Prebuild all our kernels.

* Use the tokenizer-output-stream in the llama example.

* Also use tokenizer-output-stream for llama2-c.

.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 +5,15 @@ 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: [single-gpu, nvidia-gpu, t4, ci]
+    container:
+      image: nvidia/cuda:12.3.1-devel-ubuntu22.04
+      options: --gpus 0 
+    if: ${{ github.event.pull_request.head.repo.full_name == github.event.pull_request.base.repo.full_name }}
     permissions:
       contents: write
       packages: write
@@ -56,32 +24,10 @@ jobs:
     steps:
       - name: Checkout repository
         uses: actions/checkout@v3
+      - name: Install dependencies
+        run: apt-get update && apt install curl build-essential libssl-dev protobuf-compiler pkg-config -y
       - name: Install Rust Stable
-        run: curl https://sh.rustup.rs -sSf | sh -s -- -y
+        uses: actions-rust-lang/setup-rust-toolchain@v1
       - uses: Swatinem/rust-cache@v2
-      - run: apt-get update -y && apt-get install libssl-dev -y
       - name: Test (cuda)
-        run: PATH=$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@v2
+
+      - 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

.gitignore

@@ -23,14 +23,16 @@ flamegraph.svg
 *.dylib
 *.so
 *.swp
+*.swo
 trace-*.json
 
 candle-wasm-examples/*/build
 candle-wasm-examples/*/*.bin
 candle-wasm-examples/*/*.jpeg
-candle-wasm-examples/*/*.wav
-candle-wasm-examples/*/*.safetensors
+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/*

.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

@@ -1,13 +1,84 @@
 # Changelog
 This documents the main changes to the `candle` crate.
 
-## v0.2.1 - Unreleased
+## 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
 

Cargo.toml

@@ -7,18 +7,19 @@ members = [
     "candle-nn",
     "candle-pyo3",
     "candle-transformers",
-    "candle-wasm-examples/llama2-c",
-    "candle-wasm-examples/whisper",
-    "candle-wasm-examples/yolo",
+    "candle-wasm-examples/*",
+    "candle-wasm-tests",
 ]
 exclude = [
-    "candle-flash-attn",
-    "candle-kernels",
+   "candle-flash-attn",
+   "candle-kernels",
+   "candle-metal-kernels",
+   "candle-onnx",
 ]
 resolver = "2"
 
 [workspace.package]
-version = "0.2.1"
+version = "0.4.0"
 edition = "2021"
 description = "Minimalist ML framework."
 repository = "https://github.com/huggingface/candle"
@@ -30,10 +31,18 @@ license = "MIT OR Apache-2.0"
 accelerate-src = { version = "0.3.2" }
 anyhow = { version = "1", features = ["backtrace"] }
 byteorder = "1.4.3"
+candle = { path = "./candle-core", package = "candle-core", version = "0.4.0" }
+candle-datasets = { path = "./candle-datasets", version = "0.4.0" }
+candle-flash-attn = { path = "./candle-flash-attn", version = "0.4.0" }
+candle-kernels = { path = "./candle-kernels", version = "0.4.0" }
+candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.4.0" }
+candle-nn = { path = "./candle-nn", version = "0.4.0" }
+candle-onnx = { path = "./candle-onnx", version = "0.4.0" }
+candle-transformers = { path = "./candle-transformers", version = "0.4.0" }
 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" }
+criterion = { version = "0.5.1", default-features=false }
+cudarc = { version = "0.10.0", features = ["f16"] }
+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"] }
@@ -41,24 +50,27 @@ imageproc = { version = "0.23.0", 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 = "50.0.0" }
 rand = "0.8.5"
 rand_distr = "0.4.3"
 rayon = "1.7.0"
 rusttype = { version = "0.9", default-features = false }
-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 = "0.15.0", default-features = false }
 tracing = "0.1.37"
 tracing-chrome = "0.7.1"
 tracing-subscriber = "0.3.7"
 wav = "1.0.0"
+yoke = { version = "0.7.2", features = ["derive"] }
 zip = { version = "0.6.6", default-features = false }
-parquet = { version = "45.0.0" }
+metal = { version = "0.27.0", features = ["mps"]}
 
 [profile.release-with-debug]
 inherits = "release"

README.md

@@ -8,7 +8,10 @@ Candle is a minimalist ML framework for Rust with a focus on performance (includ
 and ease of use. Try our online demos: 
 [whisper](https://huggingface.co/spaces/lmz/candle-whisper),
 [LLaMA2](https://huggingface.co/spaces/lmz/candle-llama2),
-[yolo](https://huggingface.co/spaces/lmz/candle-yolo).
+[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
 
@@ -45,40 +48,83 @@ 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 and LLaMA-v2](./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.
-- [Stable Diffusion](./candle-examples/examples/stable-diffusion/): text to
-  image generative model, support for the 1.5, 2.1, and SDXL 1.0 versions.
-- [DINOv2](./candle-examples/examples/dinov2/): computer vision model trained
-  using self-supervision (can be used for imagenet classification, depth
-  evaluation, segmentation).
+- [Phi-1, Phi-1.5, and Phi-2](./candle-examples/examples/phi/): 1.3b and 2.7b general LLMs with performance on par with LLaMA-v2 7b.
+- [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/): LLM specialized to code generation.
+- [Qwen1.5](./candle-examples/examples/qwen/): Bilingual (English/Chinese) LLMs.
+- [RWKV v5](./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, support for the 1.5, 2.1, SDXL 1.0 and Turbo versions.
+
+<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.
-  [segment-anything](./candle-examples/examples/segment-anything/): image
+
+<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.
-Run them using the following commands:
+
+<img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/segment-anything/assets/sam_merged.jpg" width="200">
+
+- [Whisper](./candle-examples/examples/whisper/): speech recognition 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.
+- [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.
+
+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 -- --prompt "a rusty robot holding a fire torch"
-cargo run --example dinov2 --release -- --image path/to/myinput.jpg
 cargo run --example quantized --release
-cargo run --example yolo-v3 --release -- myimage.jpg
-cargo run --example yolo-v8 --release -- myimage.jpg # for pose estimation, add --task pose 
-cargo run --example segment-anything --release -- --image myimage.jpg
 ```
 
 In order to use **CUDA** add `--features cuda` to the example command line. If
@@ -88,7 +134,10 @@ 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
 test server:
@@ -101,6 +150,28 @@ trunk serve --release --port 8081
 And then head over to
 [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.
+- [`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.
+
+If you have an addition to this list, please submit a pull request.
+
+<!--- ANCHOR_END: useful_libraries --->
+
 <!--- ANCHOR: features --->
 
 ## Features
@@ -113,10 +184,41 @@ And then head over to
     - CUDA backend for efficiently running on GPUs, multiple GPU distribution via NCCL.
     - WASM support, run your models in a browser.
 - Included models.
-    - LLMs: LLaMA v1 and v2, Falcon, StarCoder.
+    - Language Models.
+        - LLaMA v1 and v2 with variants such as SOLAR-10.7B.
+        - Falcon.
+        - StarCoder.
+        - Phi 1, 1.5, and 2.
+        - Mamba, Minimal Mamba
+        - 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.
+        - RWKV.
+    - 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).
     - Whisper (multi-lingual support).
-    - Stable Diffusion.
-    - Computer Vision: DINOv2, EfficientNet, yolo-v3, yolo-v8.
+    - Text to image.
+        - Stable Diffusion v1.5, v2.1, XL v1.0.
+        - Wurstchen v2.
+    - Image to text.
+        - BLIP.
+        - TrOCR.
+    - Computer Vision Models.
+        - DINOv2, ConvMixer, EfficientNet, ResNet, ViT, VGG, RepVGG, ConvNeXT,
+          ConvNeXTv2.
+        - yolo-v3, yolo-v8.
+        - Segment-Anything Model (SAM).
 - 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.
@@ -153,6 +255,7 @@ 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
 
@@ -257,6 +360,29 @@ This is a bug in gcc-11 triggered by the Cuda compiler. To fix this, install a d
 env CANDLE_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

candle-book/Cargo.toml

@@ -11,11 +11,11 @@ readme = "README.md"
 
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
-candle = { path = "../candle-core", version = "0.2.1", package = "candle-core" }
-candle-datasets = { path = "../candle-datasets", version = "0.2.1" }
-candle-nn = { path = "../candle-nn", version = "0.2.1" }
-candle-transformers = { path = "../candle-transformers", version = "0.2.1" }
-candle-flash-attn = { path = "../candle-flash-attn", version = "0.2.1", 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 }
@@ -24,9 +24,10 @@ 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]
-anyhow = { workspace = true }
 byteorder = { workspace = true }
 hf-hub = { workspace = true, features=["tokio"]}
 clap = { workspace = true }
@@ -38,7 +39,6 @@ tracing-chrome = { workspace = true }
 tracing-subscriber = { workspace = true }
 wav = { workspace = true }
 # Necessary to disambiguate with tokio in wasm examples which are 1.28.1
-tokio = "1.29.1"
 parquet = { workspace = true }
 image = { workspace = true }
 

candle-book/src/SUMMARY.md

@@ -10,10 +10,11 @@
 
 # Reference Guide
 
-- [Running a model](inference/README.md)
+- [Running a model](inference/inference.md)
     - [Using the hub](inference/hub.md)
 - [Error management](error_manage.md)
-- [Training](training/README.md)
+- [Training](training/training.md)
+    - [Simplified](training/simplified.md)
     - [MNIST](training/mnist.md)
     - [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)?;
@@ -146,7 +146,7 @@ And rewrite our examples using it
 ```rust
 # extern crate candle_core;
 # extern crate candle_nn;
-use candle_core::{DType, Device, Result, Tensor};
+use candle_core::{Device, Result, Tensor};
 use candle_nn::{Linear, Module};
 
 struct Model {
@@ -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

@@ -12,6 +12,9 @@ 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.

candle-book/src/lib.rs

@@ -1,3 +1,6 @@
+#[cfg(test)]
+pub mod simplified;
+
 #[cfg(test)]
 mod tests {
     use anyhow::Result;
@@ -25,6 +28,7 @@ let weights = candle::safetensors::load(weights_filename, &Device::Cpu).unwrap()
     #[rustfmt::skip]
     #[test]
     fn book_hub_2() {
+        {
 // ANCHOR: book_hub_2
 use candle::Device;
 use hf_hub::api::sync::Api;
@@ -42,9 +46,10 @@ let weights = candle::safetensors::load_buffer(&mmap[..], &Device::Cpu).unwrap()
         assert_eq!(weights.len(), 206);
     }
 
-    #[rustfmt::skip]
-    #[test]
-    fn book_hub_3() {
+    // #[rustfmt::skip]
+    // #[test]
+    // fn book_hub_3() {
+    {
 // ANCHOR: book_hub_3
 use candle::{DType, Device, Tensor};
 use hf_hub::api::sync::Api;
@@ -99,6 +104,7 @@ let tp_tensor = Tensor::from_raw_buffer(&raw, dtype, &tp_shape, &Device::Cpu).un
         assert_eq!(view.shape(), &[768, 768]);
         assert_eq!(tp_tensor.dims(), &[192, 768]);
     }
+}
 
     #[rustfmt::skip]
     #[test]

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/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-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 = "0.2.1", 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 +43,8 @@ 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

candle-core/benches/bench_main.rs

@@ -0,0 +1,9 @@
+mod benchmarks;
+
+use criterion::criterion_main;
+criterion_main!(
+    benchmarks::affine::benches,
+    benchmarks::matmul::benches,
+    benchmarks::random::benches,
+    benchmarks::where_cond::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/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 +1,66 @@
+pub(crate) mod affine;
+pub(crate) mod matmul;
+pub(crate) mod random;
+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/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/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, 1, 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/tensor-tools.rs

@@ -1,5 +1,5 @@
-use candle_core::quantized::{gguf_file, k_quants, QTensor};
-use candle_core::{Device, Result, Tensor};
+use candle_core::quantized::{gguf_file, GgmlDType, QTensor};
+use candle_core::{Device, Result};
 use clap::{Parser, Subcommand, ValueEnum};
 use rayon::prelude::*;
 
@@ -11,12 +11,7 @@ enum QuantizationMode {
 }
 
 impl QuantizationMode {
-    fn quantize(
-        &self,
-        name: &str,
-        tensor: QTensor,
-        default: fn(&Tensor) -> Result<QTensor>,
-    ) -> Result<QTensor> {
+    fn quantize(&self, name: &str, tensor: QTensor, dtype: GgmlDType) -> Result<QTensor> {
         match self {
             Self::Llama => {
                 // Same behavior as the llama.cpp quantization.
@@ -24,9 +19,9 @@ impl QuantizationMode {
                 if should_quantize {
                     let tensor = tensor.dequantize(&Device::Cpu)?;
                     if name == "output.weight" {
-                        QTensor::quantize::<k_quants::BlockQ6K>(&tensor)
+                        QTensor::quantize(&tensor, GgmlDType::Q6K)
                     } else {
-                        default(&tensor)
+                        QTensor::quantize(&tensor, dtype)
                     }
                 } else {
                     Ok(tensor)
@@ -60,6 +55,27 @@ enum Quantization {
     F32,
 }
 
+impl Quantization {
+    fn dtype(&self) -> GgmlDType {
+        match self {
+            Quantization::Q4_0 => GgmlDType::Q4_0,
+            Quantization::Q4_1 => GgmlDType::Q4_1,
+            Quantization::Q5_0 => GgmlDType::Q5_0,
+            Quantization::Q5_1 => GgmlDType::Q5_1,
+            Quantization::Q8_0 => GgmlDType::Q8_0,
+            Quantization::Q8_1 => GgmlDType::Q8_1,
+            Quantization::Q2k => GgmlDType::Q2K,
+            Quantization::Q3k => GgmlDType::Q3K,
+            Quantization::Q4k => GgmlDType::Q4K,
+            Quantization::Q5k => GgmlDType::Q5K,
+            Quantization::Q6k => GgmlDType::Q6K,
+            Quantization::Q8k => GgmlDType::Q8K,
+            Quantization::F16 => GgmlDType::F16,
+            Quantization::F32 => GgmlDType::F32,
+        }
+    }
+}
+
 #[derive(ValueEnum, Debug, Clone)]
 enum Format {
     Safetensors,
@@ -102,9 +118,11 @@ enum Command {
     },
 
     Quantize {
-        /// The input file, in gguf format.
-        in_file: std::path::PathBuf,
+        /// The input file(s), in safetensors format.
+        in_file: Vec<std::path::PathBuf>,
+
         /// The output file, in gguf format.
+        #[arg(long)]
         out_file: std::path::PathBuf,
 
         /// The quantization schema to apply.
@@ -115,6 +133,15 @@ enum Command {
         #[arg(long, value_enum, default_value_t = QuantizationMode::Llama)]
         mode: QuantizationMode,
     },
+
+    Dequantize {
+        /// The input file, in gguf format.
+        in_file: std::path::PathBuf,
+
+        /// The output file, in safetensors format.
+        #[arg(long)]
+        out_file: std::path::PathBuf,
+    },
 }
 
 #[derive(Parser, Debug, Clone)]
@@ -123,7 +150,12 @@ struct Args {
     command: Command,
 }
 
-fn run_ls(file: &std::path::PathBuf, format: Option<Format>, verbose: bool) -> Result<()> {
+fn run_ls(
+    file: &std::path::PathBuf,
+    format: Option<Format>,
+    verbose: bool,
+    device: &Device,
+) -> Result<()> {
     let format = match format {
         Some(format) => format,
         None => match Format::infer(file) {
@@ -150,8 +182,7 @@ fn run_ls(file: &std::path::PathBuf, format: Option<Format>, verbose: bool) -> R
             }
         }
         Format::Safetensors => {
-            let tensors = unsafe { candle_core::safetensors::MmapedFile::new(file)? };
-            let tensors = tensors.deserialize()?;
+            let tensors = unsafe { candle_core::safetensors::MmapedSafetensors::new(file)? };
             let mut tensors = tensors.tensors();
             tensors.sort_by(|a, b| a.0.cmp(&b.0));
             for (name, view) in tensors.iter() {
@@ -165,7 +196,7 @@ fn run_ls(file: &std::path::PathBuf, format: Option<Format>, verbose: bool) -> R
             }
         }
         Format::Pth => {
-            let mut tensors = candle_core::pickle::read_pth_tensor_info(file, verbose)?;
+            let mut tensors = candle_core::pickle::read_pth_tensor_info(file, verbose, None)?;
             tensors.sort_by(|a, b| a.name.cmp(&b.name));
             for tensor_info in tensors.iter() {
                 println!(
@@ -190,7 +221,7 @@ fn run_ls(file: &std::path::PathBuf, format: Option<Format>, verbose: bool) -> R
         }
         Format::Ggml => {
             let mut file = std::fs::File::open(file)?;
-            let content = candle_core::quantized::ggml_file::Content::read(&mut file)?;
+            let content = candle_core::quantized::ggml_file::Content::read(&mut file, device)?;
             let mut tensors = content.tensors.into_iter().collect::<Vec<_>>();
             tensors.sort_by(|a, b| a.0.cmp(&b.0));
             for (name, qtensor) in tensors.iter() {
@@ -218,43 +249,100 @@ fn run_ls(file: &std::path::PathBuf, format: Option<Format>, verbose: bool) -> R
     Ok(())
 }
 
-fn run_quantize(
+fn run_quantize_safetensors(
+    in_files: &[std::path::PathBuf],
+    out_file: std::path::PathBuf,
+    q: Quantization,
+) -> Result<()> {
+    let mut out_file = std::fs::File::create(out_file)?;
+    let mut tensors = std::collections::HashMap::new();
+    for in_file in in_files.iter() {
+        let in_tensors = candle_core::safetensors::load(in_file, &Device::Cpu)?;
+        tensors.extend(in_tensors)
+    }
+    println!("tensors: {}", tensors.len());
+
+    let dtype = q.dtype();
+    let block_size = dtype.block_size();
+
+    let qtensors = tensors
+        .into_par_iter()
+        .map(|(name, tensor)| {
+            let should_quantize = tensor.rank() == 2 && tensor.dim(1)? % block_size == 0;
+            println!("  quantizing {name} {tensor:?} {should_quantize}");
+            let tensor = if should_quantize {
+                QTensor::quantize(&tensor, dtype)?
+            } else {
+                QTensor::quantize(&tensor, GgmlDType::F32)?
+            };
+            Ok((name, tensor))
+        })
+        .collect::<Result<Vec<_>>>()?;
+    let qtensors = qtensors
+        .iter()
+        .map(|(k, v)| (k.as_str(), v))
+        .collect::<Vec<_>>();
+    gguf_file::write(&mut out_file, &[], &qtensors)?;
+    Ok(())
+}
+
+fn run_dequantize(
     in_file: std::path::PathBuf,
     out_file: std::path::PathBuf,
+    device: &Device,
+) -> Result<()> {
+    let mut in_file = std::fs::File::open(in_file)?;
+    let content = gguf_file::Content::read(&mut in_file)?;
+    let mut tensors = std::collections::HashMap::new();
+    for (tensor_name, _) in content.tensor_infos.iter() {
+        let tensor = content.tensor(&mut in_file, tensor_name, device)?;
+        let tensor = tensor.dequantize(device)?;
+        tensors.insert(tensor_name.to_string(), tensor);
+    }
+    candle_core::safetensors::save(&tensors, out_file)?;
+    Ok(())
+}
+
+fn run_quantize(
+    in_files: &[std::path::PathBuf],
+    out_file: std::path::PathBuf,
     q: Quantization,
     qmode: QuantizationMode,
+    device: &Device,
 ) -> Result<()> {
+    if in_files.is_empty() {
+        candle_core::bail!("no specified input files")
+    }
+    if let Some(extension) = out_file.extension() {
+        if extension == "safetensors" {
+            candle_core::bail!("the generated file cannot use the safetensors extension")
+        }
+    }
+    if let Some(extension) = in_files[0].extension() {
+        if extension == "safetensors" {
+            return run_quantize_safetensors(in_files, out_file, q);
+        }
+    }
+
+    if in_files.len() != 1 {
+        candle_core::bail!("only a single in-file can be used when quantizing gguf files")
+    }
+
     // Open the out file early so as to fail directly on missing directories etc.
     let mut out_file = std::fs::File::create(out_file)?;
-    let mut in_ = std::fs::File::open(&in_file)?;
+    let mut in_ = std::fs::File::open(&in_files[0])?;
     let content = gguf_file::Content::read(&mut in_)?;
     println!("tensors: {}", content.tensor_infos.len());
 
-    let quantize_fn = match q {
-        Quantization::Q4_0 => QTensor::quantize::<k_quants::BlockQ4_0>,
-        Quantization::Q4_1 => QTensor::quantize::<k_quants::BlockQ4_1>,
-        Quantization::Q5_0 => QTensor::quantize::<k_quants::BlockQ5_0>,
-        Quantization::Q5_1 => QTensor::quantize::<k_quants::BlockQ5_1>,
-        Quantization::Q8_0 => QTensor::quantize::<k_quants::BlockQ8_0>,
-        Quantization::Q8_1 => QTensor::quantize::<k_quants::BlockQ8_1>,
-        Quantization::Q2k => QTensor::quantize::<k_quants::BlockQ2K>,
-        Quantization::Q3k => QTensor::quantize::<k_quants::BlockQ3K>,
-        Quantization::Q4k => QTensor::quantize::<k_quants::BlockQ4K>,
-        Quantization::Q5k => QTensor::quantize::<k_quants::BlockQ5K>,
-        Quantization::Q6k => QTensor::quantize::<k_quants::BlockQ6K>,
-        Quantization::Q8k => QTensor::quantize::<k_quants::BlockQ8K>,
-        Quantization::F16 => QTensor::quantize::<half::f16>,
-        Quantization::F32 => QTensor::quantize::<f32>,
-    };
-
+    let dtype = q.dtype();
     let qtensors = content
         .tensor_infos
         .par_iter()
         .map(|(name, _)| {
             println!("  quantizing {name}");
-            let mut in_file = std::fs::File::open(&in_file)?;
-            let tensor = content.tensor(&mut in_file, name)?;
-            let tensor = qmode.quantize(name, tensor, quantize_fn)?;
+            let mut in_file = std::fs::File::open(&in_files[0])?;
+            let tensor = content.tensor(&mut in_file, name, device)?;
+            let tensor = qmode.quantize(name, tensor, dtype)?;
             Ok((name, tensor))
         })
         .collect::<Result<Vec<_>>>()?;
@@ -274,6 +362,7 @@ fn run_quantize(
 
 fn main() -> anyhow::Result<()> {
     let args = Args::parse();
+    let device = Device::Cpu;
     match args.command {
         Command::Ls {
             files,
@@ -285,7 +374,7 @@ fn main() -> anyhow::Result<()> {
                 if multiple_files {
                     println!("--- {file:?} ---");
                 }
-                run_ls(file, format.clone(), verbose)?
+                run_ls(file, format.clone(), verbose, &device)?
             }
         }
         Command::Quantize {
@@ -293,7 +382,8 @@ fn main() -> anyhow::Result<()> {
             out_file,
             quantization,
             mode,
-        } => run_quantize(in_file, out_file, quantization, mode)?,
+        } => run_quantize(&in_file, out_file, quantization, mode, &device)?,
+        Command::Dequantize { in_file, out_file } => run_dequantize(in_file, out_file, &device)?,
     }
     Ok(())
 }

candle-core/src/accelerate.rs

@@ -380,6 +380,16 @@ 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()) {
@@ -402,6 +412,28 @@ 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, $accelerate_name:ident) => {
         #[inline]

candle-core/src/backend.rs

@@ -39,6 +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,
@@ -57,6 +65,7 @@ pub trait BackendStorage: Sized {
 
     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>;
@@ -110,4 +119,6 @@ pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
     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<()>;
 }

candle-core/src/backprop.rs

@@ -15,6 +15,17 @@ 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 +47,8 @@ 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, _)
@@ -55,6 +68,11 @@ impl Tensor {
                         kernel: rhs,
                         ..
                     }
+                    | Op::ConvTranspose1D {
+                        arg: lhs,
+                        kernel: rhs,
+                        ..
+                    }
                     | Op::Conv2D {
                         arg: lhs,
                         kernel: rhs,
@@ -69,7 +87,8 @@ impl Tensor {
                     | 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);
@@ -90,15 +109,18 @@ impl Tensor {
                             nodes
                         }
                     }
+                    Op::Unary(_node, UnaryOp::Ceil)
+                    | Op::Unary(_node, UnaryOp::Floor)
+                    | Op::Unary(_node, UnaryOp::Round) => nodes,
                     Op::Reshape(node)
-                    | Op::UpsampleNearest2D(node)
+                    | Op::UpsampleNearest1D(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, _)
@@ -111,6 +133,16 @@ impl Tensor {
                         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
@@ -134,10 +166,16 @@ 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) => {
@@ -192,7 +230,44 @@ 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::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,
+                        )?;
+                        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,
@@ -213,7 +288,6 @@ impl Tensor {
                             out_padding,
                             *stride,
                             *dilation,
-                            1,
                         )?;
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&grad_arg)?;
@@ -223,8 +297,18 @@ impl Tensor {
                             .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 { .. } => Err(Error::BackwardNotSupported {
                         op: "conv-transpose2d",
                     })?,
@@ -263,9 +347,39 @@ impl Tensor {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&grad_arg)?;
                     }
-                    Op::UpsampleNearest2D { .. } => Err(Error::BackwardNotSupported {
-                        op: "upsample-nearest2d",
+                    Op::UpsampleNearest1D { .. } => Err(Error::BackwardNotSupported {
+                        op: "upsample-nearest1d",
                     })?,
+                    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)?;
@@ -357,7 +471,7 @@ 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)?;
@@ -437,13 +551,61 @@ impl Tensor {
                         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(_, UnaryOp::Floor) => {
+                        Err(Error::BackwardNotSupported { op: "floor" })?
+                    }
+                    Op::Unary(_, UnaryOp::Round) => {
+                        Err(Error::BackwardNotSupported { op: "round" })?
+                    }
+                    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)))
+                        let sigmoid_arg = (*node / arg)?;
+                        let silu_grad = (&sigmoid_arg * (1. + (arg * (1. - &sigmoid_arg)?)?)?)?;
+                        *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())?;
+                        let negative_exp_mask = ((negative_mask * arg.exp())? * *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)?;
@@ -518,6 +680,7 @@ impl Tensor {
     }
 }
 
+#[derive(Debug)]
 pub struct GradStore(HashMap<TensorId, Tensor>);
 
 impl GradStore {

candle-core/src/conv.rs

@@ -25,6 +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,
@@ -37,6 +77,7 @@ pub struct ParamsConv2D {
     pub(crate) padding: usize,
     pub(crate) stride: usize,
     pub(crate) dilation: usize,
+    pub cudnn_fwd_algo: Option<CudnnFwdAlgo>,
 }
 
 impl ParamsConv2D {
@@ -60,13 +101,12 @@ pub struct ParamsConvTranspose2D {
     pub(crate) i_w: usize,
     pub(crate) k_h: usize,
     pub(crate) k_w: usize,
-    pub(crate) c_out_per_group: 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,
-    pub(crate) groups: usize,
 }
 
 impl ParamsConvTranspose2D {
@@ -81,8 +121,7 @@ impl ParamsConvTranspose2D {
     }
 
     pub(crate) fn out_dims(&self) -> Vec<usize> {
-        let c_out = self.c_out_per_group * self.groups;
-        vec![self.b_size, c_out, self.out_h(), self.out_w()]
+        vec![self.b_size, self.c_out, self.out_h(), self.out_w()]
     }
 }
 
@@ -148,6 +187,49 @@ impl Tensor {
         }
     }
 
+    /// 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,
+    ) -> Result<Self> {
+        let (b_size, c_in, l_in) = self.dims3()?;
+        let (c_in_k, c_out, k_size) = kernel.dims3()?;
+        if c_in != c_in_k {
+            crate::bail!("in_channel mismatch between input ({c_in}) and kernel ({c_in_k})")
+        }
+        let params = ParamsConvTranspose1D {
+            b_size,
+            l_in,
+            k_size,
+            c_out,
+            c_in,
+            padding,
+            output_padding,
+            stride,
+            dilation,
+        };
+        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))
+    }
+
     fn conv2d_single_group(&self, kernel: &Self, params: &ParamsConv2D) -> Result<Self> {
         let storage =
             self.storage()
@@ -190,6 +272,7 @@ impl Tensor {
             padding,
             stride,
             dilation,
+            cudnn_fwd_algo: None,
         };
         if groups == 1 {
             self.conv2d_single_group(kernel, &params)
@@ -213,29 +296,24 @@ impl Tensor {
         output_padding: usize,
         stride: usize,
         dilation: usize,
-        groups: usize,
     ) -> Result<Self> {
         let (b_size, c_in, i_h, i_w) = self.dims4()?;
-        let (c_in_k, c_out_per_group, k_h, k_w) = kernel.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})")
         }
-        if c_in % groups != 0 {
-            crate::bail!("in_channel {c_in} must be divisible by groups {groups}")
-        }
         let params = ParamsConvTranspose2D {
             b_size,
             i_h,
             i_w,
             k_h,
             k_w,
-            c_out_per_group,
+            c_out,
             c_in,
             padding,
             output_padding,
             stride,
             dilation,
-            groups,
         };
         let storage = self.storage().conv_transpose2d(
             self.layout(),
@@ -250,7 +328,6 @@ impl Tensor {
             output_padding: params.output_padding,
             stride: params.stride,
             dilation: params.dilation,
-            groups: params.groups,
         });
         let out_dims = params.out_dims();
         Ok(crate::tensor::from_storage(storage, out_dims, op, false))

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/mod.rs

@@ -1,3 +1,4 @@
+pub mod erf;
 pub mod kernels;
 
 trait Cpu<const ARR: usize> {

candle-core/src/cpu_backend.rs

@@ -4,6 +4,9 @@ use crate::{DType, Error, IntDType, Layout, Result, Shape, WithDType};
 use half::{bf16, f16};
 use rayon::prelude::*;
 
+const USE_IM2COL_CONV1D: bool = true;
+const USE_IM2COL_CONV2D: bool = true;
+
 // TODO: Maybe we should not implement [Clone] here and instead have an explicit allocator +
 // intercept the oom errors to avoid panicking and provide a proper error.
 #[derive(Debug, Clone)]
@@ -724,6 +727,36 @@ impl Map1 for MaxPool2D {
     }
 }
 
+struct UpsampleNearest1D(usize);
+
+impl Map1 for UpsampleNearest1D {
+    fn f<T: WithDType>(&self, src: &[T], layout: &Layout) -> Result<Vec<T>> {
+        // TODO: Specialized implementation for the case 2*sz?
+        let dst_sz = self.0;
+        let (b_sz, c, src_sz) = layout.shape().dims3()?;
+        let stride = layout.stride();
+        let stride_sz = stride[2];
+        let src_index = layout.start_offset();
+        let scale_sz = src_sz as f64 / dst_sz as f64;
+        let mut dst = vec![T::zero(); b_sz * c * dst_sz];
+        let src_idxs = (0..dst_sz)
+            .map(|idx| usize::min(src_sz - 1, (idx as f64 * scale_sz) as usize))
+            .collect::<Vec<_>>();
+        for b_idx in 0..b_sz {
+            let dst = &mut dst[b_idx * c * dst_sz..];
+            let src_index = src_index + b_idx * stride[0];
+            for c_idx in 0..c {
+                let dst = &mut dst[c_idx * dst_sz..];
+                let src_index = src_index + c_idx * stride[1];
+                for (idx, src_idx) in src_idxs.iter().enumerate() {
+                    dst[idx] = src[src_index + src_idx * stride_sz]
+                }
+            }
+        }
+        Ok(dst)
+    }
+}
+
 struct UpsampleNearest2D(usize, usize);
 
 impl Map1 for UpsampleNearest2D {
@@ -771,11 +804,11 @@ impl<'a, I: IntDType> Map1 for Gather<'a, I> {
     fn f<T: WithDType>(&self, src: &[T], src_l: &Layout) -> Result<Vec<T>> {
         let ids = match self.ids_l.contiguous_offsets() {
             Some((a, b)) => &self.ids[a..b],
-            None => Err(Error::RequiresContiguous { op: "gather" })?,
+            None => Err(Error::RequiresContiguous { op: "gather" }.bt())?,
         };
         let src = match src_l.contiguous_offsets() {
             Some((a, b)) => &src[a..b],
-            None => Err(Error::RequiresContiguous { op: "gather" })?,
+            None => Err(Error::RequiresContiguous { op: "gather" }.bt())?,
         };
         let dim = self.dim;
         let ids_dims = self.ids_l.dims();
@@ -824,7 +857,7 @@ impl<'a, I: IntDType> Map1 for IndexSelect<'a, I> {
     fn f<T: WithDType>(&self, src: &[T], layout: &Layout) -> Result<Vec<T>> {
         let src = match layout.contiguous_offsets() {
             Some((a, b)) => &src[a..b],
-            None => Err(Error::RequiresContiguous { op: "index-select" })?,
+            None => Err(Error::RequiresContiguous { op: "index-select" }.bt())?,
         };
         let dim = self.dim;
         let n_ids = match self.ids_l.dims() {
@@ -880,7 +913,7 @@ impl<'a, I: IntDType> Map2 for ScatterAdd<'a, I> {
         let mut dst = vec![T::zero(); dst_len];
         copy_strided_src_(v1, &mut dst, 0, l1);
         let src = match src_l.contiguous_offsets() {
-            None => Err(Error::RequiresContiguous { op: "scatter-add" })?,
+            None => Err(Error::RequiresContiguous { op: "scatter-add" }.bt())?,
             Some((o1, o2)) => &src[o1..o2],
         };
 
@@ -896,7 +929,7 @@ impl<'a, I: IntDType> Map2 for ScatterAdd<'a, I> {
 
         let ids = match self.ids_l.contiguous_offsets() {
             Some((a, b)) => &self.ids[a..b],
-            None => Err(Error::RequiresContiguous { op: "gather" })?,
+            None => Err(Error::RequiresContiguous { op: "gather" }.bt())?,
         };
         for left_i in 0..ids_left_len {
             let start_ids_idx = left_i * ids_right_len * ids_dim_len;
@@ -938,7 +971,7 @@ impl<'a, I: IntDType> Map2 for IndexAdd<'a, I> {
         let mut dst = vec![T::zero(); dst_len];
         copy_strided_src_(v1, &mut dst, 0, l1);
         let src = match src_l.contiguous_offsets() {
-            None => Err(Error::RequiresContiguous { op: "index-add" })?,
+            None => Err(Error::RequiresContiguous { op: "index-add" }.bt())?,
             Some((o1, o2)) => &src[o1..o2],
         };
         let dim = self.dim;
@@ -1089,6 +1122,208 @@ impl<'a> Map2 for Conv1D<'a> {
     }
 }
 
+struct Im2Col1D {
+    l_k: usize,
+    stride: usize,
+    dilation: usize,
+    padding: usize,
+}
+
+impl Im2Col1D {
+    fn l_out(&self, l: usize) -> usize {
+        (l + 2 * self.padding - self.dilation * (self.l_k - 1) - 1) / self.stride + 1
+    }
+}
+
+impl Map1 for Im2Col1D {
+    fn f<T: WithDType>(&self, vs: &[T], layout: &Layout) -> Result<Vec<T>> {
+        let &Self {
+            l_k,
+            stride,
+            dilation,
+            padding,
+        } = self;
+        let (b, c, l) = layout.shape().dims3()?;
+        let l_out = self.l_out(l);
+        let src = &vs[layout.start_offset()..];
+        let mut dst = vec![T::zero(); b * l_out * c * l_k];
+        let (src_s0, src_s1, src_s2) = {
+            let s = layout.stride();
+            (s[0], s[1], s[2])
+        };
+        // TODO: provide specialized kernels for the common use cases.
+        // - l_k = 1
+        // - padding = 0
+        // - stride = 1
+        // - dilation = 1
+        for b_idx in 0..b {
+            let src_idx = b_idx * src_s0;
+            let dst_idx = b_idx * l_out * c * l_k;
+            for l_idx in 0..l_out {
+                let dst_idx = dst_idx + l_idx * c * l_k;
+                for c_idx in 0..c {
+                    let dst_idx = dst_idx + c_idx * l_k;
+                    let src_idx = c_idx * src_s1 + src_idx;
+                    for l_k_idx in 0..l_k {
+                        let src_l = l_idx * stride + l_k_idx * dilation;
+                        if padding != 0 && (src_l < padding || src_l >= l + padding) {
+                            continue;
+                        }
+                        let src_l = src_l - padding;
+                        let src_idx = src_idx + src_l * src_s2;
+                        let dst_idx = dst_idx + l_k_idx;
+                        dst[dst_idx] = src[src_idx]
+                    }
+                }
+            }
+        }
+        Ok(dst)
+    }
+}
+
+struct Im2Col {
+    h_k: usize,
+    w_k: usize,
+    stride: usize,
+    dilation: usize,
+    padding: usize,
+}
+
+impl Im2Col {
+    fn hw_out(&self, h: usize, w: usize) -> (usize, usize) {
+        let h_out = (h + 2 * self.padding - self.dilation * (self.h_k - 1) - 1) / self.stride + 1;
+        let w_out = (w + 2 * self.padding - self.dilation * (self.w_k - 1) - 1) / self.stride + 1;
+        (h_out, w_out)
+    }
+}
+
+impl Map1 for Im2Col {
+    fn f<T: WithDType>(&self, vs: &[T], layout: &Layout) -> Result<Vec<T>> {
+        let &Self {
+            h_k,
+            w_k,
+            stride,
+            dilation,
+            padding,
+        } = self;
+        let (b, c, h, w) = layout.shape().dims4()?;
+        let (h_out, w_out) = self.hw_out(h, w);
+        let src = &vs[layout.start_offset()..];
+        let mut dst = vec![T::zero(); b * h_out * w_out * c * h_k * w_k];
+        let (src_s0, src_s1, src_s2, src_s3) = {
+            let s = layout.stride();
+            (s[0], s[1], s[2], s[3])
+        };
+        // TODO: provide specialized kernels for the common use cases.
+        // - h_k = w_k = 1
+        // - padding = 0
+        // - stride = 1
+        // - dilation = 1
+        for b_idx in 0..b {
+            let src_idx = b_idx * src_s0;
+            let dst_idx = b_idx * h_out * w_out * c * h_k * w_k;
+            for h_idx in 0..h_out {
+                let dst_idx = dst_idx + h_idx * w_out * c * h_k * w_k;
+                for w_idx in 0..w_out {
+                    let dst_idx = dst_idx + w_idx * c * h_k * w_k;
+                    for c_idx in 0..c {
+                        let dst_idx = dst_idx + c_idx * h_k * w_k;
+                        let src_idx = c_idx * src_s1 + src_idx;
+                        for h_k_idx in 0..h_k {
+                            let src_h = h_idx * stride + h_k_idx * dilation;
+                            if padding != 0 && (src_h < padding || src_h >= h + padding) {
+                                continue;
+                            }
+                            let src_h = src_h - padding;
+                            let src_idx = src_idx + src_h * src_s2;
+                            let dst_idx = dst_idx + h_k_idx * w_k;
+                            for w_k_idx in 0..w_k {
+                                let src_w = w_idx * stride + w_k_idx * dilation;
+                                if padding != 0 && (src_w < padding || src_w >= w + padding) {
+                                    continue;
+                                }
+                                let src_w = src_w - padding;
+                                let src_idx = src_idx + src_w * src_s3;
+                                let dst_idx = dst_idx + w_k_idx;
+                                dst[dst_idx] = src[src_idx]
+                            }
+                        }
+                    }
+                }
+            }
+        }
+        Ok(dst)
+    }
+}
+
+struct ConvTranspose1D<'a>(&'a crate::conv::ParamsConvTranspose1D);
+
+impl<'a> Map2 for ConvTranspose1D<'a> {
+    const OP: &'static str = "conv_transpose1d";
+    fn f<T: WithDType>(&self, inp: &[T], inp_l: &Layout, k: &[T], k_l: &Layout) -> Result<Vec<T>> {
+        let p = self.0;
+        let inp = &inp[inp_l.start_offset()..];
+        let (inp_s0, inp_s1, inp_s2) = crate::shape::dims3(inp_l.stride())?;
+        let (k_s0, k_s1, k_s2) = crate::shape::dims3(k_l.stride())?;
+        let l_out = p.l_out();
+
+        // Output shape: [b_size, c_out, l_out].
+        let dst_elems = p.c_out * l_out * p.b_size;
+        let dst = vec![T::zero(); dst_elems];
+        let dst_s0 = p.c_out * l_out;
+        let dst_s1 = l_out;
+        let dst_s2 = 1;
+
+        // TODO: Avoid making this copy if `inp` already has the appropriate layout.
+        let mut inp_cont = vec![T::zero(); p.b_size * p.c_in * p.l_in];
+        let cont_s0 = p.l_in * p.c_in;
+        let cont_s1 = p.c_in;
+        for b_idx in 0..p.b_size {
+            for l_idx in 0..p.l_in {
+                for c_idx in 0..p.c_in {
+                    let src_idx = b_idx * inp_s0 + c_idx * inp_s1 + l_idx * inp_s2;
+                    let dst_idx = b_idx * cont_s0 + l_idx * cont_s1 + c_idx;
+                    inp_cont[dst_idx] = inp[src_idx]
+                }
+            }
+        }
+
+        for k_idx in 0..p.k_size {
+            (0..p.c_out).into_par_iter().for_each(|dst_c_idx| {
+                let k_cont = (0..p.c_in)
+                    .map(|c_in_idx| k[c_in_idx * k_s0 + dst_c_idx * k_s1 + k_idx * k_s2])
+                    .collect::<Vec<_>>();
+                for b_idx in 0..p.b_size {
+                    for l_idx in 0..p.l_in {
+                        let out_idx = l_idx * p.stride + k_idx * p.dilation;
+                        if out_idx < p.padding {
+                            continue;
+                        }
+                        let out_idx = out_idx - p.padding;
+                        if out_idx < l_out {
+                            let inp_cont = &inp_cont[b_idx * cont_s0 + l_idx * cont_s1..];
+                            let dst_idx = b_idx * dst_s0 + out_idx * dst_s2 + dst_c_idx * dst_s1;
+                            let mut d = T::zero();
+                            unsafe {
+                                T::vec_dot(inp_cont.as_ptr(), k_cont.as_ptr(), &mut d, p.c_in)
+                            }
+                            let dst_p = dst.as_ptr();
+                            // Safety: dst_idx are uniques per dst_c_idx which is used to
+                            // parallelise the different tasks so no two threads can try to
+                            // write at the same location.
+                            unsafe {
+                                let ptr = dst_p.add(dst_idx) as *mut T;
+                                *ptr += d
+                            }
+                        }
+                    }
+                }
+            })
+        }
+        Ok(dst)
+    }
+}
+
 struct Conv2D<'a>(&'a crate::conv::ParamsConv2D);
 
 impl<'a> Map2 for Conv2D<'a> {
@@ -1190,9 +1425,8 @@ impl<'a> Map2 for ConvTranspose2D<'a> {
         let (out_h, out_w) = (p.out_h(), p.out_w());
 
         // Output shape: [b_size, c_out, out_h, out_w].
-        let c_out = p.groups * p.c_out_per_group;
-        let dst = vec![T::zero(); p.b_size * c_out * out_h * out_w];
-        let dst_s0 = c_out * out_h * out_w;
+        let dst = vec![T::zero(); p.b_size * p.c_out * out_h * out_w];
+        let dst_s0 = p.c_out * out_h * out_w;
         let dst_s1 = out_h * out_w;
         let dst_s2 = out_w;
         let dst_s3 = 1;
@@ -1215,16 +1449,12 @@ impl<'a> Map2 for ConvTranspose2D<'a> {
             }
         }
 
-        let c_in_per_group = p.c_in / p.groups;
         for k_y in 0..p.k_h {
             for k_x in 0..p.k_w {
-                (0..c_out).into_par_iter().for_each(|dst_c_idx| {
-                    let (group_idx, dst_c_idx_in_group) =
-                        (dst_c_idx / p.c_out_per_group, dst_c_idx % p.c_out_per_group);
-                    let k_cont = (0..c_in_per_group)
+                (0..p.c_out).into_par_iter().for_each(|dst_c_idx| {
+                    let k_cont = (0..p.c_in)
                         .map(|c_in_idx| {
-                            let c_in_idx = group_idx * c_in_per_group + c_in_idx;
-                            k[c_in_idx * k_s0 + dst_c_idx_in_group * k_s1 + k_y * k_s2 + k_x * k_s3]
+                            k[c_in_idx * k_s0 + dst_c_idx * k_s1 + k_y * k_s2 + k_x * k_s3]
                         })
                         .collect::<Vec<_>>();
                     for b_idx in 0..p.b_size {
@@ -1250,7 +1480,7 @@ impl<'a> Map2 for ConvTranspose2D<'a> {
                                             inp_cont.as_ptr(),
                                             k_cont.as_ptr(),
                                             &mut d,
-                                            c_in_per_group,
+                                            p.c_in,
                                         )
                                     }
                                     let dst_p = dst.as_ptr();
@@ -1299,8 +1529,9 @@ impl Map2 for MatMul {
     ) -> Result<Vec<T>> {
         use gemm::{gemm, Parallelism};
 
-        if T::DTYPE == DType::BF16 {
-            return Err(Error::UnsupportedDTypeForOp(T::DTYPE, "matmul").bt())?;
+        match T::DTYPE {
+            DType::F16 | DType::F32 | DType::F64 => {}
+            _ => Err(Error::UnsupportedDTypeForOp(T::DTYPE, "matmul").bt())?,
         }
 
         let (b, m, n, k) = self.0;
@@ -2004,6 +2235,10 @@ impl BackendStorage for CpuStorage {
         MaxPool2D(kernel_size, stride).map(self, layout)
     }
 
+    fn upsample_nearest1d(&self, layout: &Layout, sz: usize) -> Result<Self> {
+        UpsampleNearest1D(sz).map(self, layout)
+    }
+
     fn upsample_nearest2d(&self, layout: &Layout, h: usize, w: usize) -> Result<Self> {
         UpsampleNearest2D(h, w).map(self, layout)
     }
@@ -2232,7 +2467,50 @@ impl BackendStorage for CpuStorage {
         kernel_l: &Layout,
         params: &crate::conv::ParamsConv1D,
     ) -> Result<Self> {
-        Conv1D(params).map(self, l, kernel, kernel_l)
+        if !USE_IM2COL_CONV1D {
+            return Conv1D(params).map(self, l, kernel, kernel_l);
+        }
+        let op = Im2Col1D {
+            l_k: params.k_size,
+            padding: params.padding,
+            stride: params.stride,
+            dilation: params.dilation,
+        };
+        let col = op.map(self, l)?;
+        let b = params.b_size;
+        let n = params.c_out;
+        let l_out = params.l_out();
+        let k = op.l_k * params.c_in;
+        let m = l_out;
+        let col_l = Layout::contiguous((b, m, k));
+        let res = if kernel_l.is_contiguous() {
+            let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
+                .transpose(1, 2)?
+                .broadcast_as((b, k, n))?;
+            col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
+        } else {
+            // Make the kernel contiguous if not already the case.
+            let mut kernel_c = self.device().zeros_impl(kernel_l.shape(), kernel.dtype())?;
+            kernel.copy_strided_src(&mut kernel_c, 0, kernel_l)?;
+            let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
+                .transpose(1, 2)?
+                .broadcast_as((b, k, n))?;
+            col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
+        };
+        let res_l = Layout::contiguous((b, l_out, params.c_out)).transpose(1, 2)?;
+        let mut res_t = self.device().zeros_impl(res_l.shape(), res.dtype())?;
+        res.copy_strided_src(&mut res_t, 0, &res_l)?;
+        Ok(res_t)
+    }
+
+    fn conv_transpose1d(
+        &self,
+        l: &Layout,
+        kernel: &Self,
+        kernel_l: &Layout,
+        params: &crate::conv::ParamsConvTranspose1D,
+    ) -> Result<Self> {
+        ConvTranspose1D(params).map(self, l, kernel, kernel_l)
     }
 
     fn conv2d(
@@ -2242,7 +2520,43 @@ impl BackendStorage for CpuStorage {
         kernel_l: &Layout,
         params: &crate::conv::ParamsConv2D,
     ) -> Result<Self> {
-        Conv2D(params).map(self, l, kernel, kernel_l)
+        if !USE_IM2COL_CONV2D {
+            return Conv2D(params).map(self, l, kernel, kernel_l);
+        }
+        let op = Im2Col {
+            h_k: params.k_h,
+            w_k: params.k_w,
+            padding: params.padding,
+            stride: params.stride,
+            dilation: params.dilation,
+        };
+        let col = op.map(self, l)?;
+        let b = params.b_size;
+        let n = params.c_out;
+        let (h_out, w_out) = (params.out_h(), params.out_w());
+        let k = op.h_k * op.w_k * params.c_in;
+        let m = h_out * w_out;
+        let col_l = Layout::contiguous((b, m, k));
+        let res = if kernel_l.is_contiguous() {
+            let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
+                .transpose(1, 2)?
+                .broadcast_as((b, k, n))?;
+            col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
+        } else {
+            // Make the kernel contiguous if not already the case.
+            let mut kernel_c = self.device().zeros_impl(kernel_l.shape(), kernel.dtype())?;
+            kernel.copy_strided_src(&mut kernel_c, 0, kernel_l)?;
+            let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
+                .transpose(1, 2)?
+                .broadcast_as((b, k, n))?;
+            col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
+        };
+        let res_l = Layout::contiguous((b, h_out, w_out, params.c_out))
+            .transpose(1, 2)?
+            .transpose(1, 3)?;
+        let mut res_t = self.device().zeros_impl(res_l.shape(), res.dtype())?;
+        res.copy_strided_src(&mut res_t, 0, &res_l)?;
+        Ok(res_t)
     }
 
     fn conv_transpose2d(
@@ -2303,25 +2617,25 @@ impl BackendStorage for CpuStorage {
             Self::U8(ids) => {
                 let ids = match ids_l.contiguous_offsets() {
                     Some((a, b)) => &ids[a..b],
-                    None => Err(Error::RequiresContiguous { op: "index-add" })?,
+                    None => Err(Error::RequiresContiguous { op: "index-add" }.bt())?,
                 };
                 IndexAdd { ids, dim }.map(self, l, src, src_l)
             }
             Self::U32(ids) => {
                 let ids = match ids_l.contiguous_offsets() {
                     Some((a, b)) => &ids[a..b],
-                    None => Err(Error::RequiresContiguous { op: "index-add" })?,
+                    None => Err(Error::RequiresContiguous { op: "index-add" }.bt())?,
                 };
                 IndexAdd { ids, dim }.map(self, l, src, src_l)
             }
             Self::I64(ids) => {
                 let ids = match ids_l.contiguous_offsets() {
                     Some((a, b)) => &ids[a..b],
-                    None => Err(Error::RequiresContiguous { op: "index-add" })?,
+                    None => Err(Error::RequiresContiguous { op: "index-add" }.bt())?,
                 };
                 IndexAdd { ids, dim }.map(self, l, src, src_l)
             }
-            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "index-add")),
+            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "index-add").bt()),
         }
     }
 
@@ -2367,6 +2681,10 @@ impl BackendDevice for CpuDevice {
         Ok(Self)
     }
 
+    fn set_seed(&self, _seed: u64) -> Result<()> {
+        crate::bail!("cannot seed the CPU rng with set_seed")
+    }
+
     fn rand_uniform(&self, shape: &Shape, dtype: DType, min: f64, max: f64) -> Result<CpuStorage> {
         use rand::prelude::*;
 

candle-core/src/cuda_backend.rs

@@ -223,6 +223,14 @@ impl BackendDevice for CudaDevice {
         })
     }
 
+    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(),
@@ -600,6 +608,105 @@ impl Map1 for Elu {
     }
 }
 
+struct Im2Col1D {
+    l_k: usize,
+    stride: usize,
+    dilation: usize,
+    padding: usize,
+}
+
+impl Im2Col1D {
+    fn l_out(&self, l: usize) -> usize {
+        (l + 2 * self.padding - self.dilation * (self.l_k - 1) - 1) / self.stride + 1
+    }
+}
+
+impl Map1 for Im2Col1D {
+    fn f<T: DeviceRepr + WithDType>(
+        &self,
+        src: &CudaSlice<T>,
+        dev: &CudaDevice,
+        layout: &Layout,
+    ) -> Result<CudaSlice<T>> {
+        let shape = layout.shape();
+        let dims = shape.dims();
+        let l_out = self.l_out(dims[2]);
+        let dst_el = dims[0] * l_out * dims[1] * self.l_k;
+        let cfg = LaunchConfig::for_num_elems(dst_el as u32);
+        let ds = dev.htod_copy([dims, layout.stride()].concat()).w()?;
+        let src = &src.slice(layout.start_offset()..);
+        let func = dev.get_or_load_func(&kernel_name::<T>("im2col1d"), kernels::CONV)?;
+        // SAFETY: Set later by running the kernel.
+        let dst = unsafe { dev.alloc::<T>(dst_el) }.w()?;
+        let params = (
+            dst_el,
+            l_out,
+            self.l_k,
+            self.stride,
+            self.padding,
+            self.dilation,
+            &ds,
+            src,
+            &dst,
+        );
+        // SAFETY: ffi.
+        unsafe { func.launch(cfg, params) }.w()?;
+        Ok(dst)
+    }
+}
+
+struct Im2Col {
+    h_k: usize,
+    w_k: usize,
+    stride: usize,
+    dilation: usize,
+    padding: usize,
+}
+
+impl Im2Col {
+    fn hw_out(&self, h: usize, w: usize) -> (usize, usize) {
+        let h_out = (h + 2 * self.padding - self.dilation * (self.h_k - 1) - 1) / self.stride + 1;
+        let w_out = (w + 2 * self.padding - self.dilation * (self.w_k - 1) - 1) / self.stride + 1;
+        (h_out, w_out)
+    }
+}
+
+impl Map1 for Im2Col {
+    fn f<T: DeviceRepr + WithDType>(
+        &self,
+        src: &CudaSlice<T>,
+        dev: &CudaDevice,
+        layout: &Layout,
+    ) -> Result<CudaSlice<T>> {
+        let shape = layout.shape();
+        let dims = shape.dims();
+        let (h_out, w_out) = self.hw_out(dims[2], dims[3]);
+        let dst_el = dims[0] * h_out * w_out * dims[1] * self.h_k * self.w_k;
+        let cfg = LaunchConfig::for_num_elems(dst_el as u32);
+        let ds = dev.htod_copy([dims, layout.stride()].concat()).w()?;
+        let src = &src.slice(layout.start_offset()..);
+        let func = dev.get_or_load_func(&kernel_name::<T>("im2col"), kernels::CONV)?;
+        // SAFETY: Set later by running the kernel.
+        let dst = unsafe { dev.alloc::<T>(dst_el) }.w()?;
+        let params = (
+            dst_el,
+            h_out,
+            w_out,
+            self.h_k,
+            self.w_k,
+            self.stride,
+            self.padding,
+            self.dilation,
+            &ds,
+            src,
+            &dst,
+        );
+        // SAFETY: ffi.
+        unsafe { func.launch(cfg, params) }.w()?;
+        Ok(dst)
+    }
+}
+
 struct Powf(f64);
 impl Map1 for Powf {
     fn f<T: DeviceRepr + WithDType>(
@@ -785,8 +892,6 @@ impl<'a> Map1 for IndexSelect<'a> {
         };
         let ids_shape = ids_l.shape();
         let ids_dims = ids_shape.dims();
-        let ids_el = ids_shape.elem_count();
-        let cfg = LaunchConfig::for_num_elems(ids_el as u32);
         let ds = dev.htod_copy([ids_dims, ids_l.stride()].concat()).w()?;
         let src = match src_l.contiguous_offsets() {
             Some((o1, o2)) => src.slice(o1..o2),
@@ -794,19 +899,23 @@ impl<'a> Map1 for IndexSelect<'a> {
         };
         let left_size: usize = src_l.dims()[..self.2].iter().product();
         let right_size: usize = src_l.dims()[self.2 + 1..].iter().product();
-        let dim_size = src_l.dims()[self.2];
+        let src_dim_size = src_l.dims()[self.2];
+        let ids_dim_size = ids_shape.elem_count();
+        let dst_el = ids_shape.elem_count() * left_size * right_size;
+        let cfg = LaunchConfig::for_num_elems(dst_el as u32);
         let func = dev.get_or_load_func(&kernel_name::<T>(name), kernels::INDEXING)?;
         // SAFETY: Set later by running the kernel.
-        let out = unsafe { dev.alloc::<T>(ids_el * left_size * right_size) }.w()?;
+        let out = unsafe { dev.alloc::<T>(dst_el) }.w()?;
         let params = (
-            ids_el,
+            dst_el,
             ids_dims.len(),
             &ds,
             ids,
             &src,
             &out,
             left_size,
-            dim_size,
+            src_dim_size,
+            ids_dim_size,
             right_size,
         );
         // SAFETY: ffi.
@@ -1040,6 +1149,55 @@ impl<'a> Map2 for Conv2D<'a> {
     }
 }
 
+struct ConvTranspose1D<'a>(&'a crate::conv::ParamsConvTranspose1D);
+impl<'a> Map2 for ConvTranspose1D<'a> {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        inp: &CudaSlice<T>,
+        inp_l: &Layout,
+        k: &CudaSlice<T>,
+        k_l: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<CudaSlice<T>> {
+        // Kernel shape: (c_in_k, c_out, l_k)
+        // Input shape: (b_size, c_in, l_in)
+        let p = &self.0;
+        let l_out = p.l_out();
+        let dst_el = p.c_out * l_out * p.b_size;
+        let inp = &inp.slice(inp_l.start_offset()..);
+        let k = &k.slice(k_l.start_offset()..);
+        let shape = inp_l.shape();
+        let dims = shape.dims();
+        let el = shape.elem_count();
+
+        // SAFETY: Set later by running the kernel.
+        let out = unsafe { dev.alloc::<T>(dst_el) }.w()?;
+        let cfg = LaunchConfig::for_num_elems(dst_el as u32);
+        let func = dev.get_or_load_func(&kernel_name::<T>("conv_transpose1d"), kernels::CONV)?;
+        let ds = if dims.len() == 3 {
+            [dims, inp_l.stride(), k_l.dims(), k_l.stride()].concat()
+        } else {
+            crate::bail!("unexpected input shape for conv_transpose1d {dims:?}")
+        };
+        let ds = dev.htod_copy(ds).w()?;
+        let params = (
+            el,
+            l_out,
+            p.stride,
+            p.padding,
+            p.output_padding,
+            p.dilation,
+            &ds,
+            inp,
+            k,
+            &out,
+        );
+        // SAFETY: ffi.
+        unsafe { func.launch(cfg, params) }.w()?;
+        Ok(out)
+    }
+}
+
 struct ConvTranspose2D<'a>(&'a crate::conv::ParamsConvTranspose2D);
 impl<'a> Map2 for ConvTranspose2D<'a> {
     fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
@@ -1050,11 +1208,11 @@ impl<'a> Map2 for ConvTranspose2D<'a> {
         k_l: &Layout,
         dev: &CudaDevice,
     ) -> Result<CudaSlice<T>> {
-        // Kernel shape: (c_in_k, c_out / groups, h_k, w_k)
+        // Kernel shape: (c_in_k, c_out, h_k, w_k)
         // Input shape: (b_size, c_in, h_in, w_in)
         let p = &self.0;
         let (out_w, out_h) = (p.out_w(), p.out_h());
-        let dst_el = p.c_out_per_group * p.groups * out_w * out_h * p.b_size;
+        let dst_el = p.c_out * out_w * out_h * p.b_size;
         let inp = &inp.slice(inp_l.start_offset()..);
         let k = &k.slice(k_l.start_offset()..);
         let shape = inp_l.shape();
@@ -1063,13 +1221,7 @@ impl<'a> Map2 for ConvTranspose2D<'a> {
 
         // SAFETY: Set later by running the kernel.
         let out = unsafe { dev.alloc::<T>(dst_el) }.w()?;
-        const NUM_THREADS: u32 = 512;
-        let num_blocks = (dst_el as u32 + NUM_THREADS - 1) / NUM_THREADS;
-        let mut cfg = LaunchConfig {
-            grid_dim: (num_blocks, 1, 1),
-            block_dim: (NUM_THREADS, 1, 1),
-            shared_mem_bytes: 0,
-        };
+        let cfg = LaunchConfig::for_num_elems(dst_el as u32);
         let func = dev.get_or_load_func(&kernel_name::<T>("conv_transpose2d"), kernels::CONV)?;
         let ds = if dims.len() == 4 {
             [dims, inp_l.stride(), k_l.dims(), k_l.stride()].concat()
@@ -1078,13 +1230,13 @@ impl<'a> Map2 for ConvTranspose2D<'a> {
         };
         let ds = dev.htod_copy(ds).w()?;
         let params = (
+            el,
             out_w,
             out_h,
             p.stride,
             p.padding,
             p.output_padding,
             p.dilation,
-            p.groups,
             &ds,
             inp,
             k,
@@ -1663,8 +1815,58 @@ impl BackendStorage for CudaStorage {
         kernel_l: &Layout,
         params: &crate::conv::ParamsConv1D,
     ) -> Result<Self> {
+        const USE_IM2COL_CONV1D: bool = true;
+
         let device = self.device().clone();
-        let slice = Conv1D(params).map(&self.slice, l, &kernel.slice, kernel_l, &device)?;
+        if !USE_IM2COL_CONV1D {
+            let slice = Conv1D(params).map(&self.slice, l, &kernel.slice, kernel_l, &device)?;
+            return Ok(Self { slice, device });
+        }
+
+        let col = Im2Col1D {
+            l_k: params.k_size,
+            stride: params.stride,
+            dilation: params.dilation,
+            padding: params.padding,
+        }
+        .map(&self.slice, &device, l)?;
+        let col = Self { slice: col, device };
+        let l_out = params.l_out();
+        let b = params.b_size;
+        let n = params.c_out;
+        let k = params.k_size * params.c_in;
+        let m = l_out;
+        let col_l = Layout::contiguous((b, m, k));
+        let res = if kernel_l.is_contiguous() {
+            let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
+                .transpose(1, 2)?
+                .broadcast_as((b, k, n))?;
+            col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
+        } else {
+            // Make the kernel contiguous if not already the case.
+            let mut kernel_c = self.device().zeros_impl(kernel_l.shape(), kernel.dtype())?;
+            kernel.copy_strided_src(&mut kernel_c, 0, kernel_l)?;
+            let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
+                .transpose(1, 2)?
+                .broadcast_as((b, k, n))?;
+            col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
+        };
+        let res_l = Layout::contiguous((b, l_out, n)).transpose(1, 2)?;
+        let mut res_t = self.device().zeros_impl(res_l.shape(), res.dtype())?;
+        res.copy_strided_src(&mut res_t, 0, &res_l)?;
+        Ok(res_t)
+    }
+
+    fn conv_transpose1d(
+        &self,
+        l: &Layout,
+        kernel: &Self,
+        kernel_l: &Layout,
+        params: &crate::conv::ParamsConvTranspose1D,
+    ) -> Result<Self> {
+        let device = self.device().clone();
+        let slice =
+            ConvTranspose1D(params).map(&self.slice, l, &kernel.slice, kernel_l, &device)?;
         Ok(Self { slice, device })
     }
 
@@ -1676,9 +1878,50 @@ impl BackendStorage for CudaStorage {
         kernel_l: &Layout,
         params: &crate::conv::ParamsConv2D,
     ) -> Result<Self> {
+        const USE_IM2COL_CONV2D: bool = true;
+
         let device = self.device().clone();
-        let slice = Conv2D(params).map(&self.slice, l, &kernel.slice, kernel_l, &device)?;
-        Ok(Self { slice, device })
+        if !USE_IM2COL_CONV2D {
+            let slice = Conv2D(params).map(&self.slice, l, &kernel.slice, kernel_l, &device)?;
+            return Ok(Self { slice, device });
+        }
+
+        let col = Im2Col {
+            h_k: params.k_h,
+            w_k: params.k_w,
+            stride: params.stride,
+            dilation: params.dilation,
+            padding: params.padding,
+        }
+        .map(&self.slice, &device, l)?;
+        let col = Self { slice: col, device };
+        let h_out = params.out_h();
+        let w_out = params.out_w();
+        let b = params.b_size;
+        let n = params.c_out;
+        let k = params.k_h * params.k_w * params.c_in;
+        let m = h_out * w_out;
+        let col_l = Layout::contiguous((b, m, k));
+        let res = if kernel_l.is_contiguous() {
+            let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
+                .transpose(1, 2)?
+                .broadcast_as((b, k, n))?;
+            col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
+        } else {
+            // Make the kernel contiguous if not already the case.
+            let mut kernel_c = self.device().zeros_impl(kernel_l.shape(), kernel.dtype())?;
+            kernel.copy_strided_src(&mut kernel_c, 0, kernel_l)?;
+            let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
+                .transpose(1, 2)?
+                .broadcast_as((b, k, n))?;
+            col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
+        };
+        let res_l = Layout::contiguous((b, h_out, w_out, n))
+            .transpose(1, 2)?
+            .transpose(1, 3)?;
+        let mut res_t = self.device().zeros_impl(res_l.shape(), res.dtype())?;
+        res.copy_strided_src(&mut res_t, 0, &res_l)?;
+        Ok(res_t)
     }
 
     #[cfg(feature = "cudnn")]
@@ -1783,6 +2026,10 @@ impl BackendStorage for CudaStorage {
         Ok(Self { slice, device })
     }
 
+    fn upsample_nearest1d(&self, _: &Layout, _out_sz: usize) -> Result<Self> {
+        crate::bail!("upsample-nearest1d is not supported on cuda")
+    }
+
     fn upsample_nearest2d(&self, l: &Layout, out_w: usize, out_h: usize) -> Result<Self> {
         let device = self.device().clone();
         let slice = UpsampleNearest2D(out_w, out_h).map(&self.slice, &device, l)?;
@@ -1986,7 +2233,7 @@ impl BackendStorage for CudaStorage {
                 if src_l.is_contiguous() {
                     dev.dtod_copy(&src, &mut dst).w()?
                 } else {
-                    let func = dev.get_or_load_func("ucopy_64", kernels::UNARY)?;
+                    let func = dev.get_or_load_func("ucopy_f64", kernels::UNARY)?;
                     // SAFETY: Set later by running the kernel.
                     let params = (el_count, dims.len(), &ds, &src, &mut dst);
                     // SAFETY: ffi.

candle-core/src/cudnn.rs

@@ -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) {
@@ -90,7 +93,20 @@ pub(crate) fn launch_conv2d<
         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/device.rs

@@ -8,12 +8,14 @@ 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),
 }
 
 pub trait NdArray {
@@ -128,10 +130,23 @@ impl Device {
         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,
         }
     }
@@ -140,21 +155,20 @@ 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 {
-        match self {
-            Self::Cpu => false,
-            Self::Cuda(_) => true,
-        }
+        matches!(self, Self::Cuda(_))
+    }
+
+    pub fn is_metal(&self) -> bool {
+        matches!(self, Self::Metal(_))
     }
 
     pub fn cuda_if_available(ordinal: usize) -> Result<Self> {
@@ -178,8 +192,18 @@ 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))
             }
         }
     }
@@ -206,8 +230,18 @@ 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))
             }
         }
     }
@@ -231,6 +265,10 @@ 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))
+            }
         }
     }
 
@@ -244,6 +282,10 @@ 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))
+            }
         }
     }
 
@@ -255,6 +297,11 @@ impl Device {
                 let storage = device.storage_from_cpu_storage(&storage)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = array.to_cpu_storage();
+                let storage = device.storage_from_cpu_storage(&storage)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 
@@ -266,6 +313,11 @@ impl Device {
                 let storage = device.storage_from_cpu_storage(&storage)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = S::to_cpu_storage_owned(data);
+                let storage = device.storage_from_cpu_storage(&storage)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 }

candle-core/src/display.rs

@@ -14,6 +14,9 @@ impl Tensor {
             crate::DeviceLocation::Cuda { gpu_id } => {
                 format!(", cuda:{}", gpu_id)
             }
+            crate::DeviceLocation::Metal { gpu_id } => {
+                format!(", metal:{}", gpu_id)
+            }
         };
 
         write!(f, "Tensor[")?;
@@ -476,6 +479,9 @@ impl std::fmt::Display for Tensor {
             crate::DeviceLocation::Cuda { gpu_id } => {
                 format!(", cuda:{}", gpu_id)
             }
+            crate::DeviceLocation::Metal { gpu_id } => {
+                format!(", metal:{}", gpu_id)
+            }
         };
 
         write!(

candle-core/src/dtype.rs

@@ -67,6 +67,20 @@ impl DType {
             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:

candle-core/src/dummy_cuda_backend.rs

@@ -79,6 +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,
@@ -152,6 +162,10 @@ 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)
     }
@@ -163,6 +177,10 @@ impl crate::backend::BackendDevice for CudaDevice {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn set_seed(&self, _: u64) -> Result<()> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn location(&self) -> crate::DeviceLocation {
         fail!()
     }

candle-core/src/dummy_metal_backend.rs

@@ -0,0 +1,223 @@
+#![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 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)
+    }
+
+    fn storage_from_cpu_storage(&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)
+    }
+}

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 {
@@ -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),
 

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

candle-core/src/lib.rs

@@ -49,9 +49,12 @@ mod device;
 pub mod display;
 mod dtype;
 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;
@@ -69,7 +72,7 @@ pub mod utils;
 mod variable;
 
 pub use cpu_backend::CpuStorage;
-pub use device::{Device, DeviceLocation};
+pub use device::{Device, DeviceLocation, NdArray};
 pub use dtype::{DType, FloatDType, IntDType, WithDType};
 pub use error::{Error, Result};
 pub use indexer::IndexOp;
@@ -87,6 +90,12 @@ pub use cuda_backend::{CudaDevice, CudaStorage};
 #[cfg(not(feature = "cuda"))]
 pub use dummy_cuda_backend::{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;
 
@@ -110,12 +119,24 @@ impl ToUsize2 for (usize, usize) {
 }
 
 // A simple trait defining a module with forward method using a single argument.
-pub trait Module: std::fmt::Debug {
+pub trait Module {
     fn forward(&self, xs: &Tensor) -> Result<Tensor>;
 }
 
-impl Module for quantized::QMatMul {
+impl<T: Fn(&Tensor) -> Result<Tensor>> Module for T {
     fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        self(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/mkl.rs

@@ -333,6 +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()) {
@@ -355,6 +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]

candle-core/src/npy.rs

@@ -250,8 +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)
     }
 

candle-core/src/op.rs

@@ -1,5 +1,5 @@
 #![allow(clippy::redundant_closure_call)]
-use crate::{CpuStorage, CudaStorage, Layout, Result, Shape, Tensor};
+use crate::{CpuStorage, CudaStorage, Layout, MetalStorage, Result, Shape, Tensor};
 use half::{bf16, f16};
 use num_traits::float::Float;
 
@@ -58,8 +58,14 @@ pub enum UnaryOp {
     Sqr,
     Sqrt,
     Gelu,
+    GeluErf,
+    Erf,
     Relu,
+    Silu,
     Tanh,
+    Floor,
+    Ceil,
+    Round,
 }
 
 #[derive(Clone)]
@@ -85,6 +91,16 @@ pub enum Op {
         dilation: usize,
     },
 
+    #[allow(dead_code)]
+    ConvTranspose1D {
+        arg: Tensor,
+        kernel: Tensor,
+        padding: usize,
+        output_padding: usize,
+        stride: usize,
+        dilation: usize,
+    },
+
     #[allow(dead_code)]
     Conv2D {
         arg: Tensor,
@@ -102,7 +118,6 @@ pub enum Op {
         output_padding: usize,
         stride: usize,
         dilation: usize,
-        groups: usize,
     },
 
     AvgPool2D {
@@ -117,7 +132,12 @@ pub enum Op {
         stride: (usize, usize),
     },
 
-    UpsampleNearest2D(Tensor),
+    UpsampleNearest1D(Tensor),
+    UpsampleNearest2D {
+        arg: Tensor,
+        target_h: usize,
+        target_w: usize,
+    },
 
     Cat(Vec<Tensor>, usize),
 
@@ -131,6 +151,7 @@ pub enum Op {
     Copy(Tensor),
     Broadcast(Tensor),
     Narrow(Tensor, usize, usize, usize),
+    SliceScatter0(Tensor, Tensor, usize),
     Reshape(Tensor),
     ToDevice(Tensor),
     Transpose(Tensor, usize, usize),
@@ -168,6 +189,18 @@ pub trait CustomOp1 {
         ))
     }
 
+    /// 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.
@@ -203,6 +236,20 @@ pub trait CustomOp2 {
         ))
     }
 
+    /// 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,
@@ -245,6 +292,22 @@ pub trait CustomOp3 {
         ))
     }
 
+    /// 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,
@@ -325,8 +388,14 @@ 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;
 
 macro_rules! bin_op {
     ($op:ident, $name: literal, $e: expr, $f32_vec: ident, $f64_vec: ident) => {
@@ -525,13 +594,13 @@ 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!(Tanh, "tanh", v, v.tanh(), vs_tanh, vd_tanh);
-unary_op!(Abs, "abs", v, v.abs());
 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
+/// 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";
@@ -621,6 +690,283 @@ impl UnaryOpT for Gelu {
     }
 }
 
+/// `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 {
     const NAME: &'static str = "relu";
     const KERNEL: &'static str = "urelu";
@@ -706,6 +1052,10 @@ impl BackpropOp {
         };
         Self(op)
     }
+
+    pub(crate) fn is_none(&self) -> bool {
+        self.0.is_none()
+    }
 }
 
 impl std::ops::Deref for BackpropOp {

candle-core/src/pickle.rs

@@ -193,6 +193,55 @@ impl Object {
             _ => 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 {
@@ -301,8 +350,10 @@ impl Stack {
                 module_name,
                 class_name,
             } => {
-                if module_name == "collections" && class_name == "OrderedDict" {
-                    // TODO: have a separate ordered dict.
+                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
@@ -565,6 +616,7 @@ fn rebuild_args(args: Object) -> Result<(Layout, DType, String, usize)> {
         "HalfStorage" => DType::F16,
         "BFloat16Storage" => DType::BF16,
         "ByteStorage" => DType::U8,
+        "LongStorage" => DType::I64,
         other => {
             crate::bail!("unsupported storage type {other}")
         }
@@ -582,9 +634,16 @@ pub struct TensorInfo {
     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);
@@ -606,8 +665,9 @@ pub fn read_pth_tensor_info<P: AsRef<std::path::Path>>(
         stack.read_loop(&mut reader)?;
         let obj = stack.finalize()?;
         if VERBOSE || verbose {
-            println!("{obj:?}");
+            println!("{obj:#?}");
         }
+
         let obj = match obj {
             Object::Build { callable, args } => match *callable {
                 Object::Reduce { callable, args: _ } => match *callable {
@@ -621,52 +681,30 @@ pub fn read_pth_tensor_info<P: AsRef<std::path::Path>>(
             },
             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() {
-                let name = match name.unicode() {
-                    Ok(name) => name,
-                    Err(_) => continue,
-                };
-                let (callable, args) = match value.reduce() {
-                    Ok(callable_args) => callable_args,
-                    _ => continue,
-                };
-                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)
-                    }
-                    _ => (callable, args),
-                };
-                match callable {
-                    Object::Class {
-                        module_name,
-                        class_name,
-                    } if module_name == "torch._utils" && class_name == "_rebuild_tensor_v2" => {}
-                    _ => continue,
-                };
-                match rebuild_args(args) {
-                    Ok((layout, dtype, file_path, storage_size)) => {
-                        let mut path = dir_name.clone();
-                        path.push(file_path);
-                        tensor_infos.push(TensorInfo {
-                            name,
-                            dtype,
-                            layout,
-                            path: path.to_string_lossy().into_owned(),
-                            storage_size,
-                        })
-                    }
-                    Err(err) => {
-                        eprintln!("skipping {name}: {err:?}")
-                    }
+                match value.into_tensor_info(name, &dir_name) {
+                    Ok(Some(tensor_info)) => tensor_infos.push(tensor_info),
+                    Ok(None) => {}
+                    Err(err) => eprintln!("skipping: {err:?}"),
                 }
             }
         }
@@ -683,8 +721,8 @@ pub struct PthTensors {
 }
 
 impl PthTensors {
-    pub fn new<P: AsRef<std::path::Path>>(path: P) -> Result<Self> {
-        let tensor_infos = read_pth_tensor_info(path.as_ref(), false)?;
+    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))
@@ -698,6 +736,7 @@ impl PthTensors {
     }
 
     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,
@@ -706,20 +745,70 @@ impl PthTensors {
         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, use an offset, etc.
-        // For now only support the basic case.
-        if tensor_info.layout.start_offset() != 0 || !tensor_info.layout.is_contiguous() {
+        // 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,
         )?;
-        Ok(Some(tensor))
+
+        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

@@ -50,14 +50,9 @@ pub(crate) unsafe fn mul_sum_i8_pairs_float(x: __m256i, y: __m256i) -> __m256 {
 #[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}")
     }
-    if nb % 2 != 0 {
-        crate::bail!("vec_dot_q4_0_q8_0: {nb} is not even")
-    }
-
     unsafe {
         let mut acc = _mm256_setzero_ps();
         for (x, y) in xs.iter().zip(ys.iter()) {
@@ -358,7 +353,7 @@ pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Res
                 q3 = q3.add(32);
 
                 // Prepare low and high bits
-                // We hardcode the shifts here to avoid loading them into a seperate register
+                // 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)
@@ -591,7 +586,7 @@ pub(crate) fn vec_dot_q5k_q8k(n: usize, xs: &[BlockQ5K], ys: &[BlockQ8K]) -> Res
                 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 seperate register
+                //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 {
@@ -638,3 +633,35 @@ pub(crate) fn vec_dot_q5k_q8k(n: usize, xs: &[BlockQ5K], ys: &[BlockQ8K]) -> Res
         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/dummy_metal.rs

@@ -0,0 +1,43 @@
+#![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)
+    }
+}

candle-core/src/quantized/ggml_file.rs

@@ -1,7 +1,9 @@
 //! Support for the GGML file format.
 
-use super::{k_quants, GgmlDType};
-use crate::Result;
+#[cfg(feature = "metal")]
+use super::metal::load_quantized_metal;
+use super::{k_quants, GgmlDType, QStorage};
+use crate::{Device, Result};
 use byteorder::{LittleEndian, ReadBytesExt};
 use std::collections::HashMap;
 
@@ -121,11 +123,22 @@ 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) };
-    super::QTensor::new(data.to_vec(), dims)
+    let data: QStorage = match device {
+        Device::Cpu => QStorage::Cpu(Box::new(data.to_vec())),
+        #[cfg(feature = "metal")]
+        Device::Metal(metal) => load_quantized_metal(metal, data)?,
+        #[cfg(not(feature = "metal"))]
+        Device::Metal(_metal) => {
+            crate::bail!("Metal backend requires `metal` feature")
+        }
+        device => unimplemented!("Implement quantized tensor for device {device:?}"),
+    };
+    super::QTensor::new(data, dims)
 }
 
 /// Creates a [Tensor] from a raw GGML tensor.
@@ -133,23 +146,50 @@ 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"),
     }
 }
@@ -157,6 +197,7 @@ 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>()?;
@@ -177,11 +218,11 @@ 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();
+    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}"),
     }
@@ -192,10 +233,14 @@ pub struct Content {
     pub hparams: HParams,
     pub vocab: Vocab,
     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))?;
@@ -205,14 +250,16 @@ impl Content {
         let mut tensors = HashMap::new();
 
         while reader.stream_position()? != last_position {
-            let (name, tensor) = read_one_tensor(reader, magic)?;
+            let (name, tensor) = read_one_tensor(reader, magic, device)?;
             tensors.insert(name, tensor);
         }
+        let device = device.clone();
         Ok(Self {
             magic,
             hparams,
             vocab,
             tensors,
+            device,
         })
     }
 

candle-core/src/quantized/gguf_file.rs

@@ -3,7 +3,7 @@
 //! Spec: https://github.com/philpax/ggml/blob/gguf-spec/docs/gguf.md
 
 use super::{GgmlDType, QTensor};
-use crate::Result;
+use crate::{Device, Result};
 use byteorder::{LittleEndian, ReadBytesExt, WriteBytesExt};
 use std::collections::HashMap;
 
@@ -29,6 +29,7 @@ impl TryFrom<u32> for Magic {
 pub enum VersionedMagic {
     GgufV1,
     GgufV2,
+    GgufV3,
 }
 
 impl VersionedMagic {
@@ -39,7 +40,8 @@ impl VersionedMagic {
         let versioned_magic = match (magic, version) {
             (Magic::Gguf, 1) => Self::GgufV1,
             (Magic::Gguf, 2) => Self::GgufV2,
-            _ => crate::bail!("ggml: unsupported magic/version {magic:?}/{version}"),
+            (Magic::Gguf, 3) => Self::GgufV3,
+            _ => crate::bail!("gguf: unsupported magic/version {magic:?}/{version}"),
         };
         Ok(versioned_magic)
     }
@@ -57,14 +59,25 @@ impl TensorInfo {
         &self,
         reader: &mut R,
         tensor_data_offset: u64,
+        device: &Device,
     ) -> Result<QTensor> {
         let tensor_elems = self.shape.elem_count();
-        let size_in_bytes =
-            tensor_elems * self.ggml_dtype.type_size() / self.ggml_dtype.blck_size();
+        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())
+        super::ggml_file::qtensor_from_ggml(
+            self.ggml_dtype,
+            &raw_data,
+            self.shape.dims().to_vec(),
+            device,
+        )
     }
 }
 
@@ -79,7 +92,9 @@ pub struct Content {
 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 => reader.read_u64::<LittleEndian>()? as usize,
+        VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
+            reader.read_u64::<LittleEndian>()? as usize
+        }
     };
     let mut v = vec![0u8; len];
     reader.read_exact(&mut v)?;
@@ -279,7 +294,9 @@ impl Value {
                 let value_type = ValueType::from_u32(value_type)?;
                 let len = match magic {
                     VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
-                    VersionedMagic::GgufV2 => reader.read_u64::<LittleEndian>()? as usize,
+                    VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
+                        reader.read_u64::<LittleEndian>()? as usize
+                    }
                 };
                 let mut vs = Vec::with_capacity(len);
                 for _ in 0..len {
@@ -376,11 +393,15 @@ impl Content {
 
         let tensor_count = match magic {
             VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
-            VersionedMagic::GgufV2 => reader.read_u64::<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 => reader.read_u64::<LittleEndian>()? as usize,
+            VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
+                reader.read_u64::<LittleEndian>()? as usize
+            }
         };
 
         let mut metadata = HashMap::new();
@@ -402,7 +423,7 @@ impl Content {
                     reader.read_u32_into::<LittleEndian>(&mut dimensions)?;
                     dimensions.into_iter().map(|c| c as usize).collect()
                 }
-                VersionedMagic::GgufV2 => {
+                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()
@@ -445,12 +466,13 @@ impl Content {
         &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-infor for {name}"),
+            None => crate::bail!("cannot find tensor info for {name}"),
         };
-        tensor_info.read(reader, self.tensor_data_offset)
+        tensor_info.read(reader, self.tensor_data_offset, device)
     }
 }
 
@@ -502,10 +524,9 @@ pub fn write<W: std::io::Seek + std::io::Write>(
                 "internal error, unexpected current position {tensor_start_pos} {offset} {pos}"
             )
         }
-        let data_ptr = tensor.as_ptr();
-        let size_in_bytes = tensor.storage_size_in_bytes();
-        let data = unsafe { std::slice::from_raw_parts(data_ptr, size_in_bytes) };
-        w.write_all(data)?;
+        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])?;
     }

candle-core/src/quantized/k_quants.rs

@@ -34,6 +34,9 @@ pub trait GgmlType: Sized + Clone + Send + Sync {
     /// Dot product used as a building block for quantized mat-mul.
     /// n is the number of elements to be considered.
     fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32>;
+
+    /// Generic implementation of the dot product without simd optimizations.
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32>;
 }
 
 #[derive(Debug, Clone, PartialEq)]
@@ -225,15 +228,17 @@ impl GgmlType for BlockQ4_0 {
         #[cfg(target_feature = "neon")]
         return super::neon::vec_dot_q4_0_q8_0(n, xs, ys);
 
+        #[cfg(target_feature = "simd128")]
+        return super::simd128::vec_dot_q4_0_q8_0(n, xs, ys);
+
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> 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}")
         }
-        if nb % 2 != 0 {
-            crate::bail!("vec_dot_q4_0_q8_0: {nb} is not even")
-        }
-
         // Generic implementation.
         let mut sumf = 0f32;
         for (xs, ys) in xs.iter().zip(ys.iter()) {
@@ -255,6 +260,10 @@ impl GgmlType for BlockQ4_1 {
     type VecDotType = BlockQ8_1;
 
     fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         // ggml_vec_dot_q4_1_q8_1
         let qk = QK8_1;
         if n % qk != 0 {
@@ -354,7 +363,10 @@ impl GgmlType for BlockQ5_0 {
         if nb % 2 != 0 {
             crate::bail!("vec_dot_q5_0_q8_0: {n}, nb is not divisible by 2")
         }
+        Self::vec_dot_unopt(n, xs, ys)
+    }
 
+    fn vec_dot_unopt(_n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         // Generic implementation.
         let mut sumf = 0f32;
 
@@ -445,6 +457,10 @@ impl GgmlType for BlockQ5_1 {
     type VecDotType = BlockQ8_1;
 
     fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         let qk = Self::BLCK_SIZE;
         if n % Self::BLCK_SIZE != 0 {
             crate::bail!("vec_dot_q5_1_q8_1: {n} is not divisible by {qk}")
@@ -606,6 +622,13 @@ impl GgmlType for BlockQ8_0 {
         #[cfg(target_feature = "neon")]
         return super::neon::vec_dot_q8_0_q8_0(n, xs, ys);
 
+        #[cfg(target_feature = "simd128")]
+        return super::simd128::vec_dot_q8_0_q8_0(n, xs, ys);
+
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> 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}")
@@ -631,7 +654,11 @@ impl GgmlType for BlockQ8_1 {
     const BLCK_SIZE: usize = QK8_1;
     type VecDotType = BlockQ8_1;
 
-    fn vec_dot(_n: usize, _xs: &[Self], _ys: &[Self::VecDotType]) -> Result<f32> {
+    fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(_n: usize, _xs: &[Self], _ys: &[Self::VecDotType]) -> Result<f32> {
         unimplemented!("no support for vec-dot on Q8_1")
     }
 
@@ -681,6 +708,13 @@ impl GgmlType for BlockQ2K {
         #[cfg(target_feature = "neon")]
         return super::neon::vec_dot_q2k_q8k(n, xs, ys);
 
+        #[cfg(target_feature = "simd128")]
+        return super::simd128::vec_dot_q2k_q8k(n, xs, ys);
+
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         if n % QK_K != 0 {
             crate::bail!("vec_dot_q2k_q8k: {n} is not divisible by {QK_K}")
         }
@@ -701,18 +735,17 @@ impl GgmlType for BlockQ2K {
 
             let mut isum = 0;
             let mut is = 0;
-            let mut d;
             for _ in 0..(QK_K / 128) {
                 let mut shift = 0;
                 for _ in 0..4 {
-                    d = (sc[is] & 0xF) as i32;
+                    let d = (sc[is] & 0xF) as i32;
                     is += 1;
                     let mut isuml = 0;
                     for l in 0..16 {
                         isuml += q8[l] as i32 * (((q2[l] >> shift) & 3) as i32);
                     }
                     isum += d * isuml;
-                    d = (sc[is] & 0xF) as i32;
+                    let d = (sc[is] & 0xF) as i32;
                     is += 1;
                     isuml = 0;
                     for l in 16..32 {
@@ -851,6 +884,10 @@ impl GgmlType for BlockQ3K {
         #[cfg(target_feature = "neon")]
         return super::neon::vec_dot_q3k_q8k(n, xs, ys);
 
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         if n % QK_K != 0 {
             crate::bail!("vec_dot_q3k_q8k: {n} is not divisible by {QK_K}")
         }
@@ -1077,7 +1114,6 @@ impl GgmlType for BlockQ3K {
             let d_all = block.d.to_f32();
             let mut m = 1;
             let mut is = 0;
-            let mut dl;
 
             // Dequantize both 128 long blocks
             // 32 qs values per 128 long block
@@ -1088,7 +1124,7 @@ impl GgmlType for BlockQ3K {
                     for (scale_index, scale_scoped_y) in
                         shift_scoped_y.chunks_exact_mut(16).enumerate()
                     {
-                        dl = d_all * (scales[is] as f32 - 32.0);
+                        let dl = d_all * (scales[is] as f32 - 32.0);
                         for (i, inner_y) in scale_scoped_y.iter_mut().enumerate() {
                             let new_y = dl
                                 * (((qs[i + 16 * scale_index] >> shift) & 3) as i8
@@ -1126,6 +1162,13 @@ impl GgmlType for BlockQ4K {
         #[cfg(target_feature = "neon")]
         return super::neon::vec_dot_q4k_q8k(n, xs, ys);
 
+        #[cfg(target_feature = "simd128")]
+        return super::simd128::vec_dot_q4k_q8k(n, xs, ys);
+
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         if n % QK_K != 0 {
             crate::bail!("vec_dot_q4k_q8k: {n} is not divisible by {QK_K}")
         }
@@ -1312,6 +1355,10 @@ impl GgmlType for BlockQ5K {
         #[cfg(target_feature = "neon")]
         return super::neon::vec_dot_q5k_q8k(n, xs, ys);
 
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         if n % QK_K != 0 {
             crate::bail!("vec_dot_q5k_q8k: {n} is not divisible by {QK_K}")
         }
@@ -1498,13 +1545,13 @@ impl GgmlType for BlockQ5K {
                 let d2 = d * sc as f32;
                 let m2 = min * m as f32;
                 for (ql, qh) in ql.iter().zip(qh) {
-                    let to_add = if qh & u1 != 0 { 16 } else { 1 };
-                    y[ys_index] = d1 * ((ql & 0xF) + to_add) as f32 - m1;
+                    let to_add = if qh & u1 != 0 { 16f32 } else { 0f32 };
+                    y[ys_index] = d1 * ((ql & 0xF) as f32 + to_add) - m1;
                     ys_index += 1;
                 }
                 for (ql, qh) in ql.iter().zip(qh) {
-                    let to_add = if qh & u2 != 0 { 16 } else { 1 };
-                    y[ys_index] = d2 * ((ql >> 4) + to_add) as f32 - m2;
+                    let to_add = if qh & u2 != 0 { 16f32 } else { 0f32 };
+                    y[ys_index] = d2 * ((ql >> 4) as f32 + to_add) - m2;
                     ys_index += 1;
                 }
                 is += 2;
@@ -1529,6 +1576,13 @@ impl GgmlType for BlockQ6K {
         #[cfg(target_feature = "neon")]
         return super::neon::vec_dot_q6k_q8k(n, xs, ys);
 
+        #[cfg(target_feature = "simd128")]
+        return super::simd128::vec_dot_q6k_q8k(n, xs, ys);
+
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         if n % QK_K != 0 {
             crate::bail!("vec_dot_q6k_q8k: {n} is not divisible by {QK_K}")
         }
@@ -1697,8 +1751,38 @@ impl GgmlType for BlockQ8K {
     const BLCK_SIZE: usize = QK_K;
     type VecDotType = BlockQ8K;
 
-    fn vec_dot(_n: usize, _xs: &[Self], _ys: &[Self::VecDotType]) -> Result<f32> {
-        unreachable!()
+    #[allow(unreachable_code)]
+    fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
+        #[cfg(target_feature = "avx")]
+        return super::avx::vec_dot_q8k_q8k(n, xs, ys);
+
+        #[cfg(target_feature = "neon")]
+        return super::neon::vec_dot_q8k_q8k(n, xs, ys);
+
+        #[cfg(target_feature = "simd128")]
+        return super::simd128::vec_dot_q8k_q8k(n, xs, ys);
+
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
+        let qk = QK_K;
+        if n % QK_K != 0 {
+            crate::bail!("vec_dot_q8k_q8k: {n} is not divisible by {qk}")
+        }
+
+        // Generic implementation.
+        let mut sumf = 0f32;
+        for (xs, ys) in xs.iter().zip(ys.iter()) {
+            let sum_i = xs
+                .qs
+                .iter()
+                .zip(ys.qs.iter())
+                .map(|(&x, &y)| x as i32 * y as i32)
+                .sum::<i32>();
+            sumf += sum_i as f32 * xs.d * ys.d
+        }
+        Ok(sumf)
     }
 
     fn from_float(xs: &[f32], ys: &mut [Self]) -> Result<()> {
@@ -1804,6 +1888,10 @@ impl GgmlType for f32 {
     type VecDotType = f32;
 
     fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         if xs.len() < n {
             crate::bail!("size mismatch {} < {n}", xs.len())
         }
@@ -1838,6 +1926,10 @@ impl GgmlType for f16 {
     type VecDotType = f16;
 
     fn vec_dot(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
+        Self::vec_dot_unopt(n, xs, ys)
+    }
+
+    fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         if xs.len() < n {
             crate::bail!("size mismatch {} < {n}", xs.len())
         }

candle-core/src/quantized/metal.rs

@@ -0,0 +1,234 @@
+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> {
+        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];
+        match self.dtype {
+            GgmlDType::F32 => {
+                let vec: Vec<f32> = read_to_vec(&buffer, elem_count);
+                use crate::quantized::k_quants::GgmlType;
+                f32::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::F16 => {
+                let vec: Vec<half::f16> = read_to_vec(&buffer, elem_count);
+                use crate::quantized::k_quants::GgmlType;
+                half::f16::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q4_0 => {
+                let vec: Vec<crate::quantized::BlockQ4_0> = read_to_vec(&buffer, elem_count);
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ4_0::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q4_1 => {
+                let vec: Vec<crate::quantized::BlockQ4_1> = read_to_vec(&buffer, elem_count);
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ4_1::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q5_0 => {
+                let vec: Vec<crate::quantized::BlockQ5_0> = read_to_vec(&buffer, elem_count);
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ5_0::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q5_1 => {
+                let vec: Vec<crate::quantized::BlockQ5_1> = read_to_vec(&buffer, elem_count);
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ5_1::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q8_0 => {
+                let vec: Vec<crate::quantized::BlockQ8_0> = read_to_vec(&buffer, elem_count);
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ8_0::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q8_1 => {
+                let vec: Vec<crate::quantized::BlockQ8_1> = read_to_vec(&buffer, elem_count);
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ8_1::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q2K => {
+                let vec: Vec<crate::quantized::BlockQ2K> =
+                    read_to_vec(&buffer, elem_count / self.dtype.block_size());
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ2K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q3K => {
+                let vec: Vec<crate::quantized::BlockQ3K> =
+                    read_to_vec(&buffer, elem_count / self.dtype.block_size());
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ3K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q4K => {
+                let vec: Vec<crate::quantized::BlockQ4K> =
+                    read_to_vec(&buffer, elem_count / self.dtype.block_size());
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ4K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q5K => {
+                let vec: Vec<crate::quantized::BlockQ5K> =
+                    read_to_vec(&buffer, elem_count / self.dtype.block_size());
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ5K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q6K => {
+                let vec: Vec<crate::quantized::BlockQ6K> =
+                    read_to_vec(&buffer, elem_count / self.dtype.block_size());
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ6K::to_float(&vec, &mut out)?;
+            }
+            GgmlDType::Q8K => {
+                let vec: Vec<crate::quantized::BlockQ8K> =
+                    read_to_vec(&buffer, elem_count / self.dtype.block_size());
+                use crate::quantized::k_quants::GgmlType;
+                crate::quantized::BlockQ8K::to_float(&vec, &mut out)?;
+            }
+        }
+
+        let buffer = self.device.new_buffer_with_data(&out)?;
+        Ok(MetalStorage::new(buffer, self.device.clone(), 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();
+
+        let (b, m) = match dst_shape.len() {
+            3 => (dst_shape[0], dst_shape[1]),
+            2 => (1, 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()?;
+        candle_metal_kernels::call_quantized_matmul_t(
+            device.device(),
+            &command_buffer,
+            device.kernels(),
+            self.dtype.into(),
+            (b, m, n, k),
+            storage.buffer(),
+            layout.start_offset() * storage.dtype().size_in_bytes(),
+            &self.buffer,
+            &dst,
+        )
+        .map_err(MetalError::from)?;
+        let dst_storage = crate::MetalStorage::new(dst, device, DType::F32);
+        Ok((dst_storage, dst_shape))
+    }
+}
+
+pub fn load_quantized_metal<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

@@ -1,21 +1,118 @@
-use crate::{Device, Result, Shape, Tensor};
+use crate::{CpuStorage, Device, Result, Shape, Storage, Tensor};
+use k_quants::*;
+use std::borrow::Cow;
 
 #[cfg(target_feature = "avx")]
 pub mod avx;
+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(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,
 }
 
+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) => {
+                crate::bail!("Cuda ggml quantization not supported");
+            }
+        }
+    }
+}
+
+pub enum QStorage {
+    Cpu(Box<dyn QuantizedType>),
+    Metal(metal::QMetalStorage),
+}
+
+impl QStorage {
+    fn block_size(&self) -> usize {
+        match self {
+            QStorage::Cpu(storage) => storage.block_size(),
+            QStorage::Metal(storage) => storage.dtype().block_size(),
+        }
+    }
+
+    fn dtype(&self) -> GgmlDType {
+        match self {
+            QStorage::Cpu(storage) => storage.dtype(),
+            QStorage::Metal(storage) => storage.dtype(),
+        }
+    }
+
+    fn device(&self) -> Device {
+        match self {
+            QStorage::Cpu(_storage) => Device::Cpu,
+            QStorage::Metal(storage) => Device::Metal(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(),
+        }
+    }
+
+    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)?,
+            _ => 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)?)),
+        }
+    }
+
+    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(_storage) => {
+                crate::bail!("not implemented");
+            }
+        }
+    }
+}
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
 pub enum GgmlDType {
     F32,
@@ -75,6 +172,25 @@ impl GgmlDType {
         }
     }
 
+    /// 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::*;
@@ -98,7 +214,7 @@ impl GgmlDType {
     }
 
     /// The block size, i.e. the number of elements stored in each block.
-    pub fn blck_size(&self) -> usize {
+    pub fn block_size(&self) -> usize {
         match self {
             Self::F32 => 1,
             Self::F16 => 1,
@@ -117,9 +233,13 @@ 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> {
@@ -127,12 +247,26 @@ 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 {
@@ -150,56 +284,53 @@ impl std::fmt::Debug for QTensor {
     }
 }
 
-fn check_shape<T: k_quants::GgmlType>(shape: &Shape) -> Result<()> {
+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] % T::BLCK_SIZE != 0 {
+    if dims[dims.len() - 1] % block_size != 0 {
         crate::bail!(
             "quantized tensor must have their last dim divisible by block size {shape:?} {}",
-            T::BLCK_SIZE
+            block_size
         )
     }
     Ok(())
 }
 
 impl QTensor {
-    pub fn new<S: Into<Shape>, T: k_quants::GgmlType + Send + Sync + 'static>(
-        data: Vec<T>,
-        shape: S,
-    ) -> Result<Self> {
+    pub fn new<S: Into<Shape>>(storage: QStorage, shape: S) -> Result<Self> {
         let shape = shape.into();
-        check_shape::<T>(&shape)?;
-        Ok(Self {
-            data: Box::new(data),
-            shape,
-        })
+        check_shape(&shape, storage.block_size())?;
+        Ok(Self { storage, shape })
     }
 
-    pub fn quantize<T: k_quants::GgmlType + Send + Sync + 'static>(src: &Tensor) -> Result<Self> {
+    pub fn quantize(src: &Tensor, dtype: GgmlDType) -> Result<Self> {
         let shape = src.shape();
-        check_shape::<T>(shape)?;
-        let src = src
-            .to_dtype(crate::DType::F32)?
-            .flatten_all()?
-            .to_vec1::<f32>()?;
-        if src.len() % T::BLCK_SIZE != 0 {
+        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 {}",
-                T::BLCK_SIZE
+                block_size
             )
         }
-        let mut data = vec![T::zeros(); src.len() / T::BLCK_SIZE];
-        T::from_float(&src, &mut data)?;
+        let mut storage = src.device().qzeros(elem_count, dtype)?;
+        storage.quantize(&src.storage())?;
         Ok(Self {
-            data: Box::new(data),
+            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 {
@@ -211,38 +342,56 @@ 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)
-    }
-
-    pub fn matmul_t(&self, mkn: (usize, usize, usize), lhs: &[f32], dst: &mut [f32]) -> Result<()> {
-        self.data.matmul_t(mkn, lhs, dst)
+        let storage = self.storage.dequantize(self.shape.elem_count())?;
+        let none = crate::op::BackpropOp::none();
+        let is_variable = false;
+        crate::tensor::from_storage(storage, self.shape.clone(), none, is_variable)
+            .to_device(device)
     }
 
     pub fn storage_size_in_bytes(&self) -> usize {
-        self.data.storage_size_in_bytes()
+        self.storage.size_in_bytes()
     }
 
-    pub fn as_ptr(&self) -> *const u8 {
-        self.data.as_ptr()
+    pub fn data(&self) -> Result<Cow<'_, [u8]>> {
+        self.storage.data()
     }
 }
 
-#[derive(Debug)]
-pub struct QMatMul(std::sync::Arc<QTensor>);
+#[derive(Clone, Debug)]
+pub enum QMatMul {
+    QTensor(std::sync::Arc<QTensor>),
+    Tensor(Tensor),
+}
+
+thread_local! {
+    static DEQUANTIZE_ALL: bool = {
+        match std::env::var("CANDLE_DEQUANTIZE_ALL") {
+            Ok(s) => {
+                !s.is_empty() && s != "0"
+            },
+            Err(_) => false,
+        }
+    }
+}
 
 impl QMatMul {
-    pub fn from_arc(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(&Device::Cpu)?;
+            Self::Tensor(tensor)
+        } else {
+            Self::QTensor(qtensor)
+        };
+        Ok(t)
     }
 
-    pub fn from_qtensor(qtensor: QTensor) -> Self {
-        Self(std::sync::Arc::new(qtensor))
-    }
-
-    pub fn inner(&self) -> &std::sync::Arc<QTensor> {
-        &self.0
+    pub fn from_qtensor(qtensor: QTensor) -> Result<Self> {
+        Self::from_arc(std::sync::Arc::new(qtensor))
     }
 }
 
@@ -272,21 +421,43 @@ impl crate::CustomOp1 for QTensor {
         }
         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(_) => 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.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)
+    }
 }
 
-impl QMatMul {
-    pub fn forward(&self, xs: &Tensor) -> Result<Tensor> {
-        xs.apply_op1_no_bwd(self.0.as_ref())
+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)
+            }
+        }
     }
 }

candle-core/src/quantized/neon.rs

@@ -12,6 +12,14 @@ use core::arch::arm::*;
 #[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;
@@ -19,71 +27,39 @@ pub(crate) fn vec_dot_q4_0_q8_0(n: usize, xs: &[BlockQ4_0], ys: &[BlockQ8_0]) ->
     if n % QK8_0 != 0 {
         crate::bail!("vec_dot_q4_0_q8_0: {n} is not divisible by {qk}")
     }
-    if nb % 2 != 0 {
-        crate::bail!("vec_dot_q4_0_q8_0: {nb} is not even")
-    }
 
     unsafe {
         let mut sumv0 = vdupq_n_f32(0.0f32);
-        let mut sumv1 = vdupq_n_f32(0.0f32);
-        for i in (0..nb).step_by(2) {
+        for i in 0..nb {
             let x0 = &xs[i];
-            let x1 = &xs[i + 1];
             let y0 = &ys[i];
-            let y1 = &ys[i + 1];
 
             let m4b = vdupq_n_u8(0x0F);
             let s8b = vdupq_n_s8(0x8);
 
             let v0_0 = vld1q_u8(x0.qs.as_ptr());
-            let v0_1 = vld1q_u8(x1.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));
-            let v0_1l = vreinterpretq_s8_u8(vandq_u8(v0_1, m4b));
-            let v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
 
             // sub 8
             let v0_0ls = vsubq_s8(v0_0l, s8b);
             let v0_0hs = vsubq_s8(v0_0h, s8b);
-            let v0_1ls = vsubq_s8(v0_1l, s8b);
-            let v0_1hs = vsubq_s8(v0_1h, s8b);
 
             // load y
             let v1_0l = vld1q_s8(y0.qs.as_ptr());
             let v1_0h = vld1q_s8(y0.qs.as_ptr().add(16));
-            let v1_1l = vld1q_s8(y1.qs.as_ptr());
-            let v1_1h = vld1q_s8(y1.qs.as_ptr().add(16));
-
-            // TODO: Support dotprod when it's available outside of nightly.
-            let pl0l = vmull_s8(vget_low_s8(v0_0ls), vget_low_s8(v1_0l));
-            let pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l));
-            let ph0l = vmull_s8(vget_low_s8(v0_0hs), vget_low_s8(v1_0h));
-            let ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h));
-
-            let pl1l = vmull_s8(vget_low_s8(v0_1ls), vget_low_s8(v1_1l));
-            let pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l));
-            let ph1l = vmull_s8(vget_low_s8(v0_1hs), vget_low_s8(v1_1h));
-            let ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h));
-
-            let pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
-            let ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
-            let pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
-            let ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
 
+            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(),
             );
-            sumv1 = vmlaq_n_f32(
-                sumv1,
-                vcvtq_f32_s32(vaddq_s32(pl1, ph1)),
-                x1.d.to_f32() * y1.d.to_f32(),
-            );
         }
-        Ok(vaddvq_f32(sumv0) + vaddvq_f32(sumv1))
+        Ok(vaddvq_f32(sumv0))
     }
 }
 
@@ -94,60 +70,58 @@ pub(crate) fn vec_dot_q8_0_q8_0(n: usize, xs: &[BlockQ8_0], ys: &[BlockQ8_0]) ->
         crate::bail!("vec_dot_q8_0_q8_0: {n} is not divisible by {qk}")
     }
     let nb = n / QK8_0;
-    if nb % 2 != 0 {
-        crate::bail!("vec_dot_q8_0_q8_0: {nb} is not even")
-    }
     unsafe {
         let mut sumv0 = vdupq_n_f32(0.0f32);
-        let mut sumv1 = vdupq_n_f32(0.0f32);
-        for i in (0..nb).step_by(2) {
+        for i in 0..nb {
             let x0 = &xs[i];
-            let x1 = &xs[i + 1];
             let y0 = &ys[i];
-            let y1 = &ys[i + 1];
 
             let x0_0 = vld1q_s8(x0.qs.as_ptr());
             let x0_1 = vld1q_s8(x0.qs.as_ptr().add(16));
-            let x1_0 = vld1q_s8(x1.qs.as_ptr());
-            let x1_1 = vld1q_s8(x1.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 y1_0 = vld1q_s8(y1.qs.as_ptr());
-            let y1_1 = vld1q_s8(y1.qs.as_ptr().add(16));
 
-            // TODO dotprod once this is the intrinsics are.
-            let p0_0 = vmull_s8(vget_low_s8(x0_0), vget_low_s8(y0_0));
-            let p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0));
-            let p0_2 = vmull_s8(vget_low_s8(x0_1), vget_low_s8(y0_1));
-            let p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1));
-
-            let p1_0 = vmull_s8(vget_low_s8(x1_0), vget_low_s8(y1_0));
-            let p1_1 = vmull_s8(vget_high_s8(x1_0), vget_high_s8(y1_0));
-            let p1_2 = vmull_s8(vget_low_s8(x1_1), vget_low_s8(y1_1));
-            let p1_3 = vmull_s8(vget_high_s8(x1_1), vget_high_s8(y1_1));
-
-            let p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1));
-            let p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3));
-            let p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1));
-            let p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3));
+            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(),
             );
-            sumv1 = vmlaq_n_f32(
-                sumv1,
-                vcvtq_f32_s32(vaddq_s32(p2, p3)),
-                x1.d.to_f32() * y1.d.to_f32(),
-            );
         }
-        Ok(vaddvq_f32(sumv0) + vaddvq_f32(sumv1))
+        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 {
@@ -209,30 +183,16 @@ pub(crate) fn vec_dot_q6k_q8k(n: usize, xs: &[BlockQ6K], ys: &[BlockQ8K]) -> Res
                 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));
 
-                // TODO: dotprod
-
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q6bytes_0), vget_high_s8(q8bytes.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q6bytes_1), vget_high_s8(q8bytes.1)),
-                );
+                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_s16(p0) as i32 * scale0 + vaddvq_s16(p1) as i32 * scale1;
+                isum += vaddvq_s32(p0) * scale0 + vaddvq_s32(p1) * scale1;
                 scale = scale.add(2);
 
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_2), vget_low_s8(q8bytes.2)),
-                    vmull_s8(vget_high_s8(q6bytes_2), vget_high_s8(q8bytes.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_3), vget_low_s8(q8bytes.3)),
-                    vmull_s8(vget_high_s8(q6bytes_3), vget_high_s8(q8bytes.3)),
-                );
+                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_s16(p2) as i32 * scale0 + vaddvq_s16(p3) as i32 * scale1;
+                isum += vaddvq_s32(p2) * scale0 + vaddvq_s32(p3) * scale1;
                 scale = scale.add(2);
 
                 let q8bytes = vld1q_s8_x4(q8);
@@ -252,29 +212,16 @@ pub(crate) fn vec_dot_q6k_q8k(n: usize, xs: &[BlockQ6K], ys: &[BlockQ8K]) -> Res
                 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));
 
-                // TODO: dotprod case.
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q6bytes_0), vget_high_s8(q8bytes.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q6bytes_1), vget_high_s8(q8bytes.1)),
-                );
+                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_s16(p0) as i32 * scale0 + vaddvq_s16(p1) as i32 * scale1;
+                isum += vaddvq_s32(p0) * scale0 + vaddvq_s32(p1) * scale1;
                 scale = scale.add(2);
 
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_2), vget_low_s8(q8bytes.2)),
-                    vmull_s8(vget_high_s8(q6bytes_2), vget_high_s8(q8bytes.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_3), vget_low_s8(q8bytes.3)),
-                    vmull_s8(vget_high_s8(q6bytes_3), vget_high_s8(q8bytes.3)),
-                );
+                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_s16(p2) as i32 * scale0 + vaddvq_s16(p3) as i32 * scale1;
+                isum += vaddvq_s32(p2) * scale0 + vaddvq_s32(p3) * scale1;
                 scale = scale.add(2);
             }
             sum += d_all * y.d * ((isum - 32 * isum_mins) as f32);
@@ -351,28 +298,14 @@ pub(crate) fn vec_dot_q5k_q8k(n: usize, xs: &[BlockQ5K], ys: &[BlockQ8K]) -> Res
                 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));
 
-                // TODO: dotprod
-
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q5bytes_0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q5bytes_0), vget_high_s8(q8bytes.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q5bytes_1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q5bytes_1), vget_high_s8(q8bytes.1)),
-                );
-                sumi += vaddvq_s16(vaddq_s16(p0, p1)) as i32 * *scales as i32;
+                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 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q5bytes_2), vget_low_s8(q8bytes.2)),
-                    vmull_s8(vget_high_s8(q5bytes_2), vget_high_s8(q8bytes.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q5bytes_3), vget_low_s8(q8bytes.3)),
-                    vmull_s8(vget_high_s8(q5bytes_3), vget_high_s8(q8bytes.3)),
-                );
-                sumi += vaddvq_s16(vaddq_s16(p2, p3)) as i32 * *scales as i32;
+                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;
@@ -435,22 +368,15 @@ pub(crate) fn vec_dot_q4k_q8k(n: usize, xs: &[BlockQ4K], ys: &[BlockQ8K]) -> Res
             for j in 0..QK_K / 64 {
                 let q4bits = vld1q_u8_x2(q4);
                 q4 = q4.add(32);
-                // TODO: dotprod
                 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 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q4bytes.0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q4bytes.0), vget_high_s8(q8bytes.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q4bytes.1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q4bytes.1), vget_high_s8(q8bytes.1)),
-                );
-                sumi1 += vaddvq_s16(vaddq_s16(p0, p1)) as i32 * scales[2 * j] as i32;
+                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);
@@ -458,15 +384,9 @@ pub(crate) fn vec_dot_q4k_q8k(n: usize, xs: &[BlockQ4K], ys: &[BlockQ8K]) -> Res
                     vreinterpretq_s8_u8(vshrq_n_u8(q4bits.0, 4)),
                     vreinterpretq_s8_u8(vshrq_n_u8(q4bits.1, 4)),
                 );
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q4bytes.0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q4bytes.0), vget_high_s8(q8bytes.0)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q4bytes.1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q4bytes.1), vget_high_s8(q8bytes.1)),
-                );
-                sumi2 += vaddvq_s16(vaddq_s16(p2, p3)) as i32 * scales[2 * j + 1] as i32;
+                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;
         }
@@ -544,27 +464,14 @@ pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Res
                     vreinterpretq_s8_u8(q3h_3),
                 );
 
-                // TODO: dotprod
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_0), vget_low_s8(q8bytes_1.0)),
-                    vmull_s8(vget_high_s8(q3bytes_0), vget_high_s8(q8bytes_1.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_1), vget_low_s8(q8bytes_1.1)),
-                    vmull_s8(vget_high_s8(q3bytes_1), vget_high_s8(q8bytes_1.1)),
-                );
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_2), vget_low_s8(q8bytes_1.2)),
-                    vmull_s8(vget_high_s8(q3bytes_2), vget_high_s8(q8bytes_1.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_3), vget_low_s8(q8bytes_1.3)),
-                    vmull_s8(vget_high_s8(q3bytes_3), vget_high_s8(q8bytes_1.3)),
-                );
-                isum += vaddvq_s16(p0) as i32 * *scale as i32
-                    + vaddvq_s16(p1) as i32 * *scale.add(1) as i32
-                    + vaddvq_s16(p2) as i32 * *scale.add(2) as i32
-                    + vaddvq_s16(p3) as i32 * *scale.add(3) as i32;
+                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);
@@ -589,27 +496,14 @@ pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Res
                     vreinterpretq_s8_u8(q3h_3),
                 );
 
-                // TODO: dotprod
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_0), vget_low_s8(q8bytes_2.0)),
-                    vmull_s8(vget_high_s8(q3bytes_0), vget_high_s8(q8bytes_2.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_1), vget_low_s8(q8bytes_2.1)),
-                    vmull_s8(vget_high_s8(q3bytes_1), vget_high_s8(q8bytes_2.1)),
-                );
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_2), vget_low_s8(q8bytes_2.2)),
-                    vmull_s8(vget_high_s8(q3bytes_2), vget_high_s8(q8bytes_2.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_3), vget_low_s8(q8bytes_2.3)),
-                    vmull_s8(vget_high_s8(q3bytes_3), vget_high_s8(q8bytes_2.3)),
-                );
-                isum += vaddvq_s16(p0) as i32 * *scale as i32
-                    + vaddvq_s16(p1) as i32 * *scale.add(1) as i32
-                    + vaddvq_s16(p2) as i32 * *scale.add(2) as i32
-                    + vaddvq_s16(p3) as i32 * *scale.add(3) as i32;
+                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 {
@@ -667,7 +561,6 @@ pub(crate) fn vec_dot_q2k_q8k(n: usize, xs: &[BlockQ2K], ys: &[BlockQ8K]) -> Res
             let mut is = 0usize;
 
             // TODO: dotprod
-
             for _j in 0..QK_K / 128 {
                 let q2bits = vld1q_u8_x2(q2);
                 q2 = q2.add(32);
@@ -714,14 +607,7 @@ unsafe fn multiply_accum_with_scale(
     q2bytes: int8x16x2_t,
     q8bytes: int8x16x2_t,
 ) -> i32 {
-    let p1 = vaddq_s16(
-        vmull_s8(vget_low_s8(q2bytes.0), vget_low_s8(q8bytes.0)),
-        vmull_s8(vget_high_s8(q2bytes.0), vget_high_s8(q8bytes.0)),
-    );
-    let p2 = vaddq_s16(
-        vmull_s8(vget_low_s8(q2bytes.1), vget_low_s8(q8bytes.1)),
-        vmull_s8(vget_high_s8(q2bytes.1), vget_high_s8(q8bytes.1)),
-    );
-    vaddvq_s16(p1) as i32 * aux[is + index] as i32
-        + vaddvq_s16(p2) as i32 * aux[is + 1 + index] as 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

@@ -17,7 +17,7 @@ pub(super) fn group_for_quantization<'a, 'b, T: super::k_quants::GgmlType>(
     let expected_blocks = xs.len() / block_size;
     let actual_blocks = ys.len();
 
-    //validate that the input is the right size
+    // 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!")
     }
@@ -37,12 +37,12 @@ pub(super) fn group_for_dequantization<'a, 'b, T: super::k_quants::GgmlType>(
 
     let actual_output_len = ys.len();
     let expected_output_len = xs.len() * block_size;
-    //validate that the output is the right 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
+    // Zip the blocks and outputs together
     Ok(xs.iter().zip(ys.chunks_exact_mut(block_size)).collect())
 }
 

candle-core/src/safetensors.rs

@@ -78,11 +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())?)
     }
@@ -255,6 +251,134 @@ 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 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,
@@ -267,7 +391,7 @@ 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/shape.rs

@@ -203,7 +203,7 @@ impl Shape {
 
     /// 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(crate) fn broadcast_shape_binary_op(&self, rhs: &Self, op: &'static str) -> Result<Shape> {
+    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();
@@ -444,6 +444,18 @@ impl<D1: Dim, D2: Dim, D3: Dim, D4: Dim, D5: Dim> Dims for (D1, D2, D3, D4, D5)
     }
 }
 
+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));
@@ -466,23 +478,6 @@ extract_dims!(
     (usize, usize, usize, usize, usize)
 );
 
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn stride() {
-        let shape = Shape::from(());
-        assert_eq!(shape.stride_contiguous(), Vec::<usize>::new());
-        let shape = Shape::from(42);
-        assert_eq!(shape.stride_contiguous(), [1]);
-        let shape = Shape::from((42, 1337));
-        assert_eq!(shape.stride_contiguous(), [1337, 1]);
-        let shape = Shape::from((299, 792, 458));
-        assert_eq!(shape.stride_contiguous(), [458 * 792, 458, 1]);
-    }
-}
-
 pub trait ShapeWithOneHole {
     fn into_shape(self, el_count: usize) -> Result<Shape>;
 }
@@ -499,154 +494,136 @@ impl ShapeWithOneHole for ((),) {
     }
 }
 
+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;
-        if el_count % d1 != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d1}")
-        }
-        Ok((el_count / d1, d1).into())
+        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;
-        if el_count % d1 != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d1}")
-        }
-        Ok((d1, el_count / d1).into())
+        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;
-        let d = d1 * d2;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((el_count / d, d1, d2).into())
+        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;
-        let d = d1 * d2;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((d1, el_count / d, d2).into())
+        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;
-        let d = d1 * d2;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((d1, d2, el_count / d).into())
+        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 = d1 * d2 * d3;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((el_count / d, d1, d2, d3).into())
+        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 = d1 * d2 * d3;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((d1, el_count / d, d2, d3).into())
+        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 = d1 * d2 * d3;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((d1, d2, el_count / d, d3).into())
+        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 = d1 * d2 * d3;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((d1, d2, d3, el_count / d).into())
+        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 = d1 * d2 * d3 * d4;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((el_count / d, d1, d2, d3, d4).into())
+        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 = d1 * d2 * d3 * d4;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((d1, el_count / d, d2, d3, d4).into())
+        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 = d1 * d2 * d3 * d4;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((d1, d2, el_count / d, d3, d4).into())
+        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 = d1 * d2 * d3 * d4;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((d1, d2, d3, el_count / d, d4).into())
+        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 = d1 * d2 * d3 * d4;
-        if el_count % d != 0 {
-            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
-        }
-        Ok((d1, d2, d3, d4, el_count / d).into())
+        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
+        Ok((d1, d2, d3, d4, d).into())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn stride() {
+        let shape = Shape::from(());
+        assert_eq!(shape.stride_contiguous(), Vec::<usize>::new());
+        let shape = Shape::from(42);
+        assert_eq!(shape.stride_contiguous(), [1]);
+        let shape = Shape::from((42, 1337));
+        assert_eq!(shape.stride_contiguous(), [1337, 1]);
+        let shape = Shape::from((299, 792, 458));
+        assert_eq!(shape.stride_contiguous(), [458 * 792, 458, 1]);
     }
 }

candle-core/src/storage.rs

@@ -1,6 +1,6 @@
 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::{CpuStorage, CudaStorage, DType, Device, Error, Layout, MetalStorage, Result, Shape};
 
 // 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 +8,7 @@ use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, Result, Shape
 pub enum Storage {
     Cpu(CpuStorage),
     Cuda(CudaStorage),
+    Metal(MetalStorage),
 }
 
 impl Storage {
@@ -18,6 +19,10 @@ 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 +30,7 @@ impl Storage {
         match self {
             Self::Cpu(_) => Device::Cpu,
             Self::Cuda(storage) => Device::Cuda(storage.device().clone()),
+            Self::Metal(storage) => Device::Metal(storage.device().clone()),
         }
     }
 
@@ -32,6 +38,7 @@ impl Storage {
         match self {
             Self::Cpu(storage) => storage.dtype(),
             Self::Cuda(storage) => storage.dtype(),
+            Self::Metal(storage) => storage.dtype(),
         }
     }
 
@@ -65,6 +72,10 @@ 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))
+            }
         }
     }
 
@@ -78,6 +89,10 @@ impl Storage {
                 let storage = storage.powf(layout, alpha)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.powf(layout, alpha)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -91,6 +106,10 @@ 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))
+            }
         }
     }
 
@@ -112,6 +131,10 @@ 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.
@@ -135,6 +158,10 @@ 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))
+            }
         }
     }
 
@@ -148,6 +175,10 @@ 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))
+            }
         }
     }
 
@@ -161,6 +192,10 @@ 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))
+            }
         }
     }
 
@@ -181,6 +216,10 @@ 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!(),
         }
     }
@@ -205,6 +244,10 @@ 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!(),
         }
     }
@@ -219,6 +262,10 @@ 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))
+            }
         }
     }
 
@@ -239,6 +286,10 @@ 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.
@@ -270,6 +321,10 @@ 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(),
@@ -279,6 +334,33 @@ 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))
+            }
+            (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
+                lhs: lhs.device().location(),
+                rhs: rhs.device().location(),
+                op: "conv-transpose1d",
+            }
+            .bt()),
+        }
+    }
+
     pub(crate) fn conv2d(
         &self,
         l: &Layout,
@@ -297,6 +379,10 @@ 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(),
@@ -324,6 +410,10 @@ impl Storage {
                 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(),
@@ -348,6 +438,10 @@ 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))
+            }
         }
     }
 
@@ -366,6 +460,27 @@ 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))
+            }
         }
     }
 
@@ -379,6 +494,10 @@ 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))
+            }
         }
     }
 
@@ -402,6 +521,10 @@ 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(),
@@ -428,6 +551,10 @@ 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!(),
         }
     }
@@ -452,6 +579,10 @@ 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!(),
         }
     }
@@ -476,6 +607,10 @@ 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!(),
         }
     }
@@ -497,6 +632,10 @@ 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(),
@@ -524,6 +663,10 @@ 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(),
@@ -543,6 +686,9 @@ 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(),

candle-core/src/tensor.rs

@@ -1,4 +1,4 @@
-//! Tensors are N-dimenional matrixes of elements using a single data type.
+//! Tensors are N-dimensional matrixes of elements using a single data type.
 #![allow(clippy::redundant_closure_call)]
 use crate::backend::{BackendDevice, BackendStorage};
 use crate::op::{
@@ -6,7 +6,7 @@ use crate::op::{
 };
 use crate::scalar::TensorOrScalar;
 use crate::shape::{Dim, Dims};
-use crate::{storage::Storage, DType, Device, Error, Layout, Result, Shape};
+use crate::{bail, storage::Storage, DType, Device, Error, Layout, Result, Shape};
 use std::sync::{Arc, RwLock};
 
 /// Unique identifier for tensors.
@@ -105,6 +105,28 @@ macro_rules! binary_op {
     };
 }
 
+macro_rules! binary_op_scalar {
+    ($fn_name:ident, $op_name:ident) => {
+        pub fn $fn_name<T: TensorOrScalar>(&self, rhs: T) -> Result<Self> {
+            let rhs = match rhs.to_tensor_scalar()? {
+                crate::scalar::TensorScalar::Tensor(rhs) => rhs,
+                crate::scalar::TensorScalar::Scalar(rhs) => rhs
+                    .to_dtype(self.dtype())?
+                    .to_device(self.device())?
+                    .broadcast_as(self.shape())?,
+            };
+            let shape = self.same_shape_binary_op(&rhs, stringify!($fn_name))?;
+            let storage = self.storage().binary_impl::<crate::op::$op_name>(
+                &*rhs.storage(),
+                self.layout(),
+                rhs.layout(),
+            )?;
+            let op = BackpropOp::new2(self, &rhs, |t1, t2| Op::Binary(t1, t2, BinaryOp::$op_name));
+            Ok(from_storage(storage, shape.clone(), op, false))
+        }
+    };
+}
+
 macro_rules! broadcast_binary_op {
     ($fn_name:ident, $inner_fn_name:ident) => {
         pub fn $fn_name(&self, rhs: &Self) -> Result<Self> {
@@ -155,14 +177,9 @@ impl Tensor {
         is_variable: bool,
     ) -> Result<Self> {
         let none = BackpropOp::none();
-        if is_variable {
-            let shape = shape.into();
-            let storage = device.ones(&shape, dtype)?;
-            Ok(from_storage(storage, shape, none, is_variable))
-        } else {
-            let storage = device.ones(&crate::shape::SCALAR, dtype)?;
-            from_storage(storage, crate::shape::SCALAR, none, is_variable).broadcast_as(shape)
-        }
+        let shape = shape.into();
+        let storage = device.ones(&shape, dtype)?;
+        Ok(from_storage(storage, shape, none, is_variable))
     }
 
     /// Creates a new tensor filled with ones.
@@ -200,14 +217,9 @@ impl Tensor {
         is_variable: bool,
     ) -> Result<Self> {
         let none = BackpropOp::none();
-        if is_variable {
-            let shape = shape.into();
-            let storage = device.zeros(&shape, dtype)?;
-            Ok(from_storage(storage, shape, none, is_variable))
-        } else {
-            let storage = device.zeros(&crate::shape::SCALAR, dtype)?;
-            from_storage(storage, crate::shape::SCALAR, none, is_variable).broadcast_as(shape)
-        }
+        let shape = shape.into();
+        let storage = device.zeros(&shape, dtype)?;
+        Ok(from_storage(storage, shape, none, is_variable))
     }
 
     /// Creates a new tensor filled with zeros.
@@ -349,6 +361,16 @@ impl Tensor {
         Self::new_impl(array, shape, device, false)
     }
 
+    /// Returns a new tensor with all the elements having the same specified value. Note that
+    /// the tensor is not contiguous so you would have to call `.contiguous()` on it if needed.
+    pub fn full<D: crate::WithDType, S: Into<Shape>>(
+        value: D,
+        shape: S,
+        device: &Device,
+    ) -> Result<Self> {
+        Self::from_vec_impl(vec![value], (), device, false)?.broadcast_as(shape)
+    }
+
     /// Creates a new 1D tensor from an iterator.
     pub fn from_iter<D: crate::WithDType>(
         iter: impl IntoIterator<Item = D>,
@@ -373,11 +395,21 @@ impl Tensor {
         step: D,
         device: &Device,
     ) -> Result<Self> {
+        if D::is_zero(&step) {
+            bail!("step cannot be zero")
+        }
         let mut data = vec![];
         let mut current = start;
-        while current < end {
-            data.push(current);
-            current += step;
+        if step >= D::zero() {
+            while current < end {
+                data.push(current);
+                current += step;
+            }
+        } else {
+            while current > end {
+                data.push(current);
+                current += step;
+            }
         }
         let len = data.len();
         Self::from_vec_impl(data, len, device, false)
@@ -437,7 +469,7 @@ impl Tensor {
 
     /// Returns true if the computation graph should track this op, that is if it is
     /// a variable or if it has some variable as dependencies.
-    pub(crate) fn track_op(&self) -> bool {
+    pub fn track_op(&self) -> bool {
         self.is_variable || self.op.is_some()
     }
 
@@ -447,14 +479,20 @@ impl Tensor {
     binary_op!(mul, Mul);
     binary_op!(sub, Sub);
     binary_op!(div, Div);
-    binary_op!(maximum, Maximum);
-    binary_op!(minimum, Minimum);
+    binary_op_scalar!(maximum, Maximum);
+    binary_op_scalar!(minimum, Minimum);
     broadcast_binary_op!(broadcast_add, add);
     broadcast_binary_op!(broadcast_mul, mul);
     broadcast_binary_op!(broadcast_sub, sub);
     broadcast_binary_op!(broadcast_div, div);
     broadcast_binary_op!(broadcast_maximum, maximum);
     broadcast_binary_op!(broadcast_minimum, minimum);
+    broadcast_binary_op!(broadcast_eq, eq);
+    broadcast_binary_op!(broadcast_ne, ne);
+    broadcast_binary_op!(broadcast_lt, lt);
+    broadcast_binary_op!(broadcast_le, le);
+    broadcast_binary_op!(broadcast_gt, gt);
+    broadcast_binary_op!(broadcast_ge, ge);
 
     unary_op!(recip, Recip);
     unary_op!(neg, Neg);
@@ -467,7 +505,22 @@ impl Tensor {
     unary_op!(sqr, Sqr);
     unary_op!(sqrt, Sqrt);
     unary_op!(gelu, Gelu);
+    unary_op!(gelu_erf, GeluErf);
+    unary_op!(erf, Erf);
     unary_op!(relu, Relu);
+    unary_op!(silu, Silu);
+    unary_op!(ceil, Ceil);
+    unary_op!(floor, Floor);
+    unary_op!(round, Round);
+
+    /// Round element of the input tensor to the nearest integer.
+    ///
+    /// If the number of decimals is negative, it specifies the number of positions to the left of
+    /// the decimal point.
+    pub fn round_to(&self, decimals: i32) -> Result<Self> {
+        let mult = 10f64.powi(decimals);
+        (self * mult)?.round()? * (1f64 / mult)
+    }
 
     /// Retrieves the single scalar value hold in the tensor. If the tensor contains multiple
     /// dimensions, an error is returned instead.
@@ -487,6 +540,7 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(cpu_storage) => from_cpu_storage(cpu_storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
+            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -514,6 +568,73 @@ impl Tensor {
         Ok(inp)
     }
 
+    /// Creates grids of coordinates specified by the 1D inputs.
+    ///
+    /// # Arguments
+    ///
+    /// * `args` - A slice of 1D tensors.
+    /// * `xy_indexing` - Whether to use xy indexing or ij indexing. If xy is selected, the
+    /// first dimension corresponds to the cardinality of the second input and the second
+    /// dimension corresponds to the cardinality of the first input. If ij is selected, the
+    /// dimensions are in the same order as the cardinality of the inputs.
+    ///
+    /// # Examples
+    ///
+    /// ```rust
+    /// use candle_core::{Tensor, Device, Shape};
+    /// let x = Tensor::new(&[1f32, 2., 3.], &Device::Cpu)?;
+    /// let y = Tensor::new(&[4f32, 5., 6.], &Device::Cpu)?;
+    ///
+    /// let grids_xy = Tensor::meshgrid(&[&x, &y], true)?;
+    ///
+    /// assert_eq!(grids_xy.len(), 2);
+    /// assert_eq!(grids_xy[0].dims(), &[3, 3]);
+    ///
+    /// assert_eq!(grids_xy[0].to_vec2::<f32>()?, &[[1., 2., 3.], [1., 2., 3.], [1., 2., 3.]]);
+    /// assert_eq!(grids_xy[1].to_vec2::<f32>()?, &[[4., 4., 4.], [5., 5., 5.], [6., 6., 6.]]);
+    ///
+    /// let grids_ij = Tensor::meshgrid(&[&x, &y], false)?;
+    ///
+    /// assert_eq!(grids_ij[0].to_vec2::<f32>()?, &[[1., 1., 1.], [2., 2., 2.], [3., 3., 3.]]);
+    /// assert_eq!(grids_ij[1].to_vec2::<f32>()?, &[[4., 5., 6.], [4., 5., 6.], [4., 5., 6.]]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    ///
+    /// # Errors
+    ///
+    /// * Will return `Err` if `args` contains less than 2 tensors.
+    ///
+    pub fn meshgrid<A: AsRef<Tensor>>(args: &[A], xy_indexing: bool) -> Result<Vec<Self>> {
+        if args.len() <= 1 {
+            Err(Error::OpRequiresAtLeastTwoTensors { op: "meshgrid" }.bt())?
+        }
+        let args: Vec<_> = if xy_indexing {
+            args.iter().rev().collect()
+        } else {
+            args.iter().collect()
+        };
+
+        let mut shape = Vec::with_capacity(args.len());
+        for arg in args.iter() {
+            shape.push(arg.as_ref().dims1()?)
+        }
+
+        let mut grids = Vec::with_capacity(args.len());
+        for idx in 0..args.len() {
+            let mut ones = vec![1usize; args.len()];
+            ones[idx] = shape[idx];
+            let arg = args[idx].as_ref().reshape(ones)?;
+            let mut repeats = shape.clone();
+            repeats[idx] = 1;
+            let repeated_tensor = arg.repeat(repeats)?;
+            grids.push(repeated_tensor);
+        }
+        if xy_indexing {
+            grids.reverse();
+        }
+        Ok(grids)
+    }
+
     /// This operation multiplies the input tensor by `mul` then adds `add` and return the result.
     /// The input values `mul` and `add` are casted to the appropriate type so some rounding might
     /// be performed.
@@ -559,7 +680,7 @@ impl Tensor {
     }
 
     /// Split a tensor into the specified number of chunks, this may return less chunks than
-    /// specificed.
+    /// specified.
     pub fn chunk<D: Dim>(&self, chunks: usize, dim: D) -> Result<Vec<Self>> {
         let dim = dim.to_index(self.shape(), "chunk")?;
         let size = self.dim(dim)?;
@@ -589,15 +710,23 @@ impl Tensor {
     pub fn narrow<D: Dim>(&self, dim: D, start: usize, len: usize) -> Result<Self> {
         let dims = self.dims();
         let dim = dim.to_index(self.shape(), "narrow")?;
-        if start + len > dims[dim] {
-            Err(Error::NarrowInvalidArgs {
-                shape: self.shape().clone(),
-                dim,
-                start,
-                len,
-                msg: "start + len > dim_len",
-            }
-            .bt())?
+        let err = |msg| {
+            Err::<(), _>(
+                Error::NarrowInvalidArgs {
+                    shape: self.shape().clone(),
+                    dim,
+                    start,
+                    len,
+                    msg,
+                }
+                .bt(),
+            )
+        };
+        if start > dims[dim] {
+            err("start > dim_len")?
+        }
+        if start.saturating_add(len) > dims[dim] {
+            err("start + len > dim_len")?
         }
         if start == 0 && dims[dim] == len {
             Ok(self.clone())
@@ -644,7 +773,12 @@ impl Tensor {
         let storage = self.storage().reduce_op(op, self.layout(), &[dim])?;
         let mut dims = self.dims().to_vec();
         dims[dim] = 1;
-        let op = BackpropOp::new1(self, |arg| Op::Reduce(arg, op, dims.to_vec()));
+        let op = match op {
+            ReduceOp::Sum | ReduceOp::Min | ReduceOp::Max => {
+                BackpropOp::new1(self, |arg| Op::Reduce(arg, op, dims.to_vec()))
+            }
+            ReduceOp::ArgMin | ReduceOp::ArgMax => BackpropOp::none(),
+        };
         let res = from_storage(storage, dims, op, false);
         if keepdim {
             Ok(res)
@@ -671,6 +805,35 @@ impl Tensor {
         }
     }
 
+    /// Roll the tensor input along the given dimension.
+    /// Elements that are shifted beyond the last position are re-introduced at the first position.
+    ///
+    /// ```rust
+    /// # use candle_core::{Tensor, Device};
+    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
+    /// let tensor = tensor.roll(1, 0)?;
+    /// assert_eq!(tensor.to_vec2::<f32>()?, &[[4., 5.], [0., 1.], [2., 3.]]);
+    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
+    /// let tensor = tensor.roll(-1, 0)?;
+    /// assert_eq!(tensor.to_vec2::<f32>()?, &[[2., 3.], [4., 5.], [0., 1.]]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn roll<D>(&self, shift: i32, dim: D) -> Result<Self>
+    where
+        D: Dim + Clone,
+    {
+        let dim = dim.to_index(self.shape(), "roll")?;
+        let dim_size = self.dim(dim)?;
+        let shift = shift.rem_euclid(dim_size as i32) as usize;
+        if shift == 0 {
+            Ok(self.clone())
+        } else {
+            let a = self.narrow(dim, 0, dim_size - shift)?;
+            let b = self.narrow(dim, dim_size - shift, shift)?;
+            Tensor::cat(&[&b, &a], dim)
+        }
+    }
+
     /// Returns the sum of all elements in the input tensor. The sum is performed over all the
     /// input dimensions.
     ///
@@ -733,6 +896,20 @@ impl Tensor {
         self.sum_impl(mean_dims, false)? * scale
     }
 
+    /// Returns the unbiased variance over the selected dimension.
+    pub fn var_keepdim<D: Dim>(&self, dim: D) -> Result<Self> {
+        let dim = dim.to_index(self.shape(), "var")?;
+        let mean = self.mean_keepdim(dim)?;
+        let squares = self.broadcast_sub(&mean)?.sqr()?;
+        squares.sum_impl(dim, true)? / (self.dim(dim)? - 1) as f64
+    }
+
+    /// Returns the unbiased variance over the selected dimension.
+    pub fn var<D: Dim>(&self, dim: D) -> Result<Self> {
+        let dim = dim.to_index(self.shape(), "var")?;
+        self.var_keepdim(dim)?.squeeze(dim)
+    }
+
     /// Gathers the maximum value across the selected dimension. The resulting shape has the same
     /// number of dimensions as the original tensor and the select dimension has a single element.
     pub fn max_keepdim<D: Dim>(&self, dim: D) -> Result<Self> {
@@ -827,20 +1004,52 @@ impl Tensor {
         self.cmp(rhs, CmpOp::Le)
     }
 
-    /// Upsample the input tensor to the `(target_h, target_w)` size, taking the value of the
+    /// Clamp the tensor values to be between `min` and `max`.
+    pub fn clamp<T1: TensorOrScalar, T2: TensorOrScalar>(&self, min: T1, max: T2) -> Result<Self> {
+        self.maximum(min)?.minimum(max)
+    }
+
+    /// Interpolate the input tensor to the `target_size` size, taking the value of the nearest element.
+    ///
+    /// The input tensor should have three dimensions, `(batch, channels, l)`, the returned
+    /// tensor also has three dimensions, `(batch, channels, target_size)`.
+    pub fn interpolate1d(&self, target_size: usize) -> Result<Self> {
+        let (n, c, _l) = self.dims3()?;
+        let op = BackpropOp::new1(self, Op::UpsampleNearest1D);
+        let storage = self
+            .storage()
+            .upsample_nearest1d(self.layout(), target_size)?;
+        Ok(from_storage(storage, (n, c, target_size), op, false))
+    }
+
+    /// Alias for `interpolate1d`.
+    pub fn upsample_nearest1d(&self, target_size: usize) -> Result<Self> {
+        self.interpolate1d(target_size)
+    }
+
+    /// Interpolate the input tensor to the `(target_h, target_w)` size, taking the value of the
     /// nearest element.
     ///
     /// The input tensor should have four dimensions, `(batch, channels, h, w)`, the returned
     /// tensor also has four dimensions, `(batch, channels, target_h, target_w)`.
-    pub fn upsample_nearest2d(&self, target_h: usize, target_w: usize) -> Result<Self> {
+    pub fn interpolate2d(&self, target_h: usize, target_w: usize) -> Result<Self> {
         let (n, c, _h, _w) = self.dims4()?;
-        let op = BackpropOp::new1(self, Op::UpsampleNearest2D);
+        let op = BackpropOp::new1(self, |arg| Op::UpsampleNearest2D {
+            arg,
+            target_h,
+            target_w,
+        });
         let storage = self
             .storage()
             .upsample_nearest2d(self.layout(), target_h, target_w)?;
         Ok(from_storage(storage, (n, c, target_h, target_w), op, false))
     }
 
+    /// Alias for `interpolate2d`.
+    pub fn upsample_nearest2d(&self, target_h: usize, target_w: usize) -> Result<Self> {
+        self.interpolate2d(target_h, target_w)
+    }
+
     /// 2D average pooling over an input tensor with multiple channels.
     ///
     /// The input tensor should have four dimensions, `(batch, channels, h, w)`, the returned
@@ -862,6 +1071,9 @@ impl Tensor {
         let kernel_size = kernel_size.to_usize2();
         let stride = stride.to_usize2();
         let (n, c, h, w) = self.dims4()?;
+        if h < kernel_size.0 || w < kernel_size.1 {
+            bail!("kernel-size {kernel_size:?} is larger than the input size {h},{w}")
+        }
         // https://pytorch.org/docs/stable/generated/torch.nn.AvgPool2d.html#torch.nn.AvgPool2d
         let h_out = (h - kernel_size.0) / stride.0 + 1;
         let w_out = (w - kernel_size.1) / stride.1 + 1;
@@ -897,6 +1109,9 @@ impl Tensor {
         let kernel_size = kernel_size.to_usize2();
         let stride = stride.to_usize2();
         let (n, c, h, w) = self.dims4()?;
+        if h < kernel_size.0 || w < kernel_size.1 {
+            bail!("kernel-size {kernel_size:?} is larger than the input size {h},{w}")
+        }
         // https://pytorch.org/docs/stable/generated/torch.nn.MaxPool2d.html#torch.nn.MaxPool2d
         let h_out = (h - kernel_size.0) / stride.0 + 1;
         let w_out = (w - kernel_size.1) / stride.1 + 1;
@@ -1052,14 +1267,16 @@ impl Tensor {
                 op: "scatter-add (self, src)",
                 lhs: self.shape().clone(),
                 rhs: source.shape().clone(),
-            })?
+            }
+            .bt())?
         }
         if indexes.dims() != source.dims() {
             Err(Error::ShapeMismatchBinaryOp {
                 op: "scatter-add (indexes, src)",
                 lhs: indexes.shape().clone(),
                 rhs: source.shape().clone(),
-            })?
+            }
+            .bt())?
         }
         let storage = self.storage().scatter_add(
             self.layout(),
@@ -1075,6 +1292,75 @@ impl Tensor {
         Ok(from_storage(storage, self.shape(), op, false))
     }
 
+    /// Embeds the values of the `src` tensor into the `self` tensor on the specified dimension.
+    pub fn slice_scatter<D: Dim>(&self, src: &Self, dim: D, start: usize) -> Result<Self> {
+        let dim = dim.to_index(self.shape(), "slice-scatter")?;
+        if dim == 0 {
+            self.slice_scatter0(src, start)
+        } else {
+            // TODO: Maybe we want to add a more efficient implementation at some point.
+            self.transpose(0, dim)?
+                .slice_scatter0(&src.transpose(0, dim)?, start)?
+                .transpose(0, dim)
+        }
+    }
+
+    /// Embeds the values of the `src` tensor into the `self` tensor on the first dimension.
+    pub fn slice_scatter0(&self, src: &Self, start: usize) -> Result<Self> {
+        if self.dtype() != src.dtype() {
+            Err(Error::DTypeMismatchBinaryOp {
+                lhs: self.dtype(),
+                rhs: src.dtype(),
+                op: "slice-scatter",
+            }
+            .bt())?
+        }
+        if self.device().location() != src.device.location() {
+            Err(Error::DeviceMismatchBinaryOp {
+                lhs: self.device().location(),
+                rhs: src.device().location(),
+                op: "slice-scatter",
+            }
+            .bt())?
+        }
+        if self.rank() != src.rank() {
+            Err(Error::UnexpectedNumberOfDims {
+                expected: self.rank(),
+                got: src.rank(),
+                shape: src.shape().clone(),
+            }
+            .bt())?
+        }
+        let shape_ok =
+            self.dims()
+                .iter()
+                .zip(src.dims().iter())
+                .enumerate()
+                .all(|(dim_idx, (&d1, &d2))| {
+                    if 0 == dim_idx {
+                        d2 + start <= d1
+                    } else {
+                        d1 == d2
+                    }
+                });
+        if !shape_ok {
+            Err(Error::ShapeMismatchBinaryOp {
+                op: "slice-scatter (self, src)",
+                lhs: self.shape().clone(),
+                rhs: src.shape().clone(),
+            }
+            .bt())?
+        }
+        let mut storage = self.device().zeros(self.shape(), self.dtype())?;
+        self.storage()
+            .copy_strided_src(&mut storage, 0, self.layout())?;
+        let offset = start * src.dims()[1..].iter().product::<usize>();
+        src.storage()
+            .copy_strided_src(&mut storage, offset, src.layout())?;
+        let op = BackpropOp::new2(self, src, |t1, t2| Op::SliceScatter0(t1, t2, start));
+        Ok(from_storage(storage, self.shape(), op, false))
+    }
+
     /// Accumulate element from `source` at indexes `indexes` and add them to `self`.
     pub fn index_add<D: Dim>(&self, indexes: &Self, source: &Self, dim: D) -> Result<Self> {
         let dim = dim.to_index(self.shape(), "index-add")?;
@@ -1097,7 +1383,8 @@ impl Tensor {
                 op: "index-add (self, source)",
                 lhs: self.shape().clone(),
                 rhs: source.shape().clone(),
-            })?
+            }
+            .bt())?
         }
         // The number of element in indexes must match the dimension on which the add is
         // performed on the source tensor (and the index values from `indexes` are taken from
@@ -1108,7 +1395,8 @@ impl Tensor {
                 op: "index-add (ids, source))",
                 lhs: indexes.shape().clone(),
                 rhs: source.shape().clone(),
-            })?
+            }
+            .bt())?
         }
         let storage = self.storage().index_add(
             self.layout(),
@@ -1156,7 +1444,8 @@ impl Tensor {
                 op: "gather",
                 lhs: self.shape().clone(),
                 rhs: indexes.shape().clone(),
-            })?
+            }
+            .bt())?
         }
         let storage =
             self.storage()
@@ -1230,6 +1519,7 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(storage) => from_cpu_storage(storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
+            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -1260,6 +1550,7 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(storage) => from_cpu_storage(storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
+            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -1300,6 +1591,7 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(storage) => from_cpu_storage(storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
+            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -1463,6 +1755,24 @@ impl Tensor {
         }
     }
 
+    /// Returns the sub-tensor fixing the index at `index` on the dimension `dim`.
+    ///
+    /// ```rust
+    /// use candle_core::{Tensor, Device};
+    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
+    /// let t = tensor.get_on_dim(1, 0)?;
+    /// assert_eq!(t.to_vec1::<f32>()?, &[0., 2., 4.]);
+    /// let t = tensor.get_on_dim(1, 1)?;
+    /// assert_eq!(t.to_vec1::<f32>()?, &[1., 3., 5.]);
+    /// let t = tensor.get_on_dim(0, 1)?;
+    /// assert_eq!(t.to_vec1::<f32>()?, &[2., 3.]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn get_on_dim<D: Dim>(&self, dim: D, index: usize) -> Result<Tensor> {
+        let dim = dim.to_index(self.shape(), "get_on_dim")?;
+        self.narrow(dim, index, 1)?.squeeze(dim)
+    }
+
     /// Returns a tensor that is a transposed version of the input, the two last dimensions of the
     /// input are swapped.
     ///
@@ -1491,6 +1801,9 @@ impl Tensor {
     pub fn transpose<D1: Dim, D2: Dim>(&self, dim1: D1, dim2: D2) -> Result<Tensor> {
         let dim1 = dim1.to_index(self.shape(), "transpose")?;
         let dim2 = dim2.to_index(self.shape(), "transpose")?;
+        if dim1 == dim2 {
+            return Ok(self.clone());
+        }
         let op = BackpropOp::new1(self, |t| Op::Transpose(t, dim1, dim2));
         let tensor_ = Tensor_ {
             id: TensorId::new(),
@@ -1521,7 +1834,7 @@ impl Tensor {
         let is_permutation =
             dims.len() == self.rank() && (0..dims.len()).all(|i| dims.contains(&i));
         if !is_permutation {
-            crate::bail!(
+            bail!(
                 "dimension mismatch in permute, tensor {:?}, dims: {:?}",
                 self.dims(),
                 dims
@@ -1568,17 +1881,23 @@ impl Tensor {
 
     /// Returns a new tensor detached from the current graph, gradient are not propagated through
     /// this new node. The storage of this tensor is shared with the initial tensor.
-    pub fn detach(&self) -> Result<Tensor> {
-        let tensor_ = Tensor_ {
-            id: TensorId::new(),
-            storage: self.storage.clone(),
-            layout: self.layout.clone(),
-            op: BackpropOp::none(),
-            is_variable: false,
-            dtype: self.dtype,
-            device: self.device.clone(),
-        };
-        Ok(Tensor(Arc::new(tensor_)))
+    ///
+    /// If the tensor is already detached from the computation graph, the same tensor is returned.
+    pub fn detach(&self) -> Tensor {
+        if self.op.is_none() && !self.is_variable {
+            self.clone()
+        } else {
+            let tensor_ = Tensor_ {
+                id: TensorId::new(),
+                storage: self.storage.clone(),
+                layout: self.layout.clone(),
+                op: BackpropOp::none(),
+                is_variable: false,
+                dtype: self.dtype,
+                device: self.device.clone(),
+            };
+            Tensor(Arc::new(tensor_))
+        }
     }
 
     /// If the target device is the same as the tensor device, only a shallow copy is performed.
@@ -1590,7 +1909,11 @@ impl Tensor {
                 (Storage::Cpu(storage), Device::Cuda(cuda)) => {
                     Storage::Cuda(cuda.storage_from_cpu_storage(storage)?)
                 }
+                (Storage::Cpu(storage), Device::Metal(metal)) => {
+                    Storage::Metal(metal.storage_from_cpu_storage(storage)?)
+                }
                 (Storage::Cuda(storage), Device::Cpu) => Storage::Cpu(storage.to_cpu_storage()?),
+                (Storage::Metal(storage), Device::Cpu) => Storage::Cpu(storage.to_cpu_storage()?),
                 (Storage::Cuda(storage), Device::Cuda(cuda)) => {
                     // TODO: Avoid passing through the cpu storage here, especially if the gpu ids
                     // are the same.
@@ -1598,6 +1921,9 @@ impl Tensor {
                     Storage::Cuda(cuda.storage_from_cpu_storage(&cpu_storage)?)
                 }
                 (Storage::Cpu(storage), Device::Cpu) => Storage::Cpu(storage.clone()),
+                _ => {
+                    bail!("not implemented yet")
+                }
             };
             let op = BackpropOp::new1(self, Op::ToDevice);
             let tensor_ = Tensor_ {
@@ -1852,6 +2178,34 @@ impl Tensor {
         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())?
+                }
+            }
+        }
         if dim == 0 {
             Self::cat0(args)
         } else {
@@ -1969,11 +2323,56 @@ impl Tensor {
         }
     }
 
+    /// Pad the input tensor using same values along dimension `dim`. This adds `left` elements before the
+    /// input tensor values and `right` elements after.
+    pub fn pad_with_same<D: Dim>(&self, dim: D, left: usize, right: usize) -> Result<Self> {
+        if left == 0 && right == 0 {
+            Ok(self.clone())
+        } else if self.elem_count() == 0 {
+            bail!("cannot use pad_with_same on an empty tensor")
+        } else if left == 0 {
+            let dim = dim.to_index(self.shape(), "pad_with_same")?;
+            let r = self.narrow(dim, self.dim(dim)? - 1, 1)?;
+            let mut v = vec![self];
+            for _ in 0..right {
+                v.push(&r)
+            }
+            Tensor::cat(&v, dim)
+        } else if right == 0 {
+            let dim = dim.to_index(self.shape(), "pad_with_same")?;
+            let l = self.narrow(dim, 0, 1)?;
+            let mut v = vec![];
+            for _ in 0..left {
+                v.push(&l)
+            }
+            v.push(self);
+            Tensor::cat(&v, dim)
+        } else {
+            let dim = dim.to_index(self.shape(), "pad_with_same")?;
+            let l = self.narrow(dim, 0, 1)?;
+            let r = self.narrow(dim, self.dim(dim)? - 1, 1)?;
+            let mut v = vec![];
+            for _ in 0..left {
+                v.push(&l)
+            }
+            v.push(self);
+            for _ in 0..right {
+                v.push(&r)
+            }
+            Tensor::cat(&v, dim)
+        }
+    }
+
     /// Run the `forward` method of `m` on `self`.
     pub fn apply<M: crate::Module>(&self, m: &M) -> Result<Self> {
         m.forward(self)
     }
 
+    /// Run the `forward` method of `m` on `self`.
+    pub fn apply_t<M: crate::ModuleT>(&self, m: &M, train: bool) -> Result<Self> {
+        m.forward_t(self, train)
+    }
+
     pub(crate) fn storage(&self) -> std::sync::RwLockReadGuard<'_, Storage> {
         self.storage.read().unwrap()
     }
@@ -2088,6 +2487,142 @@ impl Tensor {
     ) -> Result<Self> {
         self.apply_op3_arc(t2, t3, Arc::new(Box::new(c)))
     }
+
+    /// Normalize a 'relative' axis value: positive values are kept, negative
+    /// values means counting the dimensions from the back.
+    pub fn normalize_axis(&self, axis: i64) -> Result<usize> {
+        let rank = self.rank() as i64;
+        if rank <= axis {
+            bail!("axis {axis} is too large, tensor rank {rank}")
+        } else if 0 <= axis {
+            Ok(axis as usize)
+        } else {
+            let naxis = rank + axis;
+            if naxis < 0 {
+                bail!("axis {axis} is too small, tensor rank {rank}")
+            }
+            Ok(naxis as usize)
+        }
+    }
+
+    /// Returns a lower triangular matrix of ones of size n by n.
+    pub fn tril2(n: usize, dtype: DType, device: &Device) -> Result<Self> {
+        let t = Tensor::arange(0u32, n as u32, device)?;
+        let t1 = t.reshape((1, n))?.broadcast_as((n, n))?;
+        let t2 = t.reshape((n, 1))?.broadcast_as((n, n))?;
+        t1.le(&t2)?.to_dtype(dtype)
+    }
+
+    /// Returns an upper triangular matrix of ones of size n by n.
+    pub fn triu2(n: usize, dtype: DType, device: &Device) -> Result<Self> {
+        let t = Tensor::arange(0u32, n as u32, device)?;
+        let t1 = t.reshape((1, n))?.broadcast_as((n, n))?;
+        let t2 = t.reshape((n, 1))?.broadcast_as((n, n))?;
+        t1.ge(&t2)?.to_dtype(dtype)
+    }
+
+    /// Returns a matrix with a diagonal of ones of size n by n.
+    pub fn eye(n: usize, dtype: DType, device: &Device) -> Result<Self> {
+        let t = Tensor::arange(0u32, n as u32, device)?;
+        let t1 = t.reshape((1, n))?.broadcast_as((n, n))?;
+        let t2 = t.reshape((n, 1))?.broadcast_as((n, n))?;
+        t1.eq(&t2)?.to_dtype(dtype)
+    }
+
+    /// Returns the cumulative sum of elements of the input tensor summed over the specified
+    /// dimension.
+    ///
+    /// This operation is most efficient when dim is the last dimension of the tensor.
+    pub fn cumsum<D: Dim>(&self, dim: D) -> Result<Self> {
+        let dim = dim.to_index(self.shape(), "cumsum")?;
+        let rank = self.rank();
+        if rank == 0 {
+            return Ok(self.clone());
+        }
+        let n_axis = self.dim(dim)?;
+        let triu = Tensor::triu2(n_axis, self.dtype(), self.device())?;
+        if rank == 1 {
+            self.unsqueeze(0)?.matmul(&triu)?.squeeze(0)
+        } else {
+            let last = rank - 1;
+            let t = self.transpose(dim, last)?;
+            let t = t.broadcast_matmul(&triu)?;
+            t.transpose(dim, last)
+        }
+    }
+
+    /// Returns a copy of `self` where the values within `ranges` have been replaced with the
+    /// content of `src`.
+    pub fn slice_assign<D: std::ops::RangeBounds<usize>>(
+        &self,
+        ranges: &[D],
+        src: &Tensor,
+    ) -> Result<Self> {
+        let src_dims = src.dims();
+        let self_dims = self.dims();
+        if self_dims.len() != src_dims.len() {
+            bail!(
+                "slice-assign requires input with the same rank {} <> {}",
+                self_dims.len(),
+                src_dims.len()
+            )
+        }
+        if self_dims.len() != ranges.len() {
+            bail!(
+                "slice-assign requires input with the same rank as there are ranges {} <> {}",
+                self_dims.len(),
+                ranges.len()
+            )
+        }
+        let mut src = src.clone();
+        let mut mask = Self::ones(src.shape(), DType::U8, src.device())?;
+        for (i, range) in ranges.iter().enumerate() {
+            let start_included = match range.start_bound() {
+                std::ops::Bound::Unbounded => 0,
+                std::ops::Bound::Included(v) => *v,
+                std::ops::Bound::Excluded(v) => *v + 1,
+            };
+            let end_excluded = match range.end_bound() {
+                std::ops::Bound::Unbounded => self_dims[i],
+                std::ops::Bound::Included(v) => *v + 1,
+                std::ops::Bound::Excluded(v) => *v,
+            };
+            if end_excluded <= start_included {
+                bail!("slice-assign: empty range for dim {i}, {start_included} {end_excluded}")
+            }
+            if self_dims[i] < end_excluded {
+                bail!(
+                    "slice-assign: upper bound is out of range for dim {i}, {end_excluded} {}",
+                    self_dims[i]
+                )
+            }
+            if end_excluded - start_included != src_dims[i] {
+                bail!(
+                    "slice-assign: the range for dim {i} ({start_included}..{end_excluded}) does not match the size of src {}", src_dims[i]
+                )
+            }
+            src = src.pad_with_zeros(i, start_included, self_dims[i] - end_excluded)?;
+            mask = mask.pad_with_zeros(i, start_included, self_dims[i] - end_excluded)?
+        }
+        mask.where_cond(/* on_true= */ &src, /* on_false= */ self)
+    }
+
+    /// Returns log(sum(exp(tensor), dim)).
+    pub fn log_sum_exp<D: Dims>(&self, sum_dims: D) -> Result<Self> {
+        let exp = self.exp()?;
+        let sum = exp.sum(sum_dims)?;
+        sum.log()
+    }
+
+    /// Pointwise pow operation.
+    pub fn pow(&self, rhs: &Tensor) -> Result<Self> {
+        rhs.mul(&self.log()?)?.exp()
+    }
+
+    /// Broadcasting version of `pow`.
+    pub fn broadcast_pow(&self, rhs: &Tensor) -> Result<Self> {
+        rhs.broadcast_mul(&self.log()?)?.exp()
+    }
 }
 
 macro_rules! bin_trait {

candle-core/src/test_utils.rs

@@ -4,7 +4,7 @@ use crate::{Result, Tensor};
 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)
@@ -15,6 +15,12 @@ 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)?)
+        }
     };
 }
 

candle-core/src/utils.rs

@@ -23,6 +23,10 @@ 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")
 }

candle-core/src/variable.rs

@@ -107,6 +107,10 @@ 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

@@ -13,6 +13,11 @@ 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)
 */
 fn conv1d(dev: &Device) -> Result<()> {
     let t = Tensor::new(
@@ -45,6 +50,15 @@ fn conv1d(dev: &Device) -> Result<()> {
         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 res = t.conv_transpose1d(&w.transpose(0, 1)?, 0, 0, 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
+        ],
+    );
     Ok(())
 }
 
@@ -130,7 +144,7 @@ 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, 1)?;
+    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)?,
@@ -164,7 +178,7 @@ fn conv2d(dev: &Device) -> Result<()> {
     );
 
     // Transpose and dilations.
-    let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 2, 1)?;
+    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)?,
@@ -246,13 +260,13 @@ fn conv2d_small(dev: &Device) -> Result<()> {
             0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000
         ]
     );
-    let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 1, 1)?;
+    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, 1)?;
+    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)?,
@@ -479,17 +493,103 @@ fn conv2d_grad(dev: &Device) -> Result<()> {
             ]
         ]
     );
+
+    // 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]]
+        ]
+    );
+
     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!(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_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
 );
-test_device!(conv2d_small, conv2d_small_cpu, conv2d_small_gpu);
-test_device!(conv2d_smaller, conv2d_smaller_cpu, conv2d_smaller_gpu);
-test_device!(conv2d_grad, conv2d_grad_cpu, conv2d_grad_gpu);

candle-core/tests/grad_tests.rs

@@ -192,6 +192,257 @@ fn unary_grad(device: &Device) -> Result<()> {
         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],
+    );
+
+    // 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)?;
+
+    #[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., 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)?;
+
+    #[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., 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(())
 }
 
@@ -218,12 +469,48 @@ fn binary_grad(device: &Device) -> Result<()> {
     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);
-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);
-test_device!(binary_grad, binary_grad_cpu, binary_grad_gpu);
+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
+);

candle-core/tests/indexing_tests.rs

@@ -91,3 +91,32 @@ fn index_3d() -> Result<()> {
     assert_eq!(tensor.i((1, .., 3))?.to_vec1::<u32>()?, &[15, 19, 23]);
     Ok(())
 }
+
+#[test]
+fn slice_assign() -> Result<()> {
+    let dev = Device::Cpu;
+
+    let tensor = Tensor::arange(0u32, 4 * 5, &dev)?.reshape((4, 5))?;
+    let src = Tensor::arange(0u32, 2 * 3, &dev)?.reshape((3, 2))?;
+    let out = tensor.slice_assign(&[1..4, 3..5], &src)?;
+    assert_eq!(
+        out.to_vec2::<u32>()?,
+        &[
+            [0, 1, 2, 3, 4],
+            [5, 6, 7, 0, 1],
+            [10, 11, 12, 2, 3],
+            [15, 16, 17, 4, 5]
+        ]
+    );
+    let out = tensor.slice_assign(&[0..3, 0..2], &src)?;
+    assert_eq!(
+        out.to_vec2::<u32>()?,
+        &[
+            [0, 1, 2, 3, 4],
+            [2, 3, 7, 8, 9],
+            [4, 5, 12, 13, 14],
+            [15, 16, 17, 18, 19]
+        ]
+    );
+    Ok(())
+}

candle-core/tests/layout_tests.rs

@@ -49,7 +49,7 @@ fn contiguous(device: &Device) -> Result<()> {
     Ok(())
 }
 
-test_device!(contiguous, contiguous_cpu, contiguous_gpu);
+test_device!(contiguous, contiguous_cpu, contiguous_gpu, contiguous_metal);
 
 #[test]
 fn strided_blocks() -> Result<()> {

candle-core/tests/npy.py

@@ -0,0 +1,9 @@
+import numpy as np
+x = np.arange(10)
+
+# Write a npy file.
+np.save("test.npy", x)
+
+# Write multiple values to a npz file.
+values = { "x": x, "x_plus_one": x + 1 }
+np.savez("test.npz", **values)

candle-core/tests/pool_tests.rs

@@ -98,15 +98,17 @@ fn upsample_nearest2d(dev: &Device) -> Result<()> {
     Ok(())
 }
 
-test_device!(avg_pool2d, avg_pool2d_cpu, avg_pool2d_gpu);
+test_device!(avg_pool2d, avg_pool2d_cpu, avg_pool2d_gpu, avg_pool2d_metal);
 test_device!(
     avg_pool2d_pytorch,
     avg_pool2d_pytorch_cpu,
-    avg_pool2d_pytorch_gpu
+    avg_pool2d_pytorch_gpu,
+    avg_pool2d_pytorch_metal
 );
-test_device!(max_pool2d, max_pool2d_cpu, max_pool2d_gpu);
+test_device!(max_pool2d, max_pool2d_cpu, max_pool2d_gpu, max_pool2d_metal);
 test_device!(
     upsample_nearest2d,
     upsample_nearest2d_cpu,
-    upsample_nearest2d_gpu
+    upsample_nearest2d_gpu,
+    upsample_nearest2d_metal
 );

candle-core/tests/pth.py

@@ -0,0 +1,37 @@
+import torch
+from collections import OrderedDict
+
+# Write a trivial tensor to a pt file
+a= torch.tensor([[1,2,3,4], [5,6,7,8]])
+o = OrderedDict()
+o["test"] = a
+
+# Write a trivial tensor to a pt file
+torch.save(o, "test.pt")
+
+############################################################################################################
+# Write a trivial tensor to a pt file with a key
+torch.save({"model_state_dict": o}, "test_with_key.pt")
+
+############################################################################################################
+# Create a tensor with fortran contiguous memory layout
+import numpy as np
+
+# Step 1: Create a 3D NumPy array with Fortran order using a range of numbers
+# For example, creating a 2x3x4 array
+array_fortran = np.asfortranarray(np.arange(1, 2*3*4 + 1).reshape(2, 3, 4))
+
+# Verify the memory order
+print("Is Fortran contiguous (F order):", array_fortran.flags['F_CONTIGUOUS'])  # Should be True
+print("Is C contiguous (C order):", array_fortran.flags['C_CONTIGUOUS'])  # Should be False
+
+# Step 2: Convert the NumPy array to a PyTorch tensor
+tensor_fortran = torch.from_numpy(array_fortran)
+
+# Verify the tensor layout
+print("Tensor stride:", tensor_fortran.stride())  # Stride will reflect the Fortran memory layout
+
+# Step 3: Save the PyTorch tensor to a .pth file
+torch.save({"tensor_fortran": tensor_fortran}, 'fortran_tensor_3d.pth')
+
+print("3D Tensor saved with Fortran layout.")

candle-core/tests/pth_tests.rs

@@ -0,0 +1,31 @@
+/// Regression test for pth files not loading on Windows.
+#[test]
+fn test_pth() {
+    let tensors = candle_core::pickle::PthTensors::new("tests/test.pt", None).unwrap();
+    tensors.get("test").unwrap().unwrap();
+}
+
+#[test]
+fn test_pth_with_key() {
+    let tensors =
+        candle_core::pickle::PthTensors::new("tests/test_with_key.pt", Some("model_state_dict"))
+            .unwrap();
+    tensors.get("test").unwrap().unwrap();
+}
+
+#[test]
+fn test_pth_fortran_congiguous() {
+    let tensors =
+        candle_core::pickle::PthTensors::new("tests/fortran_tensor_3d.pth", None).unwrap();
+    let tensor = tensors.get("tensor_fortran").unwrap().unwrap();
+
+    assert_eq!(tensor.dims3().unwrap(), (2, 3, 4));
+
+    assert_eq!(
+        tensor.to_vec3::<i64>().unwrap(),
+        [
+            [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]],
+            [[13, 14, 15, 16], [17, 18, 19, 20], [21, 22, 23, 24]]
+        ]
+    );
+}

candle-core/tests/quantized_tests.rs

@@ -1,7 +1,9 @@
 use candle_core::{
+    bail,
     quantized::{self, GgmlDType},
+    test_device,
     test_utils::to_vec2_round,
-    Device, Result, Tensor,
+    Device, Module, Result, Tensor,
 };
 use quantized::{k_quants, GgmlType};
 use rand::prelude::*;
@@ -13,16 +15,48 @@ const GGML_MAX_QUANTIZATION_TOTAL_ERROR_2BITS: f32 = 0.0075;
 const GGML_MAX_QUANTIZATION_TOTAL_ERROR_3BITS: f32 = 0.0040;
 const GGML_MAX_DOT_PRODUCT_ERROR: f32 = 0.02;
 
-#[test]
-fn quantized_matmul() -> Result<()> {
-    let cpu = &Device::Cpu;
+fn test_matmul(
+    device: &Device,
+    (b, m, n, k): (usize, usize, usize, usize),
+    dtype: GgmlDType,
+) -> Result<()> {
+    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)?;
+    let rhs = Tensor::from_slice(&rhs, (k, n), device)?;
+    let mm = lhs.matmul(&rhs)?;
+    let qtensor = quantized::QTensor::quantize(&rhs.t()?, dtype)?;
+    let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
+    let res = matmul.forward(&lhs)?;
+
+    let error: f32 = ((&mm - &res)?.abs()? / &mm.abs()?)?
+        .sum_all()?
+        .to_scalar()?;
+    let error = error / (b * m * n) as f32;
+    assert!(
+        error <= 0.02,
+        "Error {error} is too big. \nExpected:\n {mm} \nFound:\n {res}\n for {dtype:?}"
+    );
+
+    Ok(())
+}
+
+fn quantized_matmul(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
     let (m, k, n) = (3, 64, 4);
     let lhs = (0..(m * k)).map(|v| v as f32).collect::<Vec<_>>();
-    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), cpu)?;
+    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), device)?;
     let mut dst = vec![42.; 3 * 4];
     let mut rhs_t = vec![k_quants::BlockQ4_0::zeros(); 8];
     let rhs = (0..(k * n)).map(|v| v as f32).collect::<Vec<_>>();
-    let tensor_rhs = Tensor::from_slice(&rhs, (n, k), cpu)?.t()?;
     k_quants::BlockQ4_0::from_float(&rhs, &mut rhs_t)?;
     k_quants::matmul((m, k, n), &lhs, &rhs_t, &mut dst)?;
     assert_eq!(
@@ -32,6 +66,7 @@ fn quantized_matmul() -> Result<()> {
             341876.0, 994283.0, 1655709.0, 2301518.0
         ]
     );
+    let tensor_rhs = Tensor::from_slice(&rhs, (n, k), device)?.t()?;
     let mm = tensor_lhs.matmul(&tensor_rhs)?;
     assert_eq!(
         mm.to_vec2::<f32>()?,
@@ -42,35 +77,49 @@ fn quantized_matmul() -> Result<()> {
         ]
     );
 
-    let qtensor = quantized::QTensor::new(rhs_t, (4, 64))?;
-    let matmul = quantized::QMatMul::from_qtensor(qtensor);
+    let qtensor = quantized::QTensor::quantize(&tensor_rhs.t()?, GgmlDType::Q4_0)?;
+    let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
     let res = matmul.forward(&tensor_lhs)?;
-    assert_eq!(
-        to_vec2_round(&res, 0)?,
-        &[
-            [85120.0, 214562.0, 345455.0, 474748.0],
-            [213475.0, 604465.0, 1000686.0, 1388317.0],
-            [341876.0, 994283.0, 1655709.0, 2301518.0]
-        ]
-    );
+    match device {
+        Device::Metal(_) => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [84946.0, 214126.0, 344757.0, 473798.0],
+                [213458.0, 604350.0, 1000469.0, 1387990.0],
+                [341970.0, 994574.0, 1656181.0, 2302182.0]
+            ]
+        ),
+        _ => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [85120.0, 214562.0, 345455.0, 474748.0],
+                [213475.0, 604465.0, 1000686.0, 1388317.0],
+                [341876.0, 994283.0, 1655709.0, 2301518.0]
+            ]
+        ),
+    }
+
+    test_matmul(device, (1, 3, 4, 256), GgmlDType::Q4_0)?;
 
     Ok(())
 }
 
-#[test]
-fn quantized_matmul_neg() -> Result<()> {
-    let cpu = &Device::Cpu;
+fn quantized_matmul_neg(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
     let (m, k, n) = (3, 64, 4);
     let lhs = (0..(m * k))
         .map(|v| v as f32 - (m * k) as f32 / 2.0)
         .collect::<Vec<_>>();
-    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), cpu)?;
+    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), device)?;
     let mut dst = vec![42.; 3 * 4];
     let mut rhs_t = vec![k_quants::BlockQ4_0::zeros(); 8];
     let rhs = (0..k * n)
         .map(|v| v as f32 - (k * n) as f32 / 3.0)
         .collect::<Vec<_>>();
-    let tensor_rhs = Tensor::from_slice(&rhs, (n, k), cpu)?.t()?;
+    let tensor_rhs = Tensor::from_slice(&rhs, (n, k), device)?.t()?;
     k_quants::BlockQ4_0::from_float(&rhs, &mut rhs_t)?;
     k_quants::matmul((m, k, n), &lhs, &rhs_t, &mut dst)?;
     assert_eq!(
@@ -90,32 +139,56 @@ fn quantized_matmul_neg() -> Result<()> {
         ]
     );
 
-    let qtensor = quantized::QTensor::new(rhs_t, (4, 64))?;
-    let matmul = quantized::QMatMul::from_qtensor(qtensor);
+    let qtensor = quantized::QTensor::quantize(&tensor_rhs.t()?, GgmlDType::Q4_0)?;
+    let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
     let res = matmul.forward(&tensor_lhs)?;
-    assert_eq!(
-        to_vec2_round(&res, 0)?,
-        &[
-            [243524.0, -19596.0, -285051.0, -549815.0],
-            [23777.0, 21651.0, 19398.0, 18367.0],
-            [-196472.0, 63012.0, 324585.0, 587902.0]
-        ]
-    );
+    match device {
+        Device::Metal(_) => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [243666.0, -19714.0, -285433.0, -550453.0],
+                [23782.0, 21654.0, 19400.0, 18369.0],
+                [-196102.0, 63022.0, 324233.0, 587191.0]
+            ]
+        ),
+        _ => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [243524.0, -19596.0, -285051.0, -549815.0],
+                [23777.0, 21651.0, 19398.0, 18367.0],
+                [-196472.0, 63012.0, 324585.0, 587902.0]
+            ]
+        ),
+    }
 
     Ok(())
 }
 
-#[test]
-fn quantize_q4_0() -> Result<()> {
-    use k_quants::BlockQ4_0;
+test_device!(
+    quantized_matmul,
+    quantized_matmul_cpu,
+    quantized_matmul_cuda,
+    quantized_matmul_metal
+);
+test_device!(
+    quantized_matmul_neg,
+    quantized_matmul_neg_cpu,
+    quantized_matmul_neg_cuda,
+    quantized_matmul_neg_metal
+);
 
+fn quantize_q4_0(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
-    let mut dst = vec![0f32; 32 * 4];
-    let mut quant = vec![BlockQ4_0::zeros(); 4];
-    BlockQ4_0::from_float(&src, &mut quant)?;
-    BlockQ4_0::to_float(&quant, dst.as_mut_slice())?;
+
+    let src = Tensor::from_slice(&src, (32 * 4,), device)?;
+    let quant = quantized::QTensor::quantize(&src, GgmlDType::Q4_0)?;
+    let dst = quant.dequantize(device)?;
     assert_eq!(
-        dst,
+        dst.to_vec1::<f32>()?,
         &[
             -0.0, -0.0, 3.875, 3.875, 3.875, 3.875, 7.75, 7.75, 7.75, 7.75, 11.625, 11.625, 11.625,
             11.625, 15.5, 15.5, 15.5, 15.5, 19.375, 19.375, 19.375, 19.375, 23.25, 23.25, 23.25,
@@ -131,21 +204,21 @@ fn quantize_q4_0() -> Result<()> {
             127.0, 127.0
         ]
     );
-    ggml_quantization_error_test::<BlockQ4_0>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(GgmlDType::Q4_0, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q4_1() -> Result<()> {
-    use k_quants::BlockQ4_1;
-
+fn quantize_q4_1(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
-    let mut dst = vec![0f32; 32 * 4];
-    let mut quant = vec![BlockQ4_1::zeros(); 4];
-    BlockQ4_1::from_float(&src, &mut quant)?;
-    BlockQ4_1::to_float(&quant, dst.as_mut_slice())?;
+    let src = Tensor::from_slice(&src, (32 * 4,), device)?;
+    let quant = quantized::QTensor::quantize(&src, GgmlDType::Q4_1)?;
+    let dst = quant.dequantize(device)?;
     assert_eq!(
-        round_vector(&dst),
+        round_vector(&dst.to_vec1::<f32>()?),
         &[
             0.0, 0.0, 2.066, 2.066, 4.133, 4.133, 6.199, 6.199, 8.266, 8.266, 10.332, 10.332,
             12.398, 12.398, 14.465, 14.465, 16.531, 16.531, 18.598, 18.598, 20.664, 20.664, 22.73,
@@ -161,21 +234,21 @@ fn quantize_q4_1() -> Result<()> {
             118.73, 118.73, 120.797, 120.797, 122.863, 122.863, 124.93, 124.93, 126.996, 126.996
         ]
     );
-    ggml_quantization_error_test::<BlockQ4_1>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(GgmlDType::Q4_1, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q5_0() -> Result<()> {
-    use k_quants::BlockQ5_0;
-
+fn quantize_q5_0(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
-    let mut dst = vec![0f32; 32 * 4];
-    let mut quant = vec![BlockQ5_0::zeros(); 4];
-    BlockQ5_0::from_float(&src, &mut quant)?;
-    BlockQ5_0::to_float(&quant, dst.as_mut_slice())?;
+    let src = Tensor::from_slice(&src, (32 * 4,), device)?;
+    let quant = quantized::QTensor::quantize(&src, GgmlDType::Q5_0)?;
+    let dst = quant.dequantize(device)?;
     assert_eq!(
-        round_vector(&dst),
+        round_vector(&dst.to_vec1::<f32>()?),
         &[
             -0.0, 1.938, 1.938, 3.875, 3.875, 5.813, 5.813, 7.75, 7.75, 9.688, 9.688, 11.625,
             11.625, 13.563, 13.563, 15.5, 15.5, 17.438, 17.438, 19.375, 19.375, 21.313, 21.313,
@@ -191,21 +264,21 @@ fn quantize_q5_0() -> Result<()> {
             119.063, 119.063, 119.063, 119.063, 127.0, 127.0, 127.0, 127.0
         ]
     );
-    ggml_quantization_error_test::<BlockQ5_0>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(GgmlDType::Q5_0, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q5_1() -> Result<()> {
-    use k_quants::BlockQ5_1;
-
+fn quantize_q5_1(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
-    let mut dst = vec![0f32; 32 * 4];
-    let mut quant = vec![BlockQ5_1::zeros(); 4];
-    BlockQ5_1::from_float(&src, &mut quant)?;
-    BlockQ5_1::to_float(&quant, dst.as_mut_slice())?;
+    let src = Tensor::from_slice(&src, (32 * 4,), device)?;
+    let quant = quantized::QTensor::quantize(&src, GgmlDType::Q5_1)?;
+    let dst = quant.dequantize(device)?;
     assert_eq!(
-        dst,
+        round_vector(&dst.to_vec1::<f32>()?),
         &[
             0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0,
             16.0, 17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0, 25.0, 26.0, 27.0, 28.0, 29.0,
@@ -219,13 +292,11 @@ fn quantize_q5_1() -> Result<()> {
             124.0, 125.0, 126.0, 127.0
         ]
     );
-
-    ggml_quantization_error_test::<BlockQ5_1>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(GgmlDType::Q5_1, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-/// Generates a small test vector ranging from -`bound` to `bound` with `size` steps
-fn get_test_vector(bound: f32, size: usize) -> (Vec<f32>, Vec<f32>) {
+fn get_test_vector2(bound: f32, size: usize, device: &Device) -> Result<Tensor> {
     assert!(
         size % crate::quantized::k_quants::QK_K == 0,
         "size must be a multiple of {}",
@@ -235,10 +306,8 @@ fn get_test_vector(bound: f32, size: usize) -> (Vec<f32>, Vec<f32>) {
     let src = (0..size)
         .map(|v| (v as f32 - size as f32 / 2.) * bound / (size as f32 / 2.))
         .collect::<Vec<_>>();
-
-    let dst = vec![0f32; size];
     assert_eq!([src[0], src[size / 2]], [-bound, 0.0]);
-    (src, dst)
+    Tensor::from_vec(src, (size,), device)
 }
 
 /// Round a vector
@@ -265,7 +334,8 @@ fn compare_with_error(values: &[f32], expected: &[f32], tolerance: f32) {
     }
 }
 
-/// Creates a vector simillarly to the one used in GGML unit tests: https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L26-L30
+/// Creates a vector similar to the ones used in GGML unit tests:
+/// https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L26-L30
 fn create_ggml_like_vector(offset: f32) -> Vec<f32> {
     (0..GGML_TEST_SIZE)
         .map(|i| 0.1 + 2.0 * (i as f32 + offset).cos())
@@ -284,14 +354,16 @@ fn calculate_rmse(a: &[f32], b: &[f32]) -> f32 {
     sum / a.len() as f32
 }
 
-/// Mirrores the GGML quanitzation unit test: https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L43-L50
-fn ggml_quantization_error_test<T: GgmlType>(max_error: f32) -> Result<()> {
+/// Similar to the GGML quantization unit test:
+/// https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L43-L50
+fn ggml_quantization_error_test(dtype: GgmlDType, device: &Device, max_error: f32) -> Result<()> {
     let src = create_ggml_like_vector(0.0);
-    let mut dst = vec![0.0; GGML_TEST_SIZE];
-    let _quant = quantize_roundtrip::<T>(src.as_slice(), dst.as_mut_slice())?;
-    let error = calculate_rmse(src.as_slice(), dst.as_slice());
+    let src = Tensor::from_slice(&src, (GGML_TEST_SIZE,), device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
+    let error = calculate_rmse(&src.to_vec1::<f32>()?, &dst.to_vec1::<f32>()?);
     if error > max_error {
-        candle_core::bail!(
+        bail!(
             "Quantization error {} exceeds max error {}",
             error,
             max_error
@@ -300,19 +372,19 @@ fn ggml_quantization_error_test<T: GgmlType>(max_error: f32) -> Result<()> {
     Ok(())
 }
 
-fn quantize_roundtrip<T: GgmlType>(src: &[f32], dst: &mut [f32]) -> Result<Vec<T>> {
-    let mut quant = vec![T::zeros(); src.len() / T::BLCK_SIZE];
-    T::from_float(src, &mut quant)?;
-    T::to_float(&quant, dst)?;
-    Ok(quant)
-}
+fn quantize_q2k(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
+    let dtype = GgmlDType::Q2K;
 
-#[test]
-fn quantize_q2k() -> Result<()> {
-    use k_quants::BlockQ2K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ2K>(src.as_slice(), dst.as_mut_slice())?;
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
     compare_with_error(dst.as_slice(), src.as_slice(), 0.1);
 
     // Test some specific values
@@ -326,20 +398,30 @@ fn quantize_q2k() -> Result<()> {
         [-0.499, -0.366, -0.249, 0.0, 0.295, 0.492]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ2K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 6.0);
 
-    ggml_quantization_error_test::<BlockQ2K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR_2BITS)?;
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR_2BITS)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q3k() -> Result<()> {
-    use k_quants::BlockQ3K;
+fn quantize_q3k(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
+    let dtype = GgmlDType::Q3K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ3K>(src.as_slice(), dst.as_mut_slice())?;
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
     compare_with_error(dst.as_slice(), src.as_slice(), 0.03);
 
     // Test some specific values
@@ -353,20 +435,30 @@ fn quantize_q3k() -> Result<()> {
         [-0.493, -0.37, -0.243, -0.0, 0.292, 0.492]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ3K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 3.5);
 
-    ggml_quantization_error_test::<BlockQ3K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR_3BITS)?;
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR_3BITS)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q4k() -> Result<()> {
-    use k_quants::BlockQ4K;
+fn quantize_q4k(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
+    let dtype = GgmlDType::Q4K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ4K>(src.as_slice(), dst.as_mut_slice())?;
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
     compare_with_error(dst.as_slice(), src.as_slice(), 0.017);
 
     // Test some specific values
@@ -380,21 +472,31 @@ fn quantize_q4k() -> Result<()> {
         [-0.5, -0.373, -0.25, 0.0, 0.288, 0.498]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ4K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 4.5);
 
-    ggml_quantization_error_test::<BlockQ4K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q5k() -> Result<()> {
-    use k_quants::BlockQ5K;
+fn quantize_q5k(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
+    let dtype = GgmlDType::Q5K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ5K>(src.as_slice(), dst.as_mut_slice())?;
-    compare_with_error(dst.as_slice(), src.as_slice(), 0.008);
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
+    compare_with_error(dst.as_slice(), src.as_slice(), 0.009);
 
     // Test some specific values
     assert_eq!(
@@ -404,24 +506,33 @@ fn quantize_q5k() -> Result<()> {
     let dst = round_vector(&dst);
     assert_eq!(
         [dst[0], dst[128], dst[256], dst[512], dst[800], dst[1023]],
-        [-0.499, -0.372, -0.249, 0.001, 0.279, 0.499]
+        [-0.5, -0.373, -0.25, 0.0, 0.279, 0.499]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ5K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 2.5);
 
-    ggml_quantization_error_test::<BlockQ5K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
-
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q6k() -> Result<()> {
-    use k_quants::BlockQ6K;
+fn quantize_q6k(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
+    let dtype = GgmlDType::Q6K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ6K>(src.as_slice(), dst.as_mut_slice())?;
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
     compare_with_error(dst.as_slice(), src.as_slice(), 0.008);
 
     // Test some specific values
@@ -435,22 +546,31 @@ fn quantize_q6k() -> Result<()> {
         [-0.497, -0.372, -0.25, -0.0, 0.284, 0.5]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ6K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 2.0);
 
-    ggml_quantization_error_test::<BlockQ6K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
-
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q8k() -> Result<()> {
-    use k_quants::BlockQ8K;
+fn quantize_q8k(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
+    let dtype = GgmlDType::Q8K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ8K>(src.as_slice(), dst.as_mut_slice())?;
-    compare_with_error(dst.as_slice(), src.as_slice(), 0.003);
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
+    compare_with_error(dst.as_slice(), src.as_slice(), 0.008);
 
     // Test some specific values
     assert_eq!(
@@ -463,15 +583,79 @@ fn quantize_q8k() -> Result<()> {
         [-0.5, -0.375, -0.25, -0.0, 0.281, 0.499]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ8K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 0.6);
 
-    ggml_quantization_error_test::<BlockQ8K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
-
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
+test_device!(
+    quantize_q4_0,
+    quantize_q4_0_cpu,
+    quantize_q4_0_cuda,
+    quantize_q4_0_metal
+);
+test_device!(
+    quantize_q4_1,
+    quantize_q4_1_cpu,
+    quantize_q4_1_cuda,
+    quantize_q4_1_metal
+);
+test_device!(
+    quantize_q5_0,
+    quantize_q5_0_cpu,
+    quantize_q5_0_cuda,
+    quantize_q5_0_metal
+);
+test_device!(
+    quantize_q5_1,
+    quantize_q5_1_cpu,
+    quantize_q5_1_cuda,
+    quantize_q5_1_metal
+);
+test_device!(
+    quantize_q2k,
+    quantize_q2k_cpu,
+    quantize_q2k_cuda,
+    quantize_q2k_metal
+);
+test_device!(
+    quantize_q3k,
+    quantize_q3k_cpu,
+    quantize_q3k_cuda,
+    quantize_q3k_metal
+);
+test_device!(
+    quantize_q4k,
+    quantize_q4k_cpu,
+    quantize_q4k_cuda,
+    quantize_q4k_metal
+);
+test_device!(
+    quantize_q5k,
+    quantize_q5k_cpu,
+    quantize_q5k_cuda,
+    quantize_q5k_metal
+);
+test_device!(
+    quantize_q6k,
+    quantize_q6k_cpu,
+    quantize_q6k_cuda,
+    quantize_q6k_metal
+);
+test_device!(
+    quantize_q8k,
+    quantize_q8k_cpu,
+    quantize_q8k_cuda,
+    quantize_q8k_metal
+);
+
 /// Very simple dot product implementation
 fn vec_dot_reference(a: &[f32], b: &[f32]) -> f32 {
     a.iter().zip(b).map(|(a, b)| a * b).sum()
@@ -487,46 +671,66 @@ fn ggml_reference_matmul_error(dtype: GgmlDType) -> Result<f32> {
         GgmlDType::Q5K => 0.000740,
         GgmlDType::Q6K => 0.000952,
         GgmlDType::Q4_0 => 0.001143,
-        GgmlDType::Q4_1 => 0.007784,
+        GgmlDType::Q4_1 => 0.008,
         GgmlDType::Q5_0 => 0.001353,
-        GgmlDType::Q5_1 => 0.001363,
+        GgmlDType::Q5_1 => 0.00149,
         GgmlDType::Q8_0 => 0.000092,
-        _ => candle_core::bail!("No GGML results for quantization type {dtype:?}",),
+
+        // Not from the ggml repo.
+        GgmlDType::Q8K => 0.00065,
+        _ => bail!("No GGML results for quantization type {dtype:?}",),
     };
     Ok(err)
 }
 
-/// Mirrores the GGML matmul unit test: https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L76-L91
+/// Similar to the GGML matmul unit test:
+/// https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L76-L91
 fn ggml_matmul_error_test<T: GgmlType>() -> Result<()> {
     let a = create_ggml_like_vector(0.0);
     let b = create_ggml_like_vector(1.0);
+    ggml_matmul_error_test_::<T>(a.as_slice(), b.as_slice(), 1.0)?;
+    // Another example that is more likely to trigger the overflow reported in #1526
+    let a = (0..GGML_TEST_SIZE)
+        .map(|i| i as f32 / GGML_TEST_SIZE as f32)
+        .collect::<Vec<_>>();
+    let b = (0..GGML_TEST_SIZE)
+        .map(|i| i as f32 / GGML_TEST_SIZE as f32)
+        .collect::<Vec<_>>();
+    ggml_matmul_error_test_::<T>(a.as_slice(), b.as_slice(), 2.0)?;
+    Ok(())
+}
+
+fn ggml_matmul_error_test_<T: GgmlType>(a: &[f32], b: &[f32], err_m: f32) -> Result<()> {
     let length = a.len();
 
     let mut a_quant = vec![T::zeros(); length / T::BLCK_SIZE];
     let mut b_quant = vec![T::VecDotType::zeros(); length / T::VecDotType::BLCK_SIZE];
-    T::from_float(&a, &mut a_quant)?;
-    T::VecDotType::from_float(&b, &mut b_quant)?;
+    T::from_float(a, &mut a_quant)?;
+    T::VecDotType::from_float(b, &mut b_quant)?;
 
     let result = T::vec_dot(length, &a_quant, &b_quant)?;
-    let reference_result = vec_dot_reference(&a, &b);
+    let result_unopt = T::vec_dot_unopt(length, &a_quant, &b_quant)?;
+    let reference_result = vec_dot_reference(a, b);
+
+    if (result - result_unopt).abs() / length as f32 > 1e-6 {
+        bail!(
+            "the opt and unopt vec-dot returned different values, opt {result}, unopt {result_unopt}"
+        )
+    }
 
     let error = (result - reference_result).abs() / length as f32;
 
-    let ggml_error = ggml_reference_matmul_error(T::DTYPE)?;
+    let ggml_error = ggml_reference_matmul_error(T::DTYPE)? * err_m;
 
-    if error > GGML_MAX_DOT_PRODUCT_ERROR {
-        candle_core::bail!(
-            "Dot product error {} exceeds max error {}",
-            error,
-            GGML_MAX_DOT_PRODUCT_ERROR
-        );
+    if !error.is_finite() || error > GGML_MAX_DOT_PRODUCT_ERROR {
+        bail!("Dot product error {error} exceeds max error {GGML_MAX_DOT_PRODUCT_ERROR}",);
     }
 
     // We diverge slightly due to different rounding behavior / f16 to f32 conversions in GGML
     // => we use a slightly higher error threshold
     const ERROR_LENIENCY: f32 = 0.00001;
     if error - ERROR_LENIENCY > ggml_error {
-        candle_core::bail!(
+        bail!(
             "Dot product error {} exceeds ggml reference error {}",
             error,
             ggml_error
@@ -535,6 +739,16 @@ fn ggml_matmul_error_test<T: GgmlType>() -> Result<()> {
     Ok(())
 }
 
+#[test]
+fn quantized_mm() -> Result<()> {
+    ggml_matmul_error_test::<k_quants::BlockQ4_0>()?;
+    ggml_matmul_error_test::<k_quants::BlockQ4_1>()?;
+    ggml_matmul_error_test::<k_quants::BlockQ5_0>()?;
+    ggml_matmul_error_test::<k_quants::BlockQ5_1>()?;
+    ggml_matmul_error_test::<k_quants::BlockQ8_0>()?;
+    Ok(())
+}
+
 /// generates random tensors of size `m x k` and `n x k` and calculates their expected matrix multiplication result.
 fn get_random_tensors(
     m: usize,
@@ -558,6 +772,112 @@ fn get_random_tensors(
     Ok((lhs, rhs, mm))
 }
 
+#[macro_export]
+macro_rules! quantized_matmul {
+    // 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, $fn_name_cpu: ident, $fn_name_cuda: ident, $fn_name_metal: ident, $dtype: expr) => {
+        fn $fn_name(device: &Device) -> Result<()> {
+            if device.is_cuda() {
+                // TODO Enable Cuda GGML sometime maybe.
+                return Ok(());
+            }
+            test_matmul(device, (1, 3, 4, 256), $dtype)?;
+            Ok(())
+        }
+
+        test_device!($fn_name, $fn_name_cpu, $fn_name_cuda, $fn_name_metal);
+    };
+}
+
+quantized_matmul!(
+    quantized_matmul_q4_0_bis,
+    quantized_matmul_q4_0_cpu,
+    quantized_matmul_q4_0_cuda,
+    quantized_matmul_q4_0_metal,
+    GgmlDType::Q4_0
+);
+quantized_matmul!(
+    quantized_matmul_q4_1_bis,
+    quantized_matmul_q4_1_cpu,
+    quantized_matmul_q4_1_cuda,
+    quantized_matmul_q4_1_metal,
+    GgmlDType::Q4_1
+);
+quantized_matmul!(
+    quantized_matmul_q5_0_bis,
+    quantized_matmul_q5_0_cpu,
+    quantized_matmul_q5_0_cuda,
+    quantized_matmul_q5_0_metal,
+    GgmlDType::Q5_0
+);
+quantized_matmul!(
+    quantized_matmul_q5_1_bis,
+    quantized_matmul_q5_1_cpu,
+    quantized_matmul_q5_1_cuda,
+    quantized_matmul_q5_1_metal,
+    GgmlDType::Q5_1
+);
+quantized_matmul!(
+    quantized_matmul_q8_0_bis,
+    quantized_matmul_q8_0_cpu,
+    quantized_matmul_q8_0_cuda,
+    quantized_matmul_q8_0_metal,
+    GgmlDType::Q8_0
+);
+// Not implemented in Ggml
+// quantized_matmul!(
+//     quantized_matmul_q8_1_bis,
+//     quantized_matmul_q8_1_cpu,
+//     quantized_matmul_q8_1_cuda,
+//     quantized_matmul_q8_1_metal,
+//     GgmlDType::Q8_1
+// );
+// TODO This is bugged (also bugged in GGML
+quantized_matmul!(
+    quantized_matmul_q2k_bis,
+    quantized_matmul_q2k_cpu,
+    quantized_matmul_q2k_cuda,
+    quantized_matmul_q2k_metal,
+    GgmlDType::Q2K
+);
+quantized_matmul!(
+    quantized_matmul_q3k_bis,
+    quantized_matmul_q3k_cpu,
+    quantized_matmul_q3k_cuda,
+    quantized_matmul_q3k_metal,
+    GgmlDType::Q3K
+);
+quantized_matmul!(
+    quantized_matmul_q4k_bis,
+    quantized_matmul_q4k_cpu,
+    quantized_matmul_q4k_cuda,
+    quantized_matmul_q4k_metal,
+    GgmlDType::Q4K
+);
+quantized_matmul!(
+    quantized_matmul_q5k_bis,
+    quantized_matmul_q5k_cpu,
+    quantized_matmul_q5k_cuda,
+    quantized_matmul_q5k_metal,
+    GgmlDType::Q5K
+);
+quantized_matmul!(
+    quantized_matmul_q6k_bis,
+    quantized_matmul_q6k_cpu,
+    quantized_matmul_q6k_cuda,
+    quantized_matmul_q6k_metal,
+    GgmlDType::Q6K
+);
+// Not implemented on metal
+// quantized_matmul!(
+//     quantized_matmul_q8k_bis,
+//     quantized_matmul_q8k_cpu,
+//     quantized_matmul_q8k_cuda,
+//     quantized_matmul_q8k_metal,
+//     GgmlDType::Q8K
+// );
+
 #[test]
 fn quantized_matmul_q2k() -> Result<()> {
     use k_quants::BlockQ2K;
@@ -570,8 +890,8 @@ fn quantized_matmul_q2k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ2K>(&rhs)?;
-    let rhs = quantized::QMatMul::from_qtensor(rhs);
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q2K)?;
+    let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
     assert_eq!(mm.dims(), [m, n]);
@@ -596,8 +916,8 @@ fn quantized_matmul_q3k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ3K>(&rhs)?;
-    let rhs = quantized::QMatMul::from_qtensor(rhs);
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q3K)?;
+    let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
     assert_eq!(mm.dims(), [m, n]);
@@ -622,8 +942,8 @@ fn quantized_matmul_q4k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ4K>(&rhs)?;
-    let rhs = quantized::QMatMul::from_qtensor(rhs);
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q4K)?;
+    let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
     assert_eq!(mm.dims(), [m, n]);
@@ -648,8 +968,8 @@ fn quantized_matmul_q5k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ5K>(&rhs)?;
-    let rhs = quantized::QMatMul::from_qtensor(rhs);
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q5K)?;
+    let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
     assert_eq!(mm.dims(), [m, n]);
@@ -675,8 +995,8 @@ fn quantized_matmul_q6k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ6K>(&rhs)?;
-    let rhs = quantized::QMatMul::from_qtensor(rhs);
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q6K)?;
+    let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
     assert_eq!(mm.dims(), [m, n]);
@@ -687,3 +1007,28 @@ fn quantized_matmul_q6k() -> Result<()> {
     ggml_matmul_error_test::<BlockQ6K>()?;
     Ok(())
 }
+
+#[test]
+fn quantized_matmul_q8k() -> Result<()> {
+    use k_quants::BlockQ8K;
+
+    let cpu = &Device::Cpu;
+    let (m, k, n) = (11, 512, 21);
+    let (lhs, rhs, mm) = get_random_tensors(m, k, n, cpu)?;
+    assert_eq!(mm.dims(), [m, n]);
+    let dst = mm.flatten_all()?.to_vec1::<f32>()?;
+    let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
+    assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
+
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q8K)?;
+    let rhs = quantized::QMatMul::from_qtensor(rhs)?;
+    let mm = rhs.forward(&lhs)?;
+
+    assert_eq!(mm.dims(), [m, n]);
+    let dst = mm.flatten_all()?.to_vec1::<f32>()?;
+    let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
+    assert_eq!(dst, [1.266, 1.504, -0.204, 1.7]);
+
+    ggml_matmul_error_test::<BlockQ8K>()?;
+    Ok(())
+}

candle-core/tests/serialization_tests.rs

@@ -0,0 +1,24 @@
+use candle_core::{DType, Result, Tensor};
+
+#[test]
+fn npy() -> Result<()> {
+    let npy = Tensor::read_npy("tests/test.npy")?;
+    assert_eq!(
+        npy.to_dtype(DType::U8)?.to_vec1::<u8>()?,
+        [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
+    );
+    Ok(())
+}
+
+#[test]
+fn npz() -> Result<()> {
+    let npz = Tensor::read_npz("tests/test.npz")?;
+    assert_eq!(npz.len(), 2);
+    assert_eq!(npz[0].0, "x");
+    assert_eq!(npz[1].0, "x_plus_one");
+    assert_eq!(
+        npz[1].1.to_dtype(DType::U8)?.to_vec1::<u8>()?,
+        [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
+    );
+    Ok(())
+}

candle-core/tests/tensor_tests.rs

@@ -1,4 +1,4 @@
-use candle_core::{test_device, DType, Device, IndexOp, Result, Tensor};
+use candle_core::{test_device, test_utils, DType, Device, IndexOp, Result, Tensor, D};
 
 fn zeros(device: &Device) -> Result<()> {
     let tensor = Tensor::zeros((5, 2), DType::F32, device)?;
@@ -8,6 +8,58 @@ fn zeros(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn ones(device: &Device) -> Result<()> {
+    assert_eq!(
+        Tensor::ones((2, 3), DType::U8, device)?.to_vec2::<u8>()?,
+        [[1, 1, 1], [1, 1, 1]],
+    );
+    assert_eq!(
+        Tensor::ones((2, 3), DType::U32, device)?.to_vec2::<u32>()?,
+        [[1, 1, 1], [1, 1, 1]],
+    );
+    assert_eq!(
+        Tensor::ones((2, 3), DType::I64, device)?.to_vec2::<i64>()?,
+        [[1, 1, 1], [1, 1, 1]],
+    );
+    assert_eq!(
+        Tensor::ones((2, 3), DType::F32, device)?.to_vec2::<f32>()?,
+        [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
+    );
+    assert_eq!(
+        Tensor::ones((2, 3), DType::F64, device)?.to_vec2::<f64>()?,
+        [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
+    );
+    Ok(())
+}
+
+fn full(device: &Device) -> Result<()> {
+    assert_eq!(
+        Tensor::full(42u32, (2, 3), device)?.to_vec2::<u32>()?,
+        [[42, 42, 42], [42, 42, 42]],
+    );
+    Ok(())
+}
+
+fn arange(device: &Device) -> Result<()> {
+    assert_eq!(
+        Tensor::arange(0u8, 5u8, device)?.to_vec1::<u8>()?,
+        [0, 1, 2, 3, 4],
+    );
+    assert_eq!(
+        Tensor::arange_step(0u8, 5u8, 2, device)?.to_vec1::<u8>()?,
+        [0, 2, 4],
+    );
+    assert_eq!(
+        Tensor::arange_step(0u8, 5u8, 3, device)?.to_vec1::<u8>()?,
+        [0, 3],
+    );
+    assert_eq!(
+        Tensor::arange_step(5i64, 0i64, -1, device)?.to_vec1::<i64>()?,
+        [5, 4, 3, 2, 1],
+    );
+    Ok(())
+}
+
 fn add_mul(device: &Device) -> Result<()> {
     let tensor = Tensor::new(&[3f32, 1., 4.], device)?;
     let dim1 = tensor.dims1()?;
@@ -33,6 +85,72 @@ fn tensor_2d(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn clamp(device: &Device) -> Result<()> {
+    let data = &[[3f32, 1., 4., 1., 5.], [2., 1., 7., 8., 2.]];
+    let tensor = Tensor::new(data, device)?;
+    let tensor = tensor.clamp(1.5, 6.2)?;
+    assert_eq!(
+        tensor.to_vec2::<f32>()?,
+        [[3.0, 1.5, 4.0, 1.5, 5.0], [2.0, 1.5, 6.2, 6.2, 2.0]],
+    );
+    Ok(())
+}
+
+fn unary_op(device: &Device) -> Result<()> {
+    let data = &[[-3f32, 1., 4., -0.1, 0.5], [2.7, -1.8, -0.28, 1.8, 2.8]];
+    let tensor = Tensor::new(data, device)?;
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.gelu()?, 4)?,
+        [
+            [-0.0036, 0.8412, 3.9999, -0.046, 0.3457],
+            [2.6911, -0.0647, -0.1091, 1.7353, 2.7933]
+        ]
+    );
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.gelu_erf()?, 4)?,
+        [
+            [-0.004, 0.8413, 3.9999, -0.046, 0.3457],
+            [2.6906, -0.0647, -0.1091, 1.7353, 2.7928]
+        ]
+    );
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.erf()?, 4)?,
+        [
+            [-1.0, 0.8427, 1.0, -0.1125, 0.5205],
+            [0.9999, -0.9891, -0.3079, 0.9891, 0.9999]
+        ]
+    );
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.silu()?, 4)?,
+        [
+            [-0.1423, 0.7311, 3.9281, -0.0475, 0.3112],
+            [2.53, -0.2553, -0.1205, 1.5447, 2.6395]
+        ]
+    );
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.ceil()?, 4)?,
+        [[-3.0, 1.0, 4.0, -0.0, 1.0], [3.0, -1.0, -0.0, 2.0, 3.0]]
+    );
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.floor()?, 4)?,
+        [[-3.0, 1.0, 4.0, -1.0, 0.0], [2.0, -2.0, -1.0, 1.0, 2.0]]
+    );
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.round()?, 4)?,
+        [[-3.0, 1.0, 4.0, -0.0, 1.0], [3.0, -2.0, -0.0, 2.0, 3.0]]
+    );
+    let tensor = Tensor::new(&[2997.9246, 314.15926f32], device)?;
+    assert_eq!(
+        test_utils::to_vec1_round(&tensor.round_to(2)?, 4)?,
+        [2997.92, 314.16]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(&tensor.round_to(-2)?, 4)?,
+        [3000.0, 300.]
+    );
+    Ok(())
+}
+
 fn binary_op(device: &Device) -> Result<()> {
     let data = &[[3f32, 1., 4., 1., 5.], [2., 1., 7., 8., 2.]];
     let tensor1 = Tensor::new(data, device)?;
@@ -77,6 +195,22 @@ fn transpose(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn var(device: &Device) -> Result<()> {
+    // Values taken from https://pytorch.org/docs/stable/generated/torch.var.html
+    let data = &[
+        [0.2035f32, 1.2959, 1.8101, -0.4644],
+        [1.5027, -0.3270, 0.5905, 0.6538],
+        [-1.5745, 1.3330, -0.5596, -0.6548],
+        [0.1264, -0.5080, 1.6420, 0.1992],
+    ];
+    let tensor = Tensor::new(data, device)?;
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.var_keepdim(1)?, 4)?,
+        &[[1.0631], [0.559], [1.4893], [0.8258]]
+    );
+    Ok(())
+}
+
 fn sum(device: &Device) -> Result<()> {
     let data = &[[[3u32, 1, 4], [1, 5, 9]], [[2, 1, 7], [8, 2, 8]]];
     let tensor = Tensor::new(data, device)?;
@@ -590,6 +724,30 @@ fn index_select(device: &Device) -> Result<()> {
         hs.to_vec2::<f32>()?,
         &[[0.0, 1.0, 2.0], [6.0, 7.0, 8.0], [3.0, 4.0, 5.0]]
     );
+    // Prior to https://github.com/huggingface/candle/pull/1022
+    // There would be a bug where the last values in the result tensor would be set to 0.
+    let ids = Tensor::new(&[0u32, 2u32, 1u32, 0u32, 2u32, 1u32], device)?;
+    let hs = t.index_select(&ids, 0)?;
+    assert_eq!(
+        hs.to_vec2::<f32>()?,
+        &[
+            [0.0, 1.0, 2.0],
+            [6.0, 7.0, 8.0],
+            [3.0, 4.0, 5.0],
+            [0.0, 1.0, 2.0],
+            [6.0, 7.0, 8.0],
+            [3.0, 4.0, 5.0],
+        ]
+    );
+
+    // Test when selecting dim > 0 with ids size different from elem count of
+    // target dim in source/input.
+    let ids = Tensor::new(&[1u32, 0u32, 1u32], device)?;
+    let t = Tensor::arange(1f32, 5f32, device)?.reshape((2, 2))?;
+    assert_eq!(t.to_vec2::<f32>()?, &[[1.0, 2.0], [3.0, 4.0]]);
+    let hs = t.index_select(&ids, 1)?;
+    assert_eq!(hs.to_vec2::<f32>()?, &[[2.0, 1.0, 2.0], [4.0, 3.0, 4.0]]);
+
     Ok(())
 }
 
@@ -636,6 +794,48 @@ fn index_add(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn slice_scatter(device: &Device) -> Result<()> {
+    let t = Tensor::arange(0f32, 12f32, device)?.reshape((4, 3))?;
+    assert_eq!(
+        t.to_vec2::<f32>()?,
+        &[
+            [0.0, 1.0, 2.0],
+            [3.0, 4.0, 5.0],
+            [6.0, 7.0, 8.0],
+            [9.0, 10.0, 11.0]
+        ]
+    );
+    let src = Tensor::arange(100f32, 106f32, device)?.reshape((2, 3))?;
+    assert_eq!(
+        t.slice_scatter0(&src, 0)?.to_vec2::<f32>()?,
+        &[
+            [100.0, 101.0, 102.0],
+            [103.0, 104.0, 105.0],
+            [6.0, 7.0, 8.0],
+            [9.0, 10.0, 11.0]
+        ]
+    );
+    assert_eq!(
+        t.slice_scatter0(&src, 1)?.to_vec2::<f32>()?,
+        &[
+            [0.0, 1.0, 2.0],
+            [100.0, 101.0, 102.0],
+            [103.0, 104.0, 105.0],
+            [9.0, 10.0, 11.0]
+        ]
+    );
+    assert_eq!(
+        t.slice_scatter0(&src, 2)?.to_vec2::<f32>()?,
+        &[
+            [0.0, 1.0, 2.0],
+            [3.0, 4.0, 5.0],
+            [100.0, 101.0, 102.0],
+            [103.0, 104.0, 105.0],
+        ]
+    );
+    Ok(())
+}
+
 fn scatter_add(device: &Device) -> Result<()> {
     let t = Tensor::arange(0f32, 12f32, device)?.reshape((4, 3))?;
     assert_eq!(
@@ -885,29 +1085,61 @@ fn randn(device: &Device) -> Result<()> {
     Ok(())
 }
 
-test_device!(zeros, zeros_cpu, zeros_gpu);
-test_device!(add_mul, add_mul_cpu, add_mul_gpu);
-test_device!(tensor_2d, tensor_2d_cpu, tensor_2d_gpu);
-test_device!(narrow, narrow_cpu, narrow_gpu);
-test_device!(broadcast, broadcast_cpu, broadcast_gpu);
-test_device!(cat, cat_cpu, cat_gpu);
-test_device!(sum, sum_cpu, sum_gpu);
-test_device!(min, min_cpu, min_gpu);
-test_device!(max, max_cpu, max_gpu);
-test_device!(argmax, argmax_cpu, argmax_gpu);
-test_device!(argmin, argmin_cpu, argmin_gpu);
-test_device!(transpose, transpose_cpu, transpose_gpu);
-test_device!(binary_op, binary_op_cpu, binary_op_gpu);
-test_device!(embeddings, embeddings_cpu, embeddings_gpu);
-test_device!(cmp, cmp_cpu, cmp_gpu);
-test_device!(matmul, matmul_cpu, matmul_gpu);
-test_device!(broadcast_matmul, broadcast_matmul_cpu, broadcast_matmul_gpu);
-test_device!(broadcasting, broadcasting_cpu, broadcasting_gpu);
-test_device!(index_select, index_select_cpu, index_select_gpu);
-test_device!(index_add, index_add_cpu, index_add_gpu);
-test_device!(gather, gather_cpu, gather_gpu);
-test_device!(scatter_add, scatter_add_cpu, scatter_add_gpu);
-test_device!(randn, randn_cpu, randn_gpu);
+test_device!(zeros, zeros_cpu, zeros_gpu, zeros_metal);
+test_device!(ones, ones_cpu, ones_gpu, ones_metal);
+test_device!(full, full_cpu, full_gpu, full_metal);
+test_device!(arange, arange_cpu, arange_gpu, arange_metal);
+test_device!(add_mul, add_mul_cpu, add_mul_gpu, add_mul_metal);
+test_device!(tensor_2d, tensor_2d_cpu, tensor_2d_gpu, tensor_2d_metal);
+test_device!(narrow, narrow_cpu, narrow_gpu, narrow_metal);
+test_device!(broadcast, broadcast_cpu, broadcast_gpu, broadcast_metal);
+test_device!(cat, cat_cpu, cat_gpu, cat_metal);
+test_device!(sum, sum_cpu, sum_gpu, sum_metal);
+test_device!(min, min_cpu, min_gpu, min_metal);
+test_device!(max, max_cpu, max_gpu, max_metal);
+test_device!(argmax, argmax_cpu, argmax_gpu, argmax_metal);
+test_device!(argmin, argmin_cpu, argmin_gpu, argmin_metal);
+test_device!(transpose, transpose_cpu, transpose_gpu, transpose_metal);
+test_device!(unary_op, unary_op_cpu, unary_op_gpu, unary_op_metal);
+test_device!(binary_op, binary_op_cpu, binary_op_gpu, binary_op_metal);
+test_device!(embeddings, embeddings_cpu, embeddings_gpu, embeddings_metal);
+test_device!(cmp, cmp_cpu, cmp_gpu, cmp_metal);
+test_device!(matmul, matmul_cpu, matmul_gpu, matmul_metal);
+test_device!(
+    broadcast_matmul,
+    broadcast_matmul_cpu,
+    broadcast_matmul_gpu,
+    broadcast_matmul_metal
+);
+test_device!(
+    broadcasting,
+    broadcasting_cpu,
+    broadcasting_gpu,
+    broadcasting_metal
+);
+test_device!(
+    index_select,
+    index_select_cpu,
+    index_select_gpu,
+    index_select_metal
+);
+test_device!(index_add, index_add_cpu, index_add_gpu, index_add_metal);
+test_device!(gather, gather_cpu, gather_gpu, gather_metal);
+test_device!(
+    scatter_add,
+    scatter_add_cpu,
+    scatter_add_gpu,
+    scatter_add_metal
+);
+test_device!(
+    slice_scatter,
+    slice_scatter_cpu,
+    slice_scatter_gpu,
+    slice_scatter_metal
+);
+test_device!(randn, randn_cpu, randn_gpu, randn_metal);
+test_device!(clamp, clamp_cpu, clamp_gpu, clamp_metal);
+test_device!(var, var_cpu, var_gpu, var_metal);
 
 // There was originally a bug on the CPU implementation for randn
 // https://github.com/huggingface/candle/issues/381
@@ -919,3 +1151,124 @@ fn randn_hasneg() -> Result<()> {
     }
     Ok(())
 }
+
+#[test]
+fn pad_with_same() -> Result<()> {
+    let t = Tensor::arange(1f32, 5f32, &Device::Cpu)?.reshape((2, 2))?;
+    let t0 = t.pad_with_same(0, 1, 2)?;
+    assert_eq!(
+        t0.to_vec2::<f32>()?,
+        [[1.0, 2.0], [1.0, 2.0], [3.0, 4.0], [3.0, 4.0], [3.0, 4.0]]
+    );
+    let t1 = t.pad_with_same(1, 1, 2)?;
+    assert_eq!(
+        t1.to_vec2::<f32>()?,
+        [[1.0, 1.0, 2.0, 2.0, 2.0], [3.0, 3.0, 4.0, 4.0, 4.0]]
+    );
+    Ok(())
+}
+
+#[test]
+fn i64_abs() -> Result<()> {
+    let t = Tensor::new(&[-42i64, 1337], &Device::Cpu)?;
+    let t = t.abs()?;
+    assert_eq!(t.to_vec1::<i64>()?, [42, 1337]);
+    Ok(())
+}
+
+#[test]
+fn tril_triu_eye() -> Result<()> {
+    let t = Tensor::tril2(4, DType::F32, &Device::Cpu)?;
+    assert_eq!(
+        t.to_vec2::<f32>()?,
+        [
+            [1.0, 0.0, 0.0, 0.0],
+            [1.0, 1.0, 0.0, 0.0],
+            [1.0, 1.0, 1.0, 0.0],
+            [1.0, 1.0, 1.0, 1.0]
+        ],
+    );
+    let t = Tensor::triu2(4, DType::F32, &Device::Cpu)?;
+    assert_eq!(
+        t.to_vec2::<f32>()?,
+        [
+            [1.0, 1.0, 1.0, 1.0],
+            [0.0, 1.0, 1.0, 1.0],
+            [0.0, 0.0, 1.0, 1.0],
+            [0.0, 0.0, 0.0, 1.0]
+        ]
+    );
+    let t = Tensor::eye(4, DType::F32, &Device::Cpu)?;
+    assert_eq!(
+        t.to_vec2::<f32>()?,
+        [
+            [1.0, 0.0, 0.0, 0.0],
+            [0.0, 1.0, 0.0, 0.0],
+            [0.0, 0.0, 1.0, 0.0],
+            [0.0, 0.0, 0.0, 1.0]
+        ]
+    );
+    Ok(())
+}
+
+#[test]
+fn cumsum() -> Result<()> {
+    let t = &[3f32, 1., 4., 1., 5.];
+    let t = Tensor::new(t, &Device::Cpu)?;
+    assert_eq!(t.cumsum(0)?.to_vec1::<f32>()?, [3., 4., 8., 9., 14.]);
+    let t = t.unsqueeze(1)?;
+    assert_eq!(
+        t.cumsum(0)?.to_vec2::<f32>()?,
+        [[3.0], [4.0], [8.0], [9.0], [14.0]]
+    );
+    assert_eq!(
+        t.cumsum(1)?.to_vec2::<f32>()?,
+        [[3.0], [1.0], [4.0], [1.0], [5.0]]
+    );
+    let t = &[[3f32, 1., 4., 1., 5.], [2., 1., 7., 8., 2.]];
+    let t = Tensor::new(t, &Device::Cpu)?;
+    assert_eq!(
+        t.cumsum(1)?.to_vec2::<f32>()?,
+        [[3.0, 4.0, 8.0, 9.0, 14.0], [2.0, 3.0, 10.0, 18.0, 20.0]],
+    );
+    assert_eq!(
+        t.cumsum(0)?.to_vec2::<f32>()?,
+        [[3.0, 1.0, 4.0, 1.0, 5.0], [5.0, 2.0, 11.0, 9.0, 7.0]]
+    );
+    Ok(())
+}
+
+/// A helper function for floating point comparison. Both a and b must be 1D Tensor and contains the same amount of data.
+/// Assertion passes if the difference of all pairs of a and b is smaller than epsilon.
+fn assert_close(a: &Tensor, b: &Tensor, epsilon: f64) -> Result<()> {
+    let a_vec: Vec<f64> = a.to_vec1()?;
+    let b_vec: Vec<f64> = b.to_vec1()?;
+
+    assert_eq!(a_vec.len(), b_vec.len());
+    for (a, b) in a_vec.iter().zip(b_vec.iter()) {
+        assert!((a - b).abs() < epsilon);
+    }
+    Ok(())
+}
+
+#[test]
+fn log_sum_exp() -> Result<()> {
+    let input = Tensor::new(&[[1f64, 2., 3.], [4., 5., 6.]], &Device::Cpu)?;
+    let output = input.log_sum_exp(D::Minus1)?;
+    // The expectations obtained from pytorch.
+    let expected = Tensor::new(&[3.4076, 6.4076], &Device::Cpu)?;
+    assert_close(&output, &expected, 0.00001)?;
+    Ok(())
+}
+
+#[test]
+fn pow() -> Result<()> {
+    let lhs = Tensor::new(&[[1f32, 2., 3.], [4., 5., 6.]], &Device::Cpu)?;
+    let rhs = (&lhs - 2.)?;
+    let res = lhs.pow(&rhs)?;
+    assert_eq!(
+        test_utils::to_vec2_round(&res, 4)?,
+        [[1.0, 1.0, 3.0], [16.0, 125.0, 1296.0001]]
+    );
+    Ok(())
+}

candle-datasets/Cargo.toml

@@ -11,8 +11,8 @@ readme = "README.md"
 
 [dependencies]
 byteorder = { workspace = true }
-candle = { path = "../candle-core", version = "0.2.1", package = "candle-core" }
-candle-nn = { path = "../candle-nn", version = "0.2.1" }
+candle = { workspace = true }
+candle-nn = { workspace = true }
 hf-hub = { workspace = true}
 intel-mkl-src = { workspace = true, optional = true }
 memmap2 = { workspace = true }

candle-datasets/src/vision/cifar.rs

@@ -4,7 +4,9 @@
 //! <https://www.cs.toronto.edu/~kriz/cifar.html>
 //! The binary version of the dataset is used.
 use crate::vision::Dataset;
-use candle::{DType, Device, Result, Tensor};
+use candle::{DType, Device, Error, Result, Tensor};
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use parquet::file::reader::{FileReader, SerializedFileReader};
 use std::fs::File;
 use std::io::{BufReader, Read};
 
@@ -60,3 +62,58 @@ pub fn load_dir<T: AsRef<std::path::Path>>(dir: T) -> Result<Dataset> {
         labels: 10,
     })
 }
+
+fn load_parquet(parquet: SerializedFileReader<std::fs::File>) -> Result<(Tensor, Tensor)> {
+    let samples = parquet.metadata().file_metadata().num_rows() as usize;
+    let mut buffer_images: Vec<u8> = Vec::with_capacity(samples * 1_024);
+    let mut buffer_labels: Vec<u8> = Vec::with_capacity(samples);
+    for row in parquet.into_iter().flatten() {
+        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();
+                        buffer_images.extend(image.to_rgb8().as_raw());
+                    }
+                }
+            } else if let parquet::record::Field::Long(label) = field {
+                buffer_labels.push(*label as u8);
+            }
+        }
+    }
+    let images = (Tensor::from_vec(buffer_images, (samples, 3, 32, 32), &Device::Cpu)?
+        .to_dtype(DType::U8)?
+        / 255.)?;
+    let labels = Tensor::from_vec(buffer_labels, (samples,), &Device::Cpu)?;
+    Ok((images, labels))
+}
+
+pub fn load() -> Result<Dataset> {
+    let api = Api::new().map_err(|e| Error::Msg(format!("Api error: {e}")))?;
+    let dataset_id = "cifar10".to_string();
+    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("plain_text/test/0000.parquet")
+        .map_err(|e| Error::Msg(format!("Api error: {e}")))?;
+    let train_parquet_filename = repo
+        .get("plain_text/train/0000.parquet")
+        .map_err(|e| Error::Msg(format!("Api error: {e}")))?;
+    let test_parquet = SerializedFileReader::new(std::fs::File::open(test_parquet_filename)?)
+        .map_err(|e| Error::Msg(format!("Parquet error: {e}")))?;
+    let train_parquet = SerializedFileReader::new(std::fs::File::open(train_parquet_filename)?)
+        .map_err(|e| Error::Msg(format!("Parquet error: {e}")))?;
+    let (test_images, test_labels) = load_parquet(test_parquet)?;
+    let (train_images, train_labels) = load_parquet(train_parquet)?;
+    Ok(crate::vision::Dataset {
+        train_images,
+        train_labels,
+        test_images,
+        test_labels,
+        labels: 10,
+    })
+}

candle-examples/Cargo.toml

@@ -11,50 +11,75 @@ readme = "README.md"
 
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
-candle = { path = "../candle-core", version = "0.2.1", package = "candle-core" }
-candle-datasets = { path = "../candle-datasets", version = "0.2.1" }
-candle-nn = { path = "../candle-nn", version = "0.2.1" }
-candle-transformers = { path = "../candle-transformers", version = "0.2.1" }
-candle-flash-attn = { path = "../candle-flash-attn", version = "0.2.1", optional = true }
+candle = { workspace = true }
+candle-datasets = { workspace = true }
+candle-nn = { workspace = true }
+candle-transformers = { workspace = true }
+candle-flash-attn = { workspace = true, optional = true }
+candle-onnx = { workspace = true, optional = true }
+
+csv = "1.3.0"
+cudarc = { workspace = true, optional = true }
+half = { workspace = true, optional = true }
+hf-hub = { workspace = true, features = ["tokio"] }
+image = { workspace = true }
+intel-mkl-src = { workspace = true, optional = true }
+num-traits = { workspace = true }
+pyo3 = { version = "0.20.0", features = ["auto-initialize"], optional = true }
+rayon = { workspace = 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 }
-rayon = { workspace = true }
+symphonia = { version = "0.5.3", features = ["all"] }
+tokenizers = { workspace = true, features = ["onig"] }
+cpal= { version = "0.15.2", optional = true }
 
 [dev-dependencies]
 anyhow = { workspace = true }
 byteorder = { workspace = true }
 clap = { workspace = true }
-hf-hub = { workspace = true, features=["tokio"]}
 imageproc = { workspace = true }
 memmap2 = { workspace = true }
 rand = { workspace = true }
 rusttype = { workspace = true }
-tokenizers = { workspace = true, features = ["onig"] }
 tracing = { workspace = true }
 tracing-chrome = { workspace = true }
 tracing-subscriber = { workspace = true }
-wav = { workspace = true }
 # Necessary to disambiguate with tokio in wasm examples which are 1.28.1
 tokio = "1.29.1"
 
 [build-dependencies]
 anyhow = { workspace = true }
+bindgen_cuda = { version = "0.1.1", optional = true }
 
 [features]
 default = []
 accelerate = ["dep:accelerate-src", "candle/accelerate", "candle-nn/accelerate", "candle-transformers/accelerate"]
-cuda = ["candle/cuda", "candle-nn/cuda", "candle-transformers/cuda"]
+cuda = ["candle/cuda", "candle-nn/cuda", "candle-transformers/cuda", "dep:bindgen_cuda"]
 cudnn = ["candle/cudnn"]
-flash-attn = ["cuda", "dep:candle-flash-attn"]
+flash-attn = ["cuda", "candle-transformers/flash-attn", "dep:candle-flash-attn"]
 mkl = ["dep:intel-mkl-src", "candle/mkl", "candle-nn/mkl", "candle-transformers/mkl"]
 nccl = ["cuda", "cudarc/nccl", "dep:half"]
+onnx = ["candle-onnx"]
+metal = ["candle/metal", "candle-nn/metal"]
+microphone = ["cpal"]
 
 [[example]]
 name = "llama_multiprocess"
 required-features = ["cuda", "nccl", "flash-attn"]
+
+[[example]]
+name = "reinforcement-learning"
+required-features = ["pyo3"]
+
+[[example]]
+name = "onnx"
+required-features = ["onnx"]
+
+[[example]]
+name = "onnx_basics"
+required-features = ["onnx"]
+
+[[example]]
+name = "whisper-microphone"
+required-features = ["microphone"]

candle-examples/build.rs

@@ -4,235 +4,28 @@ use std::io::Write;
 use std::path::PathBuf;
 
 struct KernelDirectories {
-    kernel_dir: &'static str,
+    kernel_glob: &'static str,
     rust_target: &'static str,
     include_dirs: &'static [&'static str],
 }
 
-const DIRS: [KernelDirectories; 1] = [KernelDirectories {
-    kernel_dir: "examples/custom-ops/kernels/",
+const KERNEL_DIRS: [KernelDirectories; 1] = [KernelDirectories {
+    kernel_glob: "examples/custom-ops/kernels/*.cu",
     rust_target: "examples/custom-ops/cuda_kernels.rs",
     include_dirs: &[],
 }];
 
-impl KernelDirectories {
-    fn maybe_build_ptx(
-        &self,
-        cu_file: &std::path::Path,
-        ptx_file: &std::path::Path,
-        compute_cap: usize,
-    ) -> Result<()> {
-        let should_compile = if ptx_file.exists() {
-            let ptx_modified = ptx_file.metadata()?.modified()?;
-            let cu_modified = cu_file.metadata()?.modified()?;
-            cu_modified.duration_since(ptx_modified).is_ok()
-        } else {
-            true
-        };
-        if should_compile {
-            #[cfg(feature = "cuda")]
-            {
-                let mut command = std::process::Command::new("nvcc");
-                let out_dir = ptx_file.parent().context("no parent for ptx file")?;
-                let include_dirs: Vec<String> =
-                    self.include_dirs.iter().map(|c| format!("-I{c}")).collect();
-                command
-                    .arg(format!("--gpu-architecture=sm_{compute_cap}"))
-                    .arg("--ptx")
-                    .args(["--default-stream", "per-thread"])
-                    .args(["--output-directory", out_dir.to_str().unwrap()])
-                    .arg(format!("-I/{}", self.kernel_dir))
-                    .args(include_dirs)
-                    .arg(cu_file);
-                let output = command
-                    .spawn()
-                    .context("failed spawning nvcc")?
-                    .wait_with_output()?;
-                if !output.status.success() {
-                    anyhow::bail!(
-                    "nvcc error while compiling {cu_file:?}:\n\n# stdout\n{:#}\n\n# stderr\n{:#}",
-                    String::from_utf8_lossy(&output.stdout),
-                    String::from_utf8_lossy(&output.stderr)
-                )
-                }
-            }
-            #[cfg(not(feature = "cuda"))]
-            std::fs::OpenOptions::new()
-                .create(true)
-                .write(true)
-                .open(ptx_file)?;
-        }
-        Ok(())
-    }
-    fn process(&self, out_dir: &std::path::Path, compute_cap: usize) -> Result<()> {
-        println!("cargo:rerun-if-changed={}", self.kernel_dir);
-        let kernel_dir = PathBuf::from(self.kernel_dir);
-        let out_dir = out_dir.join(self.kernel_dir);
-        if !out_dir.exists() {
-            std::fs::create_dir_all(&out_dir)?;
-        }
-        let mut cu_files = vec![];
-        let mut cuh_files = vec![];
-        for file in std::fs::read_dir(kernel_dir)?.flatten() {
-            let file = file.path();
-            match file.extension().and_then(|v| v.to_str()) {
-                Some("cu") => cu_files.push(file),
-                Some("cuh") => cuh_files.push(file),
-                _ => {}
-            }
-        }
-
-        let mut ptx_paths = vec![];
-        for cu_file in cu_files.iter() {
-            let file_stem = cu_file
-                .file_stem()
-                .with_context(|| format!("no stem {cu_file:?}"))?;
-            let file_stem = file_stem.to_string_lossy().into_owned();
-            let ptx_file = out_dir.join(&format!("{file_stem}.ptx"));
-            self.maybe_build_ptx(cu_file, &ptx_file, compute_cap)?;
-            ptx_paths.push(ptx_file);
-        }
-
-        let regenerate_rs_file = true;
-        if regenerate_rs_file {
-            let mut file = std::fs::File::create(self.rust_target)?;
-            for ptx_path in ptx_paths {
-                let name = ptx_path
-                    .file_stem()
-                    .context("empty stem")?
-                    .to_string_lossy();
-                file.write_all(b"#[rustfmt::skip]\n")?;
-                let const_definition = format!(
-                    r#"pub const {}: &str = include_str!(concat!(env!("OUT_DIR"), "/{}/{name}.ptx"));"#,
-                    name.to_uppercase().replace('.', "_"),
-                    self.kernel_dir,
-                );
-                file.write_all(const_definition.as_bytes())?;
-                file.write_all(b"\n")?;
-            }
-        }
-        Ok(())
-    }
-}
-
 fn main() -> Result<()> {
     println!("cargo:rerun-if-changed=build.rs");
 
-    let out_dir = std::env::var("OUT_DIR").context("OUT_DIR not set")?;
-    let out_dir = PathBuf::from(out_dir);
     #[cfg(feature = "cuda")]
-    set_cuda_include_dir()?;
-    #[cfg(feature = "cuda")]
-    let compute_cap = compute_cap()?;
-    #[cfg(not(feature = "cuda"))]
-    let compute_cap = 0;
-    for d in DIRS {
-        d.process(&out_dir, compute_cap)?
+    {
+        for kdir in KERNEL_DIRS.iter() {
+            let builder = bindgen_cuda::Builder::default().kernel_paths_glob(kdir.kernel_glob);
+            println!("cargo:info={builder:?}");
+            let bindings = builder.build_ptx().unwrap();
+            bindings.write(kdir.rust_target).unwrap()
+        }
     }
     Ok(())
 }
-
-fn set_cuda_include_dir() -> Result<()> {
-    // NOTE: copied from cudarc build.rs.
-    let env_vars = [
-        "CUDA_PATH",
-        "CUDA_ROOT",
-        "CUDA_TOOLKIT_ROOT_DIR",
-        "CUDNN_LIB",
-    ];
-    let env_vars = env_vars
-        .into_iter()
-        .map(std::env::var)
-        .filter_map(Result::ok)
-        .map(Into::<PathBuf>::into);
-
-    let roots = [
-        "/usr",
-        "/usr/local/cuda",
-        "/opt/cuda",
-        "/usr/lib/cuda",
-        "C:/Program Files/NVIDIA GPU Computing Toolkit",
-        "C:/CUDA",
-    ];
-    let roots = roots.into_iter().map(Into::<PathBuf>::into);
-    let root = env_vars
-        .chain(roots)
-        .find(|path| path.join("include").join("cuda.h").is_file())
-        .context("cannot find include/cuda.h")?;
-    println!(
-        "cargo:rustc-env=CUDA_INCLUDE_DIR={}",
-        root.join("include").display()
-    );
-    Ok(())
-}
-
-#[allow(unused)]
-fn compute_cap() -> Result<usize> {
-    // Grab compute code from nvidia-smi
-    let mut compute_cap = {
-        let out = std::process::Command::new("nvidia-smi")
-                    .arg("--query-gpu=compute_cap")
-                    .arg("--format=csv")
-                    .output()
-                    .context("`nvidia-smi` failed. Ensure that you have CUDA installed and that `nvidia-smi` is in your PATH.")?;
-        let out = std::str::from_utf8(&out.stdout).context("stdout is not a utf8 string")?;
-        let mut lines = out.lines();
-        assert_eq!(
-            lines.next().context("missing line in stdout")?,
-            "compute_cap"
-        );
-        let cap = lines
-            .next()
-            .context("missing line in stdout")?
-            .replace('.', "");
-        cap.parse::<usize>()
-            .with_context(|| format!("cannot parse as int {cap}"))?
-    };
-
-    // Grab available GPU codes from nvcc and select the highest one
-    let max_nvcc_code = {
-        let out = std::process::Command::new("nvcc")
-                    .arg("--list-gpu-code")
-                    .output()
-                    .expect("`nvcc` failed. Ensure that you have CUDA installed and that `nvcc` is in your PATH.");
-        let out = std::str::from_utf8(&out.stdout).unwrap();
-
-        let out = out.lines().collect::<Vec<&str>>();
-        let mut codes = Vec::with_capacity(out.len());
-        for code in out {
-            let code = code.split('_').collect::<Vec<&str>>();
-            if !code.is_empty() && code.contains(&"sm") {
-                if let Ok(num) = code[1].parse::<usize>() {
-                    codes.push(num);
-                }
-            }
-        }
-        codes.sort();
-        if !codes.contains(&compute_cap) {
-            anyhow::bail!(
-                "nvcc cannot target gpu arch {compute_cap}. Available nvcc targets are {codes:?}."
-            );
-        }
-        *codes.last().unwrap()
-    };
-
-    // If nvidia-smi compute_cap is higher than the highest gpu code from nvcc,
-    // then choose the highest gpu code in nvcc
-    if compute_cap > max_nvcc_code {
-        println!(
-            "cargo:warning=Lowering gpu arch {compute_cap} to max nvcc target {max_nvcc_code}."
-        );
-        compute_cap = max_nvcc_code;
-    }
-
-    println!("cargo:rerun-if-env-changed=CUDA_COMPUTE_CAP");
-
-    if let Ok(compute_cap_str) = std::env::var("CUDA_COMPUTE_CAP") {
-        compute_cap = compute_cap_str
-            .parse::<usize>()
-            .with_context(|| format!("cannot parse as usize '{compute_cap_str}'"))?;
-        println!("cargo:warning=Using gpu arch {compute_cap} from $CUDA_COMPUTE_CAP");
-    }
-    println!("cargo:rustc-env=CUDA_COMPUTE_CAP=sm_{compute_cap}");
-    Ok(compute_cap)
-}

candle-examples/examples/bert/README.md

@@ -0,0 +1,86 @@
+# candle-bert
+
+Bert is a general large language model. In this example it can be used for two
+different tasks:
+
+- Compute sentence embeddings for a prompt.
+- Compute similarities between a set of sentences.
+
+## Sentence embeddings
+
+Bert is used to compute the sentence embeddings for a prompt. The model weights
+are downloaded from the hub on the first run.
+
+```bash
+cargo run --example bert --release -- --prompt "Here is a test sentence"
+
+> [[[ 0.0798, -0.0665, -0.0247, ..., -0.1082, -0.1000, -0.2751],
+>   [ 0.4218,  0.2690,  0.2740, ...,  0.3889,  1.3503,  0.9908],
+>   [ 0.0466,  0.3041, -0.1143, ...,  0.4427,  0.6926, -0.1515],
+>   ...
+>   [ 0.3396,  0.4320, -0.4408, ...,  0.9212,  0.2331, -0.6777],
+>   [ 0.2789,  0.7539,  0.4306, ..., -0.0095,  0.3375, -1.7529],
+>   [ 0.6737,  0.7882,  0.0548, ...,  0.1836,  0.7299, -0.6617]]]
+> Tensor[[1, 7, 384], f32]
+```
+
+### Custom models
+
+You can specify different models, such as BGE, with the `--model-id` flag:
+
+```bash
+cargo run  --example bert --release -- \
+--model-id BAAI/bge-large-zh-v1.5 \
+--prompt "Here is a test sentence"
+Loaded and encoded 435.70775ms
+[[[ 3.0944e-1, -7.8455e-5,  -1.2768e0, ...,  1.3755e-2, -3.2371e-1,  2.3819e-1],
+  [-2.8506e-1,  1.9953e-1,  -1.3076e0, ...,  6.9819e-2,  1.0833e-2,  -1.1512e0],
+  [ 3.9892e-1,  2.0000e-1, -9.3178e-1, ..., -4.1393e-1, -4.9644e-2, -3.3786e-1],
+  ...
+  [ 6.0345e-1,  3.5744e-1,  -1.2672e0, ..., -6.9165e-1, -3.4973e-3, -8.4214e-1],
+  [ 3.9218e-1, -3.2735e-1,  -1.3123e0, ..., -4.9318e-1, -5.1334e-1, -3.6391e-1],
+  [ 3.0978e-1,  2.5662e-4,  -1.2773e0, ...,  1.3357e-2, -3.2390e-1,  2.3858e-1]]]
+Tensor[[1, 9, 1024], f32]
+Took 176.744667ms
+```
+
+### Gelu approximation
+
+You can get a speedup by using an approximation of the gelu activation, with a
+small loss of precision, by passing the `--approximate-gelu` flag:
+
+```bash
+$ cargo run  --example bert --release -- \
+--model-id BAAI/bge-large-zh-v1.5 \
+--prompt "Here is a test sentence" \
+--approximate-gelu
+Loaded and encoded 244.388042ms
+[[[ 3.1048e-1, -6.0339e-4,  -1.2758e0, ...,  1.3718e-2, -3.2362e-1,  2.3775e-1],
+  [-2.8354e-1,  1.9984e-1,  -1.3077e0, ...,  6.9390e-2,  9.9681e-3,  -1.1531e0],
+  [ 3.9947e-1,  1.9917e-1, -9.3178e-1, ..., -4.1301e-1, -5.0719e-2, -3.3955e-1],
+  ...
+  [ 6.0499e-1,  3.5664e-1,  -1.2642e0, ..., -6.9134e-1, -3.4581e-3, -8.4471e-1],
+  [ 3.9311e-1, -3.2812e-1,  -1.3105e0, ..., -4.9291e-1, -5.1270e-1, -3.6543e-1],
+  [ 3.1082e-1, -2.6737e-4,  -1.2762e0, ...,  1.3319e-2, -3.2381e-1,  2.3815e-1]]]
+Tensor[[1, 9, 1024], f32]
+Took 116.840791ms
+```
+
+## Similarities
+
+In this example, Bert is used to compute the sentence embeddings for a set of
+sentences (hardcoded in the examples). Then cosine similarities are computed for
+each sentence pair and they are reported by decreasing values, hence the first
+reported pair contains the two sentences that have the highest similarity score.
+The sentence embeddings are computed using average pooling through all the
+sentence tokens, including some potential padding.
+
+```bash
+cargo run --example bert --release
+
+> score: 0.85 'The new movie is awesome' 'The new movie is so great'
+> score: 0.61 'The cat sits outside' 'The cat plays in the garden'
+> score: 0.52 'I love pasta' 'Do you like pizza?'
+> score: 0.23 'The new movie is awesome' 'Do you like pizza?'
+> score: 0.22 'I love pasta' 'The new movie is awesome'
+```

candle-examples/examples/bert/main.rs

@@ -3,14 +3,13 @@ extern crate intel_mkl_src;
 
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
-mod model;
+use candle_transformers::models::bert::{BertModel, Config, HiddenAct, DTYPE};
 
-use anyhow::{anyhow, Error as E, Result};
+use anyhow::{Error as E, Result};
 use candle::Tensor;
 use candle_nn::VarBuilder;
 use clap::Parser;
-use hf_hub::{api::sync::Api, Cache, Repo, RepoType};
-use model::{BertModel, Config, DTYPE};
+use hf_hub::{api::sync::Api, Repo, RepoType};
 use tokenizers::{PaddingParams, Tokenizer};
 
 #[derive(Parser, Debug)]
@@ -20,10 +19,6 @@ struct Args {
     #[arg(long)]
     cpu: bool,
 
-    /// Run offline (you must have the files already cached)
-    #[arg(long)]
-    offline: bool,
-
     /// Enable tracing (generates a trace-timestamp.json file).
     #[arg(long)]
     tracing: bool,
@@ -39,6 +34,10 @@ struct Args {
     #[arg(long)]
     prompt: Option<String>,
 
+    /// Use the pytorch weights rather than the safetensors ones
+    #[arg(long)]
+    use_pth: bool,
+
     /// The number of times to run the prompt.
     #[arg(long, default_value = "1")]
     n: usize,
@@ -46,6 +45,10 @@ struct Args {
     /// L2 normalization for embeddings.
     #[arg(long, default_value = "true")]
     normalize_embeddings: bool,
+
+    /// Use tanh based approximation for Gelu instead of erf implementation.
+    #[arg(long, default_value = "false")]
+    approximate_gelu: bool,
 }
 
 impl Args {
@@ -61,35 +64,30 @@ impl Args {
         };
 
         let repo = Repo::with_revision(model_id, RepoType::Model, revision);
-        let (config_filename, tokenizer_filename, weights_filename) = if self.offline {
-            let cache = Cache::default().repo(repo);
-            (
-                cache
-                    .get("config.json")
-                    .ok_or(anyhow!("Missing config file in cache"))?,
-                cache
-                    .get("tokenizer.json")
-                    .ok_or(anyhow!("Missing tokenizer file in cache"))?,
-                cache
-                    .get("model.safetensors")
-                    .ok_or(anyhow!("Missing weights file in cache"))?,
-            )
-        } else {
+        let (config_filename, tokenizer_filename, weights_filename) = {
             let api = Api::new()?;
             let api = api.repo(repo);
-            (
-                api.get("config.json")?,
-                api.get("tokenizer.json")?,
-                api.get("model.safetensors")?,
-            )
+            let config = api.get("config.json")?;
+            let tokenizer = api.get("tokenizer.json")?;
+            let weights = if self.use_pth {
+                api.get("pytorch_model.bin")?
+            } else {
+                api.get("model.safetensors")?
+            };
+            (config, tokenizer, weights)
         };
         let config = std::fs::read_to_string(config_filename)?;
-        let config: Config = serde_json::from_str(&config)?;
+        let mut config: Config = serde_json::from_str(&config)?;
         let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
 
-        let weights = unsafe { candle::safetensors::MmapedFile::new(weights_filename)? };
-        let weights = weights.deserialize()?;
-        let vb = VarBuilder::from_safetensors(vec![weights], DTYPE, &device);
+        let vb = if self.use_pth {
+            VarBuilder::from_pth(&weights_filename, DTYPE, &device)?
+        } else {
+            unsafe { VarBuilder::from_mmaped_safetensors(&[weights_filename], DTYPE, &device)? }
+        };
+        if self.approximate_gelu {
+            config.hidden_act = HiddenAct::GeluApproximate;
+        }
         let model = BertModel::load(vb, &config)?;
         Ok((model, tokenizer))
     }

candle-examples/examples/bigcode/README.md

@@ -0,0 +1,19 @@
+# candle-starcoder: code generation model
+
+[StarCoder/BigCode](https://huggingface.co/bigcode/starcoderbase-1b) is a LLM
+model specialized to code generation. The initial model was trained on 80
+programming languages.
+
+## Running some example
+
+```bash
+cargo run --example bigcode --release -- --prompt "fn fact(n: u64) -> u64 "
+
+> fn fact(n: u64) -> u64  {
+>     if n == 0 {
+>         1
+>     } else {
+>         n * fact(n - 1)
+>     }
+> }
+```

candle-examples/examples/bigcode/main.rs

@@ -7,8 +7,7 @@ extern crate accelerate_src;
 use anyhow::{Error as E, Result};
 use clap::Parser;
 
-mod model;
-use model::{Config, GPTBigCode};
+use candle_transformers::models::bigcode::{Config, GPTBigCode};
 
 use candle::{DType, Device, Tensor};
 use candle_nn::VarBuilder;
@@ -29,9 +28,10 @@ impl TextGeneration {
         tokenizer: Tokenizer,
         seed: u64,
         temp: Option<f64>,
+        top_p: Option<f64>,
         device: &Device,
     ) -> Self {
-        let logits_processor = LogitsProcessor::new(seed, temp);
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
         Self {
             model,
             tokenizer,
@@ -95,6 +95,10 @@ struct Args {
     #[arg(long)]
     temperature: Option<f64>,
 
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
     /// The seed to use when generating random samples.
     #[arg(long, default_value_t = 299792458)]
     seed: u64,
@@ -134,23 +138,21 @@ fn main() -> Result<()> {
     println!("retrieved the files in {:?}", start.elapsed());
     let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
 
-    let weights = filenames
-        .iter()
-        .map(|f| Ok(unsafe { candle::safetensors::MmapedFile::new(f)? }))
-        .collect::<Result<Vec<_>>>()?;
-    let weights = weights
-        .iter()
-        .map(|f| Ok(f.deserialize()?))
-        .collect::<Result<Vec<_>>>()?;
-
     let start = std::time::Instant::now();
     let device = candle_examples::device(args.cpu)?;
-    let vb = VarBuilder::from_safetensors(weights, DType::F32, &device);
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, DType::F32, &device)? };
     let config = Config::starcoder_1b();
     let model = GPTBigCode::load(vb, config)?;
     println!("loaded the model in {:?}", start.elapsed());
 
-    let mut pipeline = TextGeneration::new(model, tokenizer, args.seed, args.temperature, &device);
+    let mut pipeline = TextGeneration::new(
+        model,
+        tokenizer,
+        args.seed,
+        args.temperature,
+        args.top_p,
+        &device,
+    );
     pipeline.run(&args.prompt, args.sample_len)?;
     Ok(())
 }

candle-examples/examples/blip/README.md

@@ -0,0 +1,19 @@
+# candle-blip
+
+The
+[blip-image-captioning](https://huggingface.co/Salesforce/blip-image-captioning-base)
+model can generate captions for an input image.
+
+## Running on an example
+
+```bash
+cargo run --example blip --release -- --image candle-examples/examples/yolo-v8/assets/bike.jpg
+```
+
+```
+Running on CPU, to run on GPU, build this example with `--features cuda`
+loaded image Tensor[dims 3, 384, 384; f32]
+model built
+several cyclists are riding down a road with cars behind them%
+```
+![Leading group, Giro d'Italia 2021](../yolo-v8/assets/bike.jpg)

candle-examples/examples/blip/main.rs

@@ -0,0 +1,154 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::Error as E;
+use clap::Parser;
+
+use candle::{DType, Device, Result, Tensor};
+use candle_examples::token_output_stream::TokenOutputStream;
+use candle_nn::VarBuilder;
+use candle_transformers::models::blip;
+use candle_transformers::models::quantized_blip;
+
+use tokenizers::Tokenizer;
+
+enum Model {
+    M(blip::BlipForConditionalGeneration),
+    Q(quantized_blip::BlipForConditionalGeneration),
+}
+
+impl Model {
+    fn text_decoder_forward(&mut self, xs: &Tensor, img_xs: &Tensor) -> Result<Tensor> {
+        match self {
+            Self::M(m) => m.text_decoder().forward(xs, img_xs),
+            Self::Q(m) => m.text_decoder().forward(xs, img_xs),
+        }
+    }
+}
+
+// TODO: Maybe add support for the conditional prompt.
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    tokenizer: Option<String>,
+
+    #[arg(long)]
+    image: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Use the quantized version of the model.
+    #[arg(long)]
+    quantized: bool,
+}
+
+const SEP_TOKEN_ID: u32 = 102;
+
+/// Loads an image from disk using the image crate, this returns a tensor with shape
+/// (3, 384, 384). OpenAI normalization is applied.
+pub fn load_image<P: AsRef<std::path::Path>>(p: P) -> Result<Tensor> {
+    let img = image::io::Reader::open(p)?
+        .decode()
+        .map_err(candle::Error::wrap)?
+        .resize_to_fill(384, 384, image::imageops::FilterType::Triangle);
+    let img = img.to_rgb8();
+    let data = img.into_raw();
+    let data = Tensor::from_vec(data, (384, 384, 3), &Device::Cpu)?.permute((2, 0, 1))?;
+    let mean =
+        Tensor::new(&[0.48145466f32, 0.4578275, 0.40821073], &Device::Cpu)?.reshape((3, 1, 1))?;
+    let std = Tensor::new(&[0.26862954f32, 0.261_302_6, 0.275_777_1], &Device::Cpu)?
+        .reshape((3, 1, 1))?;
+    (data.to_dtype(candle::DType::F32)? / 255.)?
+        .broadcast_sub(&mean)?
+        .broadcast_div(&std)
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    let model_file = match args.model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            if args.quantized {
+                let api = api.model("lmz/candle-blip".to_string());
+                api.get("blip-image-captioning-large-q4k.gguf")?
+            } else {
+                let api = api.repo(hf_hub::Repo::with_revision(
+                    "Salesforce/blip-image-captioning-large".to_string(),
+                    hf_hub::RepoType::Model,
+                    "refs/pr/18".to_string(),
+                ));
+                api.get("model.safetensors")?
+            }
+        }
+        Some(model) => model.into(),
+    };
+    let tokenizer = match args.tokenizer {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model("Salesforce/blip-image-captioning-large".to_string());
+            api.get("tokenizer.json")?
+        }
+        Some(file) => file.into(),
+    };
+    let tokenizer = Tokenizer::from_file(tokenizer).map_err(E::msg)?;
+    let mut tokenizer = TokenOutputStream::new(tokenizer);
+    let mut logits_processor =
+        candle_transformers::generation::LogitsProcessor::new(1337, None, None);
+
+    let config = blip::Config::image_captioning_large();
+
+    let device = candle_examples::device(args.cpu)?;
+    let (image_embeds, device, mut model) = if args.quantized {
+        let device = Device::Cpu;
+        let image = load_image(args.image)?.to_device(&device)?;
+        println!("loaded image {image:?}");
+
+        let vb = quantized_blip::VarBuilder::from_gguf(model_file, &device)?;
+        let model = quantized_blip::BlipForConditionalGeneration::new(&config, vb)?;
+        let image_embeds = image.unsqueeze(0)?.apply(model.vision_model())?;
+        (image_embeds, device, Model::Q(model))
+    } else {
+        let image = load_image(args.image)?.to_device(&device)?;
+        println!("loaded image {image:?}");
+
+        let vb =
+            unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+        let model = blip::BlipForConditionalGeneration::new(&config, vb)?;
+        let image_embeds = image.unsqueeze(0)?.apply(model.vision_model())?;
+        (image_embeds, device, Model::M(model))
+    };
+
+    let mut token_ids = vec![30522u32];
+    for index in 0..1000 {
+        let context_size = if index > 0 { 1 } else { token_ids.len() };
+        let start_pos = token_ids.len().saturating_sub(context_size);
+        let input_ids = Tensor::new(&token_ids[start_pos..], &device)?.unsqueeze(0)?;
+        let logits = model.text_decoder_forward(&input_ids, &image_embeds)?;
+        let logits = logits.squeeze(0)?;
+        let logits = logits.get(logits.dim(0)? - 1)?;
+        let token = logits_processor.sample(&logits)?;
+        if token == SEP_TOKEN_ID {
+            break;
+        }
+        token_ids.push(token);
+        if let Some(t) = tokenizer.next_token(token)? {
+            use std::io::Write;
+            print!("{t}");
+            std::io::stdout().flush()?;
+        }
+    }
+    if let Some(rest) = tokenizer.decode_rest().map_err(E::msg)? {
+        print!("{rest}");
+    }
+    println!();
+    Ok(())
+}

candle-examples/examples/chatglm/main.rs

@@ -0,0 +1,237 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::{Error as E, Result};
+use clap::Parser;
+
+use candle_transformers::models::chatglm::{Config, Model};
+
+use candle::{DType, Device, Tensor};
+use candle_nn::VarBuilder;
+use candle_transformers::generation::LogitsProcessor;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::Tokenizer;
+
+struct TextGeneration {
+    model: Model,
+    device: Device,
+    tokenizer: Tokenizer,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+    verbose_prompt: bool,
+}
+
+impl TextGeneration {
+    #[allow(clippy::too_many_arguments)]
+    fn new(
+        model: Model,
+        tokenizer: Tokenizer,
+        seed: u64,
+        temp: Option<f64>,
+        top_p: Option<f64>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        verbose_prompt: bool,
+        device: &Device,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            tokenizer,
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            verbose_prompt,
+            device: device.clone(),
+        }
+    }
+
+    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
+        use std::io::Write;
+        println!("starting the inference loop");
+        let tokens = self.tokenizer.encode(prompt, true).map_err(E::msg)?;
+        if tokens.is_empty() {
+            anyhow::bail!("Empty prompts are not supported in the chatglm model.")
+        }
+        if self.verbose_prompt {
+            for (token, id) in tokens.get_tokens().iter().zip(tokens.get_ids().iter()) {
+                let token = token.replace('▁', " ").replace("<0x0A>", "\n");
+                println!("{id:7} -> '{token}'");
+            }
+        }
+        let mut tokens = tokens.get_ids().to_vec();
+        let mut generated_tokens = 0usize;
+        let eos_token = match self.tokenizer.get_vocab(true).get("</s>") {
+            Some(token) => *token,
+            None => anyhow::bail!("cannot find the endoftext token"),
+        };
+        print!("{prompt}");
+        std::io::stdout().flush()?;
+        let start_gen = std::time::Instant::now();
+        for index in 0..sample_len {
+            let context_size = if index > 0 { 1 } else { tokens.len() };
+            let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
+            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+            let logits = self.model.forward(&input)?;
+            let logits = logits.squeeze(0)?.to_dtype(DType::F32)?;
+            let logits = if self.repeat_penalty == 1. {
+                logits
+            } else {
+                let start_at = tokens.len().saturating_sub(self.repeat_last_n);
+                candle_transformers::utils::apply_repeat_penalty(
+                    &logits,
+                    self.repeat_penalty,
+                    &tokens[start_at..],
+                )?
+            };
+
+            let next_token = self.logits_processor.sample(&logits)?;
+            tokens.push(next_token);
+            generated_tokens += 1;
+            if next_token == eos_token {
+                break;
+            }
+            let token = self.tokenizer.decode(&[next_token], true).map_err(E::msg)?;
+            print!("{token}");
+            std::io::stdout().flush()?;
+        }
+        let dt = start_gen.elapsed();
+        println!(
+            "\n{generated_tokens} tokens generated ({:.2} token/s)",
+            generated_tokens as f64 / dt.as_secs_f64(),
+        );
+        Ok(())
+    }
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    /// Display the token for the specified prompt.
+    #[arg(long)]
+    verbose_prompt: bool,
+
+    #[arg(long)]
+    prompt: String,
+
+    /// The temperature used to generate samples.
+    #[arg(long)]
+    temperature: Option<f64>,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// The seed to use when generating random samples.
+    #[arg(long, default_value_t = 299792458)]
+    seed: u64,
+
+    /// The length of the sample to generate (in tokens).
+    #[arg(long, short = 'n', default_value_t = 5000)]
+    sample_len: usize,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long)]
+    revision: Option<String>,
+
+    #[arg(long)]
+    weight_file: Option<String>,
+
+    #[arg(long)]
+    tokenizer: Option<String>,
+
+    /// Penalty to be applied for repeating tokens, 1. means no penalty.
+    #[arg(long, default_value_t = 1.1)]
+    repeat_penalty: f32,
+
+    /// The context size to consider for the repeat penalty.
+    #[arg(long, default_value_t = 64)]
+    repeat_last_n: usize,
+}
+
+fn main() -> Result<()> {
+    use tracing_chrome::ChromeLayerBuilder;
+    use tracing_subscriber::prelude::*;
+
+    let args = Args::parse();
+    let _guard = if args.tracing {
+        let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
+        tracing_subscriber::registry().with(chrome_layer).init();
+        Some(guard)
+    } else {
+        None
+    };
+    println!(
+        "avx: {}, neon: {}, simd128: {}, f16c: {}",
+        candle::utils::with_avx(),
+        candle::utils::with_neon(),
+        candle::utils::with_simd128(),
+        candle::utils::with_f16c()
+    );
+    println!(
+        "temp: {:.2} repeat-penalty: {:.2} repeat-last-n: {}",
+        args.temperature.unwrap_or(0.),
+        args.repeat_penalty,
+        args.repeat_last_n
+    );
+
+    let start = std::time::Instant::now();
+    let api = Api::new()?;
+    let model_id = match args.model_id {
+        Some(model_id) => model_id.to_string(),
+        None => "THUDM/chatglm3-6b".to_string(),
+    };
+    let revision = match args.revision {
+        Some(rev) => rev.to_string(),
+        None => "main".to_string(),
+    };
+    let repo = api.repo(Repo::with_revision(model_id, RepoType::Model, revision));
+    let tokenizer_filename = match args.tokenizer {
+        Some(file) => std::path::PathBuf::from(file),
+        None => api
+            .model("lmz/candle-chatglm".to_string())
+            .get("chatglm-tokenizer.json")?,
+    };
+    let filenames = match args.weight_file {
+        Some(weight_file) => vec![std::path::PathBuf::from(weight_file)],
+        None => candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.json")?,
+    };
+    println!("retrieved the files in {:?}", start.elapsed());
+    let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
+
+    let start = std::time::Instant::now();
+    let config = Config::glm3_6b();
+    let device = candle_examples::device(args.cpu)?;
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, DType::F32, &device)? };
+    let model = Model::new(&config, vb)?;
+
+    println!("loaded the model in {:?}", start.elapsed());
+
+    let mut pipeline = TextGeneration::new(
+        model,
+        tokenizer,
+        args.seed,
+        args.temperature,
+        args.top_p,
+        args.repeat_penalty,
+        args.repeat_last_n,
+        args.verbose_prompt,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}

candle-examples/examples/convmixer/main.rs

@@ -0,0 +1,59 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use clap::Parser;
+
+use candle::{DType, IndexOp, D};
+use candle_nn::{Module, VarBuilder};
+use candle_transformers::models::convmixer;
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    image: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    let device = candle_examples::device(args.cpu)?;
+
+    let image = candle_examples::imagenet::load_image224(args.image)?;
+    println!("loaded image {image:?}");
+
+    let model_file = match args.model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model("lmz/candle-convmixer".into());
+            api.get("convmixer_1024_20_ks9_p14.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+    let model = convmixer::c1024_20(1000, vb)?;
+    println!("model built");
+    let logits = model.forward(&image.unsqueeze(0)?)?;
+    let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
+        .i(0)?
+        .to_vec1::<f32>()?;
+    let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
+    prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
+    for &(category_idx, pr) in prs.iter().take(5) {
+        println!(
+            "{:24}: {:.2}%",
+            candle_examples::imagenet::CLASSES[category_idx],
+            100. * pr
+        );
+    }
+    Ok(())
+}

candle-examples/examples/convnext/README.md

@@ -0,0 +1,23 @@
+# candle-convnext
+
+[A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) and
+[ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808).
+
+This candle implementation uses a pre-trained ConvNeXt network for inference. The
+classification head has been trained on the ImageNet dataset and returns the
+probabilities for the top-5 classes.
+
+## Running an example
+
+```
+$ cargo run --example convnext --release -- --image candle-examples/examples/yolo-v8/assets/bike.jpg --which tiny
+
+loaded image Tensor[dims 3, 224, 224; f32]
+model built
+mountain bike, all-terrain bike, off-roader: 84.09%
+bicycle-built-for-two, tandem bicycle, tandem: 4.15%
+maillot                 : 0.74%
+crash helmet            : 0.54%
+unicycle, monocycle     : 0.44%
+
+```

candle-examples/examples/convnext/main.rs

@@ -0,0 +1,126 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use clap::{Parser, ValueEnum};
+
+use candle::{DType, IndexOp, D};
+use candle_nn::{Module, VarBuilder};
+use candle_transformers::models::convnext;
+
+#[derive(Clone, Copy, Debug, ValueEnum)]
+enum Which {
+    Atto,
+    Femto,
+    Pico,
+    Nano,
+    Tiny,
+    Small,
+    Base,
+    Large,
+    AttoV2,
+    FemtoV2,
+    PicoV2,
+    NanoV2,
+    TinyV2,
+    BaseV2,
+    LargeV2,
+    XLarge,
+    Huge,
+}
+
+impl Which {
+    fn model_filename(&self) -> String {
+        let name = match self {
+            Self::Atto => "convnext_atto.d2_in1k",
+            Self::Femto => "convnext_femto.d1_in1k",
+            Self::Pico => "convnext_pico.d1_in1k",
+            Self::Nano => "convnext_nano.d1h_in1k",
+            Self::Tiny => "convnext_tiny.fb_in1k",
+            Self::Small => "convnext_small.fb_in1k",
+            Self::Base => "convnext_base.fb_in1k",
+            Self::Large => "convnext_large.fb_in1k",
+            Self::AttoV2 => "convnextv2_atto.fcmae_ft_in1k",
+            Self::FemtoV2 => "convnextv2_femto.fcmae_ft_in1k",
+            Self::PicoV2 => "convnextv2_pico.fcmae_ft_in1k",
+            Self::NanoV2 => "convnextv2_nano.fcmae_ft_in1k",
+            Self::TinyV2 => "convnextv2_tiny.fcmae_ft_in1k",
+            Self::BaseV2 => "convnextv2_base.fcmae_ft_in1k",
+            Self::LargeV2 => "convnextv2_large.fcmae_ft_in1k",
+            Self::XLarge => "convnext_xlarge.fb_in22k_ft_in1k",
+            Self::Huge => "convnextv2_huge.fcmae_ft_in1k",
+        };
+
+        format!("timm/{name}")
+    }
+
+    fn config(&self) -> convnext::Config {
+        match self {
+            Self::Atto | Self::AttoV2 => convnext::Config::atto(),
+            Self::Femto | Self::FemtoV2 => convnext::Config::femto(),
+            Self::Pico | Self::PicoV2 => convnext::Config::pico(),
+            Self::Nano | Self::NanoV2 => convnext::Config::nano(),
+            Self::Tiny | Self::TinyV2 => convnext::Config::tiny(),
+            Self::Small => convnext::Config::small(),
+            Self::Base | Self::BaseV2 => convnext::Config::base(),
+            Self::Large | Self::LargeV2 => convnext::Config::large(),
+            Self::XLarge => convnext::Config::xlarge(),
+            Self::Huge => convnext::Config::huge(),
+        }
+    }
+}
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    image: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    #[arg(value_enum, long, default_value_t=Which::Tiny)]
+    which: Which,
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    let device = candle_examples::device(args.cpu)?;
+
+    let image = candle_examples::imagenet::load_image224(args.image)?;
+    println!("loaded image {image:?}");
+
+    let model_file = match args.model {
+        None => {
+            let model_name = args.which.model_filename();
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model(model_name);
+            api.get("model.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+    let model = convnext::convnext(&args.which.config(), 1000, vb)?;
+    println!("model built");
+    let logits = model.forward(&image.unsqueeze(0)?)?;
+    let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
+        .i(0)?
+        .to_vec1::<f32>()?;
+    let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
+    prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
+    for &(category_idx, pr) in prs.iter().take(5) {
+        println!(
+            "{:24}: {:.2}%",
+            candle_examples::imagenet::CLASSES[category_idx],
+            100. * pr
+        );
+    }
+    Ok(())
+}

candle-examples/examples/custom-ops/cuda_kernels.rs

@@ -1,2 +0,0 @@
-#[rustfmt::skip]
-pub const LAYERNORM_KERNELS: &str = include_str!(concat!(env!("OUT_DIR"), "/examples/custom-ops/kernels//layernorm_kernels.ptx"));

candle-examples/examples/custom-ops/main.rs

@@ -6,7 +6,8 @@
 #[cfg(feature = "mkl")]
 extern crate intel_mkl_src;
 
-#[allow(unused)]
+#[rustfmt::skip]
+#[cfg(feature = "cuda")]
 mod cuda_kernels;
 
 use clap::Parser;