ok(*v),
+            v => crate::bail!("not a i32 {v:?}"),
+        }
+    }
+
+    /// This will also automatically upcast any integral types which will not truncate.
+    pub fn to_u64(&self) -> Result<u64> {
+        match self {
+            Self::U64(v) => Ok(*v),
+            // Autoupcast cases here
+            Self::U8(v) => Ok(*v as u64),
+            Self::U16(v) => Ok(*v as u64),
+            Self::U32(v) => Ok(*v as u64),
+            Self::Bool(v) => Ok(*v as u64),
+            v => crate::bail!("not a u64 or upcastable to u64 {v:?}"),
+        }
+    }
+
+    pub fn to_i64(&self) -> Result<i64> {
+        match self {
+            Self::I64(v) => Ok(*v),
+            v => crate::bail!("not a i64 {v:?}"),
+        }
+    }
+
+    pub fn to_f32(&self) -> Result<f32> {
+        match self {
+            Self::F32(v) => Ok(*v),
+            v => crate::bail!("not a f32 {v:?}"),
+        }
+    }
+
+    pub fn to_f64(&self) -> Result<f64> {
+        match self {
+            Self::F64(v) => Ok(*v),
+            v => crate::bail!("not a f64 {v:?}"),
+        }
+    }
+
+    pub fn to_bool(&self) -> Result<bool> {
+        match self {
+            Self::Bool(v) => Ok(*v),
+            v => crate::bail!("not a bool {v:?}"),
+        }
+    }
+
+    pub fn to_vec(&self) -> Result<&Vec<Value>> {
+        match self {
+            Self::Array(v) => Ok(v),
+            v => crate::bail!("not a vec {v:?}"),
+        }
+    }
+
+    pub fn to_string(&self) -> Result<&String> {
+        match self {
+            Self::String(v) => Ok(v),
+            v => crate::bail!("not a string {v:?}"),
+        }
+    }
+
+    fn read<R: std::io::Read>(
+        reader: &mut R,
+        value_type: ValueType,
+        magic: &VersionedMagic,
+    ) -> Result<Self> {
+        let v = match value_type {
+            ValueType::U8 => Self::U8(reader.read_u8()?),
+            ValueType::I8 => Self::I8(reader.read_i8()?),
+            ValueType::U16 => Self::U16(reader.read_u16::<LittleEndian>()?),
+            ValueType::I16 => Self::I16(reader.read_i16::<LittleEndian>()?),
+            ValueType::U32 => Self::U32(reader.read_u32::<LittleEndian>()?),
+            ValueType::I32 => Self::I32(reader.read_i32::<LittleEndian>()?),
+            ValueType::U64 => Self::U64(reader.read_u64::<LittleEndian>()?),
+            ValueType::I64 => Self::I64(reader.read_i64::<LittleEndian>()?),
+            ValueType::F32 => Self::F32(reader.read_f32::<LittleEndian>()?),
+            ValueType::F64 => Self::F64(reader.read_f64::<LittleEndian>()?),
+            ValueType::Bool => match reader.read_u8()? {
+                0 => Self::Bool(false),
+                1 => Self::Bool(true),
+                b => crate::bail!("unexpected bool value {b}"),
+            },
+            ValueType::String => Self::String(read_string(reader, magic)?),
+            ValueType::Array => {
+                let value_type = reader.read_u32::<LittleEndian>()?;
+                let value_type = ValueType::from_u32(value_type)?;
+                let len = match magic {
+                    VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
+                    VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
+                        reader.read_u64::<LittleEndian>()? as usize
+                    }
+                };
+                let mut vs = Vec::with_capacity(len);
+                for _ in 0..len {
+                    vs.push(Value::read(reader, value_type, magic)?)
+                }
+                Self::Array(vs)
+            }
+        };
+        Ok(v)
+    }
+
+    fn write<W: std::io::Write>(&self, w: &mut W) -> Result<()> {
+        match self {
+            &Self::U8(v) => w.write_u8(v)?,
+            &Self::I8(v) => w.write_i8(v)?,
+            &Self::U16(v) => w.write_u16::<LittleEndian>(v)?,
+            &Self::I16(v) => w.write_i16::<LittleEndian>(v)?,
+            &Self::U32(v) => w.write_u32::<LittleEndian>(v)?,
+            &Self::I32(v) => w.write_i32::<LittleEndian>(v)?,
+            &Self::U64(v) => w.write_u64::<LittleEndian>(v)?,
+            &Self::I64(v) => w.write_i64::<LittleEndian>(v)?,
+            &Self::F32(v) => w.write_f32::<LittleEndian>(v)?,
+            &Self::F64(v) => w.write_f64::<LittleEndian>(v)?,
+            &Self::Bool(v) => w.write_u8(u8::from(v))?,
+            Self::String(v) => write_string(w, v.as_str())?,
+            Self::Array(v) => {
+                // The `Value` type does not enforce that all the values in an Array have the same
+                // type.
+                let value_type = if v.is_empty() {
+                    // Doesn't matter, the array is empty.
+                    ValueType::U32
+                } else {
+                    let value_type: std::collections::HashSet<_> =
+                        v.iter().map(|elem| elem.value_type()).collect();
+                    if value_type.len() != 1 {
+                        crate::bail!("multiple value-types in the same array {value_type:?}")
+                    }
+                    value_type.into_iter().next().unwrap()
+                };
+                w.write_u32::<LittleEndian>(value_type.to_u32())?;
+                w.write_u64::<LittleEndian>(v.len() as u64)?;
+                for elem in v.iter() {
+                    elem.write(w)?
+                }
+            }
+        }
+        Ok(())
+    }
+}
+
+impl ValueType {
+    fn from_u32(v: u32) -> Result<Self> {
+        let v = match v {
+            0 => Self::U8,
+            1 => Self::I8,
+            2 => Self::U16,
+            3 => Self::I16,
+            4 => Self::U32,
+            5 => Self::I32,
+            6 => Self::F32,
+            7 => Self::Bool,
+            8 => Self::String,
+            9 => Self::Array,
+            10 => Self::U64,
+            11 => Self::I64,
+            12 => Self::F64,
+            v => crate::bail!("unrecognized value-type {v:#08x}"),
+        };
+        Ok(v)
+    }
+
+    fn to_u32(self) -> u32 {
+        match self {
+            Self::U8 => 0,
+            Self::I8 => 1,
+            Self::U16 => 2,
+            Self::I16 => 3,
+            Self::U32 => 4,
+            Self::I32 => 5,
+            Self::F32 => 6,
+            Self::Bool => 7,
+            Self::String => 8,
+            Self::Array => 9,
+            Self::U64 => 10,
+            Self::I64 => 11,
+            Self::F64 => 12,
+        }
+    }
+}
+
+impl Content {
+    pub fn read<R: std::io::Seek + std::io::Read>(reader: &mut R) -> Result<Self> {
+        let magic = VersionedMagic::read(reader)?;
+
+        let tensor_count = match magic {
+            VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
+            VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
+                reader.read_u64::<LittleEndian>()? as usize
+            }
+        };
+        let metadata_kv_count = match magic {
+            VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
+            VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
+                reader.read_u64::<LittleEndian>()? as usize
+            }
+        };
+
+        let mut metadata = HashMap::new();
+        for _idx in 0..metadata_kv_count {
+            let key = read_string(reader, &magic)?;
+            let value_type = reader.read_u32::<LittleEndian>()?;
+            let value_type = ValueType::from_u32(value_type)?;
+            let value = Value::read(reader, value_type, &magic)?;
+            metadata.insert(key, value);
+        }
+        let mut tensor_infos = HashMap::new();
+        for _idx in 0..tensor_count {
+            let tensor_name = read_string(reader, &magic)?;
+            let n_dimensions = reader.read_u32::<LittleEndian>()?;
+
+            let mut dimensions: Vec<usize> = match magic {
+                VersionedMagic::GgufV1 => {
+                    let mut dimensions = vec![0; n_dimensions as usize];
+                    reader.read_u32_into::<LittleEndian>(&mut dimensions)?;
+                    dimensions.into_iter().map(|c| c as usize).collect()
+                }
+                VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
+                    let mut dimensions = vec![0; n_dimensions as usize];
+                    reader.read_u64_into::<LittleEndian>(&mut dimensions)?;
+                    dimensions.into_iter().map(|c| c as usize).collect()
+                }
+            };
+
+            dimensions.reverse();
+            let ggml_dtype = reader.read_u32::<LittleEndian>()?;
+            let ggml_dtype = GgmlDType::from_u32(ggml_dtype)?;
+            let offset = reader.read_u64::<LittleEndian>()?;
+            tensor_infos.insert(
+                tensor_name,
+                TensorInfo {
+                    shape: crate::Shape::from(dimensions),
+                    offset,
+                    ggml_dtype,
+                },
+            );
+        }
+        let position = reader.stream_position()?;
+        let alignment = match metadata.get("general.alignment") {
+            Some(Value::U8(v)) => *v as u64,
+            Some(Value::U16(v)) => *v as u64,
+            Some(Value::U32(v)) => *v as u64,
+            Some(Value::I8(v)) if *v >= 0 => *v as u64,
+            Some(Value::I16(v)) if *v >= 0 => *v as u64,
+            Some(Value::I32(v)) if *v >= 0 => *v as u64,
+            _ => DEFAULT_ALIGNMENT,
+        };
+        let tensor_data_offset = (position + alignment - 1) / alignment * alignment;
+        Ok(Self {
+            magic,
+            metadata,
+            tensor_infos,
+            tensor_data_offset,
+        })
+    }
+
+    pub fn tensor<R: std::io::Seek + std::io::Read>(
+        &self,
+        reader: &mut R,
+        name: &str,
+        device: &Device,
+    ) -> Result<QTensor> {
+        let tensor_info = match self.tensor_infos.get(name) {
+            Some(tensor_info) => tensor_info,
+            None => crate::bail!("cannot find tensor info for {name}"),
+        };
+        tensor_info.read(reader, self.tensor_data_offset, device)
+    }
+}
+
+fn write_string<W: std::io::Write>(w: &mut W, str: &str) -> Result<()> {
+    let bytes = str.as_bytes();
+    w.write_u64::<LittleEndian>(bytes.len() as u64)?;
+    w.write_all(bytes)?;
+    Ok(())
+}
+
+pub fn write<W: std::io::Seek + std::io::Write>(
+    w: &mut W,
+    metadata: &[(&str, &Value)],
+    tensors: &[(&str, &QTensor)],
+) -> Result<()> {
+    w.write_u32::<LittleEndian>(0x46554747)?;
+    w.write_u32::<LittleEndian>(2)?; // version 2.
+    w.write_u64::<LittleEndian>(tensors.len() as u64)?;
+    w.write_u64::<LittleEndian>(metadata.len() as u64)?;
+    for (name, value) in metadata.iter() {
+        write_string(w, name)?;
+        w.write_u32::<LittleEndian>(value.value_type().to_u32())?;
+        value.write(w)?;
+    }
+    let mut offset = 0usize;
+    let mut offsets = Vec::with_capacity(tensors.len());
+    for (name, tensor) in tensors.iter() {
+        write_string(w, name)?;
+        let dims = tensor.shape().dims();
+        w.write_u32::<LittleEndian>(dims.len() as u32)?;
+        for &dim in dims.iter().rev() {
+            w.write_u64::<LittleEndian>(dim as u64)?;
+        }
+        w.write_u32::<LittleEndian>(tensor.dtype().to_u32())?;
+        w.write_u64::<LittleEndian>(offset as u64)?;
+        offsets.push(offset);
+        let size_in_bytes = tensor.storage_size_in_bytes();
+        let padding = 31 - (31 + size_in_bytes) % 32;
+        offset += size_in_bytes + padding;
+    }
+    let pos = w.stream_position()? as usize;
+    let padding = 31 - (31 + pos) % 32;
+    w.write_all(&vec![0u8; padding])?;
+    let tensor_start_pos = w.stream_position()? as usize;
+    for (offset, (_name, tensor)) in offsets.iter().zip(tensors.iter()) {
+        let pos = w.stream_position()? as usize;
+        if tensor_start_pos + offset != pos {
+            crate::bail!(
+                "internal error, unexpected current position {tensor_start_pos} {offset} {pos}"
+            )
+        }
+        let data = tensor.data()?;
+        let size_in_bytes = data.len();
+        w.write_all(&data)?;
+        let padding = 31 - (31 + size_in_bytes) % 32;
+        w.write_all(&vec![0u8; padding])?;
+    }
+    Ok(())
+}

candle-core/src/quantized/metal.rs

@@ -0,0 +1,230 @@
+use super::{GgmlDType, QStorage};
+use crate::backend::BackendStorage;
+use crate::{DType, MetalDevice, MetalStorage, Result, Shape};
+use metal::Buffer;
+use std::sync::Arc;
+
+pub struct QMetalStorage {
+    dtype: GgmlDType,
+    device: MetalDevice,
+    buffer: Arc<Buffer>,
+}
+
+impl QMetalStorage {
+    pub fn zeros(device: &MetalDevice, elem_count: usize, dtype: GgmlDType) -> Result<Self> {
+        let size = elem_count * dtype.type_size() / dtype.block_size();
+        let buffer = device.allocate_zeros(size)?;
+        Ok(Self {
+            buffer,
+            device: device.clone(),
+            dtype,
+        })
+    }
+
+    pub fn dtype(&self) -> GgmlDType {
+        self.dtype
+    }
+
+    pub fn device(&self) -> &MetalDevice {
+        &self.device
+    }
+
+    pub fn buffer(&self) -> &Buffer {
+        &self.buffer
+    }
+
+    pub fn dequantize(&self, elem_count: usize) -> Result<MetalStorage> {
+        use crate::quantized::k_quants::GgmlType;
+
+        let buffer = self.device.new_buffer_managed(self.buffer.length())?;
+        let command_buffer = self.device.command_buffer()?;
+        command_buffer.set_label("to_cpu");
+        let blit = command_buffer.new_blit_command_encoder();
+        blit.set_label("blit_to_cpu");
+        blit.copy_from_buffer(&self.buffer, 0, &buffer, 0, self.buffer.length());
+        blit.end_encoding();
+        self.device.wait_until_completed()?;
+        let mut out = vec![0.0; elem_count];
+        let block_len = elem_count / self.dtype.block_size();
+        match self.dtype {
+            GgmlDType::F32 => {
+                let vec: Vec<f32> = read_to_vec(&buffer, block_len);
+                f32::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::F16 => {
+                let vec: Vec<half::f16> = read_to_vec(&buffer, block_len);
+                half::f16::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q4_0 => {
+                let vec: Vec<crate::quantized::BlockQ4_0> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ4_0::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q4_1 => {
+                let vec: Vec<crate::quantized::BlockQ4_1> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ4_1::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q5_0 => {
+                let vec: Vec<crate::quantized::BlockQ5_0> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ5_0::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q5_1 => {
+                let vec: Vec<crate::quantized::BlockQ5_1> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ5_1::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q8_0 => {
+                let vec: Vec<crate::quantized::BlockQ8_0> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ8_0::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q8_1 => {
+                let vec: Vec<crate::quantized::BlockQ8_1> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ8_1::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q2K => {
+                let vec: Vec<crate::quantized::BlockQ2K> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ2K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q3K => {
+                let vec: Vec<crate::quantized::BlockQ3K> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ3K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q4K => {
+                let vec: Vec<crate::quantized::BlockQ4K> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ4K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q5K => {
+                let vec: Vec<crate::quantized::BlockQ5K> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ5K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q6K => {
+                let vec: Vec<crate::quantized::BlockQ6K> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ6K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q8K => {
+                let vec: Vec<crate::quantized::BlockQ8K> = read_to_vec(&buffer, block_len);
+                crate::quantized::BlockQ8K::to_float(&vec, &mut out)?;
+            }
+        }
+
+        let buffer = self.device.new_buffer_with_data(&out)?;
+        Ok(MetalStorage::new(
+            buffer,
+            self.device.clone(),
+            elem_count,
+            DType::F32,
+        ))
+    }
+
+    pub fn quantize(&mut self, src: &MetalStorage) -> Result<()> {
+        // Quantization only happens on CPU for now.
+        let src = src.to_cpu::<f32>()?;
+        let elem_count = src.len();
+        let src = crate::Storage::Cpu(crate::CpuStorage::F32(src));
+        let mut qcpu_storage = crate::Device::Cpu.qzeros(elem_count, self.dtype)?;
+        qcpu_storage.quantize(&src)?;
+        let buffer = self.device.new_buffer_with_data(&qcpu_storage.data()?)?;
+        self.buffer = buffer;
+        Ok(())
+    }
+
+    pub fn storage_size_in_bytes(&self) -> usize {
+        self.buffer.length() as usize
+    }
+
+    pub fn fwd(
+        &self,
+        self_shape: &Shape,
+        storage: &MetalStorage,
+        layout: &crate::Layout,
+    ) -> Result<(MetalStorage, Shape)> {
+        use crate::MetalError;
+
+        if !layout.is_contiguous() {
+            crate::bail!("input tensor is not contiguous {layout:?}")
+        }
+        let src_shape = layout.shape();
+        // self is transposed so n is first then k.
+        if src_shape.rank() < 2 {
+            crate::bail!("input tensor has only one dimension {layout:?}")
+        }
+        let (n, k) = self_shape.dims2()?;
+        let mut dst_shape = src_shape.dims().to_vec();
+
+        // We always use a single batch dimension and stack all the tensors in the batch on the
+        // second dimension as the implementation in candle-metal-kernels doesn't handle batch
+        // properly.
+        let m = match dst_shape.len() {
+            3 => dst_shape[0] * dst_shape[1],
+            2 => dst_shape[0],
+            n => crate::bail!("Invalid rank {n} for quantized matmul metal"),
+        };
+        let last_k = dst_shape.pop().unwrap();
+        if last_k != k {
+            crate::bail!("input tensor {layout:?} incompatible with {:?}", self_shape)
+        }
+        dst_shape.push(n);
+        let dst_shape = Shape::from(dst_shape);
+        let device = storage.device().clone();
+        let dst = device.new_buffer(dst_shape.elem_count(), DType::F32, "qmatmul")?;
+        let command_buffer = device.command_buffer()?;
+        // In some cases it would be better to use the mm variant, though it has its drawbacks
+        // around memory alignemnt.
+        for batch_id in 0..m {
+            candle_metal_kernels::call_quantized_matmul_mv_t(
+                device.device(),
+                &command_buffer,
+                device.kernels(),
+                self.dtype.into(),
+                (1, 1, n, k),
+                storage.buffer(),
+                (layout.start_offset() + batch_id * k) * storage.dtype().size_in_bytes(),
+                &self.buffer,
+                batch_id * n * DType::F32.size_in_bytes(),
+                &dst,
+            )
+            .map_err(MetalError::from)?;
+        }
+        let dst_storage = crate::MetalStorage::new(dst, device, dst_shape.elem_count(), DType::F32);
+        Ok((dst_storage, dst_shape))
+    }
+}
+
+pub fn load_quantized<T: super::GgmlType + Send + Sync + 'static>(
+    device: &MetalDevice,
+    data: &[T],
+) -> Result<QStorage> {
+    let buffer = device.new_buffer_with_data(data)?;
+    let device = device.clone();
+    Ok(QStorage::Metal(QMetalStorage {
+        dtype: T::DTYPE,
+        device,
+        buffer,
+    }))
+}
+
+fn read_to_vec<T: Clone>(buffer: &Buffer, n: usize) -> Vec<T> {
+    let ptr = buffer.contents() as *const T;
+    assert!(!ptr.is_null());
+    let slice = unsafe { std::slice::from_raw_parts(ptr, n) };
+    slice.to_vec()
+}
+
+impl From<GgmlDType> for candle_metal_kernels::GgmlDType {
+    fn from(value: GgmlDType) -> Self {
+        match value {
+            GgmlDType::Q4_0 => candle_metal_kernels::GgmlDType::Q4_0,
+            GgmlDType::Q4_1 => candle_metal_kernels::GgmlDType::Q4_1,
+            GgmlDType::Q5_0 => candle_metal_kernels::GgmlDType::Q5_0,
+            GgmlDType::Q5_1 => candle_metal_kernels::GgmlDType::Q5_1,
+            GgmlDType::Q8_0 => candle_metal_kernels::GgmlDType::Q8_0,
+            GgmlDType::Q8_1 => candle_metal_kernels::GgmlDType::Q8_1,
+            GgmlDType::Q2K => candle_metal_kernels::GgmlDType::Q2K,
+            GgmlDType::Q3K => candle_metal_kernels::GgmlDType::Q3K,
+            GgmlDType::Q4K => candle_metal_kernels::GgmlDType::Q4K,
+            GgmlDType::Q5K => candle_metal_kernels::GgmlDType::Q5K,
+            GgmlDType::Q6K => candle_metal_kernels::GgmlDType::Q6K,
+            GgmlDType::Q8K => candle_metal_kernels::GgmlDType::Q8K,
+            GgmlDType::F16 => candle_metal_kernels::GgmlDType::F16,
+            GgmlDType::F32 => candle_metal_kernels::GgmlDType::F32,
+        }
+    }
+}

candle-core/src/quantized/mod.rs

@@ -1,16 +1,135 @@
-use crate::{Device, Result, Shape, Tensor};
+use crate::{CpuStorage, DType, Device, Result, Shape, Storage, Tensor};
+use k_quants::*;
+use std::borrow::Cow;
 
+#[cfg(target_feature = "avx")]
+pub mod avx;
+mod dummy_cuda;
+mod dummy_metal;
 pub mod ggml_file;
+pub mod gguf_file;
 pub mod k_quants;
+#[cfg(feature = "metal")]
+pub mod metal;
+#[cfg(not(feature = "metal"))]
+mod metal {
+    pub use super::dummy_metal::*;
+}
+#[cfg(feature = "cuda")]
+pub mod cuda;
+#[cfg(not(feature = "cuda"))]
+mod cuda {
+    pub use super::dummy_cuda::*;
+}
+
+#[cfg(target_feature = "neon")]
+pub mod neon;
+#[cfg(target_feature = "simd128")]
+pub mod simd128;
+pub mod utils;
+use half::f16;
 
 pub use k_quants::GgmlType;
 
 pub struct QTensor {
-    data: Box<dyn QuantizedType>,
+    storage: QStorage,
     shape: Shape,
 }
 
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+impl Device {
+    fn qzeros(&self, elem_count: usize, dtype: GgmlDType) -> Result<QStorage> {
+        match self {
+            Device::Cpu => {
+                let storage = dtype.cpu_zeros(elem_count);
+                Ok(QStorage::Cpu(storage))
+            }
+            Device::Metal(metal) => {
+                let storage = metal::QMetalStorage::zeros(metal, elem_count, dtype)?;
+                Ok(QStorage::Metal(storage))
+            }
+            Device::Cuda(cuda) => {
+                let storage = cuda::QCudaStorage::zeros(cuda, elem_count, dtype)?;
+                Ok(QStorage::Cuda(storage))
+            }
+        }
+    }
+}
+
+pub enum QStorage {
+    Cpu(Box<dyn QuantizedType>),
+    Metal(metal::QMetalStorage),
+    Cuda(cuda::QCudaStorage),
+}
+
+impl QStorage {
+    fn block_size(&self) -> usize {
+        match self {
+            QStorage::Cpu(storage) => storage.block_size(),
+            QStorage::Metal(storage) => storage.dtype().block_size(),
+            QStorage::Cuda(storage) => storage.dtype().block_size(),
+        }
+    }
+
+    fn dtype(&self) -> GgmlDType {
+        match self {
+            QStorage::Cpu(storage) => storage.dtype(),
+            QStorage::Metal(storage) => storage.dtype(),
+            QStorage::Cuda(storage) => storage.dtype(),
+        }
+    }
+
+    fn device(&self) -> Device {
+        match self {
+            QStorage::Cpu(_storage) => Device::Cpu,
+            QStorage::Metal(storage) => Device::Metal(storage.device().clone()),
+            QStorage::Cuda(storage) => Device::Cuda(storage.device().clone()),
+        }
+    }
+
+    fn size_in_bytes(&self) -> usize {
+        match self {
+            QStorage::Cpu(storage) => storage.storage_size_in_bytes(),
+            QStorage::Metal(storage) => storage.storage_size_in_bytes(),
+            QStorage::Cuda(storage) => storage.storage_size_in_bytes(),
+        }
+    }
+
+    fn quantize(&mut self, src: &Storage) -> Result<()> {
+        match (self, src) {
+            (QStorage::Cpu(storage), Storage::Cpu(src)) => {
+                storage.from_float(src.as_slice::<f32>()?)?;
+            }
+            (QStorage::Metal(storage), Storage::Metal(src)) => storage.quantize(src)?,
+            (QStorage::Cuda(storage), Storage::Cuda(src)) => storage.quantize(src)?,
+            _ => crate::bail!("Invalid dequantize storage locations do not match"),
+        }
+        Ok(())
+    }
+
+    fn dequantize(&self, elem_count: usize) -> Result<Storage> {
+        match self {
+            QStorage::Cpu(storage) => Ok(Storage::Cpu(storage.dequantize(elem_count)?)),
+            QStorage::Metal(storage) => Ok(Storage::Metal(storage.dequantize(elem_count)?)),
+            QStorage::Cuda(storage) => Ok(Storage::Cuda(storage.dequantize(elem_count)?)),
+        }
+    }
+
+    fn data(&self) -> Result<Cow<[u8]>> {
+        match self {
+            QStorage::Cpu(storage) => {
+                let data_ptr = storage.as_ptr();
+                let size_in_bytes = storage.storage_size_in_bytes();
+                let data = unsafe { std::slice::from_raw_parts(data_ptr, size_in_bytes) };
+                Ok(Cow::from(data))
+            }
+            QStorage::Metal(_) | QStorage::Cuda(_) => {
+                crate::bail!("not implemented");
+            }
+        }
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
 pub enum GgmlDType {
     F32,
     F16,
@@ -50,7 +169,46 @@ impl GgmlDType {
         Ok(dtype)
     }
 
-    fn type_size(&self) -> usize {
+    pub(crate) fn to_u32(self) -> u32 {
+        match self {
+            Self::F32 => 0,
+            Self::F16 => 1,
+            Self::Q4_0 => 2,
+            Self::Q4_1 => 3,
+            Self::Q5_0 => 6,
+            Self::Q5_1 => 7,
+            Self::Q8_0 => 8,
+            Self::Q8_1 => 9,
+            Self::Q2K => 10,
+            Self::Q3K => 11,
+            Self::Q4K => 12,
+            Self::Q5K => 13,
+            Self::Q6K => 14,
+            Self::Q8K => 15,
+        }
+    }
+
+    /// The block dtype
+    pub fn cpu_zeros(&self, elem_count: usize) -> Box<dyn QuantizedType> {
+        match self {
+            Self::F32 => Box::new(vec![f32::zeros(); elem_count]),
+            Self::F16 => Box::new(vec![f16::zeros(); elem_count]),
+            Self::Q4_0 => Box::new(vec![BlockQ4_0::zeros(); elem_count / BlockQ4_0::BLCK_SIZE]),
+            Self::Q4_1 => Box::new(vec![BlockQ4_1::zeros(); elem_count / BlockQ4_1::BLCK_SIZE]),
+            Self::Q5_0 => Box::new(vec![BlockQ5_0::zeros(); elem_count / BlockQ5_0::BLCK_SIZE]),
+            Self::Q5_1 => Box::new(vec![BlockQ5_1::zeros(); elem_count / BlockQ5_1::BLCK_SIZE]),
+            Self::Q8_0 => Box::new(vec![BlockQ8_0::zeros(); elem_count / BlockQ8_0::BLCK_SIZE]),
+            Self::Q8_1 => Box::new(vec![BlockQ8_1::zeros(); elem_count / BlockQ8_1::BLCK_SIZE]),
+            Self::Q2K => Box::new(vec![BlockQ2K::zeros(); elem_count / BlockQ2K::BLCK_SIZE]),
+            Self::Q3K => Box::new(vec![BlockQ3K::zeros(); elem_count / BlockQ3K::BLCK_SIZE]),
+            Self::Q4K => Box::new(vec![BlockQ4K::zeros(); elem_count / BlockQ4K::BLCK_SIZE]),
+            Self::Q5K => Box::new(vec![BlockQ5K::zeros(); elem_count / BlockQ5K::BLCK_SIZE]),
+            Self::Q6K => Box::new(vec![BlockQ6K::zeros(); elem_count / BlockQ6K::BLCK_SIZE]),
+            Self::Q8K => Box::new(vec![BlockQ8K::zeros(); elem_count / BlockQ8K::BLCK_SIZE]),
+        }
+    }
+    /// The type size for blocks in bytes.
+    pub fn type_size(&self) -> usize {
         use k_quants::*;
         match self {
             Self::F32 => 4,
@@ -71,7 +229,8 @@ impl GgmlDType {
         }
     }
 
-    fn blck_size(&self) -> usize {
+    /// The block size, i.e. the number of elements stored in each block.
+    pub fn block_size(&self) -> usize {
         match self {
             Self::F32 => 1,
             Self::F16 => 1,
@@ -90,7 +249,13 @@ impl GgmlDType {
 pub trait QuantizedType: Send + Sync {
     fn dtype(&self) -> GgmlDType;
     fn matmul_t(&self, mkn: (usize, usize, usize), lhs: &[f32], dst: &mut [f32]) -> Result<()>;
-    fn to_float(&self, ys: &mut [f32]) -> Result<()>;
+    fn dequantize(&self, elem_count: usize) -> Result<CpuStorage>;
+    fn storage_size_in_bytes(&self) -> usize;
+    fn as_ptr(&self) -> *const u8;
+    fn block_size(&self) -> usize;
+    #[allow(clippy::wrong_self_convention)]
+    fn from_float(&mut self, xs: &[f32]) -> Result<()>;
+    fn size(&self) -> usize;
 }
 
 impl<T: k_quants::GgmlType + Send + Sync> QuantizedType for Vec<T> {
@@ -98,12 +263,34 @@ impl<T: k_quants::GgmlType + Send + Sync> QuantizedType for Vec<T> {
         k_quants::matmul(mkn, lhs, self.as_slice(), dst)
     }
 
+    fn size(&self) -> usize {
+        self.len() * core::mem::size_of::<T>()
+    }
+
+    fn from_float(&mut self, xs: &[f32]) -> Result<()> {
+        T::from_float(xs, self)
+    }
+
     fn dtype(&self) -> GgmlDType {
         T::DTYPE
     }
 
-    fn to_float(&self, ys: &mut [f32]) -> Result<()> {
-        T::to_float(self.as_slice(), ys)
+    fn block_size(&self) -> usize {
+        T::BLCK_SIZE
+    }
+
+    fn dequantize(&self, elem_count: usize) -> Result<CpuStorage> {
+        let mut ys = vec![0.0f32; elem_count];
+        T::to_float(self.as_slice(), &mut ys)?;
+        Ok(CpuStorage::F32(ys))
+    }
+
+    fn storage_size_in_bytes(&self) -> usize {
+        self.len() * std::mem::size_of::<T>()
+    }
+
+    fn as_ptr(&self) -> *const u8 {
+        self.as_ptr() as *const u8
     }
 }
 
@@ -113,19 +300,57 @@ impl std::fmt::Debug for QTensor {
     }
 }
 
+fn check_shape(shape: &Shape, block_size: usize) -> Result<()> {
+    let dims = shape.dims();
+    if dims.is_empty() {
+        crate::bail!("scalar tensor cannot be quantized {shape:?}")
+    }
+    if dims[dims.len() - 1] % block_size != 0 {
+        crate::bail!(
+            "quantized tensor must have their last dim divisible by block size {shape:?} {}",
+            block_size
+        )
+    }
+    Ok(())
+}
+
 impl QTensor {
-    pub fn new<S: Into<Shape>, T: k_quants::GgmlType + Send + Sync + 'static>(
-        data: Vec<T>,
-        shape: S,
-    ) -> Self {
-        Self {
-            data: Box::new(data),
-            shape: shape.into(),
+    pub fn new<S: Into<Shape>>(storage: QStorage, shape: S) -> Result<Self> {
+        let shape = shape.into();
+        check_shape(&shape, storage.block_size())?;
+        Ok(Self { storage, shape })
+    }
+
+    pub fn quantize(src: &Tensor, dtype: GgmlDType) -> Result<Self> {
+        let shape = src.shape();
+        let block_size = dtype.block_size();
+        check_shape(shape, block_size)?;
+        let src = src.to_dtype(crate::DType::F32)?.flatten_all()?;
+        let elem_count = shape.elem_count();
+        if elem_count % block_size != 0 {
+            crate::bail!(
+                "tensor size ({shape:?}) is not divisible by block size {}",
+                block_size
+            )
         }
+        let mut storage = src.device().qzeros(elem_count, dtype)?;
+        storage.quantize(&src.storage())?;
+        Ok(Self {
+            storage,
+            shape: shape.clone(),
+        })
     }
 
     pub fn dtype(&self) -> GgmlDType {
-        self.data.dtype()
+        self.storage.dtype()
+    }
+
+    pub fn device(&self) -> Device {
+        self.storage.device()
+    }
+
+    pub fn rank(&self) -> usize {
+        self.shape.rank()
     }
 
     pub fn shape(&self) -> &Shape {
@@ -133,26 +358,109 @@ impl QTensor {
     }
 
     pub fn dequantize(&self, device: &Device) -> Result<Tensor> {
-        let mut f32_data = vec![0f32; self.shape.elem_count()];
-        self.data.to_float(&mut f32_data)?;
-        Tensor::from_vec(f32_data, &self.shape, device)
+        let storage = self.storage.dequantize(self.shape.elem_count())?;
+        let none = crate::op::BackpropOp::none();
+        crate::tensor::from_storage(storage, self.shape.clone(), none, false).to_device(device)
     }
 
-    pub fn matmul_t(&self, mkn: (usize, usize, usize), lhs: &[f32], dst: &mut [f32]) -> Result<()> {
-        self.data.matmul_t(mkn, lhs, dst)
+    pub fn dequantize_f16(&self, device: &Device) -> Result<Tensor> {
+        // In the CUDA case, we have a specialized kernel as this can be useful for volta
+        // architectures. https://github.com/huggingface/candle/issues/2136
+        match &self.storage {
+            QStorage::Cuda(s) => {
+                let s = s.dequantize_f16(self.shape.elem_count())?;
+                let none = crate::op::BackpropOp::none();
+                crate::tensor::from_storage(Storage::Cuda(s), self.shape.clone(), none, false)
+                    .to_device(device)
+            }
+            _ => {
+                let s = self.dequantize(device)?.to_dtype(crate::DType::F16)?;
+                Ok(s)
+            }
+        }
+    }
+
+    pub fn storage_size_in_bytes(&self) -> usize {
+        self.storage.size_in_bytes()
+    }
+
+    pub fn data(&self) -> Result<Cow<'_, [u8]>> {
+        self.storage.data()
     }
 }
 
-#[derive(Debug, Clone)]
-pub struct QMatMul(std::sync::Arc<QTensor>);
+#[derive(Clone, Debug)]
+pub enum QMatMul {
+    QTensor(std::sync::Arc<QTensor>),
+    Tensor(Tensor),
+    TensorF16(Tensor),
+}
+
+thread_local! {
+    static DEQUANTIZE_ALL: bool = {
+        match std::env::var("CANDLE_DEQUANTIZE_ALL") {
+            Ok(s) => {
+                !s.is_empty() && s != "0"
+            },
+            Err(_) => false,
+        }
+    }
+}
+
+thread_local! {
+    static DEQUANTIZE_ALL_F16: bool = {
+        match std::env::var("CANDLE_DEQUANTIZE_ALL_F16") {
+            Ok(s) => {
+                !s.is_empty() && s != "0"
+            },
+            Err(_) => false,
+        }
+    }
+}
 
 impl QMatMul {
-    pub fn new(qtensor: std::sync::Arc<QTensor>) -> Self {
-        Self(qtensor)
+    pub fn from_arc(qtensor: std::sync::Arc<QTensor>) -> Result<Self> {
+        let dequantize = match qtensor.dtype() {
+            GgmlDType::F32 | GgmlDType::F16 => true,
+            _ => DEQUANTIZE_ALL.with(|b| *b),
+        };
+        let t = if dequantize {
+            let tensor = qtensor.dequantize(&qtensor.device())?;
+            Self::Tensor(tensor)
+        } else if DEQUANTIZE_ALL_F16.with(|b| *b) {
+            let tensor = qtensor.dequantize_f16(&qtensor.device())?;
+            Self::TensorF16(tensor)
+        } else {
+            Self::QTensor(qtensor)
+        };
+        Ok(t)
+    }
+
+    pub fn from_qtensor(qtensor: QTensor) -> Result<Self> {
+        Self::from_arc(std::sync::Arc::new(qtensor))
+    }
+
+    pub fn dequantize_f16(&self) -> Result<Tensor> {
+        match self {
+            Self::QTensor(t) => t.dequantize_f16(&t.device()),
+            Self::Tensor(t) => t.to_dtype(DType::F16),
+            Self::TensorF16(t) => Ok(t.clone()),
+        }
+    }
+
+    pub fn forward_via_f16(&self, xs: &Tensor) -> Result<Tensor> {
+        let w = self.dequantize_f16()?;
+        let in_dtype = xs.dtype();
+        let w = match *xs.dims() {
+            [b1, b2, _, _] => w.broadcast_left((b1, b2))?.t()?,
+            [bsize, _, _] => w.broadcast_left(bsize)?.t()?,
+            _ => w.t()?,
+        };
+        xs.to_dtype(DType::F16)?.matmul(&w)?.to_dtype(in_dtype)
     }
 }
 
-impl crate::CustomOp1 for QMatMul {
+impl crate::CustomOp1 for QTensor {
     fn name(&self) -> &'static str {
         "qmatmul"
     }
@@ -166,29 +474,76 @@ impl crate::CustomOp1 for QMatMul {
             crate::bail!("input tensor is not contiguous {layout:?}")
         }
         let src_shape = layout.shape();
-        let (k, n) = self.0.shape.dims2()?;
+        // self is transposed so n is first then k.
+        let (n, k) = self.shape.dims2()?;
         if src_shape.rank() < 2 {
             crate::bail!("input tensor has only one dimension {layout:?}")
         }
         let mut dst_shape = src_shape.dims().to_vec();
         let last_k = dst_shape.pop().unwrap();
         if last_k != k {
-            crate::bail!(
-                "input tensor {layout:?} incompatible with {:?}",
-                self.0.shape
-            )
+            crate::bail!("input tensor {layout:?} incompatible with {:?}", self.shape)
         }
         dst_shape.push(n);
         let dst_shape = Shape::from(dst_shape);
-        let storage = storage.as_slice::<f32>()?;
-        let storage =
-            &storage[layout.start_offset()..layout.start_offset() + src_shape.elem_count()];
+        #[allow(clippy::infallible_destructuring_match)]
+        let self_storage = match &self.storage {
+            QStorage::Cpu(storage) => storage,
+            QStorage::Metal(_) | QStorage::Cuda(_) => crate::bail!("Invalid storage"),
+        };
+        let slice = storage.as_slice::<f32>()?;
+        let slice = &slice[layout.start_offset()..layout.start_offset() + src_shape.elem_count()];
         let mut dst_storage = vec![0f32; dst_shape.elem_count()];
-        self.0.matmul_t(
-            (dst_shape.elem_count() / n, k, n),
-            storage,
-            &mut dst_storage,
-        )?;
+        self_storage.matmul_t((dst_shape.elem_count() / n, k, n), slice, &mut dst_storage)?;
         Ok((crate::CpuStorage::F32(dst_storage), dst_shape))
     }
+
+    fn metal_fwd(
+        &self,
+        storage: &crate::MetalStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::MetalStorage, Shape)> {
+        let self_storage = match &self.storage {
+            QStorage::Metal(metal) => metal,
+            _ => unreachable!("Cannot call metal matmul on non metal QTensor"),
+        };
+        self_storage.fwd(&self.shape, storage, layout)
+    }
+
+    fn cuda_fwd(
+        &self,
+        storage: &crate::CudaStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::CudaStorage, Shape)> {
+        let self_storage = match &self.storage {
+            QStorage::Cuda(cuda) => cuda,
+            _ => unreachable!("Cannot call cuda matmul on non cuda QTensor"),
+        };
+        self_storage.fwd(&self.shape, storage, layout)
+    }
+}
+
+impl crate::Module for QMatMul {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        match self {
+            Self::QTensor(t) => xs.apply_op1_no_bwd(t.as_ref()),
+            Self::Tensor(w) => {
+                let w = match *xs.dims() {
+                    [b1, b2, _, _] => w.broadcast_left((b1, b2))?.t()?,
+                    [bsize, _, _] => w.broadcast_left(bsize)?.t()?,
+                    _ => w.t()?,
+                };
+                xs.matmul(&w)
+            }
+            Self::TensorF16(w) => {
+                let in_dtype = xs.dtype();
+                let w = match *xs.dims() {
+                    [b1, b2, _, _] => w.broadcast_left((b1, b2))?.t()?,
+                    [bsize, _, _] => w.broadcast_left(bsize)?.t()?,
+                    _ => w.t()?,
+                };
+                xs.to_dtype(DType::F16)?.matmul(&w)?.to_dtype(in_dtype)
+            }
+        }
+    }
 }

candle-core/src/quantized/neon.rs

@@ -0,0 +1,613 @@
+use super::k_quants::{
+    BlockQ2K, BlockQ3K, BlockQ4K, BlockQ4_0, BlockQ5K, BlockQ6K, BlockQ8K, BlockQ8_0, QK8_0, QK_K,
+};
+use crate::Result;
+use byteorder::{ByteOrder, LittleEndian};
+
+#[allow(unused_imports)]
+#[cfg(target_arch = "arm")]
+use core::arch::arm::*;
+
+#[allow(unused_imports)]
+#[cfg(target_arch = "aarch64")]
+use core::arch::aarch64::*;
+
+#[inline(always)]
+unsafe fn vdotq_s32(a: int8x16_t, b: int8x16_t) -> int32x4_t {
+    // TODO: dotprod
+    let p0 = vmull_s8(vget_low_s8(a), vget_low_s8(b));
+    let p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+    vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1))
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q4_0_q8_0(n: usize, xs: &[BlockQ4_0], ys: &[BlockQ8_0]) -> Result<f32> {
+    let qk = QK8_0;
+    let nb = n / qk;
+    if n % QK8_0 != 0 {
+        crate::bail!("vec_dot_q4_0_q8_0: {n} is not divisible by {qk}")
+    }
+
+    unsafe {
+        let mut sumv0 = vdupq_n_f32(0.0f32);
+        for i in 0..nb {
+            let x0 = &xs[i];
+            let y0 = &ys[i];
+
+            let m4b = vdupq_n_u8(0x0F);
+            let s8b = vdupq_n_s8(0x8);
+
+            let v0_0 = vld1q_u8(x0.qs.as_ptr());
+
+            // 4-bit -> 8-bit
+            let v0_0l = vreinterpretq_s8_u8(vandq_u8(v0_0, m4b));
+            let v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+
+            // sub 8
+            let v0_0ls = vsubq_s8(v0_0l, s8b);
+            let v0_0hs = vsubq_s8(v0_0h, s8b);
+
+            // load y
+            let v1_0l = vld1q_s8(y0.qs.as_ptr());
+            let v1_0h = vld1q_s8(y0.qs.as_ptr().add(16));
+
+            let pl0 = vdotq_s32(v0_0ls, v1_0l);
+            let ph0 = vdotq_s32(v0_0hs, v1_0h);
+            sumv0 = vmlaq_n_f32(
+                sumv0,
+                vcvtq_f32_s32(vaddq_s32(pl0, ph0)),
+                x0.d.to_f32() * y0.d.to_f32(),
+            );
+        }
+        Ok(vaddvq_f32(sumv0))
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q8_0_q8_0(n: usize, xs: &[BlockQ8_0], ys: &[BlockQ8_0]) -> Result<f32> {
+    let qk = QK8_0;
+    if n % QK8_0 != 0 {
+        crate::bail!("vec_dot_q8_0_q8_0: {n} is not divisible by {qk}")
+    }
+    let nb = n / QK8_0;
+    unsafe {
+        let mut sumv0 = vdupq_n_f32(0.0f32);
+        for i in 0..nb {
+            let x0 = &xs[i];
+            let y0 = &ys[i];
+
+            let x0_0 = vld1q_s8(x0.qs.as_ptr());
+            let x0_1 = vld1q_s8(x0.qs.as_ptr().add(16));
+
+            // load y
+            let y0_0 = vld1q_s8(y0.qs.as_ptr());
+            let y0_1 = vld1q_s8(y0.qs.as_ptr().add(16));
+
+            let p0 = vdotq_s32(x0_0, y0_0);
+            let p1 = vdotq_s32(x0_1, y0_1);
+
+            sumv0 = vmlaq_n_f32(
+                sumv0,
+                vcvtq_f32_s32(vaddq_s32(p0, p1)),
+                x0.d.to_f32() * y0.d.to_f32(),
+            );
+        }
+        Ok(vaddvq_f32(sumv0))
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q8k_q8k(n: usize, xs: &[BlockQ8K], ys: &[BlockQ8K]) -> Result<f32> {
+    let qk = QK_K;
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q8k_q8k: {n} is not divisible by {qk}")
+    }
+
+    let mut sumf = 0f32;
+    for (xs, ys) in xs.iter().zip(ys.iter()) {
+        unsafe {
+            let mut sum_i = vdupq_n_s32(0);
+            let scale = xs.d * ys.d;
+            let xs = xs.qs.as_ptr();
+            let ys = ys.qs.as_ptr();
+            for i in (0..QK_K).step_by(16) {
+                let xs = vld1q_s8(xs.add(i));
+                let ys = vld1q_s8(ys.add(i));
+                let xy = vdotq_s32(xs, ys);
+                sum_i = vaddq_s32(sum_i, xy)
+            }
+            sumf += vaddvq_s32(sum_i) as f32 * scale
+        }
+    }
+    Ok(sumf)
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q6k_q8k(n: usize, xs: &[BlockQ6K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q6k_q8k: {n} is not divisible by {QK_K}")
+    }
+    let mut sum = 0f32;
+    unsafe {
+        let m4b = vdupq_n_u8(0xF);
+
+        let mone = vdupq_n_u8(3);
+
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d_all = x.d.to_f32();
+
+            let mut q6 = x.ql.as_ptr();
+            let mut qh = x.qh.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+
+            let mut scale = x.scales.as_ptr();
+
+            let q8sums = vld1q_s16_x2(y.bsums.as_ptr());
+            let scales = vld1q_s8(scale);
+            let q6scales = int16x8x2_t(
+                vmovl_s8(vget_low_s8(scales)),
+                vmovl_s8(vget_high_s8(scales)),
+            );
+
+            let prod = vaddq_s32(
+                vaddq_s32(
+                    vmull_s16(vget_low_s16(q8sums.0), vget_low_s16(q6scales.0)),
+                    vmull_s16(vget_high_s16(q8sums.0), vget_high_s16(q6scales.0)),
+                ),
+                vaddq_s32(
+                    vmull_s16(vget_low_s16(q8sums.1), vget_low_s16(q6scales.1)),
+                    vmull_s16(vget_high_s16(q8sums.1), vget_high_s16(q6scales.1)),
+                ),
+            );
+            let isum_mins = vaddvq_s32(prod);
+
+            let mut isum = 0i32;
+
+            for _j in 0..QK_K / 128 {
+                let qhbits = vld1q_u8_x2(qh);
+                qh = qh.add(32);
+                let q6bits = vld1q_u8_x4(q6);
+                q6 = q6.add(64);
+                let q8bytes = vld1q_s8_x4(q8);
+                q8 = q8.add(64);
+
+                let q6h_0 = vshlq_n_u8(vandq_u8(mone, qhbits.0), 4);
+                let q6h_1 = vshlq_n_u8(vandq_u8(mone, qhbits.1), 4);
+                let shifted = vshrq_n_u8(qhbits.0, 2);
+                let q6h_2 = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+                let shifted = vshrq_n_u8(qhbits.1, 2);
+                let q6h_3 = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+
+                let q6bytes_0 = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.0, m4b), q6h_0));
+                let q6bytes_1 = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.1, m4b), q6h_1));
+                let q6bytes_2 = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.2, m4b), q6h_2));
+                let q6bytes_3 = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.3, m4b), q6h_3));
+
+                let p0 = vdotq_s32(q6bytes_0, q8bytes.0);
+                let p1 = vdotq_s32(q6bytes_1, q8bytes.1);
+                let (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
+                isum += vaddvq_s32(p0) * scale0 + vaddvq_s32(p1) * scale1;
+                scale = scale.add(2);
+
+                let p2 = vdotq_s32(q6bytes_2, q8bytes.2);
+                let p3 = vdotq_s32(q6bytes_3, q8bytes.3);
+                let (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
+                isum += vaddvq_s32(p2) * scale0 + vaddvq_s32(p3) * scale1;
+                scale = scale.add(2);
+
+                let q8bytes = vld1q_s8_x4(q8);
+                q8 = q8.add(64);
+
+                let shifted = vshrq_n_u8(qhbits.0, 4);
+                let q6h_0 = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+                let shifted = vshrq_n_u8(qhbits.1, 4);
+                let q6h_1 = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+                let shifted = vshrq_n_u8(qhbits.0, 6);
+                let q6h_2 = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+                let shifted = vshrq_n_u8(qhbits.1, 6);
+                let q6h_3 = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+
+                let q6bytes_0 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.0, 4), q6h_0));
+                let q6bytes_1 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.1, 4), q6h_1));
+                let q6bytes_2 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.2, 4), q6h_2));
+                let q6bytes_3 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.3, 4), q6h_3));
+
+                let p0 = vdotq_s32(q6bytes_0, q8bytes.0);
+                let p1 = vdotq_s32(q6bytes_1, q8bytes.1);
+                let (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
+                isum += vaddvq_s32(p0) * scale0 + vaddvq_s32(p1) * scale1;
+                scale = scale.add(2);
+
+                let p2 = vdotq_s32(q6bytes_2, q8bytes.2);
+                let p3 = vdotq_s32(q6bytes_3, q8bytes.3);
+                let (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
+                isum += vaddvq_s32(p2) * scale0 + vaddvq_s32(p3) * scale1;
+                scale = scale.add(2);
+            }
+            sum += d_all * y.d * ((isum - 32 * isum_mins) as f32);
+        }
+    }
+    Ok(sum)
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q5k_q8k(n: usize, xs: &[BlockQ5K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q5k_q8k: {n} is not divisible by {QK_K}")
+    }
+    let mut sumf = 0f32;
+    let mut utmp = [0u32; 4];
+    const KMASK1: u32 = 0x3f3f3f3f;
+    const KMASK2: u32 = 0x0f0f0f0f;
+    const KMASK3: u32 = 0x03030303;
+
+    unsafe {
+        let m4b = vdupq_n_u8(0xF);
+        let mone = vdupq_n_u8(1);
+        let mtwo = vdupq_n_u8(2);
+
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = y.d * x.d.to_f32();
+            let dmin = y.d * x.dmin.to_f32();
+
+            let q8sums = vpaddq_s16(
+                vld1q_s16(y.bsums.as_ptr()),
+                vld1q_s16(y.bsums.as_ptr().add(8)),
+            );
+
+            LittleEndian::read_u32_into(&x.scales, &mut utmp[0..3]);
+
+            utmp[3] = ((utmp[2] >> 4) & KMASK2) | (((utmp[1] >> 6) & KMASK3) << 4);
+            let uaux = utmp[1] & KMASK1;
+            utmp[1] = (utmp[2] & KMASK2) | (((utmp[0] >> 6) & KMASK3) << 4);
+            utmp[2] = uaux;
+            utmp[0] &= KMASK1;
+
+            let mins8 = vld1_u8((utmp.as_ptr() as *const u8).add(8));
+            let mins = vreinterpretq_s16_u16(vmovl_u8(mins8));
+            let prod = vaddq_s32(
+                vmull_s16(vget_low_s16(q8sums), vget_low_s16(mins)),
+                vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)),
+            );
+            let sumi_mins = vaddvq_s32(prod);
+
+            let mut scales = utmp.as_ptr() as *const u8;
+
+            let mut q5 = x.qs.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+
+            let mut qhbits = vld1q_u8_x2(x.qh.as_ptr());
+
+            let mut sumi = 0i32;
+
+            for _j in 0..QK_K / 64 {
+                let q5bits = vld1q_u8_x2(q5);
+                q5 = q5.add(32);
+                let q8bytes = vld1q_s8_x4(q8);
+                q8 = q8.add(64);
+
+                let q5h_0 = vshlq_n_u8(vandq_u8(mone, qhbits.0), 4);
+                let q5h_1 = vshlq_n_u8(vandq_u8(mone, qhbits.1), 4);
+                let q5h_2 = vshlq_n_u8(vandq_u8(mtwo, qhbits.0), 3);
+                let q5h_3 = vshlq_n_u8(vandq_u8(mtwo, qhbits.1), 3);
+                qhbits.0 = vshrq_n_u8(qhbits.0, 2);
+                qhbits.1 = vshrq_n_u8(qhbits.1, 2);
+
+                let q5bytes_0 = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.0, m4b), q5h_0));
+                let q5bytes_1 = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.1, m4b), q5h_1));
+                let q5bytes_2 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.0, 4), q5h_2));
+                let q5bytes_3 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.1, 4), q5h_3));
+
+                let p0 = vdotq_s32(q5bytes_0, q8bytes.0);
+                let p1 = vdotq_s32(q5bytes_1, q8bytes.1);
+                sumi += vaddvq_s32(vaddq_s32(p0, p1)) * *scales as i32;
+                scales = scales.add(1);
+
+                let p2 = vdotq_s32(q5bytes_2, q8bytes.2);
+                let p3 = vdotq_s32(q5bytes_3, q8bytes.3);
+                sumi += vaddvq_s32(vaddq_s32(p2, p3)) * *scales as i32;
+                scales = scales.add(1);
+            }
+            sumf += d * sumi as f32 - dmin * sumi_mins as f32;
+        }
+    }
+    Ok(sumf)
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q4k_q8k(n: usize, xs: &[BlockQ4K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q4k_q8k: {n} is not divisible by {QK_K}")
+    }
+    let mut sumf = 0f32;
+    let mut utmp = [0u32; 4];
+    let mut scales = [0u8; 16];
+    const KMASK1: u32 = 0x3f3f3f3f;
+    const KMASK2: u32 = 0x0f0f0f0f;
+    const KMASK3: u32 = 0x03030303;
+
+    unsafe {
+        let m4b = vdupq_n_u8(0xF);
+
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = y.d * x.d.to_f32();
+            let dmin = y.d * x.dmin.to_f32();
+
+            let q8sums = vpaddq_s16(
+                vld1q_s16(y.bsums.as_ptr()),
+                vld1q_s16(y.bsums.as_ptr().add(8)),
+            );
+
+            LittleEndian::read_u32_into(&x.scales, &mut utmp[0..3]);
+
+            let mins8 = vld1_u32(
+                [
+                    utmp[1] & KMASK1,
+                    ((utmp[2] >> 4) & KMASK2) | (((utmp[1] >> 6) & KMASK3) << 4),
+                ]
+                .as_ptr(),
+            );
+            utmp[1] = (utmp[2] & KMASK2) | (((utmp[0] >> 6) & KMASK3) << 4);
+            utmp[0] &= KMASK1;
+
+            let mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
+            let prod = vaddq_s32(
+                vmull_s16(vget_low_s16(q8sums), vget_low_s16(mins)),
+                vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)),
+            );
+            sumf -= dmin * vaddvq_s32(prod) as f32;
+
+            LittleEndian::write_u32_into(&utmp, &mut scales);
+
+            let mut q4 = x.qs.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+
+            let mut sumi1 = 0i32;
+            let mut sumi2 = 0i32;
+
+            for j in 0..QK_K / 64 {
+                let q4bits = vld1q_u8_x2(q4);
+                q4 = q4.add(32);
+                let q8bytes = vld1q_s8_x2(q8);
+                q8 = q8.add(32);
+                let q4bytes = int8x16x2_t(
+                    vreinterpretq_s8_u8(vandq_u8(q4bits.0, m4b)),
+                    vreinterpretq_s8_u8(vandq_u8(q4bits.1, m4b)),
+                );
+                let p0 = vdotq_s32(q4bytes.0, q8bytes.0);
+                let p1 = vdotq_s32(q4bytes.1, q8bytes.1);
+                sumi1 += vaddvq_s32(vaddq_s32(p0, p1)) * scales[2 * j] as i32;
+
+                let q8bytes = vld1q_s8_x2(q8);
+                q8 = q8.add(32);
+                let q4bytes = int8x16x2_t(
+                    vreinterpretq_s8_u8(vshrq_n_u8(q4bits.0, 4)),
+                    vreinterpretq_s8_u8(vshrq_n_u8(q4bits.1, 4)),
+                );
+                let p2 = vdotq_s32(q4bytes.0, q8bytes.0);
+                let p3 = vdotq_s32(q4bytes.1, q8bytes.1);
+                sumi2 += vaddvq_s32(vaddq_s32(p2, p3)) * scales[2 * j + 1] as i32;
+            }
+            sumf += d * (sumi1 + sumi2) as f32;
+        }
+    }
+    Ok(sumf)
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q3k_q8k: {n} is not divisible by {QK_K}")
+    }
+    let mut sumf = 0f32;
+    let mut utmp = [0u32; 4];
+    let mut aux = [0u32; 3];
+    const KMASK1: u32 = 0x03030303;
+    const KMASK2: u32 = 0x0f0f0f0f;
+
+    unsafe {
+        let m3b = vdupq_n_u8(0x3);
+        let m0 = vdupq_n_u8(1);
+        let m1 = vshlq_n_u8(m0, 1);
+        let m2 = vshlq_n_u8(m0, 2);
+        let m3 = vshlq_n_u8(m0, 3);
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = y.d * x.d.to_f32();
+            let mut q3 = x.qs.as_ptr();
+            let qh = x.hmask.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+
+            let mut qhbits = vld1q_u8_x2(qh);
+
+            let mut isum = 0i32;
+
+            // Set up scales
+            LittleEndian::read_u32_into(&x.scales, &mut aux);
+
+            utmp[3] = ((aux[1] >> 4) & KMASK2) | (((aux[2] >> 6) & KMASK1) << 4);
+            utmp[2] = ((aux[0] >> 4) & KMASK2) | (((aux[2] >> 4) & KMASK1) << 4);
+            utmp[1] = (aux[1] & KMASK2) | (((aux[2] >> 2) & KMASK1) << 4);
+            utmp[0] = (aux[0] & KMASK2) | ((aux[2] & KMASK1) << 4);
+
+            let mut scale = utmp.as_mut_ptr() as *mut i8;
+            for j in 0..16 {
+                *scale.add(j) -= 32i8
+            }
+
+            for j in 0..QK_K / 128 {
+                let q3bits = vld1q_u8_x2(q3);
+                q3 = q3.add(32);
+                let q8bytes_1 = vld1q_s8_x4(q8);
+                q8 = q8.add(64);
+                let q8bytes_2 = vld1q_s8_x4(q8);
+                q8 = q8.add(64);
+
+                let q3h_0 = vshlq_n_u8(vbicq_u8(m0, qhbits.0), 2);
+                let q3h_1 = vshlq_n_u8(vbicq_u8(m0, qhbits.1), 2);
+                let q3h_2 = vshlq_n_u8(vbicq_u8(m1, qhbits.0), 1);
+                let q3h_3 = vshlq_n_u8(vbicq_u8(m1, qhbits.1), 1);
+
+                let q3bytes_0 = vsubq_s8(
+                    vreinterpretq_s8_u8(vandq_u8(q3bits.0, m3b)),
+                    vreinterpretq_s8_u8(q3h_0),
+                );
+                let q3bytes_1 = vsubq_s8(
+                    vreinterpretq_s8_u8(vandq_u8(q3bits.1, m3b)),
+                    vreinterpretq_s8_u8(q3h_1),
+                );
+                let q3bytes_2 = vsubq_s8(
+                    vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.0, 2), m3b)),
+                    vreinterpretq_s8_u8(q3h_2),
+                );
+                let q3bytes_3 = vsubq_s8(
+                    vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.1, 2), m3b)),
+                    vreinterpretq_s8_u8(q3h_3),
+                );
+
+                let p0 = vdotq_s32(q3bytes_0, q8bytes_1.0);
+                let p1 = vdotq_s32(q3bytes_1, q8bytes_1.1);
+                let p2 = vdotq_s32(q3bytes_2, q8bytes_1.2);
+                let p3 = vdotq_s32(q3bytes_3, q8bytes_1.3);
+                isum += vaddvq_s32(p0) * *scale as i32
+                    + vaddvq_s32(p1) * *scale.add(1) as i32
+                    + vaddvq_s32(p2) * *scale.add(2) as i32
+                    + vaddvq_s32(p3) * *scale.add(3) as i32;
+                scale = scale.add(4);
+
+                let q3h_0 = vbicq_u8(m2, qhbits.0);
+                let q3h_1 = vbicq_u8(m2, qhbits.1);
+                let q3h_2 = vshrq_n_u8(vbicq_u8(m3, qhbits.0), 1);
+                let q3h_3 = vshrq_n_u8(vbicq_u8(m3, qhbits.1), 1);
+
+                let q3bytes_0 = vsubq_s8(
+                    vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.0, 4), m3b)),
+                    vreinterpretq_s8_u8(q3h_0),
+                );
+                let q3bytes_1 = vsubq_s8(
+                    vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.1, 4), m3b)),
+                    vreinterpretq_s8_u8(q3h_1),
+                );
+                let q3bytes_2 = vsubq_s8(
+                    vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.0, 6), m3b)),
+                    vreinterpretq_s8_u8(q3h_2),
+                );
+                let q3bytes_3 = vsubq_s8(
+                    vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.1, 6), m3b)),
+                    vreinterpretq_s8_u8(q3h_3),
+                );
+
+                let p0 = vdotq_s32(q3bytes_0, q8bytes_2.0);
+                let p1 = vdotq_s32(q3bytes_1, q8bytes_2.1);
+                let p2 = vdotq_s32(q3bytes_2, q8bytes_2.2);
+                let p3 = vdotq_s32(q3bytes_3, q8bytes_2.3);
+                isum += vaddvq_s32(p0) * *scale as i32
+                    + vaddvq_s32(p1) * *scale.add(1) as i32
+                    + vaddvq_s32(p2) * *scale.add(2) as i32
+                    + vaddvq_s32(p3) * *scale.add(3) as i32;
+                scale = scale.add(4);
+
+                if j == 0 {
+                    qhbits.0 = vshrq_n_u8(qhbits.0, 4);
+                    qhbits.1 = vshrq_n_u8(qhbits.1, 4);
+                }
+            }
+            sumf += d * isum as f32;
+        }
+    }
+    Ok(sumf)
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q2k_q8k(n: usize, xs: &[BlockQ2K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q2k_q8k: {n} is not divisible by {QK_K}")
+    }
+    let mut sumf = 0f32;
+    let mut aux = [0u8; 16];
+
+    unsafe {
+        let m3 = vdupq_n_u8(0x3);
+        let m4 = vdupq_n_u8(0xF);
+
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let d = y.d * x.d.to_f32();
+            let dmin = -y.d * x.dmin.to_f32();
+
+            let mut q2 = x.qs.as_ptr();
+            let mut q8 = y.qs.as_ptr();
+            let sc = x.scales.as_ptr();
+
+            let mins_and_scales = vld1q_u8(sc);
+            let scales = vandq_u8(mins_and_scales, m4);
+            vst1q_u8(aux.as_mut_ptr(), scales);
+
+            let mins = vshrq_n_u8(mins_and_scales, 4);
+            let q8sums = vld1q_s16_x2(y.bsums.as_ptr());
+            let mins16 = int16x8x2_t(
+                vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))),
+                vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins))),
+            );
+            let s0 = vaddq_s32(
+                vmull_s16(vget_low_s16(mins16.0), vget_low_s16(q8sums.0)),
+                vmull_s16(vget_high_s16(mins16.0), vget_high_s16(q8sums.0)),
+            );
+            let s1 = vaddq_s32(
+                vmull_s16(vget_low_s16(mins16.1), vget_low_s16(q8sums.1)),
+                vmull_s16(vget_high_s16(mins16.1), vget_high_s16(q8sums.1)),
+            );
+            sumf += dmin * vaddvq_s32(vaddq_s32(s0, s1)) as f32;
+
+            let mut isum = 0i32;
+            let mut is = 0usize;
+
+            // TODO: dotprod
+            for _j in 0..QK_K / 128 {
+                let q2bits = vld1q_u8_x2(q2);
+                q2 = q2.add(32);
+
+                let q8bytes = vld1q_s8_x2(q8);
+                q8 = q8.add(32);
+                let mut q2bytes = int8x16x2_t(
+                    vreinterpretq_s8_u8(vandq_u8(q2bits.0, m3)),
+                    vreinterpretq_s8_u8(vandq_u8(q2bits.1, m3)),
+                );
+                isum += multiply_accum_with_scale(&aux, is, 0, q2bytes, q8bytes);
+
+                let q8bytes = vld1q_s8_x2(q8);
+                q8 = q8.add(32);
+                q2bytes.0 = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.0, 2), m3));
+                q2bytes.1 = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.1, 2), m3));
+                isum += multiply_accum_with_scale(&aux, is, 2, q2bytes, q8bytes);
+
+                let q8bytes = vld1q_s8_x2(q8);
+                q8 = q8.add(32);
+                q2bytes.0 = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.0, 4), m3));
+                q2bytes.1 = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.1, 4), m3));
+                isum += multiply_accum_with_scale(&aux, is, 4, q2bytes, q8bytes);
+
+                let q8bytes = vld1q_s8_x2(q8);
+                q8 = q8.add(32);
+                q2bytes.0 = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.0, 6), m3));
+                q2bytes.1 = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.1, 6), m3));
+                isum += multiply_accum_with_scale(&aux, is, 6, q2bytes, q8bytes);
+
+                is += 8;
+            }
+            sumf += d * isum as f32;
+        }
+    }
+    Ok(sumf)
+}
+
+#[inline(always)]
+unsafe fn multiply_accum_with_scale(
+    aux: &[u8; 16],
+    is: usize,
+    index: usize,
+    q2bytes: int8x16x2_t,
+    q8bytes: int8x16x2_t,
+) -> i32 {
+    let p1 = vdotq_s32(q2bytes.0, q8bytes.0);
+    let p2 = vdotq_s32(q2bytes.1, q8bytes.1);
+    vaddvq_s32(p1) * aux[is + index] as i32 + vaddvq_s32(p2) * aux[is + 1 + index] as i32
+}

candle-core/src/quantized/simd128.rs

@@ -0,0 +1,419 @@
+use super::k_quants::{BlockQ2K, BlockQ4K, BlockQ4_0, BlockQ6K, BlockQ8K, BlockQ8_0, QK8_0, QK_K};
+use crate::Result;
+use byteorder::{ByteOrder, LittleEndian};
+use half::f16;
+
+use core::arch::wasm32::*;
+
+#[inline(always)]
+pub(crate) fn vec_dot_q4_0_q8_0(n: usize, xs: &[BlockQ4_0], ys: &[BlockQ8_0]) -> Result<f32> {
+    let qk = QK8_0;
+    if n % QK8_0 != 0 {
+        crate::bail!("vec_dot_q4_0_q8_0: {n} is not divisible by {qk}")
+    }
+    unsafe {
+        let mut acc = f32x4_splat(0.0f32);
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let x1234 = v128_load(x.qs.as_ptr() as *const v128);
+            let x12 = v128_and(x1234, u8x16_splat(0x0F));
+            let x12 = i8x16_sub(x12, i8x16_splat(8));
+            let x34 = u8x16_shr(x1234, 4);
+            let x34 = i8x16_sub(x34, i8x16_splat(8));
+
+            let x1 = i16x8_extend_low_i8x16(x12);
+            let y1 = i16x8_load_extend_i8x8(y.qs.as_ptr());
+            let sum_xy = i32x4_dot_i16x8(x1, y1);
+
+            let x2 = i16x8_extend_high_i8x16(x12);
+            let y2 = i16x8_load_extend_i8x8(y.qs.as_ptr().add(8));
+            let sum_xy = i32x4_add(sum_xy, i32x4_dot_i16x8(x2, y2));
+
+            let x3 = i16x8_extend_low_i8x16(x34);
+            let y3 = i16x8_load_extend_i8x8(y.qs.as_ptr().add(16));
+            let sum_xy = i32x4_add(sum_xy, i32x4_dot_i16x8(x3, y3));
+
+            let x4 = i16x8_extend_high_i8x16(x34);
+            let y4 = i16x8_load_extend_i8x8(y.qs.as_ptr().add(24));
+            let sum_xy = i32x4_add(sum_xy, i32x4_dot_i16x8(x4, y4));
+
+            let sum_xy = f32x4_convert_i32x4(sum_xy);
+
+            // f32x4_relaxed_madd is nightly only.
+            let d = f32x4_splat(f16::to_f32(x.d) * f16::to_f32(y.d));
+            let scaled = f32x4_mul(sum_xy, d);
+            acc = f32x4_add(acc, scaled)
+        }
+        let res = f32x4_extract_lane::<0>(acc)
+            + f32x4_extract_lane::<1>(acc)
+            + f32x4_extract_lane::<2>(acc)
+            + f32x4_extract_lane::<3>(acc);
+        Ok(res)
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q8_0_q8_0(n: usize, xs: &[BlockQ8_0], ys: &[BlockQ8_0]) -> Result<f32> {
+    let qk = QK8_0;
+    if n % QK8_0 != 0 {
+        crate::bail!("vec_dot_q8_0_q8_0: {n} is not divisible by {qk}")
+    }
+    unsafe {
+        let mut acc = f32x4_splat(0.0f32);
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let x1 = i16x8_load_extend_i8x8(x.qs.as_ptr());
+            let y1 = i16x8_load_extend_i8x8(y.qs.as_ptr());
+            let sum_xy = i32x4_dot_i16x8(x1, y1);
+
+            let x2 = i16x8_load_extend_i8x8(x.qs.as_ptr().add(8));
+            let y2 = i16x8_load_extend_i8x8(y.qs.as_ptr().add(8));
+            let sum_xy = i32x4_add(sum_xy, i32x4_dot_i16x8(x2, y2));
+
+            let x3 = i16x8_load_extend_i8x8(x.qs.as_ptr().add(16));
+            let y3 = i16x8_load_extend_i8x8(y.qs.as_ptr().add(16));
+            let sum_xy = i32x4_add(sum_xy, i32x4_dot_i16x8(x3, y3));
+
+            let x4 = i16x8_load_extend_i8x8(x.qs.as_ptr().add(24));
+            let y4 = i16x8_load_extend_i8x8(y.qs.as_ptr().add(24));
+            let sum_xy = i32x4_add(sum_xy, i32x4_dot_i16x8(x4, y4));
+
+            let sum_xy = f32x4_convert_i32x4(sum_xy);
+
+            // f32x4_relaxed_madd is nightly only.
+            let d = f32x4_splat(f16::to_f32(x.d) * f16::to_f32(y.d));
+            let scaled = f32x4_mul(sum_xy, d);
+            acc = f32x4_add(acc, scaled)
+        }
+        let res = f32x4_extract_lane::<0>(acc)
+            + f32x4_extract_lane::<1>(acc)
+            + f32x4_extract_lane::<2>(acc)
+            + f32x4_extract_lane::<3>(acc);
+        Ok(res)
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q2k_q8k(n: usize, xs: &[BlockQ2K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q2k_q8k: {n} is not divisible by {QK_K}")
+    }
+    unsafe {
+        let mut sumf = f32x4_splat(0f32);
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let mut q2: &[_] = &x.qs;
+            let mut q8: &[_] = &y.qs;
+            let sc = &x.scales;
+
+            let mut summs = i32x4_splat(0);
+            for i in (0..(QK_K / 16)).step_by(4) {
+                let bsums = i32x4_load_extend_i16x4(y.bsums.as_ptr().add(i));
+                let scales = i32x4_shr(
+                    i32x4(
+                        sc[i] as i32,
+                        sc[i + 1] as i32,
+                        sc[i + 2] as i32,
+                        sc[i + 3] as i32,
+                    ),
+                    4,
+                );
+                summs = i32x4_add(summs, i32x4_mul(bsums, scales))
+            }
+            let summs = f32x4_convert_i32x4(summs);
+
+            let dall = y.d * x.d.to_f32();
+            let dmin = y.d * x.dmin.to_f32();
+
+            let mut isum = i32x4_splat(0);
+            let mut is = 0;
+            for _ in 0..(QK_K / 128) {
+                let mut shift = 0;
+                for _ in 0..4 {
+                    let d = (sc[is] & 0xF) as i32;
+                    is += 1;
+                    let mut isuml = i16x8_splat(0);
+                    for l in (0..16).step_by(8) {
+                        let q8 = i16x8_load_extend_i8x8(q8.as_ptr().add(l));
+                        let q2 = i16x8_load_extend_u8x8(q2.as_ptr().add(l));
+                        let q2 = v128_and(i16x8_shr(q2, shift), i16x8_splat(3));
+                        isuml = i16x8_add(isuml, i16x8_mul(q2, q8))
+                    }
+                    let dd = i32x4_splat(d);
+                    isum = i32x4_add(isum, i32x4_mul(i32x4_extend_low_i16x8(isuml), dd));
+                    isum = i32x4_add(isum, i32x4_mul(i32x4_extend_high_i16x8(isuml), dd));
+                    let d = (sc[is] & 0xF) as i32;
+                    is += 1;
+                    let mut isuml = i16x8_splat(0);
+                    for l in (16..32).step_by(8) {
+                        let q8 = i16x8_load_extend_i8x8(q8.as_ptr().add(l));
+                        let q2 = i16x8_load_extend_u8x8(q2.as_ptr().add(l));
+                        let q2 = v128_and(i16x8_shr(q2, shift), i16x8_splat(3));
+                        isuml = i16x8_add(isuml, i16x8_mul(q2, q8))
+                    }
+                    let dd = i32x4_splat(d);
+                    isum = i32x4_add(isum, i32x4_mul(i32x4_extend_low_i16x8(isuml), dd));
+                    isum = i32x4_add(isum, i32x4_mul(i32x4_extend_high_i16x8(isuml), dd));
+                    shift += 2;
+                    // adjust the indexing
+                    q8 = &q8[32..];
+                }
+                // adjust the indexing
+                q2 = &q2[32..];
+            }
+            let isum = f32x4_convert_i32x4(isum);
+            sumf = f32x4_add(
+                sumf,
+                f32x4_sub(
+                    f32x4_mul(isum, f32x4_splat(dall)),
+                    f32x4_mul(summs, f32x4_splat(dmin)),
+                ),
+            );
+        }
+        let sumf = f32x4_extract_lane::<0>(sumf)
+            + f32x4_extract_lane::<1>(sumf)
+            + f32x4_extract_lane::<2>(sumf)
+            + f32x4_extract_lane::<3>(sumf);
+        Ok(sumf)
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q4k_q8k(n: usize, xs: &[BlockQ4K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q4k_q8k: {n} is not divisible by {QK_K}")
+    }
+
+    const KMASK1: u32 = 0x3f3f3f3f;
+    const KMASK2: u32 = 0x0f0f0f0f;
+    const KMASK3: u32 = 0x03030303;
+
+    let mut utmp: [u32; 4] = [0; 4];
+    let mut scales: [u8; 8] = [0; 8];
+    let mut mins: [u8; 8] = [0; 8];
+
+    let mut aux8: [u8; QK_K] = [0; QK_K];
+    let mut sums = f32x4_splat(0f32);
+    unsafe {
+        for (y, x) in ys.iter().zip(xs.iter()) {
+            let q4 = &x.qs;
+            let q8 = &y.qs;
+
+            for j in 0..QK_K / 64 {
+                let q4_1 = v128_load(q4.as_ptr().add(32 * j) as *const v128);
+                let q4_2 = v128_load(q4.as_ptr().add(32 * j + 16) as *const v128);
+                v128_store(
+                    aux8.as_mut_ptr().add(64 * j) as *mut v128,
+                    v128_and(q4_1, u8x16_splat(0x0F)),
+                );
+                v128_store(
+                    aux8.as_mut_ptr().add(64 * j + 16) as *mut v128,
+                    v128_and(q4_2, u8x16_splat(0x0F)),
+                );
+                v128_store(
+                    aux8.as_mut_ptr().add(64 * j + 32) as *mut v128,
+                    u8x16_shr(q4_1, 4),
+                );
+                v128_store(
+                    aux8.as_mut_ptr().add(64 * j + 48) as *mut v128,
+                    u8x16_shr(q4_2, 4),
+                );
+            }
+
+            LittleEndian::read_u32_into(&x.scales, &mut utmp[0..3]);
+
+            utmp[3] = ((utmp[2] >> 4) & KMASK2) | (((utmp[1] >> 6) & KMASK3) << 4);
+            let uaux = utmp[1] & KMASK1;
+            utmp[1] = (utmp[2] & KMASK2) | (((utmp[0] >> 6) & KMASK3) << 4);
+            utmp[2] = uaux;
+            utmp[0] &= KMASK1;
+
+            //extract scales and mins
+            LittleEndian::write_u32_into(&utmp[0..2], &mut scales);
+            LittleEndian::write_u32_into(&utmp[2..4], &mut mins);
+
+            let mut sumi = i32x4_splat(0);
+            for j in (0..QK_K / 16).step_by(4) {
+                let bsums = i32x4_load_extend_i16x4(y.bsums.as_ptr().add(j));
+                let (m1, m2) = (mins[j / 2] as i32, mins[j / 2 + 1] as i32);
+                let mins = i32x4(m1, m1, m2, m2);
+                sumi = i32x4_add(sumi, i32x4_mul(bsums, mins));
+            }
+
+            let mut aux32 = i32x4_splat(0i32);
+            for (scale_i, scale) in scales.iter().enumerate() {
+                let scale = i32x4_splat(*scale as i32);
+                for j in 0..4 {
+                    let i = 32 * scale_i + 8 * j;
+                    let q8 = i16x8_load_extend_i8x8(q8.as_ptr().add(i));
+                    let aux8 = i16x8_load_extend_u8x8(aux8.as_ptr().add(i));
+                    let aux16 = i16x8_mul(q8, aux8);
+                    aux32 = i32x4_add(aux32, i32x4_mul(scale, i32x4_extend_low_i16x8(aux16)));
+                    aux32 = i32x4_add(aux32, i32x4_mul(scale, i32x4_extend_high_i16x8(aux16)));
+                }
+            }
+            let aux32 = f32x4_convert_i32x4(aux32);
+            let d = f32x4_splat(x.d.to_f32() * y.d);
+            sums = f32x4_add(sums, f32x4_mul(aux32, d));
+            let dmin = x.dmin.to_f32() * y.d;
+            let dmin = f32x4_splat(dmin);
+            let sumi = f32x4_convert_i32x4(sumi);
+            sums = f32x4_sub(sums, f32x4_mul(sumi, dmin));
+        }
+        let sums = f32x4_extract_lane::<0>(sums)
+            + f32x4_extract_lane::<1>(sums)
+            + f32x4_extract_lane::<2>(sums)
+            + f32x4_extract_lane::<3>(sums);
+        Ok(sums)
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q6k_q8k(n: usize, xs: &[BlockQ6K], ys: &[BlockQ8K]) -> Result<f32> {
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q6k_q8k: {n} is not divisible by {QK_K}")
+    }
+
+    let mut aux8 = [0i8; QK_K];
+    unsafe {
+        let mut sums = f32x4_splat(0f32);
+
+        for (x, y) in xs.iter().zip(ys.iter()) {
+            let q4 = &x.ql;
+            let qh = &x.qh;
+            let q8 = &y.qs;
+            let mut aux32 = f32x4_splat(0f32);
+
+            for j in (0..QK_K).step_by(128) {
+                let aux8 = aux8.as_mut_ptr().add(j);
+                let q4 = &q4.as_ptr().add(j / 2);
+                let qh = &qh.as_ptr().add(j / 4);
+                for l in (0..32).step_by(16) {
+                    // aux8[l] = (((q4[l] & 0xF) | ((qh[l] & 3) << 4)) as i32 - 32) as i8;
+                    let a8 = v128_or(
+                        v128_and(v128_load(q4.add(l) as *const v128), u8x16_splat(0xF)),
+                        u8x16_shl(
+                            v128_and(v128_load(qh.add(l) as *const v128), u8x16_splat(3)),
+                            4,
+                        ),
+                    );
+                    let a8_low = i16x8_sub(i16x8_extend_low_u8x16(a8), i16x8_splat(32));
+                    let a8_high = i16x8_sub(i16x8_extend_high_u8x16(a8), i16x8_splat(32));
+                    v128_store(
+                        aux8.add(l) as *mut v128,
+                        i8x16_narrow_i16x8(a8_low, a8_high),
+                    );
+
+                    // aux8[l + 32] =
+                    //    (((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) as i32 - 32) as i8;
+                    let a8 = v128_or(
+                        v128_and(v128_load(q4.add(l + 32) as *const v128), u8x16_splat(0xF)),
+                        u8x16_shl(
+                            v128_and(
+                                u8x16_shr(v128_load(qh.add(l) as *const v128), 2),
+                                u8x16_splat(3),
+                            ),
+                            4,
+                        ),
+                    );
+                    let a8_low = i16x8_sub(i16x8_extend_low_u8x16(a8), i16x8_splat(32));
+                    let a8_high = i16x8_sub(i16x8_extend_high_u8x16(a8), i16x8_splat(32));
+                    v128_store(
+                        aux8.add(l + 32) as *mut v128,
+                        i8x16_narrow_i16x8(a8_low, a8_high),
+                    );
+
+                    // aux8[l + 64] = (((q4[l] >> 4) | (((qh[l] >> 4) & 3) << 4)) as i32 - 32) as i8;
+                    let a8 = v128_or(
+                        u8x16_shr(v128_load(q4.add(l) as *const v128), 4),
+                        u8x16_shl(
+                            v128_and(
+                                u8x16_shr(v128_load(qh.add(l) as *const v128), 4),
+                                u8x16_splat(3),
+                            ),
+                            4,
+                        ),
+                    );
+                    let a8_low = i16x8_sub(i16x8_extend_low_u8x16(a8), i16x8_splat(32));
+                    let a8_high = i16x8_sub(i16x8_extend_high_u8x16(a8), i16x8_splat(32));
+                    v128_store(
+                        aux8.add(l + 64) as *mut v128,
+                        i8x16_narrow_i16x8(a8_low, a8_high),
+                    );
+
+                    // aux8[l + 96] =
+                    //    (((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) as i32 - 32) as i8;
+                    let a8 = v128_or(
+                        u8x16_shr(v128_load(q4.add(l + 32) as *const v128), 4),
+                        u8x16_shl(
+                            v128_and(
+                                u8x16_shr(v128_load(qh.add(l) as *const v128), 6),
+                                u8x16_splat(3),
+                            ),
+                            4,
+                        ),
+                    );
+                    let a8_low = i16x8_sub(i16x8_extend_low_u8x16(a8), i16x8_splat(32));
+                    let a8_high = i16x8_sub(i16x8_extend_high_u8x16(a8), i16x8_splat(32));
+                    v128_store(
+                        aux8.add(l + 96) as *mut v128,
+                        i8x16_narrow_i16x8(a8_low, a8_high),
+                    );
+                }
+            }
+
+            for (j, &scale) in x.scales.iter().enumerate() {
+                let scale = f32x4_splat(scale as f32);
+                for offset in [0, 8] {
+                    let aux16 = i16x8_mul(
+                        i16x8_load_extend_i8x8(q8.as_ptr().add(16 * j + offset)),
+                        i16x8_load_extend_i8x8(aux8.as_ptr().add(16 * j + offset)),
+                    );
+                    aux32 = f32x4_add(
+                        aux32,
+                        f32x4_mul(f32x4_convert_i32x4(i32x4_extend_low_i16x8(aux16)), scale),
+                    );
+                    aux32 = f32x4_add(
+                        aux32,
+                        f32x4_mul(f32x4_convert_i32x4(i32x4_extend_high_i16x8(aux16)), scale),
+                    );
+                }
+            }
+
+            let d = f32x4_splat(x.d.to_f32() * y.d);
+            sums = f32x4_add(sums, f32x4_mul(aux32, d));
+        }
+        let sums = f32x4_extract_lane::<0>(sums)
+            + f32x4_extract_lane::<1>(sums)
+            + f32x4_extract_lane::<2>(sums)
+            + f32x4_extract_lane::<3>(sums);
+        Ok(sums)
+    }
+}
+
+#[inline(always)]
+pub(crate) fn vec_dot_q8k_q8k(n: usize, xs: &[BlockQ8K], ys: &[BlockQ8K]) -> Result<f32> {
+    let qk = QK_K;
+    if n % QK_K != 0 {
+        crate::bail!("vec_dot_q8k_q8k: {n} is not divisible by {qk}")
+    }
+
+    unsafe {
+        let mut acc = f32x4_splat(0.0f32);
+        for (xs, ys) in xs.iter().zip(ys.iter()) {
+            let x_qs = xs.qs.as_ptr();
+            let y_qs = ys.qs.as_ptr();
+            let mut sumi = i32x4_splat(0);
+            for j in (0..QK_K).step_by(8) {
+                let xs = i16x8_load_extend_i8x8(x_qs.add(j));
+                let ys = i16x8_load_extend_i8x8(y_qs.add(j));
+                let sum_xy = i32x4_dot_i16x8(xs, ys);
+                sumi = i32x4_add(sumi, sum_xy)
+            }
+            let d = f32x4_splat(xs.d * ys.d);
+            acc = f32x4_add(acc, f32x4_mul(f32x4_convert_i32x4(sumi), d))
+        }
+        let res = f32x4_extract_lane::<0>(acc)
+            + f32x4_extract_lane::<1>(acc)
+            + f32x4_extract_lane::<2>(acc)
+            + f32x4_extract_lane::<3>(acc);
+        Ok(res)
+    }
+}

candle-core/src/quantized/utils.rs

@@ -0,0 +1,326 @@
+use crate::Result;
+
+pub(super) fn nearest_int(v: f32) -> i32 {
+    v.round() as i32
+}
+
+/// Validates that the input and output are the right size and returns an iterator which maps each
+/// input region `xs` to its corresponding output block in `ys`. Each output region is guaranteed
+/// to be `T::BLCK_SIZE` long.
+pub(super) fn group_for_quantization<'a, 'b, T: super::k_quants::GgmlType>(
+    xs: &'b [f32],
+    ys: &'a mut [T],
+) -> Result<Vec<(&'a mut T, &'b [f32])>> {
+    let block_size = T::BLCK_SIZE;
+    let dtype = T::DTYPE;
+
+    let expected_blocks = xs.len() / block_size;
+    let actual_blocks = ys.len();
+
+    // Validate that the input is the right size
+    if expected_blocks != actual_blocks {
+        crate::bail!("quantize {dtype:?}: expected {expected_blocks} blocks but only {actual_blocks} were provided!")
+    }
+
+    Ok(ys.iter_mut().zip(xs.chunks_exact(block_size)).collect())
+}
+
+/// Validates that the input and output are the right size and returns an iterator which maps each
+/// input block `xs` to its corresponding output region in `ys`. Each output region is guaranteed
+/// to be `T::BLCK_SIZE` long.
+pub(super) fn group_for_dequantization<'a, 'b, T: super::k_quants::GgmlType>(
+    xs: &'a [T],
+    ys: &'b mut [f32],
+) -> Result<Vec<(&'a T, &'b mut [f32])>> {
+    let block_size = T::BLCK_SIZE;
+    let dtype = T::DTYPE;
+
+    let actual_output_len = ys.len();
+    let expected_output_len = xs.len() * block_size;
+    // Validate that the output is the right size
+    if expected_output_len != actual_output_len {
+        crate::bail!("dequantize {dtype:?}: ys (len = {actual_output_len}) does not match the expected length of {expected_output_len}!")
+    }
+
+    // Zip the blocks and outputs together
+    Ok(xs.iter().zip(ys.chunks_exact_mut(block_size)).collect())
+}
+
+pub(super) fn get_scale_min_k4(j: usize, q: &[u8]) -> (u8, u8) {
+    if j < 4 {
+        let d = q[j] & 63;
+        let m = q[j + 4] & 63;
+        (d, m)
+    } else {
+        let d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
+        let m = (q[j + 4] >> 4) | ((q[j] >> 6) << 4);
+        (d, m)
+    }
+}
+
+pub(super) unsafe fn make_qx_quants(
+    n: usize,
+    nmax: i32,
+    x: *const f32,
+    ls: *mut i8,
+    rmse_type: i32,
+) -> f32 {
+    let mut max = 0f32;
+    let mut amax = 0f32;
+    for i in 0..n {
+        let x = *x.add(i);
+        let ax = x.abs();
+        if ax > amax {
+            amax = ax;
+            max = x;
+        }
+    }
+    if amax == 0. {
+        // all zero
+        for i in 0..n {
+            *ls.add(i) = 0;
+        }
+        return 0.;
+    }
+    let mut iscale = -(nmax as f32) / max;
+    if rmse_type == 0 {
+        for i in 0..n {
+            let x = *x.add(i);
+            let l = nearest_int(iscale * x);
+            *ls.add(i) = (nmax + l.clamp(-nmax, nmax - 1)) as i8;
+        }
+        return 1.0 / iscale;
+    }
+    let weight_type = rmse_type % 2;
+    let mut sumlx = 0f32;
+    let mut suml2 = 0f32;
+    for i in 0..n {
+        let x = *x.add(i);
+        let l = nearest_int(iscale * x);
+        let l = l.clamp(-nmax, nmax - 1);
+        *ls.add(i) = (l + nmax) as i8;
+        let w = if weight_type == 1 { x * x } else { 1.0 };
+        let l = l as f32;
+        sumlx += w * x * l;
+        suml2 += w * l * l;
+    }
+    let mut scale = sumlx / suml2;
+    let mut best = scale * sumlx;
+    for _itry in 0..3 {
+        let iscale = 1.0 / scale;
+        let mut slx = 0f32;
+        let mut sl2 = 0f32;
+        let mut changed = false;
+        for i in 0..n {
+            let x = *x.add(i);
+            let l = nearest_int(iscale * x);
+            let l = l.clamp(-nmax, nmax - 1);
+            if l + nmax != *ls.add(i) as i32 {
+                changed = true;
+            }
+            let w = if weight_type == 1 { x * x } else { 1f32 };
+            let l = l as f32;
+            slx += w * x * l;
+            sl2 += w * l * l;
+        }
+        if !changed || sl2 == 0.0 || slx * slx <= best * sl2 {
+            break;
+        }
+        for i in 0..n {
+            let x = *x.add(i);
+            let l = nearest_int(iscale * x);
+            *ls.add(i) = (nmax + l.clamp(-nmax, nmax - 1)) as i8;
+        }
+        sumlx = slx;
+        suml2 = sl2;
+        scale = sumlx / suml2;
+        best = scale * sumlx;
+    }
+    for _itry in 0..5 {
+        let mut n_changed = 0;
+        for i in 0..n {
+            let x = *x.add(i);
+            let w = if weight_type == 1 { x * x } else { 1. };
+            let l = *ls.add(i) as i32 - nmax;
+            let mut slx = sumlx - w * x * l as f32;
+            if slx > 0. {
+                let mut sl2 = suml2 - w * l as f32 * l as f32;
+                let new_l = nearest_int(x * sl2 / slx);
+                let new_l = new_l.clamp(-nmax, nmax - 1);
+                if new_l != l {
+                    slx += w * x * new_l as f32;
+                    sl2 += w * new_l as f32 * new_l as f32;
+                    if sl2 > 0. && slx * slx * suml2 > sumlx * sumlx * sl2 {
+                        *ls.add(i) = (nmax + new_l) as i8;
+                        sumlx = slx;
+                        suml2 = sl2;
+                        scale = sumlx / suml2;
+                        best = scale * sumlx;
+                        n_changed += 1;
+                    }
+                }
+            }
+        }
+        if n_changed == 0 {
+            break;
+        }
+    }
+    if rmse_type < 3 {
+        return scale;
+    }
+    for is in -4..4 {
+        if is == 0 {
+            continue;
+        }
+        iscale = -(nmax as f32 + 0.1f32 * is as f32) / max;
+        let mut sumlx = 0.;
+        let mut suml2 = 0.;
+        for i in 0..n {
+            let x = *x.add(i);
+            let l = nearest_int(iscale * x);
+            let l = l.clamp(-nmax, nmax - 1);
+            let w = if weight_type == 1 { x * x } else { 1. };
+            let l = l as f32;
+            sumlx += w * x * l;
+            suml2 += w * l * l;
+        }
+        if suml2 > 0. && sumlx * sumlx > best * suml2 {
+            for i in 0..n {
+                let x = *x.add(i);
+                let l = nearest_int(iscale * x);
+                *ls.add(i) = (nmax + l.clamp(-nmax, nmax - 1)) as i8;
+            }
+            scale = sumlx / suml2;
+            best = scale * sumlx;
+        }
+    }
+    scale
+}
+
+// https://github.com/ggerganov/llama.cpp/blob/8183159cf3def112f6d1fe94815fce70e1bffa12/k_quants.c#L224
+pub(super) fn make_qkx1_quants(nmax: i32, ntry: usize, x: &[f32]) -> (f32, f32) {
+    let n = x.len();
+    let mut l = vec![0; n];
+    // Get min/max
+    let min = *x
+        .iter()
+        .take(n)
+        .min_by(|a, b| a.total_cmp(b))
+        .unwrap_or(&x[0]);
+    let max = *x.iter().max_by(|a, b| a.total_cmp(b)).unwrap_or(&x[0]);
+
+    // If min == max, all values are the same => nothing to do here
+    if max == min {
+        return (0.0, 0.0);
+    }
+
+    // Ensure min <= 0.0
+    let mut min = min.min(0.);
+
+    // Compute scale and inverse scale
+    let mut iscale = nmax as f32 / (max - min);
+    let mut scale = 1.0 / iscale;
+
+    for _ in 0..ntry {
+        let mut sumlx = 0.0;
+        let mut suml2 = 0;
+        let mut did_change = false;
+
+        for (i, value) in x.iter().enumerate().take(n) {
+            let li = nearest_int(iscale * (value - min)).clamp(0, nmax);
+            let clamped_li = li as u8;
+            if clamped_li != l[i] {
+                l[i] = clamped_li;
+                did_change = true;
+            }
+            sumlx += (value - min) * li as f32;
+            suml2 += li * li;
+        }
+        scale = sumlx / suml2 as f32;
+
+        let sum: f32 = x
+            .iter()
+            .take(n)
+            .zip(l.iter().take(n))
+            .map(|(xi, &li)| xi - scale * li as f32)
+            .sum();
+
+        min = sum / n as f32;
+        if min > 0.0 {
+            min = 0.0;
+        }
+        iscale = 1.0 / scale;
+        if !did_change {
+            break;
+        }
+    }
+    (scale, -min)
+}
+
+// https://github.com/ggerganov/llama.cpp/blob/8183159cf3def112f6d1fe94815fce70e1bffa12/k_quants.c#L165
+pub(super) fn make_q3_quants(x: &[f32], nmax: i32, do_rmse: bool) -> f32 {
+    let n = x.len();
+    let mut l = vec![0i8; n];
+
+    let mut max = 0.0;
+    let mut amax = 0.0;
+    for &xi in x.iter().take(n) {
+        let ax = xi.abs();
+        if ax > amax {
+            amax = ax;
+            max = xi;
+        }
+    }
+
+    if amax == 0.0 {
+        return 0.0;
+    }
+
+    let iscale = -(nmax as f32) / max;
+    if do_rmse {
+        let mut sumlx = 0.0;
+        let mut suml2 = 0.0;
+        for i in 0..n {
+            let li = (iscale * x[i]).round() as i32;
+            let li = li.clamp(-nmax, nmax - 1);
+            l[i] = li as i8;
+            let w = x[i] * x[i];
+            sumlx += w * x[i] * li as f32;
+            suml2 += w * (li * li) as f32;
+        }
+        for _ in 0..5 {
+            let mut n_changed = 0;
+            for i in 0..n {
+                let w = x[i] * x[i];
+                let mut slx = sumlx - w * x[i] * l[i] as f32;
+                if slx > 0.0 {
+                    let mut sl2 = suml2 - w * (l[i] as i32 * l[i] as i32) as f32;
+                    let mut new_l = (x[i] * sl2 / slx).round() as i32;
+                    new_l = new_l.clamp(-nmax, nmax - 1);
+                    if new_l != l[i] as i32 {
+                        slx += w * x[i] * new_l as f32;
+                        sl2 += w * (new_l * new_l) as f32;
+                        if sl2 > 0.0 && slx * slx * suml2 > sumlx * sumlx * sl2 {
+                            l[i] = new_l as i8;
+                            sumlx = slx;
+                            suml2 = sl2;
+                            n_changed += 1;
+                        }
+                    }
+                }
+            }
+            if n_changed == 0 {
+                break;
+            }
+        }
+        for li in l.iter_mut() {
+            *li += nmax as i8;
+        }
+        return sumlx / suml2;
+    }
+    for i in 0..n {
+        let li = (iscale * x[i]).round() as i32;
+        l[i] = (li.clamp(-nmax, nmax - 1) + nmax) as i8;
+    }
+    1.0 / iscale
+}

candle-core/src/safetensors.rs

@@ -10,6 +10,7 @@ impl From<DType> for st::Dtype {
         match value {
             DType::U8 => st::Dtype::U8,
             DType::U32 => st::Dtype::U32,
+            DType::I64 => st::Dtype::I64,
             DType::BF16 => st::Dtype::BF16,
             DType::F16 => st::Dtype::F16,
             DType::F32 => st::Dtype::F32,
@@ -24,6 +25,7 @@ impl TryFrom<st::Dtype> for DType {
         match value {
             st::Dtype::U8 => Ok(DType::U8),
             st::Dtype::U32 => Ok(DType::U32),
+            st::Dtype::I64 => Ok(DType::I64),
             st::Dtype::BF16 => Ok(DType::BF16),
             st::Dtype::F16 => Ok(DType::F16),
             st::Dtype::F32 => Ok(DType::F32),
@@ -76,11 +78,7 @@ impl st::View for &Tensor {
 }
 
 impl Tensor {
-    pub fn save_safetensors<P: AsRef<std::path::Path>>(
-        &self,
-        name: &str,
-        filename: P,
-    ) -> Result<()> {
+    pub fn save_safetensors<P: AsRef<Path>>(&self, name: &str, filename: P) -> Result<()> {
         let data = [(name, self.clone())];
         Ok(st::serialize_to_file(data, &None, filename.as_ref())?)
     }
@@ -189,6 +187,7 @@ impl Tensor {
         match dtype {
             DType::U8 => convert_slice::<u8>(data, shape, device),
             DType::U32 => convert_slice::<u32>(data, shape, device),
+            DType::I64 => convert_slice::<i64>(data, shape, device),
             DType::BF16 => convert_slice::<half::bf16>(data, shape, device),
             DType::F16 => convert_slice::<half::f16>(data, shape, device),
             DType::F32 => convert_slice::<f32>(data, shape, device),
@@ -205,24 +204,15 @@ fn convert(view: &st::TensorView<'_>, device: &Device) -> Result<Tensor> {
             convert_with_cast_::<u16, u32, _>(view, device, conv)
         }
         st::Dtype::U32 => convert_::<u32>(view, device),
+        st::Dtype::I32 => {
+            let conv = |x| Ok(i64::from(x));
+            convert_with_cast_::<i32, i64, _>(view, device, conv)
+        }
+        st::Dtype::I64 => convert_::<i64>(view, device),
         st::Dtype::BF16 => convert_::<half::bf16>(view, device),
         st::Dtype::F16 => convert_::<half::f16>(view, device),
         st::Dtype::F32 => convert_::<f32>(view, device),
         st::Dtype::F64 => convert_::<f64>(view, device),
-        st::Dtype::I32 => {
-            let conv = |x| {
-                u32::try_from(x)
-                    .map_err(|_| Error::Msg(format!("out of bounds value for u32: {x}")))
-            };
-            convert_with_cast_::<i32, u32, _>(view, device, conv)
-        }
-        st::Dtype::I64 => {
-            let conv = |x| {
-                u32::try_from(x)
-                    .map_err(|_| Error::Msg(format!("out of bounds value for u32: {x}")))
-            };
-            convert_with_cast_::<i64, u32, _>(view, device, conv)
-        }
         dtype => Err(Error::UnsupportedSafeTensorDtype(dtype)),
     }
 }
@@ -233,6 +223,7 @@ fn convert_back(tensor: &Tensor) -> Result<Vec<u8>> {
     match tensor.dtype() {
         DType::U8 => Ok(convert_back_::<u8>(tensor.to_vec1()?)),
         DType::U32 => Ok(convert_back_::<u32>(tensor.to_vec1()?)),
+        DType::I64 => Ok(convert_back_::<i64>(tensor.to_vec1()?)),
         DType::F16 => Ok(convert_back_::<half::f16>(tensor.to_vec1()?)),
         DType::BF16 => Ok(convert_back_::<half::bf16>(tensor.to_vec1()?)),
         DType::F32 => Ok(convert_back_::<f32>(tensor.to_vec1()?)),
@@ -260,6 +251,158 @@ pub fn save<K: AsRef<str> + Ord + std::fmt::Display, P: AsRef<Path>>(
     Ok(st::serialize_to_file(tensors, &None, filename.as_ref())?)
 }
 
+#[derive(yoke::Yokeable)]
+struct SafeTensors_<'a>(SafeTensors<'a>);
+
+pub struct MmapedSafetensors {
+    safetensors: Vec<yoke::Yoke<SafeTensors_<'static>, memmap2::Mmap>>,
+    routing: Option<HashMap<String, usize>>,
+}
+
+impl MmapedSafetensors {
+    /// Creates a wrapper around a memory mapped file and deserialize the safetensors header.
+    ///
+    /// # Safety
+    ///
+    /// The unsafe is inherited from [`memmap2::MmapOptions`].
+    pub unsafe fn new<P: AsRef<Path>>(p: P) -> Result<Self> {
+        let p = p.as_ref();
+        let file = std::fs::File::open(p).map_err(|e| Error::from(e).with_path(p))?;
+        let file = memmap2::MmapOptions::new()
+            .map(&file)
+            .map_err(|e| Error::from(e).with_path(p))?;
+        let safetensors = yoke::Yoke::<SafeTensors_<'static>, memmap2::Mmap>::try_attach_to_cart(
+            file,
+            |data: &[u8]| {
+                let st = safetensors::SafeTensors::deserialize(data)
+                    .map_err(|e| Error::from(e).with_path(p))?;
+                Ok::<_, Error>(SafeTensors_(st))
+            },
+        )?;
+        Ok(Self {
+            safetensors: vec![safetensors],
+            routing: None,
+        })
+    }
+
+    /// Creates a wrapper around multiple memory mapped file and deserialize the safetensors headers.
+    ///
+    /// If a tensor name appears in multiple files, the last entry is returned.
+    ///
+    /// # Safety
+    ///
+    /// The unsafe is inherited from [`memmap2::MmapOptions`].
+    pub unsafe fn multi<P: AsRef<Path>>(paths: &[P]) -> Result<Self> {
+        let mut routing = HashMap::new();
+        let mut safetensors = vec![];
+        for (index, p) in paths.iter().enumerate() {
+            let p = p.as_ref();
+            let file = std::fs::File::open(p).map_err(|e| Error::from(e).with_path(p))?;
+            let file = memmap2::MmapOptions::new()
+                .map(&file)
+                .map_err(|e| Error::from(e).with_path(p))?;
+            let data = yoke::Yoke::<SafeTensors_<'static>, memmap2::Mmap>::try_attach_to_cart(
+                file,
+                |data: &[u8]| {
+                    let st = safetensors::SafeTensors::deserialize(data)
+                        .map_err(|e| Error::from(e).with_path(p))?;
+                    Ok::<_, Error>(SafeTensors_(st))
+                },
+            )?;
+            for k in data.get().0.names() {
+                routing.insert(k.to_string(), index);
+            }
+            safetensors.push(data)
+        }
+        Ok(Self {
+            safetensors,
+            routing: Some(routing),
+        })
+    }
+
+    pub fn load(&self, name: &str, dev: &Device) -> Result<Tensor> {
+        self.get(name)?.load(dev)
+    }
+
+    pub fn tensors(&self) -> Vec<(String, st::TensorView<'_>)> {
+        let mut tensors = vec![];
+        for safetensors in self.safetensors.iter() {
+            tensors.push(safetensors.get().0.tensors())
+        }
+        tensors.into_iter().flatten().collect()
+    }
+
+    pub fn get(&self, name: &str) -> Result<st::TensorView<'_>> {
+        let index = match &self.routing {
+            None => 0,
+            Some(routing) => {
+                let index = routing.get(name).ok_or_else(|| {
+                    Error::CannotFindTensor {
+                        path: name.to_string(),
+                    }
+                    .bt()
+                })?;
+                *index
+            }
+        };
+        Ok(self.safetensors[index].get().0.tensor(name)?)
+    }
+}
+
+pub struct SliceSafetensors<'a> {
+    safetensors: SafeTensors<'a>,
+}
+
+impl<'a> SliceSafetensors<'a> {
+    /// Creates a wrapper around a binary buffer and deserialize the safetensors header.
+    pub fn new(buffer: &'a [u8]) -> Result<Self> {
+        let safetensors = safetensors::SafeTensors::deserialize(buffer)?;
+        Ok(Self { safetensors })
+    }
+
+    pub fn load(&self, name: &str, dev: &Device) -> Result<Tensor> {
+        self.safetensors.tensor(name)?.load(dev)
+    }
+
+    pub fn tensors(&self) -> Vec<(String, st::TensorView<'_>)> {
+        self.safetensors.tensors()
+    }
+
+    pub fn get(&self, name: &str) -> Result<st::TensorView<'_>> {
+        Ok(self.safetensors.tensor(name)?)
+    }
+}
+
+pub struct BufferedSafetensors {
+    safetensors: yoke::Yoke<SafeTensors_<'static>, Vec<u8>>,
+}
+
+impl BufferedSafetensors {
+    /// Creates a wrapper around a binary buffer and deserialize the safetensors header.
+    pub fn new(buffer: Vec<u8>) -> Result<Self> {
+        let safetensors = yoke::Yoke::<SafeTensors_<'static>, Vec<u8>>::try_attach_to_cart(
+            buffer,
+            |data: &[u8]| {
+                let st = safetensors::SafeTensors::deserialize(data)?;
+                Ok::<_, Error>(SafeTensors_(st))
+            },
+        )?;
+        Ok(Self { safetensors })
+    }
+
+    pub fn load(&self, name: &str, dev: &Device) -> Result<Tensor> {
+        self.get(name)?.load(dev)
+    }
+
+    pub fn tensors(&self) -> Vec<(String, st::TensorView<'_>)> {
+        self.safetensors.get().0.tensors()
+    }
+
+    pub fn get(&self, name: &str) -> Result<st::TensorView<'_>> {
+        Ok(self.safetensors.get().0.tensor(name)?)
+    }
+}
+
 pub struct MmapedFile {
     path: std::path::PathBuf,
     inner: memmap2::Mmap,
@@ -272,7 +415,7 @@ impl MmapedFile {
     /// # Safety
     ///
     /// The unsafe is inherited from [`memmap2::MmapOptions`].
-    pub unsafe fn new<P: AsRef<std::path::Path>>(p: P) -> Result<Self> {
+    pub unsafe fn new<P: AsRef<Path>>(p: P) -> Result<Self> {
         let p = p.as_ref();
         let file = std::fs::File::open(p).map_err(|e| Error::from(e).with_path(p))?;
         let inner = memmap2::MmapOptions::new()

candle-core/src/scalar.rs

@@ -0,0 +1,23 @@
+use crate::{Result, Tensor, WithDType};
+
+pub enum TensorScalar {
+    Tensor(Tensor),
+    Scalar(Tensor),
+}
+
+pub trait TensorOrScalar {
+    fn to_tensor_scalar(self) -> Result<TensorScalar>;
+}
+
+impl TensorOrScalar for &Tensor {
+    fn to_tensor_scalar(self) -> Result<TensorScalar> {
+        Ok(TensorScalar::Tensor(self.clone()))
+    }
+}
+
+impl<T: WithDType> TensorOrScalar for T {
+    fn to_tensor_scalar(self) -> Result<TensorScalar> {
+        let scalar = Tensor::new(self, &crate::Device::Cpu)?;
+        Ok(TensorScalar::Scalar(scalar))
+    }
+}

candle-core/src/shape.rs

@@ -1,3 +1,5 @@
+//! The shape of a tensor is a tuple with the size of each of its dimensions.
+#![allow(clippy::redundant_closure_call)]
 use crate::{Error, Result};
 
 #[derive(Clone, PartialEq, Eq)]
@@ -71,6 +73,14 @@ impl From<(usize, usize, usize, usize, usize)> for Shape {
     }
 }
 
+impl From<(usize, usize, usize, usize, usize, usize)> for Shape {
+    fn from(d123456: (usize, usize, usize, usize, usize, usize)) -> Self {
+        Self(vec![
+            d123456.0, d123456.1, d123456.2, d123456.3, d123456.4, d123456.5,
+        ])
+    }
+}
+
 impl From<Vec<usize>> for Shape {
     fn from(dims: Vec<usize>) -> Self {
         Self(dims)
@@ -118,6 +128,7 @@ impl Shape {
         Self(dims.to_vec())
     }
 
+    /// The rank is the number of dimensions, 0 for a scalar value, 1 for a vector, etc.
     pub fn rank(&self) -> usize {
         self.0.len()
     }
@@ -126,10 +137,12 @@ impl Shape {
         self.0
     }
 
+    /// The dimensions as a slice of `usize`.
     pub fn dims(&self) -> &[usize] {
         &self.0
     }
 
+    /// The total number of elements, this is the product of all dimension sizes.
     pub fn elem_count(&self) -> usize {
         self.0.iter().product()
     }
@@ -158,7 +171,7 @@ impl Shape {
         }
         let mut acc = 1;
         for (&stride, &dim) in stride.iter().zip(self.0.iter()).rev() {
-            if stride != acc {
+            if dim > 1 && stride != acc {
                 return false;
             }
             acc *= dim;
@@ -173,7 +186,7 @@ impl Shape {
         }
         let mut acc = 1;
         for (&stride, &dim) in stride.iter().zip(self.0.iter()) {
-            if stride != acc {
+            if dim > 1 && stride != acc {
                 return false;
             }
             acc *= dim;
@@ -181,10 +194,75 @@ impl Shape {
         true
     }
 
+    /// Modifies the shape by adding a list of additional dimensions at the end of the existing
+    /// dimensions.
     pub fn extend(mut self, additional_dims: &[usize]) -> Self {
         self.0.extend(additional_dims);
         self
     }
+
+    /// Check whether the two shapes are compatible for broadcast, and if it is the case return the
+    /// broadcasted shape. This is to be used for binary pointwise ops.
+    pub fn broadcast_shape_binary_op(&self, rhs: &Self, op: &'static str) -> Result<Shape> {
+        let lhs = self;
+        let lhs_dims = lhs.dims();
+        let rhs_dims = rhs.dims();
+        let lhs_ndims = lhs_dims.len();
+        let rhs_ndims = rhs_dims.len();
+        let bcast_ndims = usize::max(lhs_ndims, rhs_ndims);
+        let mut bcast_dims = vec![0; bcast_ndims];
+        for (idx, bcast_value) in bcast_dims.iter_mut().enumerate() {
+            let rev_idx = bcast_ndims - idx;
+            let l_value = if lhs_ndims < rev_idx {
+                1
+            } else {
+                lhs_dims[lhs_ndims - rev_idx]
+            };
+            let r_value = if rhs_ndims < rev_idx {
+                1
+            } else {
+                rhs_dims[rhs_ndims - rev_idx]
+            };
+            *bcast_value = if l_value == r_value {
+                l_value
+            } else if l_value == 1 {
+                r_value
+            } else if r_value == 1 {
+                l_value
+            } else {
+                Err(Error::ShapeMismatchBinaryOp {
+                    lhs: lhs.clone(),
+                    rhs: rhs.clone(),
+                    op,
+                }
+                .bt())?
+            }
+        }
+        Ok(Shape::from(bcast_dims))
+    }
+
+    pub(crate) fn broadcast_shape_matmul(&self, rhs: &Self) -> Result<(Shape, Shape)> {
+        let lhs = self;
+        let lhs_dims = lhs.dims();
+        let rhs_dims = rhs.dims();
+        if lhs_dims.len() < 2 || rhs_dims.len() < 2 {
+            crate::bail!("only 2d matrixes are supported {lhs:?} {rhs:?}")
+        }
+        let (m, lhs_k) = (lhs_dims[lhs_dims.len() - 2], lhs_dims[lhs_dims.len() - 1]);
+        let (rhs_k, n) = (rhs_dims[rhs_dims.len() - 2], rhs_dims[rhs_dims.len() - 1]);
+        if lhs_k != rhs_k {
+            crate::bail!("different inner dimensions in broadcast matmul {lhs:?} {rhs:?}")
+        }
+
+        let lhs_b = Self::from(&lhs_dims[..lhs_dims.len() - 2]);
+        let rhs_b = Self::from(&rhs_dims[..rhs_dims.len() - 2]);
+        let bcast = lhs_b.broadcast_shape_binary_op(&rhs_b, "broadcast_matmul")?;
+        let bcast_dims = bcast.dims();
+
+        let bcast_lhs = [bcast_dims, &[m, lhs_k]].concat();
+        let bcast_rhs = [bcast_dims, &[rhs_k, n]].concat();
+        Ok((Shape::from(bcast_lhs), Shape::from(bcast_rhs)))
+    }
 }
 
 pub trait Dim {
@@ -226,6 +304,7 @@ impl Dim for usize {
 pub enum D {
     Minus1,
     Minus2,
+    Minus(usize),
 }
 
 impl D {
@@ -233,6 +312,7 @@ impl D {
         let dim = match self {
             Self::Minus1 => -1,
             Self::Minus2 => -2,
+            Self::Minus(u) => -(*u as i32),
         };
         Error::DimOutOfRange {
             shape: shape.clone(),
@@ -249,6 +329,7 @@ impl Dim for D {
         match self {
             Self::Minus1 if rank >= 1 => Ok(rank - 1),
             Self::Minus2 if rank >= 2 => Ok(rank - 2),
+            Self::Minus(u) if *u > 0 && rank >= *u => Ok(rank - *u),
             _ => Err(self.out_of_range(shape, op)),
         }
     }
@@ -258,6 +339,7 @@ impl Dim for D {
         match self {
             Self::Minus1 => Ok(rank),
             Self::Minus2 if rank >= 1 => Ok(rank - 1),
+            Self::Minus(u) if *u > 0 && rank + 1 >= *u => Ok(rank + 1 - *u),
             _ => Err(self.out_of_range(shape, op)),
         }
     }
@@ -345,6 +427,39 @@ impl<D1: Dim, D2: Dim, D3: Dim> Dims for (D1, D2, D3) {
     }
 }
 
+impl<D1: Dim, D2: Dim, D3: Dim, D4: Dim> Dims for (D1, D2, D3, D4) {
+    fn to_indexes_internal(self, shape: &Shape, op: &'static str) -> Result<Vec<usize>> {
+        let d0 = self.0.to_index(shape, op)?;
+        let d1 = self.1.to_index(shape, op)?;
+        let d2 = self.2.to_index(shape, op)?;
+        let d3 = self.3.to_index(shape, op)?;
+        Ok(vec![d0, d1, d2, d3])
+    }
+}
+
+impl<D1: Dim, D2: Dim, D3: Dim, D4: Dim, D5: Dim> Dims for (D1, D2, D3, D4, D5) {
+    fn to_indexes_internal(self, shape: &Shape, op: &'static str) -> Result<Vec<usize>> {
+        let d0 = self.0.to_index(shape, op)?;
+        let d1 = self.1.to_index(shape, op)?;
+        let d2 = self.2.to_index(shape, op)?;
+        let d3 = self.3.to_index(shape, op)?;
+        let d4 = self.4.to_index(shape, op)?;
+        Ok(vec![d0, d1, d2, d3, d4])
+    }
+}
+
+impl<D1: Dim, D2: Dim, D3: Dim, D4: Dim, D5: Dim, D6: Dim> Dims for (D1, D2, D3, D4, D5, D6) {
+    fn to_indexes_internal(self, shape: &Shape, op: &'static str) -> Result<Vec<usize>> {
+        let d0 = self.0.to_index(shape, op)?;
+        let d1 = self.1.to_index(shape, op)?;
+        let d2 = self.2.to_index(shape, op)?;
+        let d3 = self.3.to_index(shape, op)?;
+        let d4 = self.4.to_index(shape, op)?;
+        let d5 = self.5.to_index(shape, op)?;
+        Ok(vec![d0, d1, d2, d3, d4, d5])
+    }
+}
+
 extract_dims!(dims0, 0, |_: &[usize]| (), ());
 extract_dims!(dims1, 1, |d: &[usize]| d[0], usize);
 extract_dims!(dims2, 2, |d: &[usize]| (d[0], d[1]), (usize, usize));
@@ -367,6 +482,139 @@ extract_dims!(
     (usize, usize, usize, usize, usize)
 );
 
+pub trait ShapeWithOneHole {
+    fn into_shape(self, el_count: usize) -> Result<Shape>;
+}
+
+impl<S: Into<Shape>> ShapeWithOneHole for S {
+    fn into_shape(self, _el_count: usize) -> Result<Shape> {
+        Ok(self.into())
+    }
+}
+
+impl ShapeWithOneHole for ((),) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        Ok(el_count.into())
+    }
+}
+
+fn hole_size(el_count: usize, prod_d: usize, s: &dyn std::fmt::Debug) -> Result<usize> {
+    if prod_d == 0 {
+        crate::bail!("cannot reshape tensor of {el_count} elements to {s:?}")
+    }
+    if el_count % prod_d != 0 {
+        crate::bail!("cannot reshape tensor with {el_count} elements to {s:?}")
+    }
+    Ok(el_count / prod_d)
+}
+
+impl ShapeWithOneHole for ((), usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let ((), d1) = self;
+        Ok((hole_size(el_count, d1, &self)?, d1).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, ()) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, ()) = self;
+        Ok((d1, hole_size(el_count, d1, &self)?).into())
+    }
+}
+
+impl ShapeWithOneHole for ((), usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let ((), d1, d2) = self;
+        Ok((hole_size(el_count, d1 * d2, &self)?, d1, d2).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, (), usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, (), d2) = self;
+        Ok((d1, hole_size(el_count, d1 * d2, &self)?, d2).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, ()) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, ()) = self;
+        Ok((d1, d2, hole_size(el_count, d1 * d2, &self)?).into())
+    }
+}
+
+impl ShapeWithOneHole for ((), usize, usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let ((), d1, d2, d3) = self;
+        let d = hole_size(el_count, d1 * d2 * d3, &self)?;
+        Ok((d, d1, d2, d3).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, (), usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, (), d2, d3) = self;
+        let d = hole_size(el_count, d1 * d2 * d3, &self)?;
+        Ok((d1, d, d2, d3).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, (), usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, (), d3) = self;
+        let d = hole_size(el_count, d1 * d2 * d3, &self)?;
+        Ok((d1, d2, d, d3).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, usize, ()) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, d3, ()) = self;
+        let d = hole_size(el_count, d1 * d2 * d3, &self)?;
+        Ok((d1, d2, d3, d).into())
+    }
+}
+
+impl ShapeWithOneHole for ((), usize, usize, usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let ((), d1, d2, d3, d4) = self;
+        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
+        Ok((d, d1, d2, d3, d4).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, (), usize, usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, (), d2, d3, d4) = self;
+        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
+        Ok((d1, d, d2, d3, d4).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, (), usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, (), d3, d4) = self;
+        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
+        Ok((d1, d2, d, d3, d4).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, usize, (), usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, d3, (), d4) = self;
+        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
+        Ok((d1, d2, d3, d, d4).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, usize, usize, ()) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, d3, d4, ()) = self;
+        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
+        Ok((d1, d2, d3, d4, d).into())
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

candle-core/src/sort.rs

@@ -0,0 +1,239 @@
+use crate::{Result, Tensor};
+use rayon::prelude::*;
+
+#[derive(Debug, Clone, Copy)]
+struct ArgSort {
+    asc: bool,
+    last_dim: usize,
+}
+
+impl ArgSort {
+    fn asort<T: crate::WithDType>(&self, vs: &[T], layout: &crate::Layout) -> Vec<u32> {
+        #[allow(clippy::uninit_vec)]
+        // Safety: indexes are set later in the parallelized section.
+        let mut sort_indexes = unsafe {
+            let el_count = layout.shape().elem_count();
+            let mut v = Vec::with_capacity(el_count);
+            v.set_len(el_count);
+            v
+        };
+        if self.asc {
+            sort_indexes
+                .par_chunks_exact_mut(self.last_dim)
+                .zip(vs.par_chunks_exact(self.last_dim))
+                .for_each(|(indexes, vs)| {
+                    indexes
+                        .iter_mut()
+                        .enumerate()
+                        .for_each(|(i, v)| *v = i as u32);
+                    indexes.sort_by(|&i, &j| {
+                        vs[i as usize]
+                            .partial_cmp(&vs[j as usize])
+                            .unwrap_or(std::cmp::Ordering::Greater)
+                    })
+                });
+        } else {
+            sort_indexes
+                .par_chunks_exact_mut(self.last_dim)
+                .zip(vs.par_chunks_exact(self.last_dim))
+                .for_each(|(indexes, vs)| {
+                    indexes
+                        .iter_mut()
+                        .enumerate()
+                        .for_each(|(i, v)| *v = i as u32);
+                    indexes.sort_by(|&j, &i| {
+                        vs[i as usize]
+                            .partial_cmp(&vs[j as usize])
+                            .unwrap_or(std::cmp::Ordering::Greater)
+                    })
+                });
+        }
+        sort_indexes
+    }
+}
+
+impl crate::CustomOp1 for ArgSort {
+    fn name(&self) -> &'static str {
+        "argsort"
+    }
+
+    fn cpu_fwd(
+        &self,
+        storage: &crate::CpuStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::CpuStorage, crate::Shape)> {
+        let sort_indexes = match storage {
+            crate::CpuStorage::U8(vs) => self.asort(vs, layout),
+            crate::CpuStorage::U32(vs) => self.asort(vs, layout),
+            crate::CpuStorage::I64(vs) => self.asort(vs, layout),
+            crate::CpuStorage::BF16(vs) => self.asort(vs, layout),
+            crate::CpuStorage::F16(vs) => self.asort(vs, layout),
+            crate::CpuStorage::F32(vs) => self.asort(vs, layout),
+            crate::CpuStorage::F64(vs) => self.asort(vs, layout),
+        };
+        let sort_indexes = crate::CpuStorage::U32(sort_indexes);
+        Ok((sort_indexes, layout.shape().into()))
+    }
+
+    #[cfg(feature = "cuda")]
+    fn cuda_fwd(
+        &self,
+        storage: &crate::CudaStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::CudaStorage, crate::Shape)> {
+        use crate::cuda_backend::cudarc::driver::{
+            CudaSlice, DeviceRepr, LaunchAsync, LaunchConfig, ValidAsZeroBits,
+        };
+        use crate::cuda_backend::{kernel_name, kernels, CudaStorageSlice as S, Map1Any, WrapErr};
+        use crate::{CudaDevice, WithDType};
+
+        impl Map1Any for ArgSort {
+            fn f<T: DeviceRepr + WithDType + ValidAsZeroBits, W: Fn(CudaSlice<T>) -> S>(
+                &self,
+                src: &CudaSlice<T>,
+                dev: &CudaDevice,
+                layout: &crate::Layout,
+                _wrap: W,
+            ) -> Result<S> {
+                let slice = match layout.contiguous_offsets() {
+                    None => crate::bail!("input has to be contiguous"),
+                    Some((o1, o2)) => src.slice(o1..o2),
+                };
+                let elem_count = layout.shape().elem_count();
+                let dst = unsafe { dev.alloc::<u32>(elem_count) }.w()?;
+                let func = if self.asc {
+                    dev.get_or_load_func(&kernel_name::<T>("asort_asc"), kernels::SORT)?
+                } else {
+                    dev.get_or_load_func(&kernel_name::<T>("asort_desc"), kernels::SORT)?
+                };
+                let ncols = self.last_dim;
+                let nrows = elem_count / ncols;
+                let ncols_pad = next_power_of_2(ncols);
+                let params = (&slice, &dst, ncols as i32, ncols_pad as i32);
+                let cfg = LaunchConfig {
+                    grid_dim: (1, nrows as u32, 1),
+                    block_dim: (ncols_pad as u32, 1, 1),
+                    shared_mem_bytes: (ncols_pad * std::mem::size_of::<u32>()) as u32,
+                };
+                unsafe { func.launch(cfg, params) }.w()?;
+                Ok(S::U32(dst))
+            }
+        }
+
+        use crate::backend::BackendStorage;
+        let dev = storage.device();
+        let slice = self.map(&storage.slice, dev, layout)?;
+        let dst = crate::cuda_backend::CudaStorage {
+            slice,
+            device: dev.clone(),
+        };
+        Ok((dst, layout.shape().clone()))
+    }
+
+    #[cfg(feature = "metal")]
+    fn metal_fwd(
+        &self,
+        storage: &crate::MetalStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::MetalStorage, crate::Shape)> {
+        use crate::backend::BackendStorage;
+        use crate::DType;
+
+        let name = {
+            if self.asc {
+                match storage.dtype() {
+                    DType::BF16 => "asort_asc_bf16",
+                    DType::F16 => "asort_asc_f16",
+                    DType::F32 => "asort_asc_f32",
+                    DType::F64 => "asort_asc_f64",
+                    DType::U8 => "asort_asc_u8",
+                    DType::U32 => "asort_asc_u32",
+                    DType::I64 => "asort_asc_i64",
+                }
+            } else {
+                match storage.dtype() {
+                    DType::BF16 => "asort_desc_bf16",
+                    DType::F16 => "asort_desc_f16",
+                    DType::F32 => "asort_desc_f32",
+                    DType::F64 => "asort_desc_f64",
+                    DType::U8 => "asort_desc_u8",
+                    DType::U32 => "asort_desc_u32",
+                    DType::I64 => "asort_desc_i64",
+                }
+            }
+        };
+        let device = storage.device();
+        let kernels = device.kernels();
+        let command_buffer = device.command_buffer()?;
+        let el = layout.shape().elem_count();
+        let ncols = self.last_dim;
+        let nrows = el / ncols;
+        let src = crate::metal_backend::buffer_o(storage.buffer(), layout, storage.dtype());
+        let dst = device.new_buffer(el, DType::U32, "asort")?;
+        let mut ncols_pad = 1;
+        while ncols_pad < ncols {
+            ncols_pad *= 2;
+        }
+        candle_metal_kernels::call_arg_sort(
+            device.metal_device(),
+            &command_buffer,
+            kernels,
+            name,
+            nrows,
+            ncols,
+            ncols_pad,
+            src,
+            &dst,
+        )
+        .map_err(crate::Error::wrap)?;
+        let dst = crate::MetalStorage::new(dst, device.clone(), el, DType::U32);
+        Ok((dst, layout.shape().clone()))
+    }
+}
+
+#[allow(unused)]
+fn next_power_of_2(x: usize) -> usize {
+    let mut n = 1;
+    while n < x {
+        n *= 2
+    }
+    n
+}
+
+impl Tensor {
+    /// Returns the indices that sort the tensor along the last dimension.
+    ///
+    /// If `asc` is `true`, sorting is in ascending order. Otherwise sorting is performed in
+    /// descending order. The sort is unstable so there is no guarantees on the final order when it
+    /// comes to ties.
+    pub fn arg_sort_last_dim(&self, asc: bool) -> Result<Tensor> {
+        if !self.is_contiguous() {
+            return Err(crate::Error::RequiresContiguous {
+                op: "arg_sort_last_dim",
+            });
+        }
+        let last_dim = match self.dims().last() {
+            None => crate::bail!("empty last-dim in arg-sort"),
+            Some(last_dim) => *last_dim,
+        };
+        // No need for a backward pass for arg sort.
+        self.apply_op1_no_bwd(&ArgSort { asc, last_dim })
+    }
+
+    /// Sorts the tensor along the last dimension, returns the sorted tensor together with the
+    /// sorted indexes.
+    ///
+    /// If `asc` is `true`, sorting is in ascending order. Otherwise sorting is performed in
+    /// descending order. The sort is unstable so there is no guarantees on the final order when it
+    /// comes to ties.
+    pub fn sort_last_dim(&self, asc: bool) -> Result<(Tensor, Tensor)> {
+        if !self.is_contiguous() {
+            return Err(crate::Error::RequiresContiguous {
+                op: "sort_last_dim",
+            });
+        }
+        let asort = self.arg_sort_last_dim(asc)?;
+        let sorted = self.gather(&asort, crate::D::Minus1)?;
+        Ok((sorted, asort))
+    }
+}

candle-core/src/storage.rs

@@ -1,6 +1,7 @@
 use crate::backend::BackendStorage;
-use crate::op::{self, CmpOp, CustomOp1, CustomOp2, CustomOp3, ReduceOp};
-use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, Result, Shape};
+use crate::op::{self, CmpOp, ReduceOp};
+use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, MetalStorage, Result, Shape};
+use crate::{CustomOp1, CustomOp2, CustomOp3, InplaceOp1, InplaceOp2, InplaceOp3};
 
 // We do not want to implement Clone on Storage as cloning may fail because of
 // out of memory. Instead try_clone should be used.
@@ -8,6 +9,7 @@ use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, Result, Shape
 pub enum Storage {
     Cpu(CpuStorage),
     Cuda(CudaStorage),
+    Metal(MetalStorage),
 }
 
 impl Storage {
@@ -18,6 +20,10 @@ impl Storage {
                 let storage = storage.try_clone(layout)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.try_clone(layout)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -25,6 +31,7 @@ impl Storage {
         match self {
             Self::Cpu(_) => Device::Cpu,
             Self::Cuda(storage) => Device::Cuda(storage.device().clone()),
+            Self::Metal(storage) => Device::Metal(storage.device().clone()),
         }
     }
 
@@ -32,13 +39,24 @@ impl Storage {
         match self {
             Self::Cpu(storage) => storage.dtype(),
             Self::Cuda(storage) => storage.dtype(),
+            Self::Metal(storage) => storage.dtype(),
         }
     }
 
     pub(crate) fn same_device(&self, rhs: &Self, op: &'static str) -> Result<()> {
-        let lhs = self.device().location();
-        let rhs = rhs.device().location();
-        if lhs != rhs {
+        let lhs_device = self.device();
+        let rhs_device = rhs.device();
+        let lhs = lhs_device.location();
+        let rhs = rhs_device.location();
+        let same_device = if self.device().is_metal() {
+            // On metal, we require the device to be exactly the same rather than
+            // having the same location. In cuda this is not necessary as all CudaDevice on the
+            // same GPU will use the same cuda stream.
+            lhs_device.same_device(&rhs_device)
+        } else {
+            lhs == rhs
+        };
+        if !same_device {
             Err(Error::DeviceMismatchBinaryOp { lhs, rhs, op }.bt())
         } else {
             Ok(())
@@ -65,6 +83,27 @@ impl Storage {
                 let storage = storage.affine(layout, mul, add)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.affine(layout, mul, add)?;
+                Ok(Self::Metal(storage))
+            }
+        }
+    }
+
+    pub(crate) fn powf(&self, layout: &Layout, alpha: f64) -> Result<Self> {
+        match self {
+            Storage::Cpu(storage) => {
+                let storage = storage.powf(layout, alpha)?;
+                Ok(Self::Cpu(storage))
+            }
+            Self::Cuda(storage) => {
+                let storage = storage.powf(layout, alpha)?;
+                Ok(Self::Cuda(storage))
+            }
+            Self::Metal(storage) => {
+                let storage = storage.powf(layout, alpha)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -78,6 +117,10 @@ impl Storage {
                 let storage = storage.elu(layout, alpha)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.elu(layout, alpha)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -99,6 +142,10 @@ impl Storage {
                 let storage = lhs.cmp(op, rhs, lhs_layout, rhs_layout)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(lhs), Self::Metal(rhs)) => {
+                let storage = lhs.cmp(op, rhs, lhs_layout, rhs_layout)?;
+                Ok(Self::Metal(storage))
+            }
             (lhs, rhs) => {
                 // Should not happen because of the same device check above but we're defensive
                 // anyway.
@@ -122,6 +169,10 @@ impl Storage {
                 let storage = storage.reduce_op(op, layout, s)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.reduce_op(op, layout, s)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -135,10 +186,14 @@ impl Storage {
                 let storage = storage.to_dtype(layout, dtype)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.to_dtype(layout, dtype)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
-    pub(crate) fn custom_op1(&self, l: &Layout, c: &dyn CustomOp1) -> Result<(Self, Shape)> {
+    pub(crate) fn apply_op1(&self, l: &Layout, c: &dyn CustomOp1) -> Result<(Self, Shape)> {
         match self {
             Self::Cpu(storage) => {
                 let (storage, shape) = c.cpu_fwd(storage, l)?;
@@ -148,10 +203,14 @@ impl Storage {
                 let (storage, shape) = c.cuda_fwd(storage, l)?;
                 Ok((Self::Cuda(storage), shape))
             }
+            Self::Metal(storage) => {
+                let (storage, shape) = c.metal_fwd(storage, l)?;
+                Ok((Self::Metal(storage), shape))
+            }
         }
     }
 
-    pub(crate) fn custom_op2(
+    pub(crate) fn apply_op2(
         &self,
         l1: &Layout,
         t2: &Self,
@@ -168,11 +227,15 @@ impl Storage {
                 let (s, shape) = c.cuda_fwd(s1, l1, s2, l2)?;
                 Ok((Self::Cuda(s), shape))
             }
+            (Self::Metal(s1), Self::Metal(s2)) => {
+                let (s, shape) = c.metal_fwd(s1, l1, s2, l2)?;
+                Ok((Self::Metal(s), shape))
+            }
             _ => unreachable!(),
         }
     }
 
-    pub(crate) fn custom_op3(
+    pub(crate) fn apply_op3(
         &self,
         l1: &Layout,
         t2: &Self,
@@ -192,6 +255,55 @@ impl Storage {
                 let (s, shape) = c.cuda_fwd(s1, l1, s2, l2, s3, l3)?;
                 Ok((Self::Cuda(s), shape))
             }
+            (Self::Metal(s1), Self::Metal(s2), Self::Metal(s3)) => {
+                let (s, shape) = c.metal_fwd(s1, l1, s2, l2, s3, l3)?;
+                Ok((Self::Metal(s), shape))
+            }
+            _ => unreachable!(),
+        }
+    }
+
+    pub(crate) fn inplace_op1(&mut self, l: &Layout, c: &dyn InplaceOp1) -> Result<()> {
+        match self {
+            Self::Cpu(storage) => c.cpu_fwd(storage, l),
+            Self::Cuda(storage) => c.cuda_fwd(storage, l),
+            Self::Metal(storage) => c.metal_fwd(storage, l),
+        }
+    }
+
+    pub(crate) fn inplace_op2(
+        &mut self,
+        l1: &Layout,
+        t2: &Self,
+        l2: &Layout,
+        c: &dyn InplaceOp2,
+    ) -> Result<()> {
+        self.same_device(t2, c.name())?;
+        match (self, t2) {
+            (Self::Cpu(s1), Self::Cpu(s2)) => c.cpu_fwd(s1, l1, s2, l2),
+            (Self::Cuda(s1), Self::Cuda(s2)) => c.cuda_fwd(s1, l1, s2, l2),
+            (Self::Metal(s1), Self::Metal(s2)) => c.metal_fwd(s1, l1, s2, l2),
+            _ => unreachable!(),
+        }
+    }
+
+    pub(crate) fn inplace_op3(
+        &mut self,
+        l1: &Layout,
+        t2: &Self,
+        l2: &Layout,
+        t3: &Self,
+        l3: &Layout,
+        c: &dyn InplaceOp3,
+    ) -> Result<()> {
+        self.same_device(t2, c.name())?;
+        self.same_device(t3, c.name())?;
+        match (self, t2, t3) {
+            (Self::Cpu(s1), Self::Cpu(s2), Self::Cpu(s3)) => c.cpu_fwd(s1, l1, s2, l2, s3, l3),
+            (Self::Cuda(s1), Self::Cuda(s2), Self::Cuda(s3)) => c.cuda_fwd(s1, l1, s2, l2, s3, l3),
+            (Self::Metal(s1), Self::Metal(s2), Self::Metal(s3)) => {
+                c.metal_fwd(s1, l1, s2, l2, s3, l3)
+            }
             _ => unreachable!(),
         }
     }
@@ -206,6 +318,10 @@ impl Storage {
                 let storage = storage.unary_impl::<B>(layout)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.unary_impl::<B>(layout)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -226,6 +342,10 @@ impl Storage {
                 let storage = lhs.binary_impl::<B>(rhs, lhs_layout, rhs_layout)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(lhs), Self::Metal(rhs)) => {
+                let storage = lhs.binary_impl::<B>(rhs, lhs_layout, rhs_layout)?;
+                Ok(Self::Metal(storage))
+            }
             (lhs, rhs) => {
                 // Should not happen because of the same device check above but we're defensive
                 // anyway.
@@ -257,6 +377,10 @@ impl Storage {
                 let s = inp.conv1d(l, kernel, kernel_l, params)?;
                 Ok(Self::Cuda(s))
             }
+            (Storage::Metal(inp), Storage::Metal(kernel)) => {
+                let s = inp.conv1d(l, kernel, kernel_l, params)?;
+                Ok(Self::Metal(s))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -266,6 +390,37 @@ impl Storage {
         }
     }
 
+    pub(crate) fn conv_transpose1d(
+        &self,
+        l: &Layout,
+        kernel: &Self,
+        kernel_l: &Layout,
+        params: &crate::conv::ParamsConvTranspose1D,
+    ) -> Result<Self> {
+        self.same_device(kernel, "conv-transpose1d")?;
+        self.same_dtype(kernel, "conv-transpose1d")?;
+        match (self, &kernel) {
+            (Storage::Cpu(inp), Storage::Cpu(kernel)) => {
+                let s = inp.conv_transpose1d(l, kernel, kernel_l, params)?;
+                Ok(Self::Cpu(s))
+            }
+            (Storage::Cuda(inp), Storage::Cuda(kernel)) => {
+                let s = inp.conv_transpose1d(l, kernel, kernel_l, params)?;
+                Ok(Self::Cuda(s))
+            }
+            (Storage::Metal(inp), Storage::Metal(kernel)) => {
+                let s = inp.conv_transpose1d(l, kernel, kernel_l, params)?;
+                Ok(Self::Metal(s))
+            }
+            (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
+                lhs: lhs.device().location(),
+                rhs: rhs.device().location(),
+                op: "conv-transpose1d",
+            }
+            .bt()),
+        }
+    }
+
     pub(crate) fn conv2d(
         &self,
         l: &Layout,
@@ -284,6 +439,10 @@ impl Storage {
                 let s = inp.conv2d(l, kernel, kernel_l, params)?;
                 Ok(Self::Cuda(s))
             }
+            (Storage::Metal(inp), Storage::Metal(kernel)) => {
+                let s = inp.conv2d(l, kernel, kernel_l, params)?;
+                Ok(Self::Metal(s))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -293,6 +452,37 @@ impl Storage {
         }
     }
 
+    pub(crate) fn conv_transpose2d(
+        &self,
+        l: &Layout,
+        kernel: &Self,
+        kernel_l: &Layout,
+        params: &crate::conv::ParamsConvTranspose2D,
+    ) -> Result<Self> {
+        self.same_device(kernel, "conv_transpose2d")?;
+        self.same_dtype(kernel, "conv_transpose2d")?;
+        match (self, &kernel) {
+            (Storage::Cpu(inp), Storage::Cpu(kernel)) => {
+                let s = inp.conv_transpose2d(l, kernel, kernel_l, params)?;
+                Ok(Self::Cpu(s))
+            }
+            (Storage::Cuda(inp), Storage::Cuda(kernel)) => {
+                let s = inp.conv_transpose2d(l, kernel, kernel_l, params)?;
+                Ok(Self::Cuda(s))
+            }
+            (Storage::Metal(inp), Storage::Metal(kernel)) => {
+                let s = inp.conv_transpose2d(l, kernel, kernel_l, params)?;
+                Ok(Self::Metal(s))
+            }
+            (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
+                lhs: lhs.device().location(),
+                rhs: rhs.device().location(),
+                op: "conv_transpose2d",
+            }
+            .bt()),
+        }
+    }
+
     pub(crate) fn avg_pool2d(
         &self,
         layout: &Layout,
@@ -308,6 +498,10 @@ impl Storage {
                 let storage = storage.avg_pool2d(layout, kernel_size, stride)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.avg_pool2d(layout, kernel_size, stride)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -326,6 +520,27 @@ impl Storage {
                 let storage = storage.max_pool2d(layout, kernel_size, stride)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.max_pool2d(layout, kernel_size, stride)?;
+                Ok(Self::Metal(storage))
+            }
+        }
+    }
+
+    pub(crate) fn upsample_nearest1d(&self, layout: &Layout, sz: usize) -> Result<Self> {
+        match self {
+            Storage::Cpu(storage) => {
+                let storage = storage.upsample_nearest1d(layout, sz)?;
+                Ok(Self::Cpu(storage))
+            }
+            Self::Cuda(storage) => {
+                let storage = storage.upsample_nearest1d(layout, sz)?;
+                Ok(Self::Cuda(storage))
+            }
+            Self::Metal(storage) => {
+                let storage = storage.upsample_nearest1d(layout, sz)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -339,6 +554,10 @@ impl Storage {
                 let storage = storage.upsample_nearest2d(layout, h, w)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.upsample_nearest2d(layout, h, w)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -362,6 +581,10 @@ impl Storage {
                 let storage = cond.where_cond(layout, t, layout_t, f, layout_f)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(cond), Self::Metal(t), Self::Metal(f)) => {
+                let storage = cond.where_cond(layout, t, layout_t, f, layout_f)?;
+                Ok(Self::Metal(storage))
+            }
             (_, lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -388,6 +611,10 @@ impl Storage {
                 let storage = s.gather(l, indexes, indexes_l, d)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(s), Self::Metal(indexes)) => {
+                let storage = s.gather(l, indexes, indexes_l, d)?;
+                Ok(Self::Metal(storage))
+            }
             _ => unreachable!(),
         }
     }
@@ -412,6 +639,10 @@ impl Storage {
                 let storage = s.scatter_add(l, indexes, indexes_l, source, source_l, d)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(s), Self::Metal(indexes), Self::Metal(source)) => {
+                let storage = s.scatter_add(l, indexes, indexes_l, source, source_l, d)?;
+                Ok(Self::Metal(storage))
+            }
             _ => unreachable!(),
         }
     }
@@ -436,6 +667,10 @@ impl Storage {
                 let storage = s.index_add(l, indexes, indexes_l, source, source_l, d)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(s), Self::Metal(indexes), Self::Metal(source)) => {
+                let storage = s.index_add(l, indexes, indexes_l, source, source_l, d)?;
+                Ok(Self::Metal(storage))
+            }
             _ => unreachable!(),
         }
     }
@@ -457,6 +692,10 @@ impl Storage {
                 let storage = lhs.index_select(rhs, lhs_l, rhs_l, d)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(lhs), Self::Metal(rhs)) => {
+                let storage = lhs.index_select(rhs, lhs_l, rhs_l, d)?;
+                Ok(Self::Metal(storage))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -484,6 +723,10 @@ impl Storage {
                 let storage = lhs.matmul(rhs, bmnk, lhs_layout, rhs_layout)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(lhs), Self::Metal(rhs)) => {
+                let storage = lhs.matmul(rhs, bmnk, lhs_layout, rhs_layout)?;
+                Ok(Self::Metal(storage))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -503,6 +746,9 @@ impl Storage {
         match (self, dst) {
             (Self::Cpu(src), Self::Cpu(dst)) => src.copy_strided_src(dst, dst_offset, src_l),
             (Self::Cuda(src), Self::Cuda(dst)) => Ok(src.copy_strided_src(dst, dst_offset, src_l)?),
+            (Self::Metal(src), Self::Metal(dst)) => {
+                Ok(src.copy_strided_src(dst, dst_offset, src_l)?)
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -511,4 +757,32 @@ impl Storage {
             .bt()),
         }
     }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(crate) fn copy2d(
+        &self,
+        dst: &mut Self,
+        d1: usize,
+        d2: usize,
+        src_s: usize,
+        dst_s: usize,
+        src_o: usize,
+        dst_o: usize,
+    ) -> Result<()> {
+        match (self, dst) {
+            (Self::Cpu(src), Self::Cpu(dst)) => src.copy2d(dst, d1, d2, src_s, dst_s, src_o, dst_o),
+            (Self::Cuda(src), Self::Cuda(dst)) => {
+                Ok(src.copy2d(dst, d1, d2, src_s, dst_s, src_o, dst_o)?)
+            }
+            (Self::Metal(src), Self::Metal(dst)) => {
+                Ok(src.copy2d(dst, d1, d2, src_s, dst_s, src_o, dst_o)?)
+            }
+            (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
+                lhs: lhs.device().location(),
+                rhs: rhs.device().location(),
+                op: "copy2d",
+            }
+            .bt()),
+        }
+    }
 }

candle-core/src/streaming.rs

@@ -0,0 +1,206 @@
+use crate::{Result, Shape, Tensor};
+
+pub trait Dim: crate::shape::Dim + Copy {}
+impl<T: crate::shape::Dim + Copy> Dim for T {}
+
+/// A stream tensor is used in streaming module. It can either contain an actual tensor or be
+/// empty.
+#[derive(Clone)]
+pub struct StreamTensor(Option<Tensor>);
+
+impl std::fmt::Debug for StreamTensor {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match &self.0 {
+            Some(t) => write!(f, "{:?}", t.shape()),
+            None => write!(f, "Empty"),
+        }
+    }
+}
+
+impl std::convert::From<Option<Tensor>> for StreamTensor {
+    fn from(value: Option<Tensor>) -> Self {
+        Self(value)
+    }
+}
+
+impl std::convert::From<Tensor> for StreamTensor {
+    fn from(value: Tensor) -> Self {
+        Self(Some(value))
+    }
+}
+
+impl std::convert::From<()> for StreamTensor {
+    fn from(_value: ()) -> Self {
+        Self(None)
+    }
+}
+
+impl StreamTensor {
+    pub fn empty() -> Self {
+        Self(None)
+    }
+
+    pub fn from_tensor(tensor: Tensor) -> Self {
+        Self(Some(tensor))
+    }
+
+    pub fn shape(&self) -> Option<&Shape> {
+        self.0.as_ref().map(|t| t.shape())
+    }
+
+    pub fn cat2<D: Dim>(&self, rhs: &Self, dim: D) -> Result<Self> {
+        let xs = match (&self.0, &rhs.0) {
+            (Some(lhs), Some(rhs)) => {
+                let xs = Tensor::cat(&[lhs, rhs], dim)?;
+                Some(xs)
+            }
+            (Some(xs), None) | (None, Some(xs)) => Some(xs.clone()),
+            (None, None) => None,
+        };
+        Ok(Self(xs))
+    }
+
+    pub fn seq_len<D: Dim>(&self, dim: D) -> Result<usize> {
+        match &self.0 {
+            None => Ok(0),
+            Some(v) => v.dim(dim),
+        }
+    }
+
+    pub fn reset(&mut self) {
+        self.0 = None
+    }
+
+    pub fn narrow<D: Dim>(&self, dim: D, offset: usize, len: usize) -> Result<StreamTensor> {
+        let t = match &self.0 {
+            None => None,
+            Some(t) => {
+                let seq_len = t.dim(dim)?;
+                if seq_len <= offset {
+                    None
+                } else {
+                    let t = t.narrow(dim, offset, usize::min(len, seq_len - offset))?;
+                    Some(t)
+                }
+            }
+        };
+        Ok(Self(t))
+    }
+
+    /// Splits the Streaming Tensor on the time axis `dim` with the first `lhs_len` elements
+    /// returned in the first output and the remaining in the second output.
+    pub fn split<D: Dim>(&self, dim: D, lhs_len: usize) -> Result<(Self, Self)> {
+        match &self.0 {
+            None => Ok((Self::empty(), Self::empty())),
+            Some(t) => {
+                let seq_len = t.dim(dim)?;
+                let lhs_len = usize::min(seq_len, lhs_len);
+                if lhs_len == 0 {
+                    Ok((Self::empty(), t.clone().into()))
+                } else {
+                    let lhs = Self::from_tensor(t.narrow(dim, 0, lhs_len)?);
+                    let rhs_len = seq_len - lhs_len;
+                    let rhs = if rhs_len == 0 {
+                        Self::empty()
+                    } else {
+                        Self::from_tensor(t.narrow(dim, lhs_len, rhs_len)?)
+                    };
+                    Ok((lhs, rhs))
+                }
+            }
+        }
+    }
+
+    pub fn as_option(&self) -> Option<&Tensor> {
+        self.0.as_ref()
+    }
+
+    pub fn apply<M: crate::Module>(&self, m: &M) -> Result<Self> {
+        match &self.0 {
+            None => Ok(Self::empty()),
+            Some(t) => Ok(Self::from_tensor(t.apply(m)?)),
+        }
+    }
+}
+
+/// Streaming modules take as input a stream tensor and return a stream tensor. They may perform
+/// some internal buffering so that enough data has been received for the module to be able to
+/// perform some operations.
+pub trait StreamingModule {
+    // TODO: Should we also have a flush method?
+    fn step(&mut self, xs: &StreamTensor) -> Result<StreamTensor>;
+    fn reset_state(&mut self);
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum BinOp {
+    Add,
+    Mul,
+    Sub,
+    Div,
+}
+
+#[derive(Debug, Clone)]
+pub struct StreamingBinOp {
+    prev_lhs: StreamTensor,
+    prev_rhs: StreamTensor,
+    pub op: BinOp,
+    pub dim: crate::D,
+}
+
+impl StreamingBinOp {
+    pub fn new(op: BinOp, dim: crate::D) -> Self {
+        Self {
+            prev_lhs: StreamTensor::empty(),
+            prev_rhs: StreamTensor::empty(),
+            op,
+            dim,
+        }
+    }
+
+    pub fn reset_state(&mut self) {
+        self.prev_lhs.reset();
+        self.prev_rhs.reset();
+    }
+
+    pub fn forward(&self, lhs: &Tensor, rhs: &Tensor) -> Result<Tensor> {
+        match self.op {
+            BinOp::Add => Tensor::add(lhs, rhs),
+            BinOp::Mul => Tensor::mul(lhs, rhs),
+            BinOp::Sub => Tensor::sub(lhs, rhs),
+            BinOp::Div => Tensor::div(lhs, rhs),
+        }
+    }
+
+    pub fn step(&mut self, lhs: &StreamTensor, rhs: &StreamTensor) -> Result<StreamTensor> {
+        let lhs = StreamTensor::cat2(&self.prev_lhs, lhs, self.dim)?;
+        let rhs = StreamTensor::cat2(&self.prev_rhs, rhs, self.dim)?;
+        let lhs_len = lhs.seq_len(self.dim)?;
+        let rhs_len = rhs.seq_len(self.dim)?;
+        let common_len = usize::min(lhs_len, rhs_len);
+        let (lhs, prev_lhs) = lhs.split(self.dim, common_len)?;
+        let (rhs, prev_rhs) = rhs.split(self.dim, common_len)?;
+        let ys = match (lhs.0, rhs.0) {
+            (Some(lhs), Some(rhs)) => {
+                let ys = self.forward(&lhs, &rhs)?;
+                StreamTensor::from_tensor(ys)
+            }
+            (None, None) => StreamTensor::empty(),
+            (lhs, rhs) => crate::bail!("INTERNAL ERROR inconsistent lhs and rhs {lhs:?} {rhs:?}"),
+        };
+        self.prev_lhs = prev_lhs;
+        self.prev_rhs = prev_rhs;
+        Ok(ys)
+    }
+}
+
+/// Simple wrapper that doesn't do any buffering.
+pub struct Map<T: crate::Module>(T);
+
+impl<T: crate::Module> StreamingModule for Map<T> {
+    fn reset_state(&mut self) {}
+
+    fn step(&mut self, xs: &StreamTensor) -> Result<StreamTensor> {
+        xs.apply(&self.0)
+    }
+}

candle-core/src/tensor_cat.rs

@@ -0,0 +1,300 @@
+use crate::{shape::Dim, Error, Result, Shape, Tensor};
+
+impl Tensor {
+    /// Concatenates two or more tensors along a particular dimension.
+    ///
+    /// All tensors must of the same rank, and the output will have
+    /// the same rank
+    ///
+    /// ```rust
+    /// # use candle_core::{Tensor, DType, Device};
+    /// let a = Tensor::zeros((2, 3), DType::F32, &Device::Cpu)?;
+    /// let b = Tensor::zeros((2, 3), DType::F32, &Device::Cpu)?;
+    ///
+    /// let c = Tensor::cat(&[&a, &b], 0)?;
+    /// assert_eq!(c.shape().dims(), &[4, 3]);
+    ///
+    /// let c = Tensor::cat(&[&a, &b], 1)?;
+    /// assert_eq!(c.shape().dims(), &[2, 6]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn cat<A: AsRef<Tensor>, D: Dim>(args: &[A], dim: D) -> Result<Self> {
+        if args.is_empty() {
+            Err(Error::OpRequiresAtLeastOneTensor { op: "cat" }.bt())?
+        }
+        let arg0 = args[0].as_ref();
+        if args.len() == 1 {
+            return Ok(arg0.clone());
+        }
+        let dim = dim.to_index(arg0.shape(), "cat")?;
+        for arg in args {
+            arg.as_ref().check_dim(dim, "cat")?;
+        }
+        for (arg_idx, arg) in args.iter().enumerate() {
+            let arg = arg.as_ref();
+            if arg0.rank() != arg.rank() {
+                Err(Error::UnexpectedNumberOfDims {
+                    expected: arg0.rank(),
+                    got: arg.rank(),
+                    shape: arg.shape().clone(),
+                }
+                .bt())?
+            }
+            for (dim_idx, (v1, v2)) in arg0
+                .shape()
+                .dims()
+                .iter()
+                .zip(arg.shape().dims().iter())
+                .enumerate()
+            {
+                if dim_idx != dim && v1 != v2 {
+                    Err(Error::ShapeMismatchCat {
+                        dim: dim_idx,
+                        first_shape: arg0.shape().clone(),
+                        n: arg_idx + 1,
+                        nth_shape: arg.shape().clone(),
+                    }
+                    .bt())?
+                }
+            }
+        }
+        let all_contiguous = args.iter().all(|v| v.as_ref().is_contiguous());
+        if all_contiguous {
+            Self::cat_contiguous(args, dim)
+        } else if dim == 0 {
+            Self::cat0(args)
+        } else {
+            let args: Vec<Tensor> = args
+                .iter()
+                .map(|a| a.as_ref().transpose(0, dim))
+                .collect::<Result<Vec<_>>>()?;
+            let cat = Self::cat0(&args)?;
+            cat.transpose(0, dim)
+        }
+    }
+
+    fn cat0<A: AsRef<Tensor>>(args: &[A]) -> Result<Self> {
+        if args.is_empty() {
+            Err(Error::OpRequiresAtLeastOneTensor { op: "cat" }.bt())?
+        }
+        let arg0 = args[0].as_ref();
+        if args.len() == 1 {
+            return Ok(arg0.clone());
+        }
+        let rank = arg0.rank();
+        let device = arg0.device();
+        let dtype = arg0.dtype();
+        let first_dims = arg0.shape().dims();
+        let mut cat_dims = first_dims.to_vec();
+        cat_dims[0] = 0;
+        let mut offsets = vec![0usize];
+        for (arg_idx, arg) in args.iter().enumerate() {
+            let arg = arg.as_ref();
+            if arg.dtype() != dtype {
+                Err(Error::DTypeMismatchBinaryOp {
+                    lhs: dtype,
+                    rhs: arg.dtype(),
+                    op: "cat",
+                }
+                .bt())?
+            }
+            if arg.device().location() != device.location() {
+                Err(Error::DeviceMismatchBinaryOp {
+                    lhs: device.location(),
+                    rhs: arg.device().location(),
+                    op: "cat",
+                }
+                .bt())?
+            }
+            if rank != arg.rank() {
+                Err(Error::UnexpectedNumberOfDims {
+                    expected: rank,
+                    got: arg.rank(),
+                    shape: arg.shape().clone(),
+                }
+                .bt())?
+            }
+            for (dim_idx, (v1, v2)) in arg0
+                .shape()
+                .dims()
+                .iter()
+                .zip(arg.shape().dims().iter())
+                .enumerate()
+            {
+                if dim_idx == 0 {
+                    cat_dims[0] += v2;
+                }
+                if dim_idx != 0 && v1 != v2 {
+                    Err(Error::ShapeMismatchCat {
+                        dim: dim_idx,
+                        first_shape: arg0.shape().clone(),
+                        n: arg_idx + 1,
+                        nth_shape: arg.shape().clone(),
+                    }
+                    .bt())?
+                }
+            }
+            let next_offset = offsets.last().unwrap() + arg.elem_count();
+            offsets.push(next_offset);
+        }
+        let shape = Shape::from(cat_dims);
+        let op = crate::op::BackpropOp::new(args, |args| crate::op::Op::Cat(args, 0));
+        let mut storage = unsafe { device.alloc_uninit(&shape, dtype)? };
+        for (arg, &offset) in args.iter().zip(offsets.iter()) {
+            let arg = arg.as_ref();
+            arg.storage()
+                .copy_strided_src(&mut storage, offset, arg.layout())?;
+        }
+        Ok(crate::tensor::from_storage(storage, shape, op, false))
+    }
+
+    fn cat_contiguous<A: AsRef<Tensor>>(args: &[A], dim: usize) -> Result<Self> {
+        if args.is_empty() {
+            Err(Error::OpRequiresAtLeastOneTensor { op: "cat" }.bt())?
+        }
+        let arg0 = args[0].as_ref();
+        if args.len() == 1 {
+            return Ok(arg0.clone());
+        }
+        let rank = arg0.rank();
+        let device = arg0.device();
+        let dtype = arg0.dtype();
+        let first_dims = arg0.shape().dims();
+        let mut cat_dims = first_dims.to_vec();
+        cat_dims[dim] = 0;
+        for (arg_idx, arg) in args.iter().enumerate() {
+            let arg = arg.as_ref();
+            if arg.dtype() != dtype {
+                Err(Error::DTypeMismatchBinaryOp {
+                    lhs: dtype,
+                    rhs: arg.dtype(),
+                    op: "cat",
+                }
+                .bt())?
+            }
+            if arg.device().location() != device.location() {
+                Err(Error::DeviceMismatchBinaryOp {
+                    lhs: device.location(),
+                    rhs: arg.device().location(),
+                    op: "cat",
+                }
+                .bt())?
+            }
+            if rank != arg.rank() {
+                Err(Error::UnexpectedNumberOfDims {
+                    expected: rank,
+                    got: arg.rank(),
+                    shape: arg.shape().clone(),
+                }
+                .bt())?
+            }
+            for (dim_idx, (v1, v2)) in arg0
+                .shape()
+                .dims()
+                .iter()
+                .zip(arg.shape().dims().iter())
+                .enumerate()
+            {
+                if dim_idx == dim {
+                    cat_dims[dim] += v2;
+                }
+                if dim_idx != dim && v1 != v2 {
+                    Err(Error::ShapeMismatchCat {
+                        dim: dim_idx,
+                        first_shape: arg0.shape().clone(),
+                        n: arg_idx + 1,
+                        nth_shape: arg.shape().clone(),
+                    }
+                    .bt())?
+                }
+            }
+        }
+        let cat_target_dim_len = cat_dims[dim];
+        let block_size: usize = cat_dims.iter().skip(1 + dim).product();
+        let shape = Shape::from(cat_dims);
+        let op = crate::op::BackpropOp::new(args, |args| crate::op::Op::Cat(args, dim));
+        let mut storage = unsafe { device.alloc_uninit(&shape, dtype)? };
+        let mut dst_o = 0;
+        for arg in args.iter() {
+            let arg = arg.as_ref();
+            let arg_dims = arg.shape().dims();
+            let d1: usize = arg_dims.iter().take(dim).product();
+            let d2 = block_size * arg_dims[dim];
+            let dst_s = block_size * cat_target_dim_len;
+            let src_o = arg.layout().start_offset();
+            arg.storage().copy2d(
+                &mut storage,
+                d1,
+                d2,
+                /* src_s */ d2,
+                dst_s,
+                src_o,
+                dst_o,
+            )?;
+            dst_o += d2;
+        }
+        Ok(crate::tensor::from_storage(storage, shape, op, false))
+    }
+
+    /// Set the values on `self` using values from `src`. The copy starts at the specified
+    /// `offset` for the target dimension `dim` on `self`.
+    /// `self` and `src` must have the same shape except on dimension `dim` where the `self` size
+    /// has to be greater than or equal to `offset` plus the `src` size.
+    ///
+    /// Note that this modifies `self` in place and as such is not compatibel with
+    /// back-propagation.  
+    pub fn slice_set<D: Dim>(&self, src: &Self, dim: D, offset: usize) -> Result<()> {
+        let dim = dim.to_index(self.shape(), "slice-set")?;
+        if !self.is_contiguous() || !src.is_contiguous() {
+            Err(Error::RequiresContiguous { op: "slice-set" }.bt())?
+        }
+        if self.dtype() != src.dtype() {
+            Err(Error::DTypeMismatchBinaryOp {
+                lhs: self.dtype(),
+                rhs: src.dtype(),
+                op: "slice-set",
+            }
+            .bt())?
+        }
+        if self.device().location() != src.device().location() {
+            Err(Error::DeviceMismatchBinaryOp {
+                lhs: self.device().location(),
+                rhs: src.device().location(),
+                op: "slice-set",
+            }
+            .bt())?
+        }
+        if self.rank() != src.rank() {
+            Err(Error::UnexpectedNumberOfDims {
+                expected: self.rank(),
+                got: src.rank(),
+                shape: self.shape().clone(),
+            }
+            .bt())?
+        }
+        for (dim_idx, (v1, v2)) in self.dims().iter().zip(src.dims().iter()).enumerate() {
+            if dim_idx == dim && *v2 + offset > *v1 {
+                crate::bail!("shape mismatch on target dim, dst: {v1}, src: {v2} + {offset}")
+            }
+            if dim_idx != dim && v1 != v2 {
+                crate::bail!("shape mismatch on dim {dim_idx}, {v1} <> {v2}")
+            }
+        }
+        let block_size: usize = src.dims().iter().skip(1 + dim).product();
+        let d1: usize = src.dims().iter().take(dim).product();
+        let d2 = block_size * src.dims()[dim];
+        let dst_o = self.layout().start_offset() + offset * block_size;
+        let src_o = src.layout().start_offset();
+        src.storage().copy2d(
+            &mut self.storage_mut(),
+            d1,
+            d2,
+            /* src_s */ d2,
+            /* dst_s */ block_size * self.dims()[dim],
+            src_o,
+            dst_o,
+        )?;
+
+        Ok(())
+    }
+}

candle-core/src/test_utils.rs

@@ -1,15 +1,10 @@
-#![allow(dead_code)]
-
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-
-use candle_core::{Result, Tensor};
+use crate::{Result, Tensor};
 
 #[macro_export]
 macro_rules! test_device {
     // TODO: Switch to generating the two last arguments automatically once concat_idents is
     // stable. https://github.com/rust-lang/rust/issues/29599
-    ($fn_name: ident, $test_cpu: ident, $test_cuda: ident) => {
+    ($fn_name: ident, $test_cpu: ident, $test_cuda: ident, $test_metal: ident) => {
         #[test]
         fn $test_cpu() -> Result<()> {
             $fn_name(&Device::Cpu)
@@ -20,9 +15,21 @@ macro_rules! test_device {
         fn $test_cuda() -> Result<()> {
             $fn_name(&Device::new_cuda(0)?)
         }
+
+        #[cfg(feature = "metal")]
+        #[test]
+        fn $test_metal() -> Result<()> {
+            $fn_name(&Device::new_metal(0)?)
+        }
     };
 }
 
+pub fn to_vec0_round(t: &Tensor, digits: i32) -> Result<f32> {
+    let b = 10f32.powi(digits);
+    let t = t.to_vec0::<f32>()?;
+    Ok(f32::round(t * b) / b)
+}
+
 pub fn to_vec1_round(t: &Tensor, digits: i32) -> Result<Vec<f32>> {
     let b = 10f32.powi(digits);
     let t = t.to_vec1::<f32>()?;
@@ -40,7 +47,7 @@ pub fn to_vec2_round(t: &Tensor, digits: i32) -> Result<Vec<Vec<f32>>> {
     Ok(t)
 }
 
-pub fn to_vec3_round(t: Tensor, digits: i32) -> Result<Vec<Vec<Vec<f32>>>> {
+pub fn to_vec3_round(t: &Tensor, digits: i32) -> Result<Vec<Vec<Vec<f32>>>> {
     let b = 10f32.powi(digits);
     let t = t.to_vec3::<f32>()?;
     let t = t

candle-core/src/utils.rs

@@ -22,3 +22,23 @@ pub fn has_mkl() -> bool {
 pub fn cuda_is_available() -> bool {
     cfg!(feature = "cuda")
 }
+
+pub fn metal_is_available() -> bool {
+    cfg!(feature = "metal")
+}
+
+pub fn with_avx() -> bool {
+    cfg!(target_feature = "avx")
+}
+
+pub fn with_neon() -> bool {
+    cfg!(target_feature = "neon")
+}
+
+pub fn with_simd128() -> bool {
+    cfg!(target_feature = "simd128")
+}
+
+pub fn with_f16c() -> bool {
+    cfg!(target_feature = "f16c")
+}

candle-core/src/variable.rs

@@ -34,9 +34,14 @@ impl Var {
         Ok(Self(inner))
     }
 
+    // Convert a tensor to a variable, if the tensor is already a variable then it is returned as is.
     pub fn from_tensor(t: &Tensor) -> Result<Self> {
-        let inner = t.make_var()?;
-        Ok(Self(inner))
+        if t.is_variable() {
+            Ok(Self(t.clone()))
+        } else {
+            let inner = t.make_var()?;
+            Ok(Self(inner))
+        }
     }
 
     pub fn rand_f64<S: Into<Shape>>(
@@ -107,6 +112,10 @@ impl Var {
         Ok(Self(inner))
     }
 
+    pub fn as_detached_tensor(&self) -> Tensor {
+        self.0.detach()
+    }
+
     pub fn as_tensor(&self) -> &Tensor {
         &self.0
     }

candle-core/tests/conv_tests.rs

@@ -1,6 +1,5 @@
-mod test_utils;
 use anyhow::Result;
-use candle_core::{Device, Tensor};
+use candle_core::{test_device, test_utils, Device, IndexOp, Tensor};
 
 /* This test is based on the following script.
 import torch
@@ -14,6 +13,14 @@ res = torch.nn.functional.conv1d(t, w)
 print(res.flatten())
 res = torch.nn.functional.conv1d(t, w, padding=1)
 print(res.flatten())
+
+w_t = w.transpose(0, 1)
+res = torch.nn.functional.conv_transpose1d(t, w_t)
+print(res.shape)
+print(res)
+res = torch.nn.functional.conv_transpose1d(t, w_t, groups=2)
+print(res.shape)
+print(res)
 */
 fn conv1d(dev: &Device) -> Result<()> {
     let t = Tensor::new(
@@ -33,32 +40,57 @@ fn conv1d(dev: &Device) -> Result<()> {
         dev,
     )?
     .reshape((2, 4, 3))?;
-    let res = t.conv1d(&w, 0, 1)?;
+    let res = t.conv1d(&w, 0, 1, 1, 1)?;
     assert_eq!(res.dims(), [1, 2, 3]);
     assert_eq!(
         test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
         [2.6357, -1.3336, 4.1393, -1.1784, 3.5675, 0.5069]
     );
-    let res = t.conv1d(&w, /*padding*/ 1, 1)?;
+    let res = t.conv1d(&w, /*padding*/ 1, 1, 1, 1)?;
     assert_eq!(res.dims(), [1, 2, 5]);
     // Same as pytorch default padding: use zeros.
     assert_eq!(
         test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
         [2.4509, 2.6357, -1.3336, 4.1393, 0.5657, 1.8091, -1.1784, 3.5675, 0.5069, 3.3352]
     );
+
+    let w = w.transpose(0, 1)?;
+    // The CPU kernels applied in the contiguous and non contiguous cases are different.
+    for w in [w.clone(), w.contiguous()?] {
+        let res = t.conv_transpose1d(&w, 0, 0, 1, 1, 1)?;
+        assert_eq!(res.dims(), [1, 2, 7]);
+        assert_eq!(
+            test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
+            [
+                0.0699, -1.2899, 8.3018, 5.5873, 2.4572, -2.6143, -0.0706, 1.8765, 4.8318, 1.1538,
+                4.7076, -5.9745, -0.8276, 1.621
+            ],
+        );
+        let res = t.conv_transpose1d(&w, 0, 0, 1, 1, 2)?;
+        assert_eq!(res.dims(), [1, 4, 7]);
+        assert_eq!(
+            test_utils::to_vec2_round(&res.squeeze(0)?, 4)?,
+            [
+                [-1.5596, -1.8099, 2.0407, 4.8764, -0.1743, -0.735, -0.7819],
+                [0.7816, 3.8152, -0.5926, 2.2515, -5.1844, -0.3157, 1.4721],
+                [1.6295, 0.52, 6.2611, 0.7109, 2.6315, -1.8793, 0.7113],
+                [1.0949, 1.0166, 1.7464, 2.4561, -0.79, -0.5119, 0.1488]
+            ]
+        );
+    }
     Ok(())
 }
 
 fn conv1d_small(dev: &Device) -> Result<()> {
     let t = Tensor::new(&[0.4056f32, -0.8689, -0.0773, -1.5630], dev)?.reshape((1, 1, 4))?;
     let w = Tensor::new(&[1f32, 0., 0.], dev)?.reshape((1, 1, 3))?;
-    let res = t.conv1d(&w, 0, 1)?;
+    let res = t.conv1d(&w, 0, 1, 1, 1)?;
     assert_eq!(res.dims(), [1, 1, 2]);
     assert_eq!(
         test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
         [0.4056, -0.8689]
     );
-    let res = t.conv1d(&w, /*padding*/ 1, 1)?;
+    let res = t.conv1d(&w, /*padding*/ 1, 1, 1, 1)?;
     assert_eq!(res.dims(), [1, 1, 4]);
     assert_eq!(
         test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
@@ -77,6 +109,19 @@ print(t.flatten())
 print(w.flatten())
 res = torch.nn.functional.conv2d(t, w)
 print(res.flatten())
+
+w_t = w.transpose(0, 1)
+res = torch.nn.functional.conv_transpose2d(t, w_t)
+print(res.shape)
+print(res)
+
+res = torch.nn.functional.conv2d(t, w, dilation=2)
+print(res.shape)
+print(res[0])
+
+res = torch.nn.functional.conv_transpose2d(t, w_t, dilation=2)
+print(res.shape)
+print(res)
 */
 fn conv2d(dev: &Device) -> Result<()> {
     let t = Tensor::new(
@@ -90,7 +135,7 @@ fn conv2d(dev: &Device) -> Result<()> {
             0.6466, -0.5042, -0.0603, -1.6538, -1.2429, 1.8357, 1.6052, -1.3844, 0.3323, -1.3712,
             0.9634, -0.4799, -0.6451, -0.0840, -1.4247, 0.5512, -0.1747, -0.5509, -0.3742, 0.3790,
             -0.4431, -0.4720, -0.7890, 0.2620, 0.7875, 0.5377, -0.6779, -0.8088, 1.9098, 1.2006,
-            -0.8000, -0.4983, 1.5480, 0.8265, -0.1025, 0.5138, 0.5748, 0.3821, -0.4607, 0.0085,
+            -0.8, -0.4983, 1.5480, 0.8265, -0.1025, 0.5138, 0.5748, 0.3821, -0.4607, 0.0085,
         ],
         dev,
     )?;
@@ -109,7 +154,7 @@ fn conv2d(dev: &Device) -> Result<()> {
     )?;
     let t = t.reshape((1, 4, 5, 5))?;
     let w = w.reshape((2, 4, 3, 3))?;
-    let res = t.conv2d(&w, 0, 1)?;
+    let res = t.conv2d(&w, 0, 1, 1, 1)?;
     assert_eq!(res.dims(), [1, 2, 3, 3]);
     assert_eq!(
         test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
@@ -118,6 +163,73 @@ fn conv2d(dev: &Device) -> Result<()> {
             10.389, 3.6023, -4.2808, 0.2672, 5.3646, -5.2023, -2.1955, -9.4075
         ]
     );
+
+    let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 1)?;
+
+    assert_eq!(res.dims(), [1, 2, 7, 7]);
+    assert_eq!(
+        test_utils::to_vec3_round(&res.i(0)?, 4)?,
+        [
+            [
+                [-1.9918, 2.6797, -0.4599, -1.6037, 1.4131, -2.4012, 2.9277],
+                [1.8016, -3.5361, 1.0757, 3.5395, -8.2168, -3.2023, 0.5375],
+                [0.8243, 1.8675, 7.8929, -4.0746, -6.4415, 5.1139, 1.6889],
+                [0.2722, 8.9679, 3.3477, 1.8514, -4.2896, -3.8228, -7.5632],
+                [-8.5412, -5.8142, -7.1587, -1.6095, 0.4651, 0.2748, -2.0985],
+                [2.0833, -0.6482, -12.1692, -4.1284, -2.9765, -0.0656, -4.5114],
+                [5.307, 2.6957, 2.3087, 1.0478, 0.7808, -1.1519, -0.9579]
+            ],
+            [
+                [1.089, 0.1872, -0.6408, -0.9897, 0.8503, 1.1019, -0.9211],
+                [-0.1741, -0.2915, 4.2472, 1.9417, 1.65, 0.6303, -4.7131],
+                [1.6555, 2.4026, -2.9293, 2.9953, 0.5328, 3.5873, -0.9621],
+                [-1.4289, -3.2787, 4.1747, -6.0341, -4.6341, -5.7945, 4.142],
+                [7.5973, 6.4431, 5.9872, 2.1639, -8.6566, 3.3143, -3.4059],
+                [-0.8775, -3.048, 11.6543, 0.6442, 2.3218, -0.4765, 1.1516],
+                [-5.5423, -2.5188, 1.0754, -0.0563, -2.9386, -1.1504, 1.0171]
+            ]
+        ]
+    );
+
+    // Dilations.
+    let res = t.conv2d(&w, 0, 1, 2, 1)?;
+    assert_eq!(res.dims(), [1, 2, 1, 1]);
+    assert_eq!(
+        test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
+        [2.45, -2.3504],
+    );
+
+    // Transpose and dilations.
+    let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 2)?;
+    assert_eq!(res.dims(), [1, 2, 9, 9]);
+    assert_eq!(
+        test_utils::to_vec3_round(&res.i(0)?, 4)?,
+        [
+            [
+                [-1.9918, 3.1652, -0.6778, -4.3442, 4.4351, 0.6652, -3.0124, -0.6031, 2.9277],
+                [2.7036, -1.7156, -0.3969, 1.0516, 1.6381, -2.8886, -0.205, 2.4682, -1.0499],
+                [-0.9459, 3.1631, 3.707, -4.8369, -8.5166, -1.4496, -2.7559, -3.2698, 1.4376],
+                [-0.2157, 3.7786, -2.0252, -4.2633, 3.6731, -1.5142, 5.9391, -0.2622, -0.141],
+                [-6.8121, -3.1744, 1.5945, 3.0637, -9.6088, 1.4446, 2.9489, -3.0082, -7.3822],
+                [0.2371, 3.3303, 0.3861, 2.2646, -4.6784, 4.1235, -0.0109, 0.3176, -0.03],
+                [-2.5339, -2.9564, -3.4518, -4.4594, -9.1873, -1.9709, -0.4676, 0.51, -3.5024],
+                [4.007, 0.3067, -2.2954, 1.1105, -0.1992, 1.6372, -2.9268, 0.2807, -1.2787],
+                [5.307, 1.1317, 1.3518, 0.9049, 3.8116, -0.4075, -0.8874, -0.2241, -0.9579]
+            ],
+            [
+                [1.089, -0.6483, 0.0726, -0.4752, -1.3283, 1.7103, 1.0703, 0.1076, -0.9211],
+                [-0.8629, 0.1376, 0.3202, 2.0955, 0.9696, 2.8988, -1.0012, 1.5049, -0.1278],
+                [1.9286, -1.5255, -2.9563, 2.4589, 3.3611, -0.6951, 0.3525, -1.7724, -5.9861],
+                [1.1226, 2.1561, 3.6417, 4.7546, -0.692, 4.4126, -5.1902, 6.0805, 2.3185],
+                [1.0111, 0.3604, 0.6432, -3.6605, 7.9517, -9.2955, -5.2988, -3.7803, -2.0642],
+                [3.3172, -1.7967, -3.6576, -2.0942, 1.3158, 0.112, -1.7405, 2.9167, 0.7957],
+                [5.1001, 1.8995, -1.8639, 1.1262, 9.9629, 2.683, -3.6319, -1.1607, 0.5856],
+                [-4.8445, -0.5642, 4.2317, 0.0856, 1.2267, -0.5712, 1.736, 1.0997, 0.6908],
+                [-5.5423, -1.1831, -1.2176, 0.0843, 0.0446, -0.7545, -2.4798, -0.0827, 1.0171]
+            ]
+        ]
+    );
+
     Ok(())
 }
 
@@ -131,6 +243,16 @@ print(t.flatten())
 print(w.flatten())
 res = torch.nn.functional.conv2d(t, w)
 print(res.flatten())
+
+w_t = w.transpose(0, 1)
+res = torch.nn.functional.conv_transpose2d(t, w_t)
+print(res.shape)
+print(res.flatten())
+
+t_t = w.transpose(0, 1)
+res = torch.nn.functional.conv_transpose2d(t_t, w)
+print(res.shape)
+print(res.flatten())
 */
 fn conv2d_small(dev: &Device) -> Result<()> {
     let t = Tensor::new(
@@ -143,12 +265,41 @@ fn conv2d_small(dev: &Device) -> Result<()> {
     let w = Tensor::new(&[-0.9259f32, 1.3017], dev)?;
     let t = t.reshape((1, 2, 3, 3))?;
     let w = w.reshape((1, 2, 1, 1))?;
-    let res = t.conv2d(&w, 0, 1)?;
+    let res = t.conv2d(&w, 0, 1, 1, 1)?;
     assert_eq!(res.dims(), [1, 1, 3, 3]);
     assert_eq!(
         test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
         [0.164, -0.0111, -0.1742, 2.6437, -2.0268, 1.1823, 3.2855, -1.0324, 0.2539]
     );
+    let res = t.conv2d(&w, 2, 1, 1, 1)?;
+    assert_eq!(res.dims(), [1, 1, 7, 7]);
+    assert_eq!(
+        test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
+        [
+            0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1640,
+            -0.0111, -0.1742, 0.0, 0.0, 0.0, 0.0, 2.6437, -2.0268, 1.1823, 0.0, 0.0, 0.0, 0.0,
+            3.2855, -1.0324, 0.2539, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
+            0.0, 0.0, 0.0, 0.0
+        ]
+    );
+
+    let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 1)?;
+    assert_eq!(res.dims(), [1, 1, 3, 3]);
+    assert_eq!(
+        test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
+        [0.164, -0.0111, -0.1742, 2.6437, -2.0268, 1.1823, 3.2855, -1.0324, 0.2539],
+    );
+    let res = t.transpose(0, 1)?.conv_transpose2d(&w, 0, 0, 1, 1)?;
+    assert_eq!(res.dims(), [2, 2, 3, 3]);
+    assert_eq!(
+        test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
+        [
+            -0.3755, 0.8045, -0.6336, -0.2218, -1.1369, 0.8599, 1.5768, -0.1268, -0.1728, 0.528,
+            -1.131, 0.8908, 0.3118, 1.5984, -1.2089, -2.2168, 0.1783, 0.2429, -0.3838, 0.5802,
+            -0.3268, -2.0382, 0.6329, -0.2293, -1.2154, 0.6441, -0.3035, 0.5396, -0.8156, 0.4594,
+            2.8654, -0.8898, 0.3224, 1.7087, -0.9056, 0.4267
+        ]
+    );
     Ok(())
 }
 
@@ -162,7 +313,7 @@ fn conv2d_smaller(dev: &Device) -> Result<()> {
     let w = Tensor::new(&[1f32, 1., 1., 1., 1., 1., 1., 1., 1.], dev)?;
     let t = t.reshape((1, 1, 3, 3))?;
     let w = w.reshape((1, 1, 3, 3))?;
-    let res = t.conv2d(&w, 0, 1)?;
+    let res = t.conv2d(&w, 0, 1, 1, 1)?;
     assert_eq!(res.dims(), [1, 1, 1, 1]);
     assert_eq!(
         test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
@@ -171,8 +322,543 @@ fn conv2d_smaller(dev: &Device) -> Result<()> {
     Ok(())
 }
 
-test_device!(conv1d, conv1d_cpu, conv1d_gpu);
-test_device!(conv1d_small, conv1d_small_cpu, conv1d_small_gpu);
-test_device!(conv2d, conv2d_cpu, conv2d_gpu);
-test_device!(conv2d_small, conv2d_small_cpu, conv2d_small_gpu);
-test_device!(conv2d_smaller, conv2d_smaller_cpu, conv2d_smaller_gpu);
+/* This test is based on the following script.
+import torch
+torch.manual_seed(4242)
+
+t = torch.randn((1, 2, 4, 2))
+w = torch.randn((1, 2, 1, 1))
+print(t.flatten())
+print(w.flatten())
+res = torch.nn.functional.conv2d(t, w)
+print(res.flatten())
+*/
+fn conv2d_non_square(dev: &Device) -> Result<()> {
+    let t = Tensor::new(
+        &[
+            0.4056f32, -0.8689, -0.0773, -1.5630, -2.8012, -1.5059, 0.3972, 1.0852, 0.4997, 3.0616,
+            1.6541, 0.0964, -0.8338, -1.6523, -0.8323, -0.1699,
+        ],
+        dev,
+    )?;
+    let w = Tensor::new(&[-1.1351f32, 1.3841], dev)?;
+    let t = t.reshape((1, 2, 4, 2))?;
+    let w = w.reshape((1, 2, 1, 1))?;
+    let res = t.conv2d(&w, 0, 1, 1, 1)?;
+    assert_eq!(res.dims(), [1, 1, 4, 2]);
+    assert_eq!(
+        test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
+        [0.2312, 5.2238, 2.3772, 1.9076, 2.0256, -0.5776, -1.6028, -1.467]
+    );
+    Ok(())
+}
+
+/*
+import torch
+torch.manual_seed(4242)
+
+t = torch.randn((1, 4, 5, 5), requires_grad=True)
+w = torch.randn((2, 4, 3, 3), requires_grad=True)
+print(t.flatten())
+print(w.flatten())
+res = torch.nn.functional.conv2d(t, w)
+print(res.flatten())
+loss = (res ** 2).sum()
+print(loss)
+loss.backward()
+print(t.grad.shape)
+print(t.grad.flatten())
+print(w.grad.shape)
+print(w.grad.flatten())
+
+t.grad.zero_()
+w.grad.zero_()
+res = torch.nn.functional.conv2d(t, w, stride=2)
+print(res.flatten())
+loss = (res ** 2).sum()
+print(loss)
+loss.backward()
+print(t.grad.shape)
+print(t.grad[0])
+print(w.grad.shape)
+print(w.grad[0])
+*/
+fn conv2d_grad(dev: &Device) -> Result<()> {
+    // conv-transposes are not implemented for metal
+    use candle_core::Var;
+    let t = Var::from_slice(
+        &[
+            0.4056f32, -0.8689, -0.0773, -1.5630, -2.8012, -1.5059, 0.3972, 1.0852, 0.4997, 3.0616,
+            1.6541, 0.0964, -0.8338, -1.6523, -0.8323, -0.1699, 0.0823, 0.3526, 0.6843, 0.2395,
+            1.2279, -0.9287, -1.7030, 0.1370, 0.6047, 0.3770, -0.6266, 0.3529, 2.2013, -0.6836,
+            0.2477, 1.3127, -0.2260, 0.2622, -1.2974, -0.8140, -0.8404, -0.3490, 0.0130, 1.3123,
+            1.7569, -0.3956, -1.8255, 0.1727, -0.3538, 2.6941, 1.0529, 0.4219, -0.2071, 1.1586,
+            0.4717, 0.3865, -0.5690, -0.5010, -0.1310, 0.7796, 0.6630, -0.2021, 2.6090, 0.2049,
+            0.6466, -0.5042, -0.0603, -1.6538, -1.2429, 1.8357, 1.6052, -1.3844, 0.3323, -1.3712,
+            0.9634, -0.4799, -0.6451, -0.0840, -1.4247, 0.5512, -0.1747, -0.5509, -0.3742, 0.3790,
+            -0.4431, -0.4720, -0.7890, 0.2620, 0.7875, 0.5377, -0.6779, -0.8088, 1.9098, 1.2006,
+            -0.8, -0.4983, 1.5480, 0.8265, -0.1025, 0.5138, 0.5748, 0.3821, -0.4607, 0.0085,
+        ],
+        (1, 4, 5, 5),
+        dev,
+    )?;
+    let w = Var::from_slice(
+        &[
+            -0.9325f32, 0.6451, -0.8537, 0.2378, 0.8764, -0.1832, 0.2987, -0.6488, -0.2273,
+            -2.4184, -0.1192, -0.4821, -0.5079, -0.5766, -2.4729, 1.6734, 0.4558, 0.2851, 1.1514,
+            -0.9013, 1.0662, -0.1817, -0.0259, 0.1709, 0.5367, 0.7513, 0.8086, -2.2586, -0.5027,
+            0.9141, -1.3086, -1.3343, -1.5669, -0.1657, 0.7958, 0.1432, 0.3896, -0.4501, 0.1667,
+            0.0714, -0.0952, 1.2970, -0.1674, -0.3178, 1.0677, 0.3060, 0.7080, 0.1914, 1.1679,
+            -0.3602, 1.9265, -1.8626, -0.5112, -0.0982, 0.2621, 0.6565, 0.5908, 1.0089, -0.1646,
+            1.8032, -0.6286, 0.2016, -0.3370, 1.2555, 0.8009, -0.6488, -0.4652, -1.5685, 1.5860,
+            0.5583, 0.4623, 0.6026,
+        ],
+        (2, 4, 3, 3),
+        dev,
+    )?;
+    let res = t.conv2d(&w, 0, 1, 1, 1)?;
+    let loss = res.sqr()?.sum_all()?;
+    assert_eq!(test_utils::to_vec0_round(&loss, 2)?, 741.12f32);
+    let grads = loss.backward()?;
+    let grad_t = grads.get(&t).unwrap();
+    let grad_w = grads.get(&w).unwrap();
+    assert_eq!(grad_t.dims(), [1, 4, 5, 5]);
+    assert_eq!(grad_w.dims(), [2, 4, 3, 3]);
+    assert_eq!(
+        test_utils::to_vec1_round(&grad_t.flatten_all()?, 2)?,
+        [
+            9.29, -2.84, -5.71, 3.38, -7.71, -19.15, 7.02, 29.1, 9.34, 34.73, -22.87, 24.35,
+            -39.88, -14.01, 21.08, 9.94, 13.63, -34.68, 11.21, -6.26, 7.72, -6.32, -16.64, -1.08,
+            -20.22, 21.73, -0.37, -4.06, 5.82, -3.65, -30.73, 14.55, 87.7, 31.6, 4.53, -89.78,
+            -75.37, -57.43, -7.56, 92.96, 18.79, -4.63, -159.75, -42.47, -47.26, 52.88, 37.32,
+            49.0, 12.82, 2.01, -8.98, 20.18, 16.62, 12.06, 15.38, 20.0, 2.57, -15.22, 72.62,
+            -10.75, 2.25, -31.2, 3.75, -0.2, 9.76, -0.68, 5.21, -40.44, -22.59, -61.61, 17.28,
+            20.41, 37.55, 5.23, 6.81, 23.54, 23.62, -9.99, -9.13, 4.87, -35.06, -26.1, 63.48,
+            25.81, -39.21, -70.68, -46.96, 2.33, 41.81, 82.42, -28.63, -11.78, -35.33, -10.28,
+            -28.57, -9.13, 7.21, -9.05, -9.62, -11.25
+        ]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(&grad_w.flatten_all()?, 2)?,
+        [
+            -28.92, -22.88, -141.23, 73.35, 61.07, 47.81, -20.0, -73.71, -41.82, -13.59, 21.5,
+            28.72, 28.57, -46.85, -90.19, 143.61, 16.68, 7.43, 18.88, -90.81, -20.29, 54.79, 82.63,
+            22.94, 77.81, -16.39, -13.2, 9.34, -40.39, -26.62, 5.33, -60.91, 9.09, -59.37, 7.08,
+            58.64, 5.55, 20.52, 2.5, -17.25, -6.8, 22.21, 30.15, -7.52, -37.46, 5.67, 22.58, 9.03,
+            47.05, 17.61, 37.31, -98.13, -14.61, -4.8, -6.36, 44.69, 23.34, 8.37, -13.52, 80.05,
+            -34.24, -16.36, -12.31, 1.92, -33.62, -14.1, -49.23, -7.39, 11.5, -9.98, 9.66, 29.6
+        ]
+    );
+
+    // Same as before but with stride.
+    let res = t.conv2d(&w, 0, 2, 1, 1)?;
+    let loss = res.sqr()?.sum_all()?;
+    assert_eq!(test_utils::to_vec0_round(&loss, 2)?, 277.16f32);
+    let grads = loss.backward()?;
+    let grad_t = grads.get(&t).unwrap();
+    let grad_w = grads.get(&w).unwrap();
+    assert_eq!(grad_t.dims(), [1, 4, 5, 5]);
+    assert_eq!(grad_w.dims(), [2, 4, 3, 3]);
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_t.i(0)?, 2)?,
+        [
+            [
+                [9.29, -7.03, 0.94, 3.49, -7.71],
+                [-1.8, -7.82, 8.9, 8.46, 7.43],
+                [-25.84, 22.09, -19.27, -0.22, 1.69],
+                [4.02, 18.53, -18.37, 2.3, -24.51],
+                [7.72, -9.68, -12.34, 5.6, -20.22]
+            ],
+            [
+                [21.73, 3.39, -18.27, 3.86, -3.65],
+                [8.25, 3.73, 30.73, -8.61, -11.93],
+                [-72.15, -15.36, -17.53, -12.32, -1.61],
+                [-22.32, -7.79, -91.82, 6.44, -37.69],
+                [52.88, 14.44, 42.75, 9.88, 2.01]
+            ],
+            [
+                [-8.98, 9.91, 6.75, -4.68, 15.38],
+                [4.93, -0.33, 9.94, -1.46, 14.78],
+                [13.62, -30.63, 3.96, -3.58, -4.48],
+                [-14.13, 1.19, -34.43, 3.08, -33.83],
+                [17.28, 12.94, 31.83, -3.35, 6.81]
+            ],
+            [
+                [23.54, 6.98, -24.52, 0.52, 4.87],
+                [9.65, 6.18, 1.71, -25.23, -4.93],
+                [-54.99, -23.66, 3.19, -3.73, 18.58],
+                [-21.35, -10.39, -39.88, 28.73, -30.76],
+                [-9.13, 11.12, -14.0, -8.23, -11.25]
+            ]
+        ]
+    );
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_w.i(0)?, 2)?,
+        [
+            [
+                [28.34, -7.91, -45.75],
+                [21.03, 3.86, 29.86],
+                [0.72, -36.58, -35.28]
+            ],
+            [
+                [-16.04, 11.53, -16.38],
+                [29.62, -16.32, -48.35],
+                [57.5, 28.29, 25.81]
+            ],
+            [
+                [2.93, -19.6, 1.57],
+                [27.15, 53.88, -24.64],
+                [12.74, -22.6, -26.2]
+            ],
+            [
+                [-0.18, -14.86, -6.82],
+                [-19.55, -2.72, 45.9],
+                [-2.54, 36.97, 27.11]
+            ]
+        ]
+    );
+
+    // Replicate the issue from https://github.com/huggingface/candle/issues/1212
+    let res = t.i((.., .., 0..4, 0..4))?.conv2d(&w, 0, 2, 1, 1)?;
+    let loss = res.sqr()?.sum_all()?;
+    assert_eq!(test_utils::to_vec0_round(&loss, 2)?, 21.12f32);
+    let grads = loss.backward()?;
+    let grad_t = grads.get(&t).unwrap();
+    let grad_w = grads.get(&w).unwrap();
+    assert_eq!(grad_t.dims(), [1, 4, 5, 5]);
+    assert_eq!(grad_w.dims(), [2, 4, 3, 3]);
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_t.i(0)?, 2)?,
+        [
+            [
+                [9.29, -7.03, 7.87, 0.0, 0.0],
+                [-1.8, -7.82, 5.9, 0.0, 0.0],
+                [-3.12, 4.49, 5.52, 0.0, 0.0],
+                [0.0, 0.0, 0.0, 0.0, 0.0],
+                [0.0, 0.0, 0.0, 0.0, 0.0]
+            ],
+            [
+                [21.73, 3.39, 4.77, 0.0, 0.0],
+                [8.25, 3.73, 27.61, 0.0, 0.0],
+                [-20.55, -5.61, -2.77, 0.0, 0.0],
+                [0.0, 0.0, 0.0, 0.0, 0.0],
+                [0.0, 0.0, 0.0, 0.0, 0.0]
+            ],
+            [
+                [-8.98, 9.91, -7.15, 0.0, 0.0],
+                [4.93, -0.33, 4.56, 0.0, 0.0],
+                [-6.7, -5.76, -8.05, 0.0, 0.0],
+                [0.0, 0.0, 0.0, 0.0, 0.0],
+                [0.0, 0.0, 0.0, 0.0, 0.0]
+            ],
+            [
+                [23.54, 6.98, -10.0, 0.0, 0.0],
+                [9.65, 6.18, 18.72, 0.0, 0.0],
+                [3.29, -5.27, 0.79, 0.0, 0.0],
+                [0.0, 0.0, 0.0, 0.0, 0.0],
+                [0.0, 0.0, 0.0, 0.0, 0.0]
+            ]
+        ]
+    );
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_w.i(0)?, 2)?,
+        [
+            [
+                [-3.47, 7.44, 0.66],
+                [12.89, -3.4, -9.29],
+                [-14.16, -0.83, 7.14]
+            ],
+            [
+                [-3.23, 5.37, -3.02],
+                [-2.12, -11.24, 1.94],
+                [6.97, 7.2, 2.99]
+            ],
+            [
+                [-4.04, -3.31, 4.87],
+                [-6.68, -5.68, 1.73],
+                [-5.54, 4.32, 0.52]
+            ],
+            [[-4.72, 1.5, 4.72], [3.79, 4.04, 6.76], [-4.6, 5.8, 6.93]]
+        ]
+    );
+
+    // Conv Transpose 2d Test
+    //tested against following python
+
+    // import torch
+    // torch.manual_seed(4242)
+    // padding = 4
+    // outpadding = 2
+    // dilation = 3
+    // stride = 3
+    // input = torch.randn((1, 4, 7, 5), requires_grad=True)
+    // kernel = torch.randn((4, 2, 3, 5), requires_grad=True)
+    // print("input", input.flatten())
+    // print("kernel", kernel.flatten())
+    // res = torch.nn.functional.conv_transpose2d(
+    //     input,
+    //     kernel,
+    //     stride=stride,
+    //     padding=padding,
+    //     dilation=dilation,
+    //     output_padding=outpadding,
+    // )
+    // res.retain_grad()
+    // print(res.shape)
+    // loss = (res**2).sum()
+    // print(loss)
+    // loss.backward()
+    // print(input.grad.shape)
+    // print("input grad", torch.round(input.grad, decimals=1))
+    // print(kernel.grad.shape)
+    // print("kernel grad", torch.round(kernel.grad.flatten(), decimals=1))
+
+    let padding = 4;
+    let outpadding = 2;
+    let dilation = 3;
+    let stride = 3;
+
+    let t = Var::from_slice(
+        &[
+            0.4056_f32, -0.8689, -0.0773, -1.5630, -2.8012, -1.5059, 0.3972, 1.0852, 0.4997,
+            3.0616, 1.6541, 0.0964, -0.8338, -1.6523, -0.8323, -0.1699, 0.0823, 0.3526, 0.6843,
+            0.2395, 1.2279, -0.9287, -1.7030, 0.1370, 0.6047, 0.3770, -0.6266, 0.3529, 2.2013,
+            -0.6836, 0.2477, 1.3127, -0.2260, 0.2622, -1.2974, -0.8140, -0.8404, -0.3490, 0.0130,
+            1.3123, 1.7569, -0.3956, -1.8255, 0.1727, -0.3538, 2.6941, 1.0529, 0.4219, -0.2071,
+            1.1586, 0.4717, 0.3865, -0.5690, -0.5010, -0.1310, 0.7796, 0.6630, -0.2021, 2.6090,
+            0.2049, 0.6466, -0.5042, -0.0603, -1.6538, -1.2429, 1.8357, 1.6052, -1.3844, 0.3323,
+            -1.3712, 0.9634, -0.4799, -0.6451, -0.0840, -1.4247, 0.5512, -0.1747, -0.5509, -0.3742,
+            0.3790, -0.4431, -0.4720, -0.7890, 0.2620, 0.5411, -1.1715, -2.4997, 2.3249, -0.8912,
+            -0.4733, -0.5701, -2.8888, -1.4112, -0.5471, -0.9234, -1.1660, 0.4189, -0.7465,
+            -0.6473, 0.1402, 0.7875, 0.5377, -0.6779, -0.8088, -0.4864, -0.2312, 0.9279, 0.1264,
+            1.5480, 0.8265, -0.1025, 0.5138, -0.2512, 0.1576, 1.2705, 0.3641, -0.9325, 0.6451,
+            -0.8537, 0.2378, 0.1794, 0.2752, -0.3687, -1.1149, -0.1410, -0.5829, -0.0892, 1.4258,
+            -2.2789, 0.5270, 0.1825, 1.7007, -0.5263, -0.2954, 0.4440, 0.5537, 0.3492, 0.6186,
+            1.6475, 0.2219,
+        ],
+        (1, 4, 7, 5),
+        dev,
+    )?;
+
+    #[rustfmt::skip]
+    let w = Var::from_slice(
+        &[
+            -1.1744_f32, 0.3266, 2.5893, 1.0142, 0.1763, 0.7752, 0.6604, 0.2029, -0.2145, 0.7234,
+            -0.3441, -1.5400, -0.6333, 0.6613, 0.2083, 0.6230, -1.7002, 0.3393, 0.4049, 1.0762,
+            0.2723, 1.4181, 0.0029, -0.2122, 1.7668, 1.4168, 0.3320, -0.2719, 0.7932, -0.7204,
+            0.4447, 0.1211, 0.5908, 1.0089, -0.1646, 1.8033, -0.6286, 0.2016, -0.3370, 1.2555,
+            0.8009, -0.6488, -0.4652, -1.5685, 1.5860, 0.5583, 0.4623, 0.6026, 0.8828, 2.4990,
+            0.6811, -0.3369, 1.3320, 1.7669, -1.1067, 1.2958, -0.9415, -0.9655, -0.4462, 0.7181,
+            0.5181, -1.1658, -1.8467, -0.7763, 1.2769, 0.8651, 0.9890, 1.5092, 0.7207, -0.8481,
+            0.7417, 0.3375, -1.2685, 1.4572, 1.0915, 0.1093, -0.8550, -0.5831, -0.6309, -0.2509,
+            0.5220, -0.0914, 0.7900, 0.1096, 0.3258, 0.2723, -1.0942, -0.3393, -0.1653, 0.5732,
+            -0.8014, 1.8194, -1.9023, 0.2127, 1.8636, -0.8979, 0.1927, -0.2778, 0.3105, 0.0071,
+            -1.1823, 0.2476, -0.7178, -1.3821, 1.0769, -0.4376, -0.9967, -0.1227, 1.6197, -1.0604,
+            0.1372, 0.8141, -0.6163, 0.7304, -0.8285, 2.0636, -0.7176, 0.2495, -0.2581, -0.4478,
+        ],
+        (4, 2, 3, 5),
+        dev,
+    )?;
+    let res = t.conv_transpose2d(&w, padding, outpadding, stride, dilation)?;
+    let loss = res.sqr()?.sum_all()?;
+    assert_eq!(test_utils::to_vec0_round(&loss, 0)?, 2904.0);
+    let grads = loss.backward()?;
+
+    let grad_t = grads.get(&t).unwrap();
+    let grad_w = grads.get(&w).unwrap();
+    assert_eq!(grad_t.dims(), [1, 4, 7, 5]);
+    assert_eq!(grad_w.dims(), [4, 2, 3, 5]);
+
+    assert_eq!(
+        test_utils::to_vec1_round(&grad_w.flatten_all()?, 1)?,
+        [
+            // torch gets 89.1
+            -89.0, -135.3, 136.7, 102.0, -53.4, 117.9, 118.6, -43.9, -218.0, -58.5, -114.3, -150.0,
+            -15.6, 172.1, 66.3, -64.3, -27.9, -19.8, 31.7, 62.1, 5.5, 92.6, 28.2, -29.6, 55.9,
+            52.7, -72.7, -119.8, 53.8, -25.5, 128.8, 19.3, 68.0, 190.9, -64.1, -86.2, -111.2,
+            106.6, -67.7, 37.8, 115.9, 50.4, -77.7, -54.9, 22.3, -4.6, 89.8, 61.7, 122.4, 192.6,
+            -27.8, -104.6, 57.0, 166.4, 27.1, 6.1, 18.7, -93.2, 31.5, 168.2, -3.7, -99.5, -55.5,
+            -10.8, 17.5, 20.8, 16.9, 43.8, 42.0, -89.2, 18.8, -9.6, -84.1, 212.6, 19.7, -50.0,
+            -52.0, -40.0, -166.6, -73.2, -10.8, -73.3, 31.5, -23.4, -79.3, -27.0, -84.4, -42.9,
+            -20.3, 51.8, -16.7, 76.3, -120.5, -65.8, 96.5, -10.7, -45.9, -88.1, 65.4, -7.0, -1.5,
+            92.8, -25.1, -114.2, -5.8, -14.8, -51.2, -20.7, 54.2, -79.8, 47.7, -29.2, -8.8, 53.5,
+            -28.4, 85.0, -18.3, 107.0, 28.3, -71.8
+        ]
+    );
+
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_t.i(0)?, 1)?,
+        [
+            [
+                [32.3, -41.6, -24.0, 14.1, 17.6],
+                [-11.8, 72.5, 87.6, 46.4, 61.5],
+                [115.0, 108.5, -48.6, -63.4, -50.0],
+                [51.3, 5.4, 31.3, 91.1, -30.9],
+                [52.7, 92.8, -68.0, -47.0, 83.0],
+                // pytorch gets -107.1
+                [-10.2, -107.0, -5.4, 213.1, -31.4],
+                [-2.4, 65.1, 9.2, -146.2, -24.2]
+            ],
+            [
+                [-72.6, -63.9, -61.9, 45.3, 33.0],
+                [79.3, -0.5, -26.2, 78.2, 42.7],
+                [90.9, 141.6, 40.1, -62.7, 37.0],
+                [32.8, 198.2, -0.8, -31.1, 27.3],
+                // torch gets 48.0
+                [34.5, 34.9, -47.9, 127.6, -12.3],
+                [-61.4, -3.2, -2.9, -10.9, -16.6],
+                [74.6, 60.1, -68.9, 34.5, -50.4]
+            ],
+            [
+                [37.5, -56.9, -43.6, -13.5, -9.9],
+                [40.0, 97.3, 28.6, 14.2, -30.1],
+                [-22.3, -126.3, -68.8, -8.2, 26.1],
+                [-32.9, 37.3, 108.5, -54.8, 29.6],
+                [34.9, -176.9, -125.0, -28.3, -13.9],
+                [-54.9, 142.6, 62.1, -80.4, -65.6],
+                [7.4, -91.1, -67.6, 35.0, 39.7]
+            ],
+            [
+                [-57.2, -40.9, -10.1, 32.6, 29.4],
+                [18.7, -18.0, 29.5, -1.2, 59.2],
+                [-14.0, -74.4, 19.8, -117.0, 58.2],
+                [-21.8, 163.5, -71.1, -99.0, 80.9],
+                [-58.9, -10.9, 93.8, -139.6, 98.0],
+                // torch gets 54.5
+                [-54.4, 135.3, 6.0, -79.1, 134.6],
+                [27.5, -76.0, 43.4, -2.8, -7.8]
+            ]
+        ]
+    );
+
+    // Test the same, but then with the following properties, t & w are unmodified.
+    let padding = 1;
+    let outpadding = 1;
+    let dilation = 1;
+    let stride = 2;
+
+    let res = t.conv_transpose2d(&w, padding, outpadding, stride, dilation)?;
+    let loss = res.sqr()?.sum_all()?;
+    assert_eq!(test_utils::to_vec0_round(&loss, 0)?, 3627.0); // torch gives 3626.8560
+
+    let grads = loss.backward()?;
+
+    let grad_t = grads.get(&t).unwrap();
+    let grad_w = grads.get(&w).unwrap();
+    assert_eq!(grad_t.dims(), [1, 4, 7, 5]);
+    assert_eq!(grad_w.dims(), [4, 2, 3, 5]);
+
+    #[rustfmt::skip]
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_t.i(0)?, 1)?,
+        [
+            [
+                [  13.2,  -40.7,   -9.7,  -47.3,  -82.7],
+                [ -98.2,    9.7,   57.7,   -6.2,  180.7],
+                [ 100.2,   24.1,    3.7, -100.5,  -48.1],
+                [  -0.3,   13.5,   -2.9,   80.0,  -49.8],
+                [  47.2,  -25.6,  -74.4,   61.2,  -18.4],
+                [   4.6,  -69.5,   27.9,   66.5,  -88.1],
+                 // 4th column on next row; torch is 4.2
+                [ -12.0,   79.2,  -40.0,    4.1,  -97.1],
+            ],
+            [
+                [ -42.2,  -36.5,  -51.1,    7.5,   32.3],
+                [  74.1,  -44.6,  -68.8,   19.5,    7.7],
+                [ 137.1,   54.2,  153.8,  -58.0,   45.5],
+                [  24.4,  -56.8,    9.7,  -41.0,  -14.5],
+                [  -3.7,   72.6,    8.3,  134.8,   40.5],
+                [  43.2,  -56.9,  -47.5,  -89.4,  -95.4],
+                [  68.2,  108.1,  -80.0,   57.0, -121.1]
+            ],
+            [
+                [  31.1,  -11.4,  -34.8,   33.1,  -44.2],
+                [  29.4,  -31.6,  -40.2,   13.7,   13.1],
+                [  -0.8,  -83.8,   -7.8,  -17.3,   78.2],
+                [  12.0, -118.7,  137.5,  -76.7,   50.8],
+                [ -28.7, -114.2,   -3.7,  -96.3,  -13.8],
+                [ -31.8,   28.5,  -14.3,    4.6,   13.4],
+                [  28.0,   -0.2,  -38.9,  -29.7,  -59.0]
+            ],
+            [
+                [ -16.8,   38.5,   15.5,   26.6,   48.9],
+                [  14.5,   49.6,  -24.8,   65.6,   61.7],
+                [  22.1,  -64.7,   -4.3,  -51.0,   36.3],
+                [  31.0,  -88.9,   47.1, -123.5,   -3.8],
+                [ -14.8,  -39.8,  128.2, -110.3,   42.6],
+                // 1st column on next row; torch is -7.2
+                [  -7.1,   95.3,  -21.3,  -58.7,  -13.9], 
+                [  26.9,   21.3,   16.1,   70.3,   32.1]
+            ]
+        ]
+    );
+
+    #[rustfmt::skip]
+    assert_eq!(
+        test_utils::to_vec1_round(&grad_w.flatten_all()?, 1)?,
+        [
+            // 2nd value; torch gets -3.2, 3rd value; torch gets 221.8
+           -2.460e+01, -3.100e+00,  2.219e+02,  7.400e+00,  5.620e+01,
+            7.420e+01,  7.830e+01,  8.900e+00,  1.050e+01,  2.810e+01,
+            5.100e+00, -1.046e+02, -1.572e+02,  8.710e+01, -9.840e+01,
+           -4.230e+01, -1.898e+02,  1.860e+01, -3.570e+01,  9.810e+01,
+            4.680e+01,  1.182e+02,  4.020e+01, -1.900e+00,  1.508e+02,
+            1.094e+02,  1.018e+02, -4.620e+01,  1.591e+02, -2.320e+01,
+            // 5th value; torch gets 7.1
+           -8.450e+01, -4.600e+00,  6.330e+01,  1.123e+02, -7.000e+00,
+            1.101e+02, -6.620e+01,  2.090e+01, -5.120e+01,  8.990e+01,
+            9.050e+01, -6.990e+01,  6.800e+01, -9.250e+01,  1.380e+02,
+            4.720e+01,  4.710e+01,  6.210e+01,  8.870e+01,  2.098e+02,
+            3.870e+01, -1.390e+01,  6.270e+01,  1.484e+02, -9.920e+01,
+           -4.200e+01, -1.505e+02, -1.480e+01, -2.620e+01,  8.220e+01,
+           -3.350e+01, -2.260e+01, -1.198e+02, -5.080e+01,  1.259e+02,
+            5.600e+01,  9.270e+01,  1.209e+02,  6.590e+01, -8.330e+01,
+            7.000e+00, -2.600e+01, -1.133e+02,  3.870e+01,  4.020e+01,
+           -6.300e+00, -8.710e+01, -5.150e+01, -8.510e+01,  2.000e-01,
+            3.640e+01, -6.100e+00,  6.590e+01, -2.700e+00,  6.550e+01,
+            // 4th value; torch gets 3.8
+            5.300e+00, -6.760e+01, -4.270e+01, -3.900e+00,  2.880e+01,
+            5.260e+01,  6.170e+01, -1.203e+02, -1.610e+01,  7.740e+01,
+           -1.008e+02, -1.070e+01, -9.900e+00,  3.300e+00, -2.620e+01,
+           -4.440e+01,  2.580e+01, -6.920e+01, -4.220e+01,  1.108e+02,
+            1.240e+01, -3.440e+01, -2.800e+00,  7.880e+01, -6.690e+01,
+            1.480e+01,  2.310e+01, -4.260e+01, -1.500e+00, -4.760e+01,
+            5.350e+01, -2.260e+01,  8.000e-01, -3.840e+01, -2.500e+00
+        ]
+    );
+
+    Ok(())
+}
+
+test_device!(conv1d, conv1d_cpu, conv1d_gpu, conv1d_metal);
+test_device!(
+    conv1d_small,
+    conv1d_small_cpu,
+    conv1d_small_gpu,
+    conv1d_small_metal
+);
+test_device!(conv2d, conv2d_cpu, conv2d_gpu, conv2d_metal);
+test_device!(
+    conv2d_non_square,
+    conv2d_non_square_cpu,
+    conv2d_non_square_gpu,
+    conv2d_non_square_metal
+);
+test_device!(
+    conv2d_small,
+    conv2d_small_cpu,
+    conv2d_small_gpu,
+    conv2d_small_metal
+);
+test_device!(
+    conv2d_smaller,
+    conv2d_smaller_cpu,
+    conv2d_smaller_gpu,
+    conv2d_smaller_metal
+);
+test_device!(
+    conv2d_grad,
+    conv2d_grad_cpu,
+    conv2d_grad_gpu,
+    conv2_grad_metal
+);

candle-core/tests/custom_op_tests.rs

@@ -1,10 +1,8 @@
 use candle_core::backend::BackendStorage;
 use candle_core::cpu_backend;
+use candle_core::test_utils::to_vec1_round;
 use candle_core::{CpuStorage, CustomOp1, DType, Device, Error, Layout, Result, Shape, Tensor};
 
-mod test_utils;
-use test_utils::to_vec1_round;
-
 fn fwd<T: num_traits::Float>(v: T, alpha: f64) -> T {
     if v.is_sign_positive() {
         v
@@ -39,7 +37,7 @@ fn custom_op1_no_backward() -> Result<()> {
     let cpu = &Device::Cpu;
     let t = Tensor::arange(0u32, 12u32, cpu)?.to_dtype(DType::F32)?;
     let t = (t - 5.)?;
-    let elu_t = t.custom_op1(Elu { alpha: 1. })?;
+    let elu_t = t.apply_op1_no_bwd(&Elu { alpha: 1. })?;
     assert_eq!(
         to_vec1_round(&elu_t, 4)?,
         &[-0.9933, -0.9817, -0.9502, -0.8647, -0.6321, 0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
@@ -96,7 +94,7 @@ impl CustomOp1 for EluWithBackward {
 
     fn bwd(&self, arg: &Tensor, _res: &Tensor, grad_res: &Tensor) -> Result<Option<Tensor>> {
         let alpha = self.0.alpha;
-        let bwd = arg.custom_op1(EluBackward { alpha })?;
+        let bwd = arg.apply_op1(EluBackward { alpha })?;
         Ok(Some(grad_res.mul(&bwd)?))
     }
 }
@@ -105,7 +103,7 @@ impl CustomOp1 for EluWithBackward {
 fn custom_op1_with_backward() -> Result<()> {
     let cpu = &Device::Cpu;
     let t = candle_core::Var::new(&[-2f32, 0f32, 2f32], cpu)?;
-    let elu_t = t.custom_op1(EluWithBackward::new(2.))?;
+    let elu_t = t.apply_op1(EluWithBackward::new(2.))?;
     assert_eq!(to_vec1_round(&elu_t, 4)?, &[-1.7293, 0.0, 2.0]);
 
     let grads = elu_t.backward()?;
@@ -114,3 +112,34 @@ fn custom_op1_with_backward() -> Result<()> {
 
     Ok(())
 }
+
+impl candle_core::InplaceOp1 for Elu {
+    fn name(&self) -> &'static str {
+        "elu"
+    }
+
+    fn cpu_fwd(&self, s: &mut CpuStorage, _l: &Layout) -> Result<()> {
+        let alpha = self.alpha;
+        match s {
+            CpuStorage::BF16(s) => s.iter_mut().for_each(|v| *v = fwd(*v, alpha)),
+            CpuStorage::F16(s) => s.iter_mut().for_each(|v| *v = fwd(*v, alpha)),
+            CpuStorage::F32(s) => s.iter_mut().for_each(|v| *v = fwd(*v, alpha)),
+            CpuStorage::F64(s) => s.iter_mut().for_each(|v| *v = fwd(*v, alpha)),
+            _ => candle_core::bail!("unsupported dtype for inplace elu"),
+        }
+        Ok(())
+    }
+}
+
+#[test]
+fn inplace_op1() -> Result<()> {
+    let cpu = &Device::Cpu;
+    let t = Tensor::arange(0u32, 12u32, cpu)?.to_dtype(DType::F32)?;
+    let t = (t - 5.)?;
+    t.inplace_op1(&Elu { alpha: 1. })?;
+    assert_eq!(
+        to_vec1_round(&t, 4)?,
+        &[-0.9933, -0.9817, -0.9502, -0.8647, -0.6321, 0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
+    );
+    Ok(())
+}

candle-core/tests/grad_tests.rs

@@ -1,6 +1,6 @@
+#![allow(clippy::approx_constant)]
 use anyhow::{Context, Result};
-use candle_core::{Device, Shape, Tensor, Var};
-mod test_utils;
+use candle_core::{test_device, test_utils, Device, Shape, Tensor, Var};
 
 fn simple_grad(device: &Device) -> Result<()> {
     let x = Var::new(&[3f32, 1., 4.], device)?;
@@ -97,24 +97,24 @@ fn unary_grad(device: &Device) -> Result<()> {
     let grads = y.backward()?;
     let grad_x = grads.get(x).context("no grad for x")?;
     assert_eq!(
-        y.to_vec1::<f32>()?,
-        [20.085537, 2.7182817, 54.59815, 1.1618342]
+        test_utils::to_vec1_round(&y, 4)?,
+        [20.0855, 2.7183, 54.5982, 1.1618]
     );
     assert_eq!(
-        grad_x.to_vec1::<f32>()?,
-        [20.085537, 2.7182817, 54.59815, 1.1618342]
+        test_utils::to_vec1_round(grad_x, 4)?,
+        [20.0855, 2.7183, 54.5982, 1.1618]
     );
     let y = x.exp()?.sqr()?;
     let grads = y.backward()?;
     let grad_x = grads.get(x).context("no grad for x")?;
     assert_eq!(
-        y.to_vec1::<f32>()?,
-        [403.4288, 7.3890557, 2980.9578, 1.3498588]
+        test_utils::to_vec1_round(&y, 3)?,
+        [403.429, 7.389, 2980.958, 1.35]
     );
     // exp(x)^2 = exp(2*x)
     assert_eq!(
-        grad_x.to_vec1::<f32>()?,
-        [806.8576, 14.778111, 5961.9155, 2.6997175]
+        test_utils::to_vec1_round(grad_x, 2)?,
+        [806.86, 14.78, 5961.92, 2.7]
     );
     let y = x.sin()?;
     let grads = y.backward()?;
@@ -174,11 +174,360 @@ fn unary_grad(device: &Device) -> Result<()> {
     let grad_x = grads.get(x).context("no grad for x")?;
     assert_eq!(y.to_vec1::<f32>()?, [6., 2., 8., 0.3]);
     assert_eq!(grad_x.to_vec1::<f32>()?, [2., 2., 2., 2.]);
+
+    let x = Var::new(&[3f32, 1., 4., 0.15], device)?;
+    let y = x.powf(2.5)?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(test_utils::to_vec1_round(&y, 2)?, [15.59, 1.0, 32.0, 0.01]);
+    assert_eq!(
+        test_utils::to_vec1_round(grad_x, 2)?,
+        [12.99, 2.5, 20.0, 0.15]
+    );
+
+    let y = x.tanh()?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(test_utils::to_vec1_round(&y, 2)?, [1.0, 0.76, 1.0, 0.15]);
+    assert_eq!(
+        test_utils::to_vec1_round(grad_x, 2)?,
+        [0.01, 0.42, 0.0, 0.98],
+    );
+
+    // testing compared to pytorch nn.GELU(approximate = 'tanh')
+    let y = x.gelu()?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [2.9964, 0.8412, 3.9999, 0.0839]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(grad_x, 4)?,
+        [1.0116, 1.0830, 1.0003, 0.6188],
+    );
+
+    // Testing compared to pytorch torch.erf
+    //
+    // import torch
+    // x = torch.tensor([3.0, 1.0, 4.0, 0.15], requires_grad=True)
+    // y = x.erf()
+    // print(y)
+    // loss = y.sum()
+    // loss.backward()
+    // print(x.grad)
+    let y = x.erf()?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(test_utils::to_vec1_round(&y, 4)?, [1.0, 0.8427, 1.0, 0.168]);
+    assert_eq!(
+        test_utils::to_vec1_round(grad_x, 4)?,
+        [0.0001, 0.4151, 0.0, 1.1033],
+    );
+
+    // Testing compared to pytorch nn.GELU(approximate = 'none')
+    //
+    // import torch
+    // import torch.nn.functional as F
+    // x = torch.tensor([3.0, 1.0, 4.0, 0.15], requires_grad=True)
+    // y = F.gelu(x, approximate='none')
+    // print(y)
+    // loss = y.sum()
+    // loss.backward()
+    // print(x.grad)
+    let y = x.gelu_erf()?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [2.9960, 0.8413, 3.9999, 0.0839]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(grad_x, 4)?,
+        [1.0119, 1.0833, 1.0005, 0.6188],
+    );
+
+    // Testing compared to pytorch elu
+    //
+    // import torch
+    // import torch.nn.functional as F
+    // x = torch.tensor([-1.0, 0.0, -2.0, 3.0], requires_grad=True)
+    // y = F.elu(x, alpha=2.0)
+    // print(y)
+    // loss = y.min
+    // loss = y.sum()
+    // loss.backward()
+    // print(x.grad)
+    let elu_x = Var::new(&[-1.0f32, 0., -2., 3.], device)?;
+    let y = elu_x.elu(2.)?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(&elu_x).context("no grad for x")?;
+
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [-1.2642, 0.0000, -1.7293, 3.0000]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(grad_x, 4)?,
+        [0.7358, 2.0000, 0.2707, 1.0000]
+    );
+
+    // testing compared to pytorch nn.Silu()
+    let y = x.silu()?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [2.8577, 0.7311, 3.9281, 0.0806]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(grad_x, 4)?,
+        [1.0881, 0.9277, 1.0527, 0.5747],
+    );
+
+    if device.is_cpu() {
+        let x = Var::new(&[[[1f32, 2., 3.], [4., 5., 6.], [7., 8., 9.]]], device)?;
+        let y = x.interpolate1d(12)?.reshape(36)?;
+
+        let z = Tensor::new(
+            &[
+                1_f32, 02., 03., 04., 05., 06., 07., 08., 09., 10., 11., 12., 13., 14., 15., 16.,
+                17., 18., 19., 20., 21., 22., 23., 24., 25., 26., 27., 28., 29., 30., 31., 32.,
+                33., 34., 35., 36.,
+            ],
+            device,
+        )?;
+
+        let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+        let grads = loss.backward()?;
+        let grad_x = grads.get(&x).context("no grad for x")?;
+
+        assert_eq!(
+            test_utils::to_vec3_round(grad_x, 4)?,
+            [[[10_f32, 26., 42.], [58., 74., 90.], [106., 122., 138.]]]
+        );
+    }
+
+    // manually checked: see comments
+    let x = Var::new(&[[[[1f32, 2., 3.], [4., 5., 6.], [7., 8., 9.]]]], device)?;
+    let y = x.interpolate2d(6, 6)?.reshape(36)?;
+
+    let z = Tensor::new(
+        &[
+            1_f32, 02., 03., 04., 05., 06., 07., 08., 09., 10., 11., 12., 13., 14., 15., 16., 17.,
+            18., 19., 20., 21., 22., 23., 24., 25., 26., 27., 28., 29., 30., 31., 32., 33., 34.,
+            35., 36.,
+        ],
+        device,
+    )?;
+    // gradient should be
+    // row 1
+    // 1+2+7+8 = 18
+    // 3+4+9+10 = 26
+    // 5+6+11+12 = 34
+    // row 2
+    // 13+14+19+20 = 66
+    // 15+16+21+22 = 74
+    // 17+18+23+24 = 82
+    // row 3
+    // 25+26+31+32 = 114
+    // 27+28+33+34 = 122
+    // 29+30+35+36 = 130
+    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+
+    let grads = loss.backward()?;
+
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(
+        test_utils::to_vec2_round(&grad_x.flatten(0, 2)?, 4)?,
+        [[18_f32, 26., 34.], [66., 74., 82.], [114., 122., 130.]]
+    );
+
+    // manually checked: see comments
+    let x = Var::new(&[[[[1f32, 2.], [4., 5.]]]], device)?;
+    let y = x.interpolate2d(6, 6)?.reshape(36)?;
+
+    let z = Tensor::new(
+        &[
+            1_f32, 02., 03., 04., 05., 06., 07., 08., 09., 10., 11., 12., 13., 14., 15., 16., 17.,
+            18., 19., 20., 21., 22., 23., 24., 25., 26., 27., 28., 29., 30., 31., 32., 33., 34.,
+            35., 36.,
+        ],
+        device,
+    )?;
+    // gradient should be
+    // row 1
+    // 1+2+3+7+8+9+13+14+15 = 72
+    // 4+5+6+10+11+12+16+17+18 = 99
+    // row 2
+    // 19+20+21+25+26+27+31+32+33 = 234
+    // 22+23+24+28+29+30+34+35+36 = 243
+    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+
+    let grads = loss.backward()?;
+
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(
+        test_utils::to_vec2_round(&grad_x.flatten(0, 2)?, 4)?,
+        [[72_f32, 99.], [234., 261.]]
+    );
+
+    // manually checked: see comments
+    let x = Var::new(&[[[[1f32, 2.], [4., 5.]], [[6f32, 7.], [8., 9.]]]], device)?;
+
+    let y = x.interpolate2d(4, 4)?.reshape(32)?;
+
+    #[rustfmt::skip]
+    let z = Tensor::new(
+        &[
+            1_f32, 02., 03., 04.,
+            05.,   06., 07., 08.,
+            09.,   10., 11., 12.,
+            13.,   14., 15., 16.,
+            17.,   18., 19., 20.,
+            21.,   22., 23., 24.,
+            25.,   26., 27., 28.,
+            29.,   30., 31., 32.
+        ],
+        device,
+    )?;
+    // gradient should be
+    // m1r1
+    // 1+2+5+6=14
+    // 3+4+7+8=22
+    // m1r2
+    // 9+10+13+14=46
+    // 11+12+15+16=54
+    // m2r1
+    // 17+18+21+22=78
+    // 19+20+23+24=86
+    // m2r2
+    // 25+26+29+30=110
+    // 27+28+31+32=118
+    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+
+    let grads = loss.backward()?;
+
+    let grad_x = grads.get(&x).context("no grad for x")?;
+
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_x.flatten(0, 1)?, 4)?,
+        [[[14_f32, 22.], [46., 54.]], [[78., 86.], [110., 118.]]]
+    );
+
+    // manually checked: see comments
+    let x = Var::new(
+        &[[[[1f32, 2.], [4., 5.]]], [[[6f32, 7.], [8., 9.]]]],
+        device,
+    )?;
+
+    let y = x.interpolate2d(4, 4)?.reshape(32)?;
+
+    #[rustfmt::skip]
+       let z = Tensor::new(
+           &[
+               1_f32, 02., 03., 04.,
+               05.,   06., 07., 08.,
+               09.,   10., 11., 12.,
+               13.,   14., 15., 16.,
+               17.,   18., 19., 20.,
+               21.,   22., 23., 24.,
+               25.,   26., 27., 28.,
+               29.,   30., 31., 32.
+           ],
+           device,
+       )?;
+    // gradient should be
+    // m1r1
+    // 1+2+5+6=14
+    // 3+4+7+8=22
+    // m1r2
+    // 9+10+13+14=46
+    // 11+12+15+16=54
+    // m2r1
+    // 17+18+21+22=78
+    // 19+20+23+24=86
+    // m2r2
+    // 25+26+29+30=110
+    // 27+28+31+32=118
+    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+
+    let grads = loss.backward()?;
+
+    let grad_x = grads.get(&x).context("no grad for x")?;
+
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_x.flatten(0, 1)?, 4)?,
+        [[[14_f32, 22.], [46., 54.]], [[78., 86.], [110., 118.]]]
+    );
     Ok(())
 }
 
-test_device!(simple_grad, simple_grad_cpu, simple_grad_gpu);
-test_device!(sum_grad, sum_grad_cpu, sum_grad_gpu);
-test_device!(matmul_grad, matmul_grad_cpu, matmul_grad_gpu);
-test_device!(grad_descent, grad_descent_cpu, grad_descent_gpu);
-test_device!(unary_grad, unary_grad_cpu, unary_grad_gpu);
+fn binary_grad(device: &Device) -> Result<()> {
+    let x = Var::new(&[3f32, 1., -4., -1.], device)?;
+    let x = x.as_tensor();
+    // leaky relu
+    let y = x.maximum(&(x * 0.1)?)?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(x).context("no grad for x")?;
+    assert_eq!(x.to_vec1::<f32>()?, [3., 1., -4., -1.]);
+    assert_eq!(y.to_vec1::<f32>()?, [3., 1., -0.4, -0.1]);
+    assert_eq!(grad_x.to_vec1::<f32>()?, [1., 1., 0.1, 0.1]);
+
+    let y = x.minimum(&(x * 0.1)?)?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(x).context("no grad for x")?;
+    assert_eq!(y.to_vec1::<f32>()?, [0.3, 0.1, -4., -1.]);
+    assert_eq!(grad_x.to_vec1::<f32>()?, [0.1, 0.1, 1., 1.]);
+
+    // This one is easy to mess up, we want the gradient to be one as it is the identity function.
+    let y = x.minimum(x)?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(x).context("no grad for x")?;
+    assert_eq!(y.to_vec1::<f32>()?, [3., 1., -4., -1.]);
+    assert_eq!(grad_x.to_vec1::<f32>()?, [1., 1., 1., 1.]);
+
+    let x_var = Var::new(&[3f32, 1., -4., -1., 5., 9.], device)?;
+    let x = x_var.as_tensor();
+    let y_var = Var::new(&[2f32, 7., 1.], device)?;
+    let y = y_var.as_tensor();
+
+    let ss = x
+        .reshape((2, 3))?
+        .slice_scatter0(&y.reshape((1, 3))?, 1)?
+        .sqr()?;
+    let grads = ss.backward()?;
+    let grad_x = grads.get(x).context("no grad for x")?;
+    let grad_y = grads.get(y).context("no grad for y")?;
+    assert_eq!(ss.to_vec2::<f32>()?, [[9., 1., 16.], [4., 49., 1.]]);
+    assert_eq!(grad_x.to_vec1::<f32>()?, [6.0, 2.0, -8.0, 0.0, 0.0, 0.0]);
+    assert_eq!(grad_y.to_vec1::<f32>()?, [4.0, 14.0, 2.0]);
+    Ok(())
+}
+
+test_device!(
+    simple_grad,
+    simple_grad_cpu,
+    simple_grad_gpu,
+    simple_grad_metal
+);
+test_device!(sum_grad, sum_grad_cpu, sum_grad_gpu, sum_grad_metal);
+test_device!(
+    matmul_grad,
+    matmul_grad_cpu,
+    matmul_grad_gpu,
+    matmul_grad_metal
+);
+test_device!(
+    grad_descent,
+    grad_descent_cpu,
+    grad_descent_gpu,
+    grad_descent_metal
+);
+test_device!(unary_grad, unary_grad_cpu, unary_grad_gpu, unary_grad_metal);
+test_device!(
+    binary_grad,
+    binary_grad_cpu,
+    binary_grad_gpu,
+    binary_grad_metal
+);