Add a scattered kv cache. (#2936)

* Add a scattered kv cache.

* Update some comments.

.github/workflows/book-cd.yml

@@ -1,40 +0,0 @@
-name: Deploy Rust book
-on:
-  push:
-    branches:
-      - main
-
-jobs:
-  deploy:
-    runs-on: ubuntu-latest
-    permissions:
-      contents: write  # To push a branch 
-      pull-requests: write  # To create a PR from that branch
-    steps:
-    - uses: actions/checkout@v3
-      with:
-        fetch-depth: 0
-    - name: Install latest mdbook
-      run: |
-        tag=$(curl 'https://api.github.com/repos/rust-lang/mdbook/releases/latest' | jq -r '.tag_name')
-        url="https://github.com/rust-lang/mdbook/releases/download/${tag}/mdbook-${tag}-x86_64-unknown-linux-gnu.tar.gz"
-        mkdir mdbook
-        curl -sSL $url | tar -xz --directory=./mdbook
-        echo `pwd`/mdbook >> $GITHUB_PATH
-    - name: Deploy GitHub Pages
-      run: |
-        # This assumes your book is in the root of your repository.
-        # Just add a `cd` here if you need to change to another directory.
-        cd candle-book
-        mdbook build
-        git worktree add gh-pages
-        git config user.name "Deploy from CI"
-        git config user.email ""
-        cd gh-pages
-        # Delete the ref to avoid keeping history.
-        git update-ref -d refs/heads/gh-pages
-        rm -rf *
-        mv ../book/* .
-        git add .
-        git commit -m "Deploy $GITHUB_SHA to gh-pages"
-        git push --force --set-upstream origin gh-pages

.github/workflows/book.yml

@@ -1,29 +0,0 @@
-name: CI
-on: 
-  pull_request:
-
-jobs:
-  test:
-    name: Test candle-book
-    runs-on: ubuntu-latest
-    permissions:
-      contents: write  # To push a branch 
-      pull-requests: write  # To create a PR from that branch
-    steps:
-    - uses: actions/checkout@master
-    - name: Install Rust
-      run: |
-        rustup set profile minimal
-        rustup toolchain install stable
-        rustup default stable
-    - name: Install latest mdbook
-      run: |
-        tag=$(curl 'https://api.github.com/repos/rust-lang/mdbook/releases/latest' | jq -r '.tag_name')
-        url="https://github.com/rust-lang/mdbook/releases/download/${tag}/mdbook-${tag}-x86_64-unknown-linux-gnu.tar.gz"
-        mkdir bin
-        curl -sSL $url | tar -xz --directory=bin
-        echo "$(pwd)/bin" >> $GITHUB_PATH
-    - name: Run tests
-      run: cd candle-book && cargo build && mdbook test -L ../target/debug/deps/
-
-

.github/workflows/ci_cuda.yaml

@@ -9,7 +9,8 @@ jobs:
     concurrency:
       group: ${{ github.workflow }}-${{ github.job }}-${{ github.head_ref || github.run_id }}
       cancel-in-progress: true
-    runs-on: [single-gpu, nvidia-gpu, t4, ci]
+    runs-on:
+      group: aws-g4dn-2xlarge
     container:
       image: nvidia/cuda:12.3.1-devel-ubuntu22.04
       options: --gpus 0 

.github/workflows/python.yml

@@ -18,9 +18,9 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-latest] # For now, only test on Linux
-    steps: 
+    steps:
       - name: Checkout repository
-        uses: actions/checkout@v2
+        uses: actions/checkout@v4
 
       - name: Install Rust
         uses: actions-rs/toolchain@v1
@@ -65,4 +65,4 @@ jobs:
         working-directory: ./candle-pyo3
         run: |
           source .env/bin/activate
-          python -m pytest -s -v tests
+          python -m pytest -s -v tests

.github/workflows/rust-ci.yml

@@ -1,6 +1,6 @@
-on: 
+on:
   push:
-    branches: 
+    branches:
       - main
   pull_request:
 
@@ -15,7 +15,10 @@ jobs:
         os: [ubuntu-latest, windows-latest, macOS-latest]
         rust: [stable]
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal
@@ -34,7 +37,13 @@ jobs:
         os: [ubuntu-latest, windows-latest, macOS-latest]
         rust: [stable]
     steps:
-      - uses: actions/checkout@v2
+      - name: Delete huge unnecessary tools folder
+        if: runner.os == 'Linux'
+        run: rm -rf /opt/hostedtoolcache
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal
@@ -49,7 +58,7 @@ jobs:
     name: Rustfmt
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal
@@ -65,7 +74,7 @@ jobs:
     name: Clippy
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal

.github/workflows/trufflehog.yml

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

.gitignore

@@ -9,6 +9,10 @@ target/
 # More information here https://doc.rust-lang.org/cargo/guide/cargo-toml-vs-cargo-lock.html
 Cargo.lock
 
+# editor config
+.helix
+.vscode
+
 # These are backup files generated by rustfmt
 **/*.rs.bk
 
@@ -36,3 +40,9 @@ candle-wasm-examples/*/package-lock.json
 candle-wasm-examples/**/config*.json
 .DS_Store
 .idea/*
+__pycache__
+out.safetensors
+out.wav
+bria.mp3
+bria.safetensors
+bria.wav

Cargo.toml

@@ -3,7 +3,6 @@ members = [
     "candle-core",
     "candle-datasets",
     "candle-examples",
-    "candle-book",
     "candle-nn",
     "candle-pyo3",
     "candle-transformers",
@@ -12,6 +11,7 @@ members = [
     "tensor-tools",
 ]
 exclude = [
+   "candle-book",
    "candle-flash-attn",
    "candle-kernels",
    "candle-metal-kernels",
@@ -20,7 +20,7 @@ exclude = [
 resolver = "2"
 
 [workspace.package]
-version = "0.5.0"
+version = "0.9.1"
 edition = "2021"
 description = "Minimalist ML framework."
 repository = "https://github.com/huggingface/candle"
@@ -33,22 +33,23 @@ ab_glyph = "0.2.23"
 accelerate-src = { version = "0.3.2" }
 anyhow = { version = "1", features = ["backtrace"] }
 byteorder = "1.4.3"
-candle = { path = "./candle-core", package = "candle-core", version = "0.5.0" }
-candle-datasets = { path = "./candle-datasets", version = "0.5.0" }
-candle-flash-attn = { path = "./candle-flash-attn", version = "0.5.0" }
-candle-kernels = { path = "./candle-kernels", version = "0.5.0" }
-candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.5.0" }
-candle-nn = { path = "./candle-nn", version = "0.5.0" }
-candle-onnx = { path = "./candle-onnx", version = "0.5.0" }
-candle-transformers = { path = "./candle-transformers", version = "0.5.0" }
+candle = { path = "./candle-core", package = "candle-core", version = "0.9.1" }
+candle-datasets = { path = "./candle-datasets", version = "0.9.1" }
+candle-flash-attn = { path = "./candle-flash-attn", version = "0.9.1" }
+candle-kernels = { path = "./candle-kernels", version = "0.9.1" }
+candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.9.1" }
+candle-nn = { path = "./candle-nn", version = "0.9.1" }
+candle-onnx = { path = "./candle-onnx", version = "0.9.1" }
+candle-transformers = { path = "./candle-transformers", version = "0.9.1" }
 clap = { version = "4.2.4", features = ["derive"] }
 criterion = { version = "0.5.1", default-features=false }
-cudarc = { version = "0.10.0", features = ["f16"] }
+cudarc = { version = "0.16.1", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16", "cuda-version-from-build-system", "dynamic-linking"], default-features=false }
 fancy-regex = "0.13.0"
 gemm = { version = "0.17.0", features = ["wasm-simd128-enable"] }
-hf-hub = "0.3.0"
-half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
-image = { version = "0.25.0", default-features = false, features = ["jpeg", "png"] }
+hf-hub = "0.4.1"
+half = { version = "2.5.0", features = ["num-traits", "use-intrinsics", "rand_distr"] }
+hound = "3.5.1"
+image = { version = "0.25.2", default-features = false, features = ["jpeg", "png"] }
 imageproc = { version = "0.24.0", default-features = false }
 intel-mkl-src = { version = "0.8.1", features = ["mkl-static-lp64-iomp"] }
 libc = { version = "0.2.147" }
@@ -57,21 +58,23 @@ memmap2 = { version = "0.9.3", features = ["stable_deref_trait"] }
 num_cpus = "1.15.0"
 num-traits = "0.2.15"
 parquet = { version = "51.0.0" }
-rand = "0.8.5"
-rand_distr = "0.4.3"
+rand = "0.9.0"
+rand_distr = "0.5.1"
 rayon = "1.7.0"
 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 = "0.15.0", default-features = false }
+tokenizers = { version = "0.21.0", default-features = false }
 tracing = "0.1.37"
 tracing-chrome = "0.7.1"
 tracing-subscriber = "0.3.7"
-wav = "1.0.0"
+ug = "0.4.0"
+ug-cuda = "0.4.0"
+ug-metal = "0.4.0"
 yoke = { version = "0.7.2", features = ["derive"] }
-zip = { version = "0.6.6", default-features = false }
+zip = { version = "1.1.1", default-features = false }
 metal = { version = "0.27.0", features = ["mps"]}
 
 [profile.release-with-debug]

README.md

@@ -2,7 +2,8 @@
 [![discord server](https://dcbadge.vercel.app/api/server/hugging-face-879548962464493619)](https://discord.gg/hugging-face-879548962464493619)
 [![Latest version](https://img.shields.io/crates/v/candle-core.svg)](https://crates.io/crates/candle-core)
 [![Documentation](https://docs.rs/candle-core/badge.svg)](https://docs.rs/candle-core)
-![License](https://img.shields.io/crates/l/candle-core.svg)
+[![License](https://img.shields.io/github/license/base-org/node?color=blue)](https://github.com/huggingface/candle/blob/main/LICENSE-MIT)
+[![License](https://img.shields.io/badge/license-Apache%202.0-blue?style=flat-square)](https://github.com/huggingface/candle/blob/main/LICENSE-APACHE)
 
 Candle is a minimalist ML framework for Rust with a focus on performance (including GPU support) 
 and ease of use. Try our online demos: 
@@ -60,12 +61,16 @@ These online demos run entirely in your browser:
 
 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
+- [LLaMA v1, v2, and v3](./candle-examples/examples/llama/): general LLM, includes
   the SOLAR-10.7B variant.
 - [Falcon](./candle-examples/examples/falcon/): general LLM.
-- [Gemma](./candle-examples/examples/gemma/): 2b and 7b general LLMs from Google
-  Deepmind.
-- [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.
+- [Codegeex4](./candle-examples/examples/codegeex4-9b/): Code completion,code interpreter,web search,fuction calling,repository-level
+- [GLM4](./candle-examples/examples/glm4/): Open Multilingual Multimodal Chat LMs by THUDM
+- [Gemma v1 and v2](./candle-examples/examples/gemma/): 2b and 7b+/9b general LLMs from Google Deepmind.
+- [RecurrentGemma](./candle-examples/examples/recurrent-gemma/): 2b and 7b
+  Griffin based models from Google that mix attention with a RNN like state.
+- [Phi-1, Phi-1.5, Phi-2, and Phi-3](./candle-examples/examples/phi/): 1.3b,
+  2.7b, and 3.8b general LLMs with performance on par with 7b models.
 - [StableLM-3B-4E1T](./candle-examples/examples/stable-lm/): a 3b general LLM
   pre-trained on 1T tokens of English and code datasets. Also supports
   StableLM-2, a 1.6b LLM trained on 2T tokens, as well as the code variants.
@@ -110,12 +115,14 @@ We also provide a some command line based examples using state of the art models
 
 <img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/segment-anything/assets/sam_merged.jpg" width="200">
 
-- [SegFormer](./candle-examples/examples/segformer/): transformer based semantic segmantation model.
+- [SegFormer](./candle-examples/examples/segformer/): transformer based semantic segmentation model.
 - [Whisper](./candle-examples/examples/whisper/): speech recognition model.
 - [EnCodec](./candle-examples/examples/encodec/): high-quality audio compression
   model using residual vector quantization.
 - [MetaVoice](./candle-examples/examples/metavoice/): foundational model for
   text-to-speech.
+- [Parler-TTS](./candle-examples/examples/parler-tts/): large text-to-speech
+  model.
 - [T5](./candle-examples/examples/t5), [Bert](./candle-examples/examples/bert/),
   [JinaBert](./candle-examples/examples/jina-bert/) : useful for sentence embeddings.
 - [DINOv2](./candle-examples/examples/dinov2/): computer vision model trained
@@ -181,6 +188,8 @@ And then head over to
 - [`candle-sampling`](https://github.com/EricLBuehler/candle-sampling): Sampling techniques for Candle.
 - [`gpt-from-scratch-rs`](https://github.com/jeroenvlek/gpt-from-scratch-rs): A port of Andrej Karpathy's _Let's build GPT_ tutorial on YouTube showcasing the Candle API on a toy problem.
 - [`candle-einops`](https://github.com/tomsanbear/candle-einops): A pure rust implementation of the python [einops](https://github.com/arogozhnikov/einops) library.
+- [`atoma-infer`](https://github.com/atoma-network/atoma-infer): A Rust library for fast inference at scale, leveraging FlashAttention2 for efficient attention computation, PagedAttention for efficient KV-cache memory management, and multi-GPU support. It is OpenAI api compatible.
+- [`llms-from-scratch-rs`](https://github.com/nerdai/llms-from-scratch-rs): A comprehensive Rust translation of the code from Sebastian Raschka's Build an LLM from Scratch book.
 
 If you have an addition to this list, please submit a pull request.
 
@@ -199,12 +208,12 @@ If you have an addition to this list, please submit a pull request.
     - WASM support, run your models in a browser.
 - Included models.
     - Language Models.
-        - LLaMA v1 and v2 with variants such as SOLAR-10.7B.
+        - LLaMA v1, v2, and v3 with variants such as SOLAR-10.7B.
         - Falcon.
         - StarCoder, StarCoder2.
-        - Phi 1, 1.5, and 2.
+        - Phi 1, 1.5, 2, and 3.
         - Mamba, Minimal Mamba
-        - Gemma 2b and 7b.
+        - Gemma v1 2b and 7b+, v2 2b and 9b.
         - Mistral 7b v0.1.
         - Mixtral 8x7b v0.1.
         - StableLM-3B-4E1T, StableLM-2-1.6B, Stable-Code-3B.
@@ -232,9 +241,10 @@ If you have an addition to this list, please submit a pull request.
         - Whisper, multi-lingual speech-to-text.
         - EnCodec, audio compression model.
         - MetaVoice-1B, text-to-speech model.
+        - Parler-TTS, text-to-speech model.
     - Computer Vision Models.
         - DINOv2, ConvMixer, EfficientNet, ResNet, ViT, VGG, RepVGG, ConvNeXT,
-          ConvNeXTv2, MobileOne, EfficientVit (MSRA).
+          ConvNeXTv2, MobileOne, EfficientVit (MSRA), MobileNetv4, Hiera, FastViT.
         - yolo-v3, yolo-v8.
         - Segment-Anything Model (SAM).
         - SegFormer.
@@ -280,6 +290,8 @@ Cheatsheet:
 
 ### Why should I use Candle?
 
+<!--- ANCHOR: goals --->
+
 Candle's core goal is to *make serverless inference possible*. Full machine learning frameworks like PyTorch
 are very large, which makes creating instances on a cluster slow. Candle allows deployment of lightweight
 binaries.
@@ -289,6 +301,7 @@ and the [GIL](https://www.backblaze.com/blog/the-python-gil-past-present-and-fut
 
 Finally, Rust is cool! A lot of the HF ecosystem already has Rust crates, like [safetensors](https://github.com/huggingface/safetensors) and [tokenizers](https://github.com/huggingface/tokenizers).
 
+<!--- ANCHOR_END: goals --->
 
 ### Other ML frameworks
 
@@ -374,9 +387,9 @@ git submodule update --init
 /usr/include/c++/11/bits/std_function.h:530:146: error: parameter packs not expanded with ‘...’:
 ```
 
-This is a bug in gcc-11 triggered by the Cuda compiler. To fix this, install a different, supported gcc version - for example gcc-10, and specify the path to the compiler in the CANDLE_NVCC_CCBIN environment variable.
+This is a bug in gcc-11 triggered by the Cuda compiler. To fix this, install a different, supported gcc version - for example gcc-10, and specify the path to the compiler in the NVCC_CCBIN environment variable.
 ```
-env CANDLE_NVCC_CCBIN=/usr/lib/gcc/x86_64-linux-gnu/10 cargo ...
+env NVCC_CCBIN=/usr/lib/gcc/x86_64-linux-gnu/10 cargo ...
 ```
 
 #### Linking error on windows when running rustdoc or mdbook tests
@@ -406,3 +419,10 @@ This may be caused by the models being loaded from `/mnt/c`, more details on
 
 You can set `RUST_BACKTRACE=1` to be provided with backtraces when a candle
 error is generated.
+
+#### CudaRC error
+
+If you encounter an error like this one `called `Result::unwrap()` on an `Err` value: LoadLibraryExW { source: Os { code: 126, kind: Uncategorized, message: "The specified module could not be found." } }` on windows. To fix copy and rename these 3 files (make sure they are in path). The paths depend on your cuda version.
+`c:\Windows\System32\nvcuda.dll` -> `cuda.dll`
+`c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin\cublas64_12.dll` -> `cublas.dll`
+`c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin\curand64_10.dll` -> `curand.dll`

candle-book/CONTRIBUTING.md

@@ -0,0 +1,13 @@
+# Candle Book
+
+The book uses [mdBook](https://github.com/rust-lang/mdBook) for building.
+
+## Installation
+
+To install mdBook, run `cargo install mdbook`. More instructions can be found [here](https://rust-lang.github.io/mdBook/guide/installation.html).
+
+## Viewing the book
+
+To view the book, run `mdbook serve --open candle-book`. More instructions can be found [here](https://rust-lang.github.io/mdBook/guide/creating.html). 
+
+The book is built automatically in github CI.

candle-book/Cargo.toml

@@ -25,7 +25,7 @@ cudarc = { workspace = true, optional = true }
 half = { workspace = true, optional = true }
 image = { workspace = true, optional = true }
 anyhow = { workspace = true }
-tokio = "1.29.1"
+tokio = "1.43.0"
 
 [dev-dependencies]
 byteorder = { workspace = true }
@@ -37,7 +37,6 @@ 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
 parquet = { workspace = true }
 image = { workspace = true }

candle-book/src/README.md

@@ -1,6 +1,7 @@
 # Introduction
 
+{{#include ../../README.md:goals}}
+
 {{#include ../../README.md:features}}
 
-
-This book will introduce step by step how to use `candle`.
+This book will introduce step by step how to use `candle`.

candle-book/src/SUMMARY.md

@@ -5,7 +5,10 @@
 # User Guide
 
 - [Installation](guide/installation.md)
-- [Hello World - MNIST](guide/hello_world.md)
+- [Tutorial - MNIST](guide/mnist/intro.md)
+  - [Modeling](guide/mnist/modeling.md)
+  - [Training](guide/mnist/training.md)
+  - [Saving And Loading](guide/mnist/saving_loading.md)
 - [PyTorch cheatsheet](guide/cheatsheet.md)
 
 # Reference Guide
@@ -13,6 +16,7 @@
 - [Running a model](inference/inference.md)
     - [Using the hub](inference/hub.md)
 - [Error management](error_manage.md)
+- [Tracing](tracing.md)
 - [Training](training/training.md)
     - [Simplified](training/simplified.md)
     - [MNIST](training/mnist.md)

candle-book/src/guide/installation.md

@@ -1,8 +1,23 @@
 # Installation
 
-**With Cuda support**:
+## 1. Create a new rust app or library
 
-1. First, make sure that Cuda is correctly installed.
+```bash
+cargo new myapp
+cd myapp
+```
+
+## 2. Add the correct candle version
+
+### Standard
+
+```bash
+cargo add --git https://github.com/huggingface/candle.git candle-core
+```
+
+### CUDA
+
+First, make sure that Cuda is correctly installed.
 - `nvcc --version` should print information about your Cuda compiler driver.
 - `nvidia-smi --query-gpu=compute_cap --format=csv` should print your GPUs compute capability, e.g. something
 like:
@@ -17,43 +32,36 @@ You can also compile the Cuda kernels for a specific compute cap using the
 
 If any of the above commands errors out, please make sure to update your Cuda version.
 
-2. Create a new app and add [`candle-core`](https://github.com/huggingface/candle/tree/main/candle-core) with Cuda support.
-
-Start by creating a new cargo:
-
-```bash
-cargo new myapp
-cd myapp
-```
-
-Make sure to add the `candle-core` crate with the cuda feature:
+Add the `candle-core` crate with the cuda feature:
 
 ```bash
 cargo add --git https://github.com/huggingface/candle.git candle-core --features "cuda"
 ```
 
+### MKL
+
+You can also see the `mkl` feature which can get faster inference on CPU.
+
+Add the `candle-core` crate with the mkl feature:
+
+```bash
+cargo add --git https://github.com/huggingface/candle.git candle-core --features "mkl"
+```
+
+### Metal
+
+Metal is exclusive to MacOS.
+
+Add the `candle-core` crate with the metal feature:
+
+```bash
+cargo add --git https://github.com/huggingface/candle.git candle-core --features "metal"
+```
+
+## 3. Building
+
 Run `cargo build` to make sure everything can be correctly built.
 
 ```bash
 cargo build
 ```
-
-**Without Cuda support**:
-
-Create a new app and add [`candle-core`](https://github.com/huggingface/candle/tree/main/candle-core) as follows:
-
-```bash
-cargo new myapp
-cd myapp
-cargo add --git https://github.com/huggingface/candle.git candle-core
-```
-
-Finally, run `cargo build` to make sure everything can be correctly built.
-
-```bash
-cargo build
-```
-
-**With mkl support**
-
-You can also see the `mkl` feature which could be interesting to get faster inference on CPU. [Using mkl](./advanced/mkl.md)

candle-book/src/guide/mnist/intro.md

@@ -0,0 +1,17 @@
+# Candle MNIST Tutorial
+
+## Introduction
+
+This tutorial provides an introduction to Candle by implementing and training a neural network for MNIST digit classification from scratch. 
+
+Throughout this tutorial, you will learn the basics of:
+
+- Tensor operations and model construction
+- Creating and implementing neural network layers
+- Parameter initialization
+- Training loop implementation
+- Saving and loading trained models
+
+## Getting Started
+
+Before proceeding, please ensure that you have properly installed Candle by following the instructions in the [Installation](../installation.md) guide.

candle-book/src/guide/mnist/modeling.md

@@ -0,0 +1,172 @@
+# Candle MNIST Tutorial
+
+## Modeling
+
+Open `src/main.rs` in your project folder and insert the following code:
+
+```rust
+use candle_core::{Device, Result, Tensor};
+
+struct Model {
+    first: Tensor,
+    second: Tensor,
+}
+
+impl Model {
+    fn forward(&self, image: &Tensor) -> Result<Tensor> {
+        let x = image.matmul(&self.first)?;
+        let x = x.relu()?;
+        x.matmul(&self.second)
+    }
+}
+
+fn main() -> Result<()> {
+    // Use Device::new_cuda(0)?; to utilize GPU acceleration.
+    let device = Device::Cpu;
+
+    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::randn(0f32, 1.0, (1, 784), &device)?;
+
+    let digit = model.forward(&dummy_image)?;
+    println!("Digit {digit:?} digit");
+    Ok(())
+}
+```
+
+Execute the program with:
+
+```bash
+$ cargo run --release
+
+> Digit Tensor[dims 1, 10; f32] digit
+```
+
+Since random inputs are provided, expect an incoherent output.
+
+## Implementing a `Linear` Layer
+
+To create a more sophisticated layer type, add a `bias` to the weight to construct the standard `Linear` layer.
+
+Replace the entire content of `src/main.rs` with:
+
+```rust
+use candle_core::{Device, Result, Tensor};
+
+struct Linear {
+    weight: Tensor,
+    bias: Tensor,
+}
+
+impl Linear {
+    fn forward(&self, x: &Tensor) -> Result<Tensor> {
+        let x = x.matmul(&self.weight)?;
+        x.broadcast_add(&self.bias)
+    }
+}
+
+struct Model {
+    first: Linear,
+    second: Linear,
+}
+
+impl Model {
+    fn forward(&self, image: &Tensor) -> Result<Tensor> {
+        let x = self.first.forward(image)?;
+        let x = x.relu()?;
+        self.second.forward(&x)
+    }
+}
+
+fn main() -> Result<()> {
+    // Use Device::new_cuda(0)?; for GPU acceleration.
+    // Use Device::Cpu; for CPU computation.
+    let device = Device::cuda_if_available(0)?;
+
+    // Initialize model parameters
+    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::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::randn(0f32, 1.0, (1, 784), &device)?;
+
+    // Perform inference
+    let digit = model.forward(&dummy_image)?;
+    println!("Digit {digit:?} digit");
+    Ok(())
+}
+```
+
+Execute again with:
+
+```bash
+$ cargo run --release
+
+> Digit Tensor[dims 1, 10; f32] digit
+```
+
+## Utilizing `candle_nn`
+
+Many classical layers (such as [Linear](https://github.com/huggingface/candle/blob/main/candle-nn/src/linear.rs)) are already implemented in [candle-nn](https://github.com/huggingface/candle/tree/main/candle-nn).
+
+This `Linear` implementation follows PyTorch conventions for improved compatibility with existing models, utilizing the transpose of weights rather than direct weights.
+
+Let's simplify our implementation. First, add `candle-nn` as a dependency:
+
+```bash
+$ cargo add --git https://github.com/huggingface/candle.git candle-nn
+```
+
+Now, replace the entire content of `src/main.rs` with:
+
+```rust
+use candle_core::{Device, Result, Tensor};
+use candle_nn::{Linear, Module};
+
+struct Model {
+    first: Linear,
+    second: Linear,
+}
+
+impl Model {
+    fn forward(&self, image: &Tensor) -> Result<Tensor> {
+        let x = self.first.forward(image)?;
+        let x = x.relu()?;
+        self.second.forward(&x)
+    }
+}
+
+fn main() -> Result<()> {
+    // Use Device::new_cuda(0)?; for GPU acceleration.
+    let device = Device::Cpu;
+
+    // Note the dimension change: (784, 100) -> (100, 784)
+    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::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::randn(0f32, 1.0, (1, 784), &device)?;
+
+    let digit = model.forward(&dummy_image)?;
+    println!("Digit {digit:?} digit");
+    Ok(())
+}
+```
+
+Execute the final version:
+
+```bash
+$ cargo run --release
+
+> Digit Tensor[dims 1, 10; f32] digit
+```

candle-book/src/guide/mnist/saving_loading.md

@@ -0,0 +1,158 @@
+# Candle MNIST Tutorial
+
+## Saving and Loading Models
+
+After training a model, it is useful to save and subsequently load the model parameters. In Candle, this functionality is managed through the `VarMap` data structure, with parameters stored on disk using the [safetensors](https://huggingface.co/docs/safetensors/index) format.
+
+### Saving Model Parameters
+
+Let's modify our `training_loop` function to include functionality for saving weights:
+
+```rust
+fn training_loop(
+    m: candle_datasets::vision::Dataset,
+) -> anyhow::Result<()> {
+    let dev = Device::cuda_if_available(0)?;
+
+    let train_labels = m.train_labels;
+    let train_images = m.train_images.to_device(&dev)?;
+    let train_labels = train_labels.to_dtype(DType::U32)?.to_device(&dev)?;
+
+    // Initialize a VarMap for trainable parameters
+    let varmap = VarMap::new();
+    let vs = VarBuilder::from_varmap(&varmap, DType::F32, &dev);
+    let model = Model::new(vs.clone())?;
+
+    let learning_rate = 0.05;
+    let epochs = 10;
+
+    // Initialize stochastic gradient descent optimizer
+    let mut sgd = candle_nn::SGD::new(varmap.all_vars(), learning_rate)?;
+    let test_images = m.test_images.to_device(&dev)?;
+    let test_labels = m.test_labels.to_dtype(DType::U32)?.to_device(&dev)?;
+    
+    for epoch in 1..epochs {
+        // Standard MNIST forward pass
+        let logits = model.forward(&train_images)?;
+        let log_sm = ops::log_softmax(&logits, D::Minus1)?;
+        
+        // Compute Negative Log Likelihood loss
+        let loss = loss::nll(&log_sm, &train_labels)?;
+
+        // Perform backward pass and update weights
+        sgd.backward_step(&loss)?;
+
+        // Evaluate model on test set
+        let test_logits = model.forward(&test_images)?;
+        let sum_ok = test_logits
+            .argmax(D::Minus1)?
+            .eq(&test_labels)?
+            .to_dtype(DType::F32)?
+            .sum_all()?
+            .to_scalar::<f32>()?;
+        let test_accuracy = sum_ok / test_labels.dims1()? as f32;
+        println!(
+            "{epoch:4} train loss: {:8.5} test acc: {:5.2}%",
+            loss.to_scalar::<f32>()?,
+            test_accuracy
+        );
+    }
+    
+    // Save model weights to disk
+    varmap.save("model_weights.safetensors")?;
+    Ok(())
+}
+```
+
+```bash
+$ cargo run --release
+
+> 1 train loss:  2.40485 test acc:  0.11%
+> 2 train loss:  2.34161 test acc:  0.14%
+> 3 train loss:  2.28841 test acc:  0.17%
+> 4 train loss:  2.24158 test acc:  0.19%
+> 5 train loss:  2.19898 test acc:  0.23%
+> 6 train loss:  2.15927 test acc:  0.26%
+> 7 train loss:  2.12161 test acc:  0.29%
+> 8 train loss:  2.08549 test acc:  0.32%
+> 9 train loss:  2.05053 test acc:  0.35%
+```
+
+### Loading Model Parameters
+
+Now that we have saved our model parameters, we can modify the code to load them. The primary change required is to make the `varmap` variable mutable:
+
+```rust
+fn training_loop(
+    m: candle_datasets::vision::Dataset,
+) -> anyhow::Result<()> {
+    let dev = Device::cuda_if_available(0)?;
+
+    let train_labels = m.train_labels;
+    let train_images = m.train_images.to_device(&dev)?;
+    let train_labels = train_labels.to_dtype(DType::U32)?.to_device(&dev)?;
+
+    // Create a mutable VarMap for trainable parameters
+    let mut varmap = VarMap::new();
+    let vs = VarBuilder::from_varmap(&varmap, DType::F32, &dev);
+    let model = Model::new(vs.clone())?;
+
+    // Load pre-trained weights from file
+    varmap.load("model_weights.safetensors")?;
+
+    let learning_rate = 0.05;
+    let epochs = 10;
+
+    // Initialize stochastic gradient descent optimizer
+    let mut sgd = candle_nn::SGD::new(varmap.all_vars(), learning_rate)?;
+    let test_images = m.test_images.to_device(&dev)?;
+    let test_labels = m.test_labels.to_dtype(DType::U32)?.to_device(&dev)?;
+    
+    for epoch in 1..epochs {
+        // Standard MNIST forward pass
+        let logits = model.forward(&train_images)?;
+        let log_sm = ops::log_softmax(&logits, D::Minus1)?;
+        
+        // Compute Negative Log Likelihood loss
+        let loss = loss::nll(&log_sm, &train_labels)?;
+
+        // Perform backward pass and update weights
+        sgd.backward_step(&loss)?;
+
+        // Evaluate model on test set
+        let test_logits = model.forward(&test_images)?;
+        let sum_ok = test_logits
+            .argmax(D::Minus1)?
+            .eq(&test_labels)?
+            .to_dtype(DType::F32)?
+            .sum_all()?
+            .to_scalar::<f32>()?;
+        let test_accuracy = sum_ok / test_labels.dims1()? as f32;
+        println!(
+            "{epoch:4} train loss: {:8.5} test acc: {:5.2}%",
+            loss.to_scalar::<f32>()?,
+            test_accuracy
+        );
+    }
+    
+    // Save updated weights back to disk
+    varmap.save("model_weights.safetensors")?;
+    Ok(())
+}
+```
+
+```bash
+$ cargo run --release
+
+> 1 train loss:  2.01645 test acc:  0.38%
+> 2 train loss:  1.98300 test acc:  0.41%
+> 3 train loss:  1.95008 test acc:  0.44%
+> 4 train loss:  1.91754 test acc:  0.47%
+> 5 train loss:  1.88534 test acc:  0.50%
+> 6 train loss:  1.85349 test acc:  0.53%
+> 7 train loss:  1.82198 test acc:  0.56%
+> 8 train loss:  1.79077 test acc:  0.59%
+> 9 train loss:  1.75989 test acc:  0.61%
+```
+
+Note that loading the weights will fail if the specified file does not exist or is incompatible with the current model architecture. Implementing file existence checks and appropriate error handling is left to the user.

candle-book/src/guide/mnist/training.md

@@ -0,0 +1,134 @@
+# Candle MNIST Tutorial
+
+## Training Implementation
+
+First, let's create a utility function `make_linear` that accepts a `VarBuilder` and returns an initialized linear layer. The `VarBuilder` constructs a `VarMap`, which is the data structure that stores our trainable parameters.
+
+```rust
+use candle_core::{Device, Result, Tensor};
+use candle_nn::{Linear, Module, VarBuilder, VarMap};
+
+fn make_linear(vs: VarBuilder, in_dim: usize, out_dim: usize) -> Result<Linear> {
+    let ws = vs.get_with_hints(
+        (out_dim, in_dim),
+        "weight",
+        candle_nn::init::DEFAULT_KAIMING_NORMAL,
+    )?;
+    let bound = 1. / (in_dim as f64).sqrt();
+    let bs = vs.get_with_hints(
+        out_dim,
+        "bias",
+        candle_nn::Init::Uniform {
+            lo: -bound,
+            up: bound,
+        },
+    )?;
+    Ok(Linear::new(ws, Some(bs)))
+}
+```
+
+Next, let's implement a `new` method for our model class to accept a `VarBuilder` and initialize the model. We use `VarBuilder::pp` to "push prefix" so that the parameter names are organized hierarchically: the first layer weights as `first.weight` and `first.bias`, and the second layer weights as `second.weight` and `second.bias`.
+
+```rust
+impl Model {
+    fn new(vs: VarBuilder) -> Result<Self> {
+        const IMAGE_DIM: usize = 784;
+        const HIDDEN_DIM: usize = 100;
+        const LABELS: usize = 10;
+
+        let first = make_linear(vs.pp("first"), IMAGE_DIM, HIDDEN_DIM)?;
+        let second = make_linear(vs.pp("second"), HIDDEN_DIM, LABELS)?;
+
+        Ok(Self { first, second })
+    }
+
+    fn forward(&self, image: &Tensor) -> Result<Tensor> {
+        let x = self.first.forward(image)?;
+        let x = x.relu()?;
+        self.second.forward(&x)
+    }
+}
+```
+
+Now, let's add the `candle-datasets` package to our project to access the MNIST dataset:
+
+```bash
+$ cargo add --git https://github.com/huggingface/candle.git candle-datasets
+```
+
+With the dataset available, we can implement our training loop:
+
+```rust
+use candle_core::{DType, Device, Result, Tensor, D};
+use candle_nn::{loss, ops, Linear, Module, Optimizer, VarBuilder, VarMap};
+
+fn training_loop(
+    m: candle_datasets::vision::Dataset,
+) -> anyhow::Result<()> {
+    let dev = Device::cuda_if_available(0)?;
+
+    let train_labels = m.train_labels;
+    let train_images = m.train_images.to_device(&dev)?;
+    let train_labels = train_labels.to_dtype(DType::U32)?.to_device(&dev)?;
+
+    // Initialize a VarMap to store trainable parameters
+    let varmap = VarMap::new();
+    let vs = VarBuilder::from_varmap(&varmap, DType::F32, &dev);
+    let model = Model::new(vs.clone())?;
+
+    let learning_rate = 0.05;
+    let epochs = 10;
+
+    // Initialize a stochastic gradient descent optimizer to update parameters
+    let mut sgd = candle_nn::SGD::new(varmap.all_vars(), learning_rate)?;
+    let test_images = m.test_images.to_device(&dev)?;
+    let test_labels = m.test_labels.to_dtype(DType::U32)?.to_device(&dev)?;
+    
+    for epoch in 1..epochs {
+        // Perform forward pass on MNIST data
+        let logits = model.forward(&train_images)?;
+        let log_sm = ops::log_softmax(&logits, D::Minus1)?;
+        
+        // Compute Negative Log Likelihood loss
+        let loss = loss::nll(&log_sm, &train_labels)?;
+
+        // Perform backward pass and update weights
+        sgd.backward_step(&loss)?;
+
+        // Evaluate model on test set
+        let test_logits = model.forward(&test_images)?;
+        let sum_ok = test_logits
+            .argmax(D::Minus1)?
+            .eq(&test_labels)?
+            .to_dtype(DType::F32)?
+            .sum_all()?
+            .to_scalar::<f32>()?;
+        let test_accuracy = sum_ok / test_labels.dims1()? as f32;
+        println!(
+            "{epoch:4} train loss: {:8.5} test acc: {:5.2}%",
+            loss.to_scalar::<f32>()?,
+            test_accuracy
+        );
+    }
+    Ok(())
+}
+```
+
+Finally, let's implement our main function:
+
+```rust
+pub fn main() -> anyhow::Result<()> {
+    let m = candle_datasets::vision::mnist::load()?;
+    return training_loop(m);
+}
+```
+
+Let's execute the training process:
+
+```bash
+$ cargo run --release
+
+> 1 train loss:  2.35449 test acc:  0.12%
+> 2 train loss:  2.30760 test acc:  0.15%
+> ...
+```

candle-book/src/lib.rs

@@ -81,7 +81,7 @@ let mut tp_shape = view.shape().to_vec();
 let size = tp_shape[0];
 
 if size % world_size != 0 {
-    panic!("The dimension is not divisble by `world_size`");
+    panic!("The dimension is not divisible by `world_size`");
 }
 let block_size = size / world_size;
 let start = rank * block_size;
@@ -106,8 +106,8 @@ let tp_tensor = Tensor::from_raw_buffer(&raw, dtype, &tp_shape, &Device::Cpu).un
     }
 }
 
+    #[allow(unused)]
     #[rustfmt::skip]
-    #[test]
     fn book_training_1() -> Result<()>{
 // ANCHOR: book_training_1
 use hf_hub::{api::sync::Api, Repo, RepoType};

candle-book/src/tracing.md

@@ -0,0 +1,68 @@
+# Tracing
+
+Tracing is a powerful tool for identifying performance issues and bottlenecks in code.
+
+> Profiling on GPUs is trickier due to asynchronous execution, see the [GPU section](#gpu).
+
+## Overview
+
+Candle uses the [tracing](https://docs.rs/tracing/latest/tracing/) crate for instrumentation.
+
+To try it out, run an example in `candle-examples` with the `--tracing` flag. 
+This generates a trace file, typically named `trace-<timestamp>.json`. 
+You can view the trace in Chrome by navigating to `chrome://tracing/`, clicking **Load**, and selecting the generated trace file.
+
+## Adding Tracing
+
+Candle includes built-in tracing for many internal operations, using [spans](https://docs.rs/tracing/latest/tracing/struct.Span.html) to mark key points of execution.
+
+To add custom tracing in your code, you can define a span like this:
+
+```rust
+let span = tracing::span!(tracing::Level::TRACE, name);
+```
+
+Then, to record the span during execution, create a guard:
+
+```rust
+let _enter = span.enter();
+```
+
+This guard will record the span's duration, from when it is created to when it is dropped, into a global data structure managed by the tracing crate.
+
+## Recording and Saving a Trace
+
+To capture and save trace data, you need to configure the tracing system with an output format. Candle uses the [tracing_subscriber](https://docs.rs/tracing-subscriber/latest/tracing_subscriber/) and [tracing_chrome](https://docs.rs/tracing-chrome/latest/tracing_chrome/) crates.
+
+The snippet below sets up a Chrome compatible recorder that logs all tracing activity between creation and drop of the guard:
+
+```rust
+use tracing_chrome::ChromeLayerBuilder;
+use tracing_subscriber::prelude::*;
+
+let _guard = {
+    let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
+    tracing_subscriber::registry().with(chrome_layer).init();
+    guard
+};
+```
+
+## GPU
+
+When using CUDA, Metal, or other asynchronous GPU backends, tracing may produce misleading timing data because operations are queued rather than executed immediately.
+
+### CUDA
+
+For CUDA-specific profiling, you have two options:
+
+1. Set the environment variable `CUDA_LAUNCH_BLOCKING=1` which forces synchronous execution. This makes trace timings more accurate, at the cost of reduced performance.
+2. Use [NVIDIA's Nsight Systems](https://developer.nvidia.com/nsight-systems) (`nsys profile` and `nsys-ui`) which are designed specifically for profiling asynchronous CUDA executions.
+
+We recommend using NVIDIA's Nsight Systems when possible, as it offers accurate performance data without altering typical execution patterns. In contrast, setting the `CUDA_LAUNCH_BLOCKING` environment variable forces synchronous execution, which can significantly alter execution behavior.
+
+#### Performance Profiling with NVIDIA Nsight Systems
+
+1. Generate an `.nsys-rep` file containing performance data ([docs](https://docs.nvidia.com/nsight-systems/UserGuide/index.html#example-single-command-lines))
+   - Run `nsys profile --trace cuda,nvtx,osrt --gpu-metrics-device=all --output profile_run ./target/debug/... --prompt "whatever "`
+1. Open the generated `.nsys-rep` report file in Nsight Systems GUI
+    - File > Open

candle-core/Cargo.toml

@@ -14,7 +14,7 @@ accelerate-src = { workspace = true, optional = true }
 byteorder = { workspace = true }
 candle-kernels = { workspace = true, optional = true }
 candle-metal-kernels = { workspace = true, optional = true }
-metal = { workspace = true, optional = true}
+metal = { workspace = true, optional = true }
 cudarc = { workspace = true, optional = true }
 gemm = { workspace = true }
 half = { workspace = true }
@@ -28,23 +28,35 @@ rand_distr = { workspace = true }
 rayon = { workspace = true }
 safetensors = { workspace = true }
 thiserror = { workspace = true }
+ug-cuda = { workspace = true, optional = true }
+ug-metal = { workspace = true, optional = true }
 yoke = { workspace = true }
 zip = { workspace = true }
 
+[target.'cfg(not(target_arch = "wasm32"))'.dependencies]
+ug = { workspace = true }
+
 [dev-dependencies]
 anyhow = { workspace = true }
 clap = { workspace = true }
 criterion = { workspace = true }
 
-
 [features]
 default = []
-cuda = ["cudarc", "dep:candle-kernels"]
+cuda = ["cudarc", "dep:candle-kernels", "dep:ug-cuda"]
 cudnn = ["cuda", "cudarc/cudnn"]
 mkl = ["dep:libc", "dep:intel-mkl-src"]
 accelerate = ["dep:libc", "dep:accelerate-src"]
-metal = ["dep:metal", "dep:candle-metal-kernels"]
+metal = ["dep:metal", "dep:candle-metal-kernels", "dep:ug-metal"]
 
 [[bench]]
 name = "bench_main"
 harness = false
+
+[[example]]
+name = "metal_basics"
+required-features = ["metal"]
+
+[[example]]
+name = "cuda_basics"
+required-features = ["cuda"]

candle-core/benches/bench_main.rs

@@ -1,10 +1,15 @@
 mod benchmarks;
 
 use criterion::criterion_main;
+
 criterion_main!(
     benchmarks::affine::benches,
-    benchmarks::matmul::benches,
-    benchmarks::random::benches,
-    benchmarks::where_cond::benches,
+    benchmarks::copy::benches,
     benchmarks::conv_transpose2d::benches,
+    benchmarks::matmul::benches,
+    benchmarks::qmatmul::benches,
+    benchmarks::random::benches,
+    benchmarks::reduce::benches,
+    benchmarks::unary::benches,
+    benchmarks::where_cond::benches,
 );

candle-core/benches/benchmarks/affine.rs

@@ -12,7 +12,7 @@ fn run_affine_benchmark(c: &mut Criterion, device: &Device, dtype: DType, name:
     let m = 1024;
     let k = 1024;
 
-    let tensor = Tensor::zeros((b, m, k), dtype, &device).unwrap();
+    let tensor = Tensor::zeros((b, m, k), dtype, device).unwrap();
 
     let flops = b * m * k * dtype.size_in_bytes();
 

candle-core/benches/benchmarks/copy.rs

@@ -0,0 +1,38 @@
+use crate::benchmarks::{BenchDevice, BenchDeviceHandler};
+use candle_core::{Device, Tensor, WithDType};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use std::time::Instant;
+
+fn run_copy_mask_benchmark<D: WithDType>(c: &mut Criterion, device: &Device, name: &str) {
+    let batch_size = 128;
+    let in_seq_len = 1;
+    let kv_seq_len = 1024;
+
+    let attn_mask = vec![vec![vec![D::zero(); kv_seq_len]; in_seq_len]; batch_size];
+    let size_in_bytes = batch_size * in_seq_len * kv_seq_len * D::DTYPE.size_in_bytes();
+
+    let mut group = c.benchmark_group(device.bench_name(name));
+    group.throughput(Throughput::Bytes(size_in_bytes as u64));
+    group.bench_function("iter", move |b| {
+        b.iter_custom(|iters| {
+            let attn_masks = vec![attn_mask.clone(); iters as usize];
+            let start = Instant::now();
+            for attn_mask in attn_masks.into_iter() {
+                let tensor = Tensor::new(black_box(attn_mask), device).unwrap();
+                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_copy_mask_benchmark::<f32>(c, &device, "copy_mask");
+    }
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/benches/benchmarks/mod.rs

@@ -1,7 +1,11 @@
 pub(crate) mod affine;
 pub(crate) mod conv_transpose2d;
+pub(crate) mod copy;
 pub(crate) mod matmul;
+pub(crate) mod qmatmul;
 pub(crate) mod random;
+pub(crate) mod reduce;
+pub(crate) mod unary;
 pub(crate) mod where_cond;
 
 use candle_core::{Device, Result};
@@ -18,7 +22,9 @@ impl BenchDevice for Device {
             Device::Cpu => Ok(()),
             Device::Cuda(device) => {
                 #[cfg(feature = "cuda")]
-                return Ok(device.synchronize()?);
+                return Ok(device
+                    .synchronize()
+                    .map_err(|e| candle_core::Error::Cuda(Box::new(e)))?);
                 #[cfg(not(feature = "cuda"))]
                 panic!("Cuda device without cuda feature enabled: {:?}", device)
             }

candle-core/benches/benchmarks/qmatmul.rs

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

candle-core/benches/benchmarks/reduce.rs

@@ -0,0 +1,158 @@
+use crate::benchmarks::{BenchDevice, BenchDeviceHandler};
+use candle_core::{DType, Device, Tensor};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use half::{bf16, f16};
+use std::time::Instant;
+
+fn run_sum(a: &Tensor) {
+    a.sum_keepdim(2).unwrap();
+}
+fn run_arg_min(a: &Tensor) {
+    a.argmin_keepdim(2).unwrap();
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    let handler = BenchDeviceHandler::new().unwrap();
+    let (lo, up) = (-1000.0f32, 1000.0f32);
+    for device in handler.devices {
+        run_reduce(c, &device, (lo, up), false);
+        run_reduce(c, &device, (f16::from_f32(lo), f16::from_f32(up)), false);
+        run_reduce(c, &device, (bf16::from_f32(lo), bf16::from_f32(up)), false);
+
+        run_arg_reduce(c, &device, (lo, up), false);
+        run_arg_reduce(c, &device, (f16::from_f32(lo), f16::from_f32(up)), false);
+        run_arg_reduce(c, &device, (bf16::from_f32(lo), bf16::from_f32(up)), false);
+
+        run_reduce(c, &device, (lo, up), true);
+        run_reduce(c, &device, (f16::from_f32(lo), f16::from_f32(up)), true);
+        run_reduce(c, &device, (bf16::from_f32(lo), bf16::from_f32(up)), true);
+
+        run_arg_reduce(c, &device, (lo, up), true);
+        run_arg_reduce(c, &device, (f16::from_f32(lo), f16::from_f32(up)), true);
+        run_arg_reduce(c, &device, (bf16::from_f32(lo), bf16::from_f32(up)), true);
+    }
+}
+
+fn run_reduce<T: candle_core::FloatDType>(
+    c: &mut Criterion,
+    device: &Device,
+    (lo, up): (T, T),
+    strided: bool,
+) {
+    let b = 1;
+    let m = 1024;
+    let k = 1024;
+
+    let a = if strided {
+        Tensor::rand(lo, up, (b, m, k), &device)
+            .unwrap()
+            .transpose(0, 2)
+            .unwrap()
+    } else {
+        Tensor::rand(lo, up, (b, m, k), &device).unwrap()
+    };
+
+    let flops = b * m * k * T::DTYPE.size_in_bytes();
+
+    let name = match T::DTYPE {
+        DType::F32 => {
+            if strided {
+                "reduce_f32_strided"
+            } else {
+                "reduce_f32"
+            }
+        }
+        DType::F16 => {
+            if strided {
+                "reduce_f16_strided"
+            } else {
+                "reduce_f16"
+            }
+        }
+        DType::BF16 => {
+            if strided {
+                "reduce_bf16_strided"
+            } else {
+                "reduce_bf16"
+            }
+        }
+        _ => "unknown",
+    };
+
+    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_sum(black_box(&a));
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+fn run_arg_reduce<T: candle_core::FloatDType>(
+    c: &mut Criterion,
+    device: &Device,
+    (lo, up): (T, T),
+    strided: bool,
+) {
+    let b = 1;
+    let m = 1024;
+    let k = 1024;
+
+    let a = if strided {
+        Tensor::rand(lo, up, (b, m, k), &device)
+            .unwrap()
+            .transpose(0, 2)
+            .unwrap()
+    } else {
+        Tensor::rand(lo, up, (b, m, k), &device).unwrap()
+    };
+
+    let flops = b * m * k * T::DTYPE.size_in_bytes();
+
+    let name = match T::DTYPE {
+        DType::F32 => {
+            if strided {
+                "arg_reduce_f32_strided"
+            } else {
+                "arg_reduce_f32"
+            }
+        }
+        DType::F16 => {
+            if strided {
+                "arg_reduce_f16_strided"
+            } else {
+                "arg_reduce_f16"
+            }
+        }
+        DType::BF16 => {
+            if strided {
+                "arg_reduce_bf16_strided"
+            } else {
+                "arg_reduce_bf16"
+            }
+        }
+        _ => "unknown",
+    };
+
+    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_arg_min(black_box(&a));
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/benches/benchmarks/unary.rs

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

candle-core/benches/benchmarks/where_cond.rs

@@ -25,9 +25,9 @@ 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 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

candle-core/examples/cuda_basics.rs

@@ -5,32 +5,19 @@ extern crate accelerate_src;
 extern crate intel_mkl_src;
 
 use anyhow::Result;
-use candle_core::{Device, Module, Tensor};
-
-use candle_core::quantized::{QMatMul, QTensor};
-
+use candle_core::{Device, Tensor};
+// xs: [1024, 64, 1924], c Tensor[dims 128, 64, 8; f32, cuda:0] Conv1dConfig { padding: 0, stride: 4, dilation: 1, groups: 1 }
 fn main() -> Result<()> {
     let device = Device::new_cuda(0)?;
-    let q = Tensor::randn(0f32, 1.0, (72, 256), &device)?;
-    let q_cpu = q.to_device(&Device::Cpu)?;
-    let q = QTensor::quantize(&q, candle_core::quantized::GgmlDType::Q8K)?;
-    let q = QMatMul::from_qtensor(q)?;
-    let x = Tensor::randn(0f32, 1.0, (5, 256), &device)?;
-    let res_q_cuda = q.forward(&x)?;
-    println!("{res_q_cuda}");
-
-    let q_cpu = QTensor::quantize(&q_cpu, candle_core::quantized::GgmlDType::Q8K)?;
-    let q_cpu_tensor = q_cpu.dequantize(&Device::Cpu)?;
-    let q_cpu = QMatMul::from_qtensor(q_cpu)?;
-    let x_cpu = x.to_device(&Device::Cpu)?;
-    let res_q_cpu = q_cpu.forward(&x_cpu)?;
-    println!("{res_q_cpu}");
-
-    let res_mm = x_cpu.matmul(&q_cpu_tensor.t()?)?;
-    let diff = (res_mm - res_q_cuda.to_device(&Device::Cpu))?
-        .abs()?
-        .flatten_all()?
-        .max(0)?;
-    println!("{diff}");
+    let x = Tensor::randn(0f32, 1.0, (1024, 64, 1924), &device)?;
+    let c = Tensor::randn(0f32, 1.0, (128, 64, 8), &device)?;
+    let _x1 = x.conv1d(&c, 0, 4, 1, 1)?;
+    drop(_x1);
+    for _ in 0..20 {
+        let start_time = std::time::Instant::now();
+        let _x1 = x.conv1d(&c, 0, 4, 1, 1)?;
+        device.synchronize()?;
+        println!("conv1d: {:?}", start_time.elapsed());
+    }
     Ok(())
 }

candle-core/examples/metal_basics.rs

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

candle-core/src/backend.rs

@@ -1,3 +1,5 @@
+//! Traits to Define Backend Behavior
+//!
 use crate::op::{BinaryOpT, CmpOp, ReduceOp, UnaryOpT};
 use crate::{CpuStorage, DType, Layout, Result, Shape};
 
@@ -69,15 +71,27 @@ pub trait BackendStorage: Sized {
     fn upsample_nearest2d(&self, _: &Layout, _: usize, _: usize) -> Result<Self>;
 
     fn gather(&self, _: &Layout, _: &Self, _: &Layout, _: usize) -> Result<Self>;
-    fn scatter_add(
-        &self,
+
+    fn scatter_set(
+        &mut self,
         _: &Layout,
         _: &Self,
         _: &Layout,
         _: &Self,
         _: &Layout,
         _: usize,
-    ) -> Result<Self>;
+    ) -> Result<()>;
+
+    fn scatter_add_set(
+        &mut self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: usize,
+    ) -> Result<()>;
+
     fn index_select(&self, _: &Self, _: &Layout, _: &Layout, _: usize) -> Result<Self>;
     fn index_add(
         &self,
@@ -111,6 +125,8 @@ pub trait BackendStorage: Sized {
         _src_offset: usize,
         _dst_offset: usize,
     ) -> Result<()>;
+
+    fn const_set(&mut self, _: crate::scalar::Scalar, _: &Layout) -> Result<()>;
 }
 
 pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
@@ -125,14 +141,14 @@ pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
 
     fn zeros_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;
 
-    fn ones_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;
-
     /// # Safety
     /// This function is unsafe as it doesn't initialize the underlying data store.
     /// The caller should ensure that the data is properly initialized as early as possible
     /// after this call.
     unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;
 
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage>;
+
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage>;
 
     fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage>;
@@ -142,4 +158,7 @@ pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;
 
     fn set_seed(&self, _: u64) -> Result<()>;
+
+    /// Synchronize should block until all the operations on the device are completed.
+    fn synchronize(&self) -> Result<()>;
 }

candle-core/src/backprop.rs

@@ -1,4 +1,4 @@
-/// Methods for backpropagation of gradients.
+//! Methods for backpropagation of gradients.
 use crate::op::{BinaryOp, Op, ReduceOp, UnaryOp};
 use crate::{Error, Result, Tensor, TensorId};
 use std::collections::HashMap;
@@ -32,7 +32,7 @@ impl Tensor {
     /// elements having dependencies on the latter ones, e.g. the first element if any is the
     /// argument.
     /// This assumes that the op graph is a DAG.
-    fn sorted_nodes(&self) -> Vec<&Tensor> {
+    pub fn sorted_nodes(&self) -> Vec<&Tensor> {
         // The vec of sorted nodes is passed as an owned value rather than a mutable reference
         // to get around some lifetime limitations.
         fn walk<'a>(
@@ -53,6 +53,7 @@ impl Tensor {
             } else if let Some(op) = node.op() {
                 match op {
                     Op::IndexAdd(t1, t2, t3, _)
+                    | Op::Scatter(t1, t2, t3, _)
                     | Op::ScatterAdd(t1, t2, t3, _)
                     | Op::CustomOp3(t1, t2, t3, _)
                     | Op::WhereCond(t1, t2, t3) => {
@@ -320,13 +321,13 @@ impl Tensor {
                         dilation,
                         output_padding: _output_padding,
                     } => {
-                        let grad_arg = grad.conv2d(kernel, *padding, *dilation, *stride, 1)?;
+                        let grad_arg = grad.conv2d(kernel, *padding, *stride, *dilation, 1)?;
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&grad_arg)?;
 
                         let grad_kernel = grad
                             .transpose(0, 1)?
-                            .conv2d(&arg.transpose(0, 1)?, *padding, *stride, *dilation, 1)?
+                            .conv2d(&arg.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
                             .transpose(0, 1)?;
                         let sum_grad = grads.or_insert(kernel)?;
                         let (_, _, k0, k1) = kernel.dims4()?;
@@ -419,7 +420,7 @@ impl Tensor {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.scatter_add(indexes, &grad, *dim)?;
                     }
-                    Op::ScatterAdd(init, indexes, src, dim) => {
+                    Op::Scatter(init, indexes, src, dim) => {
                         let init_sum_grad = grads.or_insert(init)?;
                         *init_sum_grad = init_sum_grad.add(&grad)?;
 
@@ -427,6 +428,16 @@ impl Tensor {
                         let src_sum_grad = grads.or_insert(src)?;
                         *src_sum_grad = src_sum_grad.add(&src_grad)?;
                     }
+                    Op::ScatterAdd(init, indexes, src, dim) => {
+                        let init_sum_grad = grads.or_insert(init)?;
+                        let mask = init.ones_like()?;
+                        let mask = mask.scatter(indexes, &mask.zeros_like()?, *dim)?;
+                        *init_sum_grad = init_sum_grad.add(&grad.mul(&mask)?)?;
+
+                        let src_grad = grad.gather(indexes, *dim)?;
+                        let src_sum_grad = grads.or_insert(src)?;
+                        *src_sum_grad = src_sum_grad.add(&src_grad)?;
+                    }
                     Op::IndexAdd(init, indexes, src, dim) => {
                         let init_sum_grad = grads.or_insert(init)?;
                         *init_sum_grad = init_sum_grad.add(&grad)?;
@@ -623,9 +634,9 @@ impl Tensor {
                     }
                     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)?)?)?)?;
+                        // d/dx silu = sigmoid(x) * (1 + x * (1 - sigmoid(x))) = sigmoid(x) * (1 - node) + node
+                        let sigmoid_arg = (arg.neg()?.exp()? + 1.)?.recip()?;
+                        let silu_grad = &sigmoid_arg * (1. - *node) + *node;
                         *sum_grad = sum_grad.add(&(&grad * silu_grad)?)?
                     }
                     Op::Elu(arg, alpha) => {
@@ -634,7 +645,8 @@ impl Tensor {
                         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)?;
+                        // node == alpha * (e^x - 1) for x <= 0, reuse it
+                        let negative_exp_mask = (negative_mask * (*node + *alpha))?;
                         let combined_mask = (positive_mask + negative_exp_mask)?;
                         *sum_grad = sum_grad.add(&(grad * combined_mask)?)?
                     }
@@ -755,4 +767,9 @@ impl GradStore {
         };
         Ok(grad)
     }
+
+    /// Get the tensor ids of the stored gradient tensors
+    pub fn get_ids(&self) -> impl Iterator<Item = &TensorId> {
+        self.0.keys()
+    }
 }

candle-core/src/conv.rs

@@ -1,3 +1,5 @@
+//! 1D and 2D Convolutions
+//!
 use crate::{op::BackpropOp, op::Op, Error, Result, Tensor};
 
 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -12,6 +14,7 @@ pub struct ParamsConv1D {
     pub(crate) padding: usize,
     pub(crate) stride: usize,
     pub(crate) dilation: usize,
+    pub(crate) cudnn_fwd_algo: Option<CudnnFwdAlgo>,
 }
 
 impl ParamsConv1D {
@@ -52,7 +55,7 @@ impl ParamsConvTranspose1D {
     }
 }
 
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
 pub enum CudnnFwdAlgo {
     ImplicitGemm,
     ImplicitPrecompGemm,
@@ -149,6 +152,19 @@ impl Tensor {
         stride: usize,
         dilation: usize,
         groups: usize,
+    ) -> Result<Self> {
+        self.conv1d_with_algo(kernel, padding, stride, dilation, groups, None)
+    }
+
+    /// Applies a 1D convolution over the input tensor.
+    pub fn conv1d_with_algo(
+        &self,
+        kernel: &Self,
+        padding: usize,
+        stride: usize,
+        dilation: usize,
+        groups: usize,
+        cudnn_fwd_algo: Option<CudnnFwdAlgo>,
     ) -> Result<Self> {
         let (c_out, c_in_k, k_size) = kernel.dims3()?;
         let (b_size, c_in, l_in) = self.dims3()?;
@@ -172,6 +188,7 @@ impl Tensor {
             padding,
             stride,
             dilation,
+            cudnn_fwd_algo,
         };
         if groups == 1 {
             self.conv1d_single_group(kernel, &params)
@@ -276,6 +293,18 @@ impl Tensor {
         stride: usize,
         dilation: usize,
         groups: usize,
+    ) -> Result<Self> {
+        self.conv2d_with_algo(kernel, padding, stride, dilation, groups, None)
+    }
+
+    pub fn conv2d_with_algo(
+        &self,
+        kernel: &Self,
+        padding: usize,
+        stride: usize,
+        dilation: usize,
+        groups: usize,
+        cudnn_fwd_algo: Option<CudnnFwdAlgo>,
     ) -> Result<Self> {
         let (b_size, c_in, i_h, i_w) = self.dims4()?;
         let (c_out, c_in_k, k_h, k_w) = kernel.dims4()?;
@@ -295,7 +324,7 @@ impl Tensor {
             padding,
             stride,
             dilation,
-            cudnn_fwd_algo: None,
+            cudnn_fwd_algo,
         };
         if groups == 1 {
             self.conv2d_single_group(kernel, &params)

candle-core/src/cpu/mod.rs

@@ -1,6 +1,9 @@
+//! Traits and methods for CPU-backed Tensors
+
 pub mod erf;
 pub mod kernels;
 
+#[allow(unused)]
 trait Cpu<const ARR: usize> {
     type Unit;
     type Array;
@@ -18,6 +21,7 @@ trait Cpu<const ARR: usize> {
     unsafe fn vec_store(mem_addr: *mut f32, a: Self::Unit);
 }
 
+#[allow(unused)]
 trait CpuF16<const ARR: usize> {
     type Unit;
     type Array;

candle-core/src/cpu_backend/mod.rs

@@ -1,3 +1,4 @@
+//! Implementation of Backend Fns for CPU
 use crate::backend::{BackendDevice, BackendStorage};
 use crate::op::{BinaryOpT, CmpOp, ReduceOp, UnaryOpT};
 use crate::{DType, Error, IntDType, Layout, Result, Shape, WithDType};
@@ -6,11 +7,11 @@ use rayon::prelude::*;
 
 mod utils;
 pub use utils::{
-    binary_map, binary_map_vec, unary_map, unary_map_vec, Map1, Map1Any, Map2, Map2U8,
+    binary_map, binary_map_vec, unary_map, unary_map_vec, Map1, Map1Any, Map2, Map2InPlace, Map2U8,
 };
 
 const USE_IM2COL_CONV1D: bool = true;
-const USE_IM2COL_CONV1D_TR: bool = true;
+const USE_COL2IM_CONV1D_TR: bool = true;
 const USE_IM2COL_CONV2D: bool = true;
 
 // TODO: Maybe we should not implement [Clone] here and instead have an explicit allocator +
@@ -26,6 +27,17 @@ pub enum CpuStorage {
     F64(Vec<f64>),
 }
 
+#[derive(Debug, Clone)]
+pub enum CpuStorageRef<'a> {
+    U8(&'a [u8]),
+    U32(&'a [u32]),
+    I64(&'a [i64]),
+    BF16(&'a [bf16]),
+    F16(&'a [f16]),
+    F32(&'a [f32]),
+    F64(&'a [f64]),
+}
+
 #[derive(Debug, Clone)]
 pub struct CpuDevice;
 
@@ -54,7 +66,7 @@ impl Map2U8 for Cmp {
 
 struct WCond<'a, T: IntDType>(&'a [T], &'a Layout);
 
-impl<'a, I: IntDType> Map2 for WCond<'a, I> {
+impl<I: IntDType> Map2 for WCond<'_, I> {
     const OP: &'static str = "where";
     #[inline(always)]
     fn f<T: WithDType>(&self, t: &[T], t_l: &Layout, f: &[T], f_l: &Layout) -> Result<Vec<T>> {
@@ -110,7 +122,8 @@ impl ReduceIndex {
         let dst_len = src_l.shape().elem_count() / reduce_dim_size;
         let mut dst: Vec<U> = Vec::with_capacity(dst_len);
         let dst_to_set = dst.spare_capacity_mut();
-        let dst_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(dst_to_set) };
+        let dst_to_set =
+            unsafe { std::mem::transmute::<&mut [std::mem::MaybeUninit<U>], &mut [U]>(dst_to_set) };
         match src_l.contiguous_offsets() {
             Some((o1, o2)) => {
                 let src = &src[o1..o2];
@@ -203,7 +216,7 @@ struct ReduceSum<'a> {
     reduce_dims_and_stride: Vec<(usize, usize)>,
 }
 
-impl<'a> ReduceSum<'a> {
+impl ReduceSum<'_> {
     #[inline(always)]
     fn fold_impl<T>(&self, src: &[T], src_l: &Layout, start_elt: T) -> Result<Vec<T>>
     where
@@ -268,7 +281,7 @@ impl<'a> ReduceSum<'a> {
     }
 }
 
-impl<'a> Map1 for ReduceSum<'a> {
+impl Map1 for ReduceSum<'_> {
     #[inline(always)]
     fn f<T: WithDType>(&self, src: &[T], src_l: &Layout) -> Result<Vec<T>> {
         self.fold_impl(src, src_l, T::zero())
@@ -441,7 +454,7 @@ struct Gather<'a, I: IntDType> {
     dim: usize,
 }
 
-impl<'a, I: IntDType> Map1 for Gather<'a, I> {
+impl<I: IntDType> Map1 for Gather<'_, 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],
@@ -494,7 +507,7 @@ struct IndexSelect<'a, T: IntDType> {
     dim: usize,
 }
 
-impl<'a, I: IntDType> Map1 for IndexSelect<'a, I> {
+impl<I: IntDType> Map1 for IndexSelect<'_, 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],
@@ -541,26 +554,65 @@ impl<'a, I: IntDType> Map1 for IndexSelect<'a, I> {
     }
 }
 
-struct ScatterAdd<'a, I: IntDType> {
+trait ElemUpdate {
+    fn f<T: WithDType>(dst: &mut T, src: T);
+}
+
+struct Set;
+struct Add;
+
+impl ElemUpdate for Set {
+    fn f<T: WithDType>(dst: &mut T, src: T) {
+        *dst = src
+    }
+}
+
+impl ElemUpdate for Add {
+    fn f<T: WithDType>(dst: &mut T, src: T) {
+        *dst += src
+    }
+}
+
+struct Scatter<'a, I: IntDType, M: ElemUpdate> {
     ids: &'a [I],
     ids_l: &'a Layout,
     dim: usize,
+    _phantom: std::marker::PhantomData<M>,
 }
 
-impl<'a, I: IntDType> Map2 for ScatterAdd<'a, I> {
-    const OP: &'static str = "scatter-add";
-    fn f<T: WithDType>(&self, v1: &[T], l1: &Layout, src: &[T], src_l: &Layout) -> Result<Vec<T>> {
-        let dst_len = l1.shape().elem_count();
-        let mut dst = vec![T::zero(); dst_len];
-        copy_strided_src_(v1, &mut dst, 0, l1);
+impl<'a, I: IntDType, M: ElemUpdate> Scatter<'a, I, M> {
+    fn new(ids: &'a [I], ids_l: &'a Layout, dim: usize) -> Self {
+        Self {
+            ids,
+            ids_l,
+            dim,
+            _phantom: Default::default(),
+        }
+    }
+}
+
+impl<I: IntDType, M: ElemUpdate> Map2InPlace for Scatter<'_, I, M> {
+    const OP: &'static str = "scatter";
+    fn f<T: WithDType>(
+        &self,
+        dst: &mut [T],
+        dst_l: &Layout,
+        src: &[T],
+        src_l: &Layout,
+    ) -> Result<()> {
+        let dst = match dst_l.contiguous_offsets() {
+            None => Err(Error::RequiresContiguous { op: "scatter" }.bt())?,
+            Some((o1, o2)) => &mut dst[o1..o2],
+        };
+
         let src = match src_l.contiguous_offsets() {
-            None => Err(Error::RequiresContiguous { op: "scatter-add" }.bt())?,
+            None => Err(Error::RequiresContiguous { op: "scatter" }.bt())?,
             Some((o1, o2)) => &src[o1..o2],
         };
 
         let dim = self.dim;
         let ids_dims = self.ids_l.dims();
-        let dst_dims = l1.dims();
+        let dst_dims = dst_l.dims();
         let dst_dim_len = dst_dims[dim];
         let dst_right_len: usize = dst_dims[dim + 1..].iter().product();
 
@@ -589,12 +641,12 @@ impl<'a, I: IntDType> Map2 for ScatterAdd<'a, I> {
                         .bt())?
                     }
                     let dst_idx = start_dst_idx + index * dst_right_len + right_i;
-                    dst[dst_idx] += src[ids_idx]
+                    M::f(&mut dst[dst_idx], src[ids_idx])
                 }
             }
         }
 
-        Ok(dst)
+        Ok(())
     }
 }
 
@@ -603,7 +655,7 @@ struct IndexAdd<'a, I: IntDType> {
     dim: usize,
 }
 
-impl<'a, I: IntDType> Map2 for IndexAdd<'a, I> {
+impl<I: IntDType> Map2 for IndexAdd<'_, I> {
     const OP: &'static str = "index-add";
     // https://pytorch.org/docs/stable/generated/torch.Tensor.index_add_.html#torch.Tensor.index_add_
     // v1, l1 -> self
@@ -723,7 +775,7 @@ fn copy_strided_src_<T: Copy>(src: &[T], dst: &mut [T], dst_offset: usize, src_l
 
 struct Conv1D<'a>(&'a crate::conv::ParamsConv1D);
 
-impl<'a> Map2 for Conv1D<'a> {
+impl Map2 for Conv1D<'_> {
     const OP: &'static str = "conv1d";
     fn f<T: WithDType>(&self, inp: &[T], inp_l: &Layout, k: &[T], k_l: &Layout) -> Result<Vec<T>> {
         let p = self.0;
@@ -947,7 +999,7 @@ impl Map1 for Col2Im1D {
 
 struct ConvTranspose1D<'a>(&'a crate::conv::ParamsConvTranspose1D);
 
-impl<'a> Map2 for ConvTranspose1D<'a> {
+impl Map2 for ConvTranspose1D<'_> {
     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;
@@ -1016,7 +1068,7 @@ impl<'a> Map2 for ConvTranspose1D<'a> {
 
 struct Conv2D<'a>(&'a crate::conv::ParamsConv2D);
 
-impl<'a> Map2 for Conv2D<'a> {
+impl Map2 for Conv2D<'_> {
     const OP: &'static str = "conv2d";
     fn f<T: WithDType>(&self, inp: &[T], inp_l: &Layout, k: &[T], k_l: &Layout) -> Result<Vec<T>> {
         let p = self.0;
@@ -1104,7 +1156,7 @@ impl<'a> Map2 for Conv2D<'a> {
 
 struct ConvTranspose2D<'a>(&'a crate::conv::ParamsConvTranspose2D);
 
-impl<'a> Map2 for ConvTranspose2D<'a> {
+impl Map2 for ConvTranspose2D<'_> {
     const OP: &'static str = "conv_transpose2d";
     fn f<T: WithDType>(&self, inp: &[T], inp_l: &Layout, k: &[T], k_l: &Layout) -> Result<Vec<T>> {
         let p = self.0;
@@ -1276,6 +1328,15 @@ impl Map2 for MatMul {
         } else {
             Parallelism::None
         };
+        let (b, m, n, k) = if b_skip == 0 && a_skip == m * k {
+            // a_skip and c_skip should be updated but step is always 0 so
+            // it wouldn't matter.
+            (1, b * m, n, k)
+        } else if a_skip == 0 && b_skip == n * k {
+            (1, m, b * n, k)
+        } else {
+            (b, m, n, k)
+        };
         for step in 0..b {
             let lhs_p = &lhs[step * a_skip..];
             let rhs_p = &rhs[step * b_skip..];
@@ -2238,7 +2299,7 @@ impl BackendStorage for CpuStorage {
             && params.dilation == 1
             && params.padding == 0
             && params.output_padding == 0;
-        if USE_IM2COL_CONV1D_TR && can_use_col2im {
+        if USE_COL2IM_CONV1D_TR && can_use_col2im {
             let (b_size, c_in, l_in) = l.shape().dims3()?;
             let (c_in2, c_out, k_size) = kernel_l.shape().dims3()?;
             if !kernel_l.is_contiguous() {
@@ -2359,19 +2420,36 @@ impl BackendStorage for CpuStorage {
         }
     }
 
-    fn scatter_add(
-        &self,
+    fn scatter_set(
+        &mut self,
         l: &Layout,
         ids: &Self,
         ids_l: &Layout,
         src: &Self,
         src_l: &Layout,
         dim: usize,
-    ) -> Result<Self> {
+    ) -> Result<()> {
         match ids {
-            Self::U8(ids) => ScatterAdd { ids, ids_l, dim }.map(self, l, src, src_l),
-            Self::U32(ids) => ScatterAdd { ids, ids_l, dim }.map(self, l, src, src_l),
-            Self::I64(ids) => ScatterAdd { ids, ids_l, dim }.map(self, l, src, src_l),
+            Self::U8(ids) => Scatter::<_, Set>::new(ids, ids_l, dim).map(self, l, src, src_l),
+            Self::U32(ids) => Scatter::<_, Set>::new(ids, ids_l, dim).map(self, l, src, src_l),
+            Self::I64(ids) => Scatter::<_, Set>::new(ids, ids_l, dim).map(self, l, src, src_l),
+            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "scatter").bt()),
+        }
+    }
+
+    fn scatter_add_set(
+        &mut self,
+        l: &Layout,
+        ids: &Self,
+        ids_l: &Layout,
+        src: &Self,
+        src_l: &Layout,
+        dim: usize,
+    ) -> Result<()> {
+        match ids {
+            Self::U8(ids) => Scatter::<_, Add>::new(ids, ids_l, dim).map(self, l, src, src_l),
+            Self::U32(ids) => Scatter::<_, Add>::new(ids, ids_l, dim).map(self, l, src, src_l),
+            Self::I64(ids) => Scatter::<_, Add>::new(ids, ids_l, dim).map(self, l, src, src_l),
             _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "scatter-add").bt()),
         }
     }
@@ -2432,6 +2510,48 @@ impl BackendStorage for CpuStorage {
     fn to_cpu_storage(&self) -> Result<CpuStorage> {
         Ok(self.clone())
     }
+
+    fn const_set(&mut self, s: crate::scalar::Scalar, l: &Layout) -> Result<()> {
+        use crate::scalar::Scalar;
+        fn set<T: crate::WithDType>(src: &mut [T], l: &Layout, s: T) {
+            match l.strided_blocks() {
+                crate::StridedBlocks::SingleBlock { start_offset, len } => {
+                    src[start_offset..start_offset + len].fill(s)
+                }
+                crate::StridedBlocks::MultipleBlocks {
+                    block_start_index,
+                    block_len: 1,
+                } => {
+                    for src_index in block_start_index {
+                        src[src_index] = s
+                    }
+                }
+                crate::StridedBlocks::MultipleBlocks {
+                    block_start_index,
+                    block_len,
+                } => {
+                    for src_index in block_start_index {
+                        src[src_index..src_index + block_len].fill(s)
+                    }
+                }
+            }
+        }
+        match (self, s) {
+            (Self::BF16(storage), Scalar::BF16(v)) => set(storage, l, v),
+            (Self::F16(storage), Scalar::F16(v)) => set(storage, l, v),
+            (Self::F32(storage), Scalar::F32(v)) => set(storage, l, v),
+            (Self::F64(storage), Scalar::F64(v)) => set(storage, l, v),
+            (Self::U8(storage), Scalar::U8(v)) => set(storage, l, v),
+            (Self::U32(storage), Scalar::U32(v)) => set(storage, l, v),
+            (Self::I64(storage), Scalar::I64(v)) => set(storage, l, v),
+            (st, s) => crate::bail!(
+                "const_set dtype mismatch, expected {:?} but got {:?}",
+                st.dtype(),
+                s
+            ),
+        }
+        Ok(())
+    }
 }
 
 impl BackendDevice for CpuDevice {
@@ -2445,6 +2565,10 @@ impl BackendDevice for CpuDevice {
         true
     }
 
+    fn storage_from_slice<T: crate::WithDType>(&self, s: &[T]) -> Result<Self::Storage> {
+        Ok(T::to_cpu_storage(s))
+    }
+
     fn storage_from_cpu_storage(&self, s: &CpuStorage) -> Result<Self::Storage> {
         Ok(s.clone())
     }
@@ -2465,15 +2589,15 @@ impl BackendDevice for CpuDevice {
         use rand::prelude::*;
 
         let elem_count = shape.elem_count();
-        let mut rng = rand::thread_rng();
+        let mut rng = rand::rng();
         match dtype {
             DType::U8 | DType::U32 | DType::I64 => {
                 Err(Error::UnsupportedDTypeForOp(dtype, "rand_uniform").bt())
             }
             DType::BF16 => {
                 let mut data = Vec::with_capacity(elem_count);
-                let uniform =
-                    rand::distributions::Uniform::new(bf16::from_f64(min), bf16::from_f64(max));
+                let uniform = rand::distr::Uniform::new(bf16::from_f64(min), bf16::from_f64(max))
+                    .map_err(Error::wrap)?;
                 for _i in 0..elem_count {
                     data.push(rng.sample::<bf16, _>(uniform))
                 }
@@ -2481,8 +2605,8 @@ impl BackendDevice for CpuDevice {
             }
             DType::F16 => {
                 let mut data = Vec::with_capacity(elem_count);
-                let uniform =
-                    rand::distributions::Uniform::new(f16::from_f64(min), f16::from_f64(max));
+                let uniform = rand::distr::Uniform::new(f16::from_f64(min), f16::from_f64(max))
+                    .map_err(Error::wrap)?;
                 for _i in 0..elem_count {
                     data.push(rng.sample::<f16, _>(uniform))
                 }
@@ -2490,7 +2614,8 @@ impl BackendDevice for CpuDevice {
             }
             DType::F32 => {
                 let mut data = Vec::with_capacity(elem_count);
-                let uniform = rand::distributions::Uniform::new(min as f32, max as f32);
+                let uniform =
+                    rand::distr::Uniform::new(min as f32, max as f32).map_err(Error::wrap)?;
                 for _i in 0..elem_count {
                     data.push(rng.sample::<f32, _>(uniform))
                 }
@@ -2498,7 +2623,7 @@ impl BackendDevice for CpuDevice {
             }
             DType::F64 => {
                 let mut data = Vec::with_capacity(elem_count);
-                let uniform = rand::distributions::Uniform::new(min, max);
+                let uniform = rand::distr::Uniform::new(min, max).map_err(Error::wrap)?;
                 for _i in 0..elem_count {
                     data.push(rng.sample::<f64, _>(uniform))
                 }
@@ -2511,7 +2636,7 @@ impl BackendDevice for CpuDevice {
         use rand::prelude::*;
 
         let elem_count = shape.elem_count();
-        let mut rng = rand::thread_rng();
+        let mut rng = rand::rng();
         match dtype {
             DType::U8 | DType::U32 | DType::I64 => {
                 Err(Error::UnsupportedDTypeForOp(dtype, "rand_normal").bt())
@@ -2601,20 +2726,6 @@ impl BackendDevice for CpuDevice {
         Ok(storage)
     }
 
-    fn ones_impl(&self, shape: &Shape, dtype: DType) -> Result<CpuStorage> {
-        let elem_count = shape.elem_count();
-        let storage = match dtype {
-            DType::U8 => CpuStorage::U8(vec![1u8; elem_count]),
-            DType::U32 => CpuStorage::U32(vec![1u32; elem_count]),
-            DType::I64 => CpuStorage::I64(vec![1i64; elem_count]),
-            DType::BF16 => CpuStorage::BF16(vec![bf16::ONE; elem_count]),
-            DType::F16 => CpuStorage::F16(vec![f16::ONE; elem_count]),
-            DType::F32 => CpuStorage::F32(vec![1f32; elem_count]),
-            DType::F64 => CpuStorage::F64(vec![1f64; elem_count]),
-        };
-        Ok(storage)
-    }
-
     fn zeros_impl(&self, shape: &Shape, dtype: DType) -> Result<CpuStorage> {
         let elem_count = shape.elem_count();
         let storage = match dtype {
@@ -2628,6 +2739,10 @@ impl BackendDevice for CpuDevice {
         };
         Ok(storage)
     }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
 }
 
 #[macro_export]

candle-core/src/cpu_backend/utils.rs

@@ -58,6 +58,30 @@ pub trait Map2 {
     }
 }
 
+pub trait Map2InPlace {
+    const OP: &'static str;
+    fn f<T: WithDType>(&self, v1: &mut [T], l1: &Layout, v2: &[T], l2: &Layout) -> Result<()>;
+
+    fn map(&self, v1: &mut C, l1: &Layout, v2: &C, l2: &Layout) -> Result<()> {
+        match (v1, v2) {
+            (C::U8(v1), C::U8(v2)) => self.f(v1, l1, v2, l2)?,
+            (C::U32(v1), C::U32(v2)) => self.f(v1, l1, v2, l2)?,
+            (C::I64(v1), C::I64(v2)) => self.f(v1, l1, v2, l2)?,
+            (C::BF16(v1), C::BF16(v2)) => self.f(v1, l1, v2, l2)?,
+            (C::F16(v1), C::F16(v2)) => self.f(v1, l1, v2, l2)?,
+            (C::F32(v1), C::F32(v2)) => self.f(v1, l1, v2, l2)?,
+            (C::F64(v1), C::F64(v2)) => self.f(v1, l1, v2, l2)?,
+            (v1, v2) => Err(Error::DTypeMismatchBinaryOp {
+                lhs: v1.dtype(),
+                rhs: v2.dtype(),
+                op: Self::OP,
+            }
+            .bt())?,
+        };
+        Ok(())
+    }
+}
+
 pub trait Map2U8 {
     const OP: &'static str;
     fn f<T: WithDType>(&self, v1: &[T], l1: &Layout, v2: &[T], l2: &Layout) -> Result<Vec<u8>>;
@@ -174,7 +198,9 @@ pub fn binary_map_vec<T: Copy, F: FnMut(T, T) -> T, FV: FnMut(&[T], &[T], &mut [
         (Some((o_l1, o_l2)), Some((o_r1, o_r2))) => {
             let mut ys: Vec<T> = Vec::with_capacity(el_count);
             let ys_to_set = ys.spare_capacity_mut();
-            let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
+            let ys_to_set = unsafe {
+                std::mem::transmute::<&mut [std::mem::MaybeUninit<T>], &mut [T]>(ys_to_set)
+            };
             f_vec(&lhs[o_l1..o_l2], &rhs[o_r1..o_r2], ys_to_set);
             // SAFETY: values are all set by f_vec.
             unsafe { ys.set_len(el_count) };
@@ -185,7 +211,9 @@ pub fn binary_map_vec<T: Copy, F: FnMut(T, T) -> T, FV: FnMut(&[T], &[T], &mut [
                 let rhs = &rhs[ob.start..ob.start + ob.len];
                 let mut ys: Vec<T> = Vec::with_capacity(el_count);
                 let ys_to_set = ys.spare_capacity_mut();
-                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
+                let ys_to_set = unsafe {
+                    std::mem::transmute::<&mut [std::mem::MaybeUninit<T>], &mut [T]>(ys_to_set)
+                };
                 let mut dst_i = 0;
                 for src_i in (o_l1..o_l2).step_by(ob.len) {
                     f_vec(
@@ -224,7 +252,9 @@ pub fn binary_map_vec<T: Copy, F: FnMut(T, T) -> T, FV: FnMut(&[T], &[T], &mut [
                 let lhs = &lhs[ob.start..ob.start + ob.len];
                 let mut ys: Vec<T> = Vec::with_capacity(el_count);
                 let ys_to_set = ys.spare_capacity_mut();
-                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
+                let ys_to_set = unsafe {
+                    std::mem::transmute::<&mut [std::mem::MaybeUninit<T>], &mut [T]>(ys_to_set)
+                };
                 let mut dst_i = 0;
                 for src_i in (o_r1..o_r2).step_by(ob.len) {
                     f_vec(
@@ -311,7 +341,9 @@ pub fn unary_map_vec<T: Copy, U: Copy, F: FnMut(T) -> U, FV: FnMut(&[T], &mut [U
         crate::StridedBlocks::SingleBlock { start_offset, len } => {
             let mut ys: Vec<U> = Vec::with_capacity(len);
             let ys_to_set = ys.spare_capacity_mut();
-            let ys_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(ys_to_set) };
+            let ys_to_set = unsafe {
+                std::mem::transmute::<&mut [std::mem::MaybeUninit<U>], &mut [U]>(ys_to_set)
+            };
             f_vec(&vs[start_offset..start_offset + len], ys_to_set);
             // SAFETY: values are all set by f_vec.
             unsafe { ys.set_len(len) };
@@ -333,7 +365,9 @@ pub fn unary_map_vec<T: Copy, U: Copy, F: FnMut(T) -> U, FV: FnMut(&[T], &mut [U
             } else {
                 let mut ys: Vec<U> = Vec::with_capacity(el_count);
                 let ys_to_set = ys.spare_capacity_mut();
-                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(ys_to_set) };
+                let ys_to_set = unsafe {
+                    std::mem::transmute::<&mut [std::mem::MaybeUninit<U>], &mut [U]>(ys_to_set)
+                };
                 let mut dst_index = 0;
                 for src_index in block_start_index {
                     let vs = &vs[src_index..src_index + block_len];

candle-core/src/cuda_backend/cudnn.rs

@@ -1,6 +1,6 @@
 use crate::WithDType;
 use cudarc;
-use cudarc::cudnn::safe::{Conv2dForward, Cudnn};
+use cudarc::cudnn::safe::{ConvForward, Cudnn};
 use cudarc::driver::{CudaSlice, CudaView, DeviceRepr, ValidAsZeroBits};
 use std::cell::RefCell;
 use std::collections::HashMap;
@@ -26,6 +26,7 @@ impl From<cudarc::driver::DriverError> for crate::Error {
 
 pub(crate) fn launch_conv2d<
     T: DeviceRepr + WithDType + ValidAsZeroBits + cudarc::cudnn::CudnnDataType,
+    Y: cudarc::cudnn::CudnnDataType,
 >(
     src: &CudaView<T>,
     src_l: &crate::Layout,
@@ -42,13 +43,13 @@ pub(crate) fn launch_conv2d<
         if let Some(cudnn) = cudnn.borrow().get(&device_id) {
             return Ok(cudnn.clone());
         }
-        let c = Cudnn::new(dev.cuda_device());
+        let c = Cudnn::new(dev.cuda_stream());
         if let Ok(c) = &c {
             cudnn.borrow_mut().insert(device_id, c.clone());
         }
         c
     })?;
-    let conv = cudnn.create_conv2d::<T>(
+    let conv = cudnn.create_conv2d::<Y>(
         /* pad */ [params.padding as i32, params.padding as i32],
         /* stride */ [params.stride as i32, params.stride as i32],
         /* dilation */ [params.dilation as i32, params.dilation as i32],
@@ -62,18 +63,18 @@ pub(crate) fn launch_conv2d<
     ];
     // Note that `src` already starts at the proper offset.
     let x = if src_l.is_contiguous() {
-        cudnn.create_4d_tensor(
+        cudnn.create_4d_tensor::<T>(
             cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
             x_shape,
         )?
     } else {
         let s = src_l.stride();
-        cudnn.create_4d_tensor_ex(
+        cudnn.create_4d_tensor_ex::<T>(
             x_shape,
             [s[0] as i32, s[1] as i32, s[2] as i32, s[3] as i32],
         )?
     };
-    let w = cudnn.create_4d_filter(
+    let w = cudnn.create_4d_filter::<T>(
         cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
         [
             params.c_out as i32,
@@ -83,11 +84,11 @@ pub(crate) fn launch_conv2d<
         ],
     )?;
     let (w_out, h_out) = (params.out_w() as i32, params.out_h() as i32);
-    let y = cudnn.create_4d_tensor(
+    let y = cudnn.create_4d_tensor::<T>(
         cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
         [params.b_size as i32, params.c_out as i32, h_out, w_out],
     )?;
-    let conv2d = Conv2dForward {
+    let conv2d = ConvForward {
         conv: &conv,
         x: &x,
         w: &w,
@@ -108,7 +109,7 @@ pub(crate) fn launch_conv2d<
         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)?;
+    let mut workspace = dev.cuda_stream().alloc_zeros::<u8>(workspace_size)?;
     unsafe {
         conv2d.launch::<CudaSlice<u8>, _, _, _>(
             alg,
@@ -121,3 +122,104 @@ pub(crate) fn launch_conv2d<
     }
     Ok(())
 }
+
+pub(crate) fn launch_conv1d<
+    T: DeviceRepr + WithDType + ValidAsZeroBits + cudarc::cudnn::CudnnDataType,
+    Y: cudarc::cudnn::CudnnDataType,
+>(
+    src: &CudaView<T>,
+    src_l: &crate::Layout,
+    filter: &CudaView<T>,
+    dst: &mut CudaSlice<T>,
+    params: &crate::conv::ParamsConv1D,
+    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) {
+            return Ok(cudnn.clone());
+        }
+        let c = Cudnn::new(dev.cuda_stream());
+        if let Ok(c) = &c {
+            cudnn.borrow_mut().insert(device_id, c.clone());
+        }
+        c
+    })?;
+    let conv = cudnn.create_conv2d::<Y>(
+        /* pad */ [params.padding as i32, 0],
+        /* stride */ [params.stride as i32, 1],
+        /* dilation */ [params.dilation as i32, 1],
+        cudarc::cudnn::sys::cudnnConvolutionMode_t::CUDNN_CROSS_CORRELATION,
+    )?;
+    // https://docs.nvidia.com/deeplearning/cudnn/backend/latest/api/cudnn-ops-library.html#cudnnsettensornddescriptor
+    // > Tensors are restricted to having at least 4 dimensions, and at most CUDNN_DIM_MAX
+    // > dimensions (defined in cudnn.h). When working with lower dimensional data, it is
+    // > recommended that the user create a 4D tensor, and set the size along unused dimensions
+    // > to 1.
+    let x_shape = [
+        params.b_size as i32,
+        params.c_in as i32,
+        params.l_in as i32,
+        1,
+    ];
+    // Note that `src` already starts at the proper offset.
+    let x = if src_l.is_contiguous() {
+        cudnn.create_4d_tensor::<T>(
+            cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
+            x_shape,
+        )?
+    } else {
+        let s = src_l.stride();
+        cudnn.create_4d_tensor_ex::<T>(x_shape, [s[0] as i32, s[1] as i32, s[2] as i32, 1i32])?
+    };
+    let w = cudnn.create_4d_filter::<T>(
+        cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
+        [
+            params.c_out as i32,
+            params.c_in as i32,
+            params.k_size as i32,
+            1,
+        ],
+    )?;
+    let l_out = params.l_out() as i32;
+    let y = cudnn.create_4d_tensor::<T>(
+        cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
+        [params.b_size as i32, params.c_out as i32, l_out, 1],
+    )?;
+    let conv1d = ConvForward {
+        conv: &conv,
+        x: &x,
+        w: &w,
+        y: &y,
+    };
+    let alg = match params.cudnn_fwd_algo {
+        None => conv1d.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 = conv1d.get_workspace_size(alg)?;
+    let mut workspace = dev.cuda_stream().alloc_zeros::<u8>(workspace_size)?;
+    unsafe {
+        conv1d.launch::<CudaSlice<u8>, _, _, _>(
+            alg,
+            Some(&mut workspace),
+            (T::one(), T::zero()),
+            src,
+            filter,
+            dst,
+        )?;
+    }
+    Ok(())
+}

candle-core/src/cuda_backend/device.rs

@@ -1,9 +1,10 @@
 use crate::backend::BackendDevice;
-use crate::{CpuStorage, DType, Layout, Result, Shape};
+use crate::{CpuStorage, CpuStorageRef, DType, Layout, Result, Shape};
 pub use candle_kernels as kernels;
 pub use cudarc;
-use cudarc::driver::{CudaFunction, LaunchAsync, LaunchConfig};
+use cudarc::driver::CudaFunction;
 use half::{bf16, f16};
+use std::collections::HashMap;
 use std::sync::{Arc, Mutex};
 
 use super::{CudaError, CudaStorage, CudaStorageSlice, WrapErr};
@@ -24,10 +25,17 @@ impl DeviceId {
 struct CudaRng(cudarc::curand::CudaRng);
 unsafe impl Send for CudaRng {}
 
+pub struct ModuleStore {
+    mdls: [Option<Arc<cudarc::driver::CudaModule>>; kernels::ALL_IDS.len()],
+}
+
 #[derive(Clone)]
 pub struct CudaDevice {
     id: DeviceId,
-    device: Arc<cudarc::driver::CudaDevice>,
+    context: Arc<cudarc::driver::CudaContext>,
+    modules: Arc<std::sync::RwLock<ModuleStore>>,
+    custom_modules: Arc<std::sync::RwLock<HashMap<String, Arc<cudarc::driver::CudaModule>>>>,
+    stream: Arc<cudarc::driver::CudaStream>,
     pub(crate) blas: Arc<cudarc::cublas::CudaBlas>,
     curand: Arc<Mutex<CudaRng>>,
 }
@@ -38,109 +46,212 @@ impl std::fmt::Debug for CudaDevice {
     }
 }
 
-impl std::ops::Deref for CudaDevice {
-    type Target = Arc<cudarc::driver::CudaDevice>;
+impl CudaDevice {
+    #[allow(clippy::missing_safety_doc)]
+    pub unsafe fn alloc<T: cudarc::driver::DeviceRepr>(
+        &self,
+        len: usize,
+    ) -> Result<cudarc::driver::CudaSlice<T>> {
+        self.stream.alloc::<T>(len).w()
+    }
+
+    pub fn alloc_zeros<T: cudarc::driver::DeviceRepr + cudarc::driver::ValidAsZeroBits>(
+        &self,
+        len: usize,
+    ) -> Result<cudarc::driver::CudaSlice<T>> {
+        self.stream.alloc_zeros::<T>(len).w()
+    }
+
+    pub fn memcpy_htod<
+        T: cudarc::driver::DeviceRepr,
+        Src: cudarc::driver::HostSlice<T> + ?Sized,
+        Dst: cudarc::driver::DevicePtrMut<T>,
+    >(
+        &self,
+        src: &Src,
+        dst: &mut Dst,
+    ) -> Result<()> {
+        self.stream.memcpy_htod(src, dst).w()
+    }
+
+    pub fn memcpy_dtov<T: cudarc::driver::DeviceRepr, Src: cudarc::driver::DevicePtr<T>>(
+        &self,
+        src: &Src,
+    ) -> Result<Vec<T>> {
+        self.stream.memcpy_dtov(src).w()
+    }
+
+    pub fn memcpy_dtod<
+        T,
+        Src: cudarc::driver::DevicePtr<T>,
+        Dst: cudarc::driver::DevicePtrMut<T>,
+    >(
+        &self,
+        src: &Src,
+        dst: &mut Dst,
+    ) -> Result<()> {
+        self.stream.memcpy_dtod(src, dst).w()
+    }
+
+    pub fn memcpy_stod<
+        T: cudarc::driver::DeviceRepr,
+        Src: cudarc::driver::HostSlice<T> + ?Sized,
+    >(
+        &self,
+        src: &Src,
+    ) -> Result<cudarc::driver::CudaSlice<T>> {
+        self.stream.memcpy_stod(src).w()
+    }
+}
+
+pub struct CudaFunc {
+    func: CudaFunction,
+    stream: Arc<cudarc::driver::CudaStream>,
+}
+
+impl std::ops::Deref for CudaFunc {
+    type Target = CudaFunction;
 
     fn deref(&self) -> &Self::Target {
-        &self.device
+        &self.func
+    }
+}
+
+impl CudaFunc {
+    pub fn into_cuda_function(self) -> CudaFunction {
+        self.func
+    }
+}
+
+#[macro_export]
+macro_rules! builder_arg {
+    ($b:ident, $($arg:expr),*) => {
+        $(
+            let __arg = $arg;
+            $b.arg(&__arg);
+        )*
+    };
+}
+
+impl CudaFunc {
+    pub fn builder(&self) -> cudarc::driver::LaunchArgs<'_> {
+        self.stream.launch_builder(&self.func)
     }
 }
 
 impl CudaDevice {
-    pub fn cuda_device(&self) -> Arc<cudarc::driver::CudaDevice> {
-        self.device.clone()
+    pub fn cuda_stream(&self) -> Arc<cudarc::driver::CudaStream> {
+        self.stream.clone()
+    }
+
+    /// When turned on, all cuda tensors **created after calling this function** will
+    /// not track uses via cuda events.
+    ///
+    /// # Safety
+    ///
+    /// It is up to the user to ensure proper synchronization between multiple streams:
+    /// - Ensure that no tensor is freed before a use on another stream is finished.
+    /// - Ensure that a tensor is not used on another stream before allocation on the
+    ///   allocating stream finishes.
+    /// - Ensure that a tensor is not written two concurrently by multiple streams.
+    pub unsafe fn disable_event_tracking(&self) {
+        self.context.disable_event_tracking()
+    }
+
+    pub fn is_event_tracking(&self) -> bool {
+        self.context.is_event_tracking()
+    }
+
+    #[cfg(not(target_arch = "wasm32"))]
+    pub fn compile(
+        &self,
+        func_name: &'static str,
+        kernel: ug::lang::ssa::Kernel,
+    ) -> Result<CudaFunc> {
+        let mut buf = vec![];
+        ug_cuda::code_gen::gen(&mut buf, func_name, &kernel)?;
+        let cuda_code = String::from_utf8(buf)?;
+        let opts = cudarc::nvrtc::CompileOptions {
+            use_fast_math: Some(true),
+            ..Default::default()
+        };
+        let ptx = cudarc::nvrtc::safe::compile_ptx_with_opts(cuda_code, opts).w()?;
+        let module = self.context.load_module(ptx).w()?;
+        let func = module.load_function(func_name).w()?;
+        Ok(CudaFunc {
+            func,
+            stream: self.stream.clone(),
+        })
     }
 
     pub fn id(&self) -> DeviceId {
         self.id
     }
 
-    fn const_impl(&self, v: f64, shape: &Shape, dtype: DType) -> Result<CudaStorage> {
-        let elem_count = shape.elem_count();
-        let cfg = LaunchConfig::for_num_elems(elem_count as u32);
-        let slice = match dtype {
-            DType::U8 => {
-                // SAFETY: Set later by running the fill kernel.
-                let data = unsafe { self.alloc::<u8>(elem_count) }.w()?;
-                let func = self.get_or_load_func("fill_u8", kernels::FILL)?;
-                let params = (&data, v as u8, elem_count);
-                unsafe { func.launch(cfg, params) }.w()?;
-                CudaStorageSlice::U8(data)
-            }
-            DType::U32 => {
-                // SAFETY: Set later by running the fill kernel.
-                let data = unsafe { self.alloc::<u32>(elem_count) }.w()?;
-                let func = self.get_or_load_func("fill_u32", kernels::FILL)?;
-                let params = (&data, v as u32, elem_count);
-                unsafe { func.launch(cfg, params) }.w()?;
-                CudaStorageSlice::U32(data)
-            }
-            DType::I64 => {
-                // SAFETY: Set later by running the fill kernel.
-                let data = unsafe { self.alloc::<i64>(elem_count) }.w()?;
-                let func = self.get_or_load_func("fill_i64", kernels::FILL)?;
-                let params = (&data, v as i64, elem_count);
-                unsafe { func.launch(cfg, params) }.w()?;
-                CudaStorageSlice::I64(data)
-            }
-            DType::BF16 => {
-                // SAFETY: Set later by running the fill kernel.
-                let data = unsafe { self.alloc::<bf16>(elem_count) }.w()?;
-                let func = self.get_or_load_func("fill_bf16", kernels::FILL)?;
-                let params = (&data, bf16::from_f64(v), elem_count);
-                unsafe { func.launch(cfg, params) }.w()?;
-                CudaStorageSlice::BF16(data)
-            }
-            DType::F16 => {
-                // SAFETY: Set later by running the fill kernel.
-                let data = unsafe { self.alloc::<f16>(elem_count) }.w()?;
-                let func = self.get_or_load_func("fill_f16", kernels::FILL)?;
-                let params = (&data, f16::from_f64(v), elem_count);
-                unsafe { func.launch(cfg, params) }.w()?;
-                CudaStorageSlice::F16(data)
-            }
-            DType::F32 => {
-                // SAFETY: Set later by running the fill kernel.
-                let data = unsafe { self.alloc::<f32>(elem_count) }.w()?;
-                let func = self.get_or_load_func("fill_f32", kernels::FILL)?;
-                let params = (&data, v as f32, elem_count);
-                unsafe { func.launch(cfg, params) }.w()?;
-                CudaStorageSlice::F32(data)
-            }
-            DType::F64 => {
-                // SAFETY: Set later by running the fill kernel.
-                let data = unsafe { self.alloc::<f64>(elem_count) }.w()?;
-                let func = self.get_or_load_func("fill_f64", kernels::FILL)?;
-                let params = (&data, v, elem_count);
-                unsafe { func.launch(cfg, params) }.w()?;
-                CudaStorageSlice::F64(data)
-            }
-        };
-        Ok(CudaStorage {
-            slice,
-            device: self.clone(),
+    pub fn get_or_load_custom_func(
+        &self,
+        fn_name: &str,
+        module_name: &str,
+        ptx: &str,
+    ) -> Result<CudaFunc> {
+        let ms = self.custom_modules.read().unwrap();
+        if let Some(mdl) = ms.get(module_name).as_ref() {
+            let func = mdl.load_function(fn_name).w()?;
+            return Ok(CudaFunc {
+                func,
+                stream: self.stream.clone(),
+            });
+        }
+        drop(ms);
+        let mut ms = self.custom_modules.write().unwrap();
+        let cuda_module = self.context.load_module(ptx.into()).w()?;
+        ms.insert(module_name.to_string(), cuda_module.clone());
+        let func = cuda_module.load_function(fn_name).w()?;
+        Ok(CudaFunc {
+            func,
+            stream: self.stream.clone(),
         })
     }
 
-    pub fn get_or_load_func(&self, module_name: &str, ptx: &'static str) -> Result<CudaFunction> {
-        if !self.has_func(module_name, module_name) {
-            // Leaking the string here is a bit sad but we need a &'static str and this is only
-            // done once per kernel name.
-            let static_module_name = Box::leak(module_name.to_string().into_boxed_str());
-            self.load_ptx(ptx.into(), module_name, &[static_module_name])
-                .map_err(|cuda| CudaError::Load {
-                    cuda,
-                    module_name: module_name.to_string(),
-                })
-                .w()?;
+    pub fn get_or_load_func(&self, fn_name: &str, mdl: &kernels::Module) -> Result<CudaFunc> {
+        let ms = self.modules.read().unwrap();
+        if let Some(mdl) = ms.mdls[mdl.index()].as_ref() {
+            let func = mdl.load_function(fn_name).w()?;
+            return Ok(CudaFunc {
+                func,
+                stream: self.stream.clone(),
+            });
         }
-        self.get_func(module_name, module_name)
-            // Clippy recommends this `ok_or` rather than `ok_or_else` so hopefully the compiler is
-            // able to only build the error value if needed.
-            .ok_or(CudaError::MissingKernel {
-                module_name: module_name.to_string(),
-            })
-            .w()
+        drop(ms);
+        let mut ms = self.modules.write().unwrap();
+        let cuda_module = self.context.load_module(mdl.ptx().into()).w()?;
+        ms.mdls[mdl.index()] = Some(cuda_module.clone());
+        let func = cuda_module.load_function(fn_name).w()?;
+        Ok(CudaFunc {
+            func,
+            stream: self.stream.clone(),
+        })
+    }
+}
+
+impl CudaDevice {
+    pub fn new_with_stream(ordinal: usize) -> Result<Self> {
+        let context = cudarc::driver::CudaContext::new(ordinal).w()?;
+        let stream = context.new_stream().w()?;
+        let blas = cudarc::cublas::CudaBlas::new(stream.clone()).w()?;
+        let curand = cudarc::curand::CudaRng::new(299792458, stream.clone()).w()?;
+        let module_store = ModuleStore {
+            mdls: [const { None }; kernels::ALL_IDS.len()],
+        };
+        Ok(Self {
+            id: DeviceId::new(),
+            context,
+            stream,
+            blas: Arc::new(blas),
+            curand: Arc::new(Mutex::new(CudaRng(curand))),
+            modules: Arc::new(std::sync::RwLock::new(module_store)),
+            custom_modules: Arc::new(std::sync::RwLock::new(HashMap::new())),
+        })
     }
 }
 
@@ -148,14 +259,21 @@ impl BackendDevice for CudaDevice {
     type Storage = CudaStorage;
 
     fn new(ordinal: usize) -> Result<Self> {
-        let device = cudarc::driver::CudaDevice::new(ordinal).w()?;
-        let blas = cudarc::cublas::CudaBlas::new(device.clone()).w()?;
-        let curand = cudarc::curand::CudaRng::new(299792458, device.clone()).w()?;
+        let context = cudarc::driver::CudaContext::new(ordinal).w()?;
+        let stream = context.default_stream();
+        let blas = cudarc::cublas::CudaBlas::new(stream.clone()).w()?;
+        let curand = cudarc::curand::CudaRng::new(299792458, stream.clone()).w()?;
+        let module_store = ModuleStore {
+            mdls: [const { None }; kernels::ALL_IDS.len()],
+        };
         Ok(Self {
             id: DeviceId::new(),
-            device,
+            context,
+            stream,
             blas: Arc::new(blas),
             curand: Arc::new(Mutex::new(CudaRng(curand))),
+            modules: Arc::new(std::sync::RwLock::new(module_store)),
+            custom_modules: Arc::new(std::sync::RwLock::new(HashMap::new())),
         })
     }
 
@@ -163,13 +281,13 @@ impl BackendDevice for CudaDevice {
         // 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()?;
+        curand.0 = cudarc::curand::CudaRng::new(seed, self.stream.clone()).w()?;
         Ok(())
     }
 
     fn location(&self) -> crate::DeviceLocation {
         crate::DeviceLocation::Cuda {
-            gpu_id: self.device.ordinal(),
+            gpu_id: self.context.ordinal(),
         }
     }
 
@@ -181,31 +299,31 @@ impl BackendDevice for CudaDevice {
         let elem_count = shape.elem_count();
         let slice = match dtype {
             DType::U8 => {
-                let data = self.alloc_zeros::<u8>(elem_count).w()?;
+                let data = self.alloc_zeros::<u8>(elem_count)?;
                 CudaStorageSlice::U8(data)
             }
             DType::U32 => {
-                let data = self.alloc_zeros::<u32>(elem_count).w()?;
+                let data = self.alloc_zeros::<u32>(elem_count)?;
                 CudaStorageSlice::U32(data)
             }
             DType::I64 => {
-                let data = self.alloc_zeros::<i64>(elem_count).w()?;
+                let data = self.alloc_zeros::<i64>(elem_count)?;
                 CudaStorageSlice::I64(data)
             }
             DType::BF16 => {
-                let data = self.alloc_zeros::<bf16>(elem_count).w()?;
+                let data = self.alloc_zeros::<bf16>(elem_count)?;
                 CudaStorageSlice::BF16(data)
             }
             DType::F16 => {
-                let data = self.alloc_zeros::<f16>(elem_count).w()?;
+                let data = self.alloc_zeros::<f16>(elem_count)?;
                 CudaStorageSlice::F16(data)
             }
             DType::F32 => {
-                let data = self.alloc_zeros::<f32>(elem_count).w()?;
+                let data = self.alloc_zeros::<f32>(elem_count)?;
                 CudaStorageSlice::F32(data)
             }
             DType::F64 => {
-                let data = self.alloc_zeros::<f64>(elem_count).w()?;
+                let data = self.alloc_zeros::<f64>(elem_count)?;
                 CudaStorageSlice::F64(data)
             }
         };
@@ -229,12 +347,12 @@ impl BackendDevice for CudaDevice {
                 .w()?
             }
             DType::F32 => {
-                let mut data = unsafe { self.alloc::<f32>(elem_count) }.w()?;
+                let mut data = unsafe { self.alloc::<f32>(elem_count)? };
                 curand.0.fill_with_uniform(&mut data).w()?;
                 CudaStorageSlice::F32(data)
             }
             DType::F64 => {
-                let mut data = unsafe { self.alloc::<f64>(elem_count) }.w()?;
+                let mut data = unsafe { self.alloc::<f64>(elem_count)? };
                 curand.0.fill_with_uniform(&mut data).w()?;
                 CudaStorageSlice::F64(data)
             }
@@ -273,7 +391,7 @@ impl BackendDevice for CudaDevice {
                 .w()?
             }
             DType::F32 => {
-                let mut data = unsafe { self.alloc::<f32>(elem_count_round) }.w()?;
+                let mut data = unsafe { self.alloc::<f32>(elem_count_round)? };
                 curand
                     .0
                     .fill_with_normal(&mut data, mean as f32, std as f32)
@@ -281,7 +399,7 @@ impl BackendDevice for CudaDevice {
                 CudaStorageSlice::F32(data)
             }
             DType::F64 => {
-                let mut data = unsafe { self.alloc::<f64>(elem_count_round) }.w()?;
+                let mut data = unsafe { self.alloc::<f64>(elem_count_round)? };
                 curand.0.fill_with_normal(&mut data, mean, std).w()?;
                 CudaStorageSlice::F64(data)
             }
@@ -292,39 +410,72 @@ impl BackendDevice for CudaDevice {
         })
     }
 
-    fn ones_impl(&self, shape: &Shape, dtype: DType) -> Result<CudaStorage> {
-        self.const_impl(1., shape, dtype)
-    }
-
     unsafe fn alloc_uninit(&self, shape: &Shape, dtype: DType) -> Result<Self::Storage> {
         let elem_count = shape.elem_count();
         let slice = match dtype {
             DType::U8 => {
-                let data = self.alloc::<u8>(elem_count).w()?;
+                let data = self.alloc::<u8>(elem_count)?;
                 CudaStorageSlice::U8(data)
             }
             DType::U32 => {
-                let data = self.alloc::<u32>(elem_count).w()?;
+                let data = self.alloc::<u32>(elem_count)?;
                 CudaStorageSlice::U32(data)
             }
             DType::I64 => {
-                let data = self.alloc::<i64>(elem_count).w()?;
+                let data = self.alloc::<i64>(elem_count)?;
                 CudaStorageSlice::I64(data)
             }
             DType::BF16 => {
-                let data = self.alloc::<bf16>(elem_count).w()?;
+                let data = self.alloc::<bf16>(elem_count)?;
                 CudaStorageSlice::BF16(data)
             }
             DType::F16 => {
-                let data = self.alloc::<f16>(elem_count).w()?;
+                let data = self.alloc::<f16>(elem_count)?;
                 CudaStorageSlice::F16(data)
             }
             DType::F32 => {
-                let data = self.alloc::<f32>(elem_count).w()?;
+                let data = self.alloc::<f32>(elem_count)?;
                 CudaStorageSlice::F32(data)
             }
             DType::F64 => {
-                let data = self.alloc::<f64>(elem_count).w()?;
+                let data = self.alloc::<f64>(elem_count)?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
+    fn storage_from_slice<T: crate::WithDType>(&self, s: &[T]) -> Result<Self::Storage> {
+        let slice = match T::cpu_storage_ref(s) {
+            CpuStorageRef::U8(storage) => {
+                let data = self.memcpy_stod(storage)?;
+                CudaStorageSlice::U8(data)
+            }
+            CpuStorageRef::U32(storage) => {
+                let data = self.memcpy_stod(storage)?;
+                CudaStorageSlice::U32(data)
+            }
+            CpuStorageRef::I64(storage) => {
+                let data = self.memcpy_stod(storage)?;
+                CudaStorageSlice::I64(data)
+            }
+            CpuStorageRef::BF16(storage) => {
+                let data = self.memcpy_stod(storage)?;
+                CudaStorageSlice::BF16(data)
+            }
+            CpuStorageRef::F16(storage) => {
+                let data = self.memcpy_stod(storage)?;
+                CudaStorageSlice::F16(data)
+            }
+            CpuStorageRef::F32(storage) => {
+                let data = self.memcpy_stod(storage)?;
+                CudaStorageSlice::F32(data)
+            }
+            CpuStorageRef::F64(storage) => {
+                let data = self.memcpy_stod(storage)?;
                 CudaStorageSlice::F64(data)
             }
         };
@@ -337,31 +488,31 @@ impl BackendDevice for CudaDevice {
     fn storage_from_cpu_storage(&self, storage: &CpuStorage) -> Result<CudaStorage> {
         let slice = match storage {
             CpuStorage::U8(storage) => {
-                let data = self.htod_sync_copy(storage).w()?;
+                let data = self.memcpy_stod(storage)?;
                 CudaStorageSlice::U8(data)
             }
             CpuStorage::U32(storage) => {
-                let data = self.htod_sync_copy(storage).w()?;
+                let data = self.memcpy_stod(storage)?;
                 CudaStorageSlice::U32(data)
             }
             CpuStorage::I64(storage) => {
-                let data = self.htod_sync_copy(storage).w()?;
+                let data = self.memcpy_stod(storage)?;
                 CudaStorageSlice::I64(data)
             }
             CpuStorage::BF16(storage) => {
-                let data = self.htod_sync_copy(storage).w()?;
+                let data = self.memcpy_stod(storage)?;
                 CudaStorageSlice::BF16(data)
             }
             CpuStorage::F16(storage) => {
-                let data = self.htod_sync_copy(storage).w()?;
+                let data = self.memcpy_stod(storage)?;
                 CudaStorageSlice::F16(data)
             }
             CpuStorage::F32(storage) => {
-                let data = self.htod_sync_copy(storage).w()?;
+                let data = self.memcpy_stod(storage)?;
                 CudaStorageSlice::F32(data)
             }
             CpuStorage::F64(storage) => {
-                let data = self.htod_sync_copy(storage).w()?;
+                let data = self.memcpy_stod(storage)?;
                 CudaStorageSlice::F64(data)
             }
         };
@@ -374,31 +525,31 @@ impl BackendDevice for CudaDevice {
     fn storage_from_cpu_storage_owned(&self, storage: CpuStorage) -> Result<CudaStorage> {
         let slice = match storage {
             CpuStorage::U8(storage) => {
-                let data = self.htod_copy(storage).w()?;
+                let data = self.memcpy_stod(&storage)?;
                 CudaStorageSlice::U8(data)
             }
             CpuStorage::U32(storage) => {
-                let data = self.htod_copy(storage).w()?;
+                let data = self.memcpy_stod(&storage)?;
                 CudaStorageSlice::U32(data)
             }
             CpuStorage::I64(storage) => {
-                let data = self.htod_copy(storage).w()?;
+                let data = self.memcpy_stod(&storage)?;
                 CudaStorageSlice::I64(data)
             }
             CpuStorage::BF16(storage) => {
-                let data = self.htod_copy(storage).w()?;
+                let data = self.memcpy_stod(&storage)?;
                 CudaStorageSlice::BF16(data)
             }
             CpuStorage::F16(storage) => {
-                let data = self.htod_copy(storage).w()?;
+                let data = self.memcpy_stod(&storage)?;
                 CudaStorageSlice::F16(data)
             }
             CpuStorage::F32(storage) => {
-                let data = self.htod_copy(storage).w()?;
+                let data = self.memcpy_stod(&storage)?;
                 CudaStorageSlice::F32(data)
             }
             CpuStorage::F64(storage) => {
-                let data = self.htod_copy(storage).w()?;
+                let data = self.memcpy_stod(&storage)?;
                 CudaStorageSlice::F64(data)
             }
         };
@@ -407,4 +558,9 @@ impl BackendDevice for CudaDevice {
             device: self.clone(),
         })
     }
+
+    fn synchronize(&self) -> Result<()> {
+        self.stream.synchronize().map_err(crate::Error::wrap)?;
+        Ok(())
+    }
 }

candle-core/src/cuda_backend/utils.rs

@@ -1,5 +1,5 @@
 /// Helper functions to plug cuda kernels in candle.
-use crate::{Layout, Result, Shape, WithDType};
+use crate::{Layout, Result, WithDType};
 pub use cudarc;
 use cudarc::driver::{CudaSlice, DeviceRepr, ValidAsZeroBits};
 
@@ -54,11 +54,49 @@ pub trait Map2 {
     }
 }
 
+pub trait Map3 {
+    #[allow(clippy::too_many_arguments)]
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        src1: &CudaSlice<T>,
+        layout1: &Layout,
+        src2: &CudaSlice<T>,
+        layout2: &Layout,
+        src3: &CudaSlice<T>,
+        layout3: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<CudaSlice<T>>;
+
+    #[allow(clippy::too_many_arguments)]
+    fn map(
+        &self,
+        s1: &S,
+        l1: &Layout,
+        s2: &S,
+        l2: &Layout,
+        s3: &S,
+        l3: &Layout,
+        d: &CudaDevice,
+    ) -> Result<S> {
+        let out = match (s1, s2, s3) {
+            (S::U8(s1), S::U8(s2), S::U8(s3)) => S::U8(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::U32(s1), S::U32(s2), S::U32(s3)) => S::U32(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::I64(s1), S::I64(s2), S::I64(s3)) => S::I64(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::BF16(s1), S::BF16(s2), S::BF16(s3)) => S::BF16(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::F16(s1), S::F16(s2), S::F16(s3)) => S::F16(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::F32(s1), S::F32(s2), S::F32(s3)) => S::F32(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::F64(s1), S::F64(s2), S::F64(s3)) => S::F64(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            _ => Err(CudaError::InternalError("dtype mismatch in ternary op"))?,
+        };
+        Ok(out)
+    }
+}
+
 pub trait Map2InPlace {
     fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
         &self,
         dst: &mut CudaSlice<T>,
-        dst_shape: &Shape,
+        dst_l: &Layout,
         src: &CudaSlice<T>,
         src_l: &Layout,
         dev: &CudaDevice,
@@ -67,19 +105,19 @@ pub trait Map2InPlace {
     fn map(
         &self,
         dst: &mut S,
-        dst_s: &Shape,
+        dst_l: &Layout,
         src: &S,
         src_l: &Layout,
         d: &CudaDevice,
     ) -> Result<()> {
         match (dst, src) {
-            (S::U8(dst), S::U8(src)) => self.f(dst, dst_s, src, src_l, d),
-            (S::U32(dst), S::U32(src)) => self.f(dst, dst_s, src, src_l, d),
-            (S::I64(dst), S::I64(src)) => self.f(dst, dst_s, src, src_l, d),
-            (S::BF16(dst), S::BF16(src)) => self.f(dst, dst_s, src, src_l, d),
-            (S::F16(dst), S::F16(src)) => self.f(dst, dst_s, src, src_l, d),
-            (S::F32(dst), S::F32(src)) => self.f(dst, dst_s, src, src_l, d),
-            (S::F64(dst), S::F64(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::U8(dst), S::U8(src)) => self.f(dst, dst_l, src, src_l, d),
+            (S::U32(dst), S::U32(src)) => self.f(dst, dst_l, src, src_l, d),
+            (S::I64(dst), S::I64(src)) => self.f(dst, dst_l, src, src_l, d),
+            (S::BF16(dst), S::BF16(src)) => self.f(dst, dst_l, src, src_l, d),
+            (S::F16(dst), S::F16(src)) => self.f(dst, dst_l, src, src_l, d),
+            (S::F32(dst), S::F32(src)) => self.f(dst, dst_l, src, src_l, d),
+            (S::F64(dst), S::F64(src)) => self.f(dst, dst_l, src, src_l, d),
             _ => Err(CudaError::InternalError("dtype mismatch in binary op"))?,
         }
     }

candle-core/src/custom_op.rs

@@ -375,3 +375,116 @@ impl Tensor {
         )
     }
 }
+
+pub struct UgIOp1 {
+    name: &'static str,
+    #[cfg(feature = "cuda")]
+    func: cudarc::driver::CudaFunction,
+    #[cfg(feature = "metal")]
+    func: metal::ComputePipelineState,
+}
+
+impl UgIOp1 {
+    #[allow(unused)]
+    #[cfg(not(target_arch = "wasm32"))]
+    pub fn new(
+        name: &'static str,
+        kernel: ug::lang::ssa::Kernel,
+        device: &crate::Device,
+    ) -> Result<Self> {
+        #[cfg(feature = "cuda")]
+        {
+            let device = device.as_cuda_device()?;
+            let func = device.compile(name, kernel)?;
+            Ok(Self {
+                name,
+                func: func.into_cuda_function(),
+            })
+        }
+        #[cfg(feature = "metal")]
+        {
+            let device = device.as_metal_device()?;
+            let func = device.compile(name, kernel)?;
+            Ok(Self { name, func })
+        }
+        #[cfg(not(any(feature = "cuda", feature = "metal")))]
+        {
+            Ok(Self { name })
+        }
+    }
+}
+
+impl InplaceOp1 for UgIOp1 {
+    fn name(&self) -> &'static str {
+        self.name
+    }
+
+    fn cpu_fwd(&self, _: &mut CpuStorage, _: &Layout) -> Result<()> {
+        crate::bail!("ug ops are only supported on metal/cuda at the moment")
+    }
+
+    #[cfg(feature = "metal")]
+    fn metal_fwd(&self, sto: &mut MetalStorage, layout: &Layout) -> Result<()> {
+        use crate::backend::BackendStorage;
+        use candle_metal_kernels::utils::EncoderProvider;
+
+        let elem_count = layout.shape().elem_count();
+        if sto.dtype() != crate::DType::F32 {
+            // TODO: support more dtypes.
+            crate::bail!("input is not a f32 tensor")
+        }
+        let device = sto.device();
+        println!("here");
+        let command_buffer = device.command_buffer()?;
+        let command_buffer = &command_buffer;
+        let encoder = command_buffer.encoder();
+        let encoder = encoder.as_ref();
+        encoder.set_compute_pipeline_state(&self.func);
+        let (g, b) = if elem_count % 32 == 0 {
+            (elem_count / 32, 32)
+        } else {
+            (elem_count, 1)
+        };
+        let grid_dims = metal::MTLSize {
+            width: g as u64,
+            height: 1,
+            depth: 1,
+        };
+        let group_dims = candle_metal_kernels::utils::get_block_dims(b as u64, 1, 1);
+        candle_metal_kernels::utils::set_param(encoder, 0, (sto.buffer(), 0usize));
+
+        encoder.use_resource(sto.buffer(), metal::MTLResourceUsage::Write);
+        encoder.dispatch_threads(grid_dims, group_dims);
+
+        Ok(())
+    }
+
+    #[cfg(feature = "cuda")]
+    fn cuda_fwd(&self, sto: &mut CudaStorage, layout: &Layout) -> Result<()> {
+        use crate::cuda_backend::WrapErr;
+        use cudarc::driver::PushKernelArg;
+
+        let elem_count = layout.shape().elem_count();
+        let stream = sto.device.cuda_stream();
+        // TODO: support more dtypes.
+        let sto = sto.as_cuda_slice::<f32>()?;
+        let sto = match layout.contiguous_offsets() {
+            None => crate::bail!("input has to be contiguous"),
+            Some((o1, o2)) => sto.slice(o1..o2),
+        };
+        let (g, b) = if elem_count % 32 == 0 {
+            (elem_count / 32, 32)
+        } else {
+            (elem_count, 1)
+        };
+        let cfg = cudarc::driver::LaunchConfig {
+            grid_dim: (g as u32, 1, 1),
+            block_dim: (b as u32, 1, 1),
+            shared_mem_bytes: 0,
+        };
+        let mut builder = stream.launch_builder(&self.func);
+        builder.arg(&sto);
+        unsafe { builder.launch(cfg) }.w()?;
+        Ok(())
+    }
+}

candle-core/src/device.rs

@@ -11,6 +11,7 @@ pub enum DeviceLocation {
     Metal { gpu_id: usize },
 }
 
+/// Cpu, Cuda, or Metal
 #[derive(Debug, Clone)]
 pub enum Device {
     Cpu,
@@ -102,7 +103,63 @@ impl<S: WithDType, const N1: usize, const N2: usize, const N3: usize, const N4:
     }
 }
 
-impl<S: NdArray> NdArray for Vec<S> {
+impl<S: WithDType> NdArray for Vec<S> {
+    fn shape(&self) -> Result<Shape> {
+        Ok(Shape::from(self.len()))
+    }
+
+    fn to_cpu_storage(&self) -> CpuStorage {
+        S::to_cpu_storage(self.as_slice())
+    }
+}
+
+impl<S: WithDType> NdArray for Vec<&[S]> {
+    fn shape(&self) -> Result<Shape> {
+        if self.is_empty() {
+            crate::bail!("empty array")
+        }
+        let n = self.len();
+        let m = self[0].len();
+        for v in self.iter() {
+            if v.len() != m {
+                crate::bail!("two elements have different len {m} {}", v.len())
+            }
+        }
+        Ok(Shape::from((n, m)))
+    }
+
+    fn to_cpu_storage(&self) -> CpuStorage {
+        let data = self.iter().copied().flatten().copied().collect::<Vec<_>>();
+        S::to_cpu_storage_owned(data)
+    }
+}
+
+impl<S: WithDType> NdArray for Vec<Vec<S>> {
+    fn shape(&self) -> Result<Shape> {
+        if self.is_empty() {
+            crate::bail!("empty array")
+        }
+        let n = self.len();
+        let m = self[0].len();
+        for v in self.iter() {
+            if v.len() != m {
+                crate::bail!("two elements have different len {m} {}", v.len())
+            }
+        }
+        Ok(Shape::from((n, m)))
+    }
+
+    fn to_cpu_storage(&self) -> CpuStorage {
+        let len: usize = self.iter().map(|v| v.len()).sum();
+        let mut dst = Vec::with_capacity(len);
+        for v in self.iter() {
+            dst.extend(v.iter().copied());
+        }
+        S::to_cpu_storage_owned(dst)
+    }
+}
+
+impl<S: WithDType> NdArray for Vec<Vec<Vec<S>>> {
     fn shape(&self) -> Result<Shape> {
         if self.is_empty() {
             crate::bail!("empty array")
@@ -119,9 +176,57 @@ impl<S: NdArray> NdArray for Vec<S> {
     }
 
     fn to_cpu_storage(&self) -> CpuStorage {
-        // This allocates intermediary memory and shouldn't be necessary.
-        let storages = self.iter().map(|v| v.to_cpu_storage()).collect::<Vec<_>>();
-        CpuStorage::concat(storages.as_slice()).unwrap()
+        if self.is_empty() {
+            return S::to_cpu_storage_owned(vec![]);
+        }
+        let len: usize = self
+            .iter()
+            .map(|v| v.iter().map(|v| v.len()).sum::<usize>())
+            .sum();
+        let mut dst = Vec::with_capacity(len);
+        for v1 in self.iter() {
+            for v2 in v1.iter() {
+                dst.extend(v2.iter().copied());
+            }
+        }
+        S::to_cpu_storage_owned(dst)
+    }
+}
+
+impl<S: WithDType> NdArray for Vec<Vec<Vec<Vec<S>>>> {
+    fn shape(&self) -> Result<Shape> {
+        if self.is_empty() {
+            crate::bail!("empty array")
+        }
+        let shape0 = self[0].shape()?;
+        let n = self.len();
+        for v in self.iter() {
+            let shape = v.shape()?;
+            if shape != shape0 {
+                crate::bail!("two elements have different shapes {shape:?} {shape0:?}")
+            }
+        }
+        Ok(Shape::from([[n].as_slice(), shape0.dims()].concat()))
+    }
+
+    fn to_cpu_storage(&self) -> CpuStorage {
+        let len: usize = self
+            .iter()
+            .map(|v| {
+                v.iter()
+                    .map(|v| v.iter().map(|v| v.len()).sum::<usize>())
+                    .sum::<usize>()
+            })
+            .sum();
+        let mut dst = Vec::with_capacity(len);
+        for v1 in self.iter() {
+            for v2 in v1.iter() {
+                for v3 in v2.iter() {
+                    dst.extend(v3.iter().copied());
+                }
+            }
+        }
+        S::to_cpu_storage_owned(dst)
     }
 }
 
@@ -130,6 +235,26 @@ impl Device {
         Ok(Self::Cuda(crate::CudaDevice::new(ordinal)?))
     }
 
+    pub fn as_cuda_device(&self) -> Result<&crate::CudaDevice> {
+        match self {
+            Self::Cuda(d) => Ok(d),
+            Self::Cpu => crate::bail!("expected a cuda device, got cpu"),
+            Self::Metal(_) => crate::bail!("expected a cuda device, got Metal"),
+        }
+    }
+
+    pub fn as_metal_device(&self) -> Result<&crate::MetalDevice> {
+        match self {
+            Self::Cuda(_) => crate::bail!("expected a metal device, got cuda"),
+            Self::Cpu => crate::bail!("expected a metal device, got cpu"),
+            Self::Metal(d) => Ok(d),
+        }
+    }
+
+    pub fn new_cuda_with_stream(ordinal: usize) -> Result<Self> {
+        Ok(Self::Cuda(crate::CudaDevice::new_with_stream(ordinal)?))
+    }
+
     pub fn new_metal(ordinal: usize) -> Result<Self> {
         Ok(Self::Metal(crate::MetalDevice::new(ordinal)?))
     }
@@ -171,6 +296,22 @@ impl Device {
         matches!(self, Self::Metal(_))
     }
 
+    pub fn supports_bf16(&self) -> bool {
+        match self {
+            Self::Cuda(_) | Self::Metal(_) => true,
+            Self::Cpu => false,
+        }
+    }
+
+    /// Return `BF16` for devices that support it, otherwise default to `F32`.
+    pub fn bf16_default_to_f32(&self) -> DType {
+        if self.supports_bf16() {
+            DType::BF16
+        } else {
+            DType::F32
+        }
+    }
+
     pub fn cuda_if_available(ordinal: usize) -> Result<Self> {
         if crate::utils::cuda_is_available() {
             Self::new_cuda(ordinal)
@@ -255,23 +396,6 @@ impl Device {
         self.rand_normal_f64(mean.to_f64(), std.to_f64(), shape, T::DTYPE)
     }
 
-    pub(crate) fn ones(&self, shape: &Shape, dtype: DType) -> Result<Storage> {
-        match self {
-            Device::Cpu => {
-                let storage = CpuDevice.ones_impl(shape, dtype)?;
-                Ok(Storage::Cpu(storage))
-            }
-            Device::Cuda(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))
-            }
-        }
-    }
-
     pub(crate) fn zeros(&self, shape: &Shape, dtype: DType) -> Result<Storage> {
         match self {
             Device::Cpu => {
@@ -306,6 +430,20 @@ impl Device {
         }
     }
 
+    pub(crate) fn storage_from_slice<D: WithDType>(&self, data: &[D]) -> Result<Storage> {
+        match self {
+            Device::Cpu => Ok(Storage::Cpu(data.to_cpu_storage())),
+            Device::Cuda(device) => {
+                let storage = device.storage_from_slice(data)?;
+                Ok(Storage::Cuda(storage))
+            }
+            Device::Metal(device) => {
+                let storage = device.storage_from_slice(data)?;
+                Ok(Storage::Metal(storage))
+            }
+        }
+    }
+
     pub(crate) fn storage<A: NdArray>(&self, array: A) -> Result<Storage> {
         match self {
             Device::Cpu => Ok(Storage::Cpu(array.to_cpu_storage())),
@@ -337,4 +475,12 @@ impl Device {
             }
         }
     }
+
+    pub fn synchronize(&self) -> Result<()> {
+        match self {
+            Self::Cpu => Ok(()),
+            Self::Cuda(d) => d.synchronize(),
+            Self::Metal(d) => d.synchronize(),
+        }
+    }
 }

candle-core/src/display.rs

@@ -1,6 +1,7 @@
-/// Pretty printing of tensors
-/// This implementation should be in line with the PyTorch version.
-/// https://github.com/pytorch/pytorch/blob/7b419e8513a024e172eae767e24ec1b849976b13/torch/_tensor_str.py
+//! Pretty printing of tensors
+//!
+//! This implementation should be in line with the [PyTorch version](https://github.com/pytorch/pytorch/blob/7b419e8513a024e172eae767e24ec1b849976b13/torch/_tensor_str.py).
+//!
 use crate::{DType, Result, Tensor, WithDType};
 use half::{bf16, f16};
 

candle-core/src/dtype.rs

@@ -1,7 +1,7 @@
 //! Types for elements that can be stored and manipulated using tensors.
 #![allow(clippy::redundant_closure_call)]
 use crate::backend::BackendStorage;
-use crate::{CpuStorage, Error, Result};
+use crate::{CpuStorage, CpuStorageRef, Error, Result};
 
 /// The different types of elements allowed in tensors.
 #[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
@@ -100,12 +100,15 @@ pub trait WithDType:
     + 'static
     + Send
     + Sync
+    + std::any::Any
     + crate::cpu::kernels::VecOps
 {
     const DTYPE: DType;
 
     fn from_f64(v: f64) -> Self;
     fn to_f64(self) -> f64;
+    fn to_scalar(self) -> crate::scalar::Scalar;
+    fn cpu_storage_ref(data: &[Self]) -> CpuStorageRef<'_>;
     fn to_cpu_storage_owned(data: Vec<Self>) -> CpuStorage;
 
     fn to_cpu_storage(data: &[Self]) -> CpuStorage {
@@ -129,6 +132,14 @@ macro_rules! with_dtype {
                 $to_f64(self)
             }
 
+            fn to_scalar(self) -> crate::scalar::Scalar {
+                crate::scalar::Scalar::$dtype(self)
+            }
+
+            fn cpu_storage_ref(data: &[Self]) -> CpuStorageRef<'_> {
+                CpuStorageRef::$dtype(data)
+            }
+
             fn to_cpu_storage_owned(data: Vec<Self>) -> CpuStorage {
                 CpuStorage::$dtype(data)
             }

candle-core/src/dummy_cuda_backend.rs

@@ -1,3 +1,5 @@
+//! Implementation of the Cuda backend when Cuda support has not been compiled in.
+//!
 #![allow(dead_code)]
 use crate::op::{BinaryOpT, CmpOp, ReduceOp, UnaryOpT};
 use crate::{CpuStorage, DType, Error, Layout, Result, Shape};
@@ -14,6 +16,12 @@ macro_rules! fail {
     };
 }
 
+impl CudaDevice {
+    pub fn new_with_stream(_: usize) -> Result<Self> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+}
+
 impl crate::backend::BackendStorage for CudaStorage {
     type Device = CudaDevice;
 
@@ -29,6 +37,10 @@ impl crate::backend::BackendStorage for CudaStorage {
         fail!()
     }
 
+    fn const_set(&mut self, _: crate::scalar::Scalar, _: &Layout) -> Result<()> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn to_cpu_storage(&self) -> Result<CpuStorage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
@@ -116,15 +128,27 @@ impl crate::backend::BackendStorage for CudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
-    fn scatter_add(
-        &self,
+    fn scatter_set(
+        &mut self,
         _: &Layout,
         _: &Self,
         _: &Layout,
         _: &Self,
         _: &Layout,
         _: usize,
-    ) -> Result<Self> {
+    ) -> Result<()> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    fn scatter_add_set(
+        &mut self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: usize,
+    ) -> Result<()> {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
@@ -206,11 +230,11 @@ impl crate::backend::BackendDevice for CudaDevice {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
-    fn ones_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
-    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
@@ -229,4 +253,38 @@ impl crate::backend::BackendDevice for CudaDevice {
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
 }
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn gemm_reduced_precision_f16() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn set_gemm_reduced_precision_f16(_: bool) {}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn gemm_reduced_precision_bf16() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn set_gemm_reduced_precision_bf16(_: bool) {}
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn gemm_reduced_precision_f32() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn set_gemm_reduced_precision_f32(_b: bool) {}

candle-core/src/dummy_metal_backend.rs

@@ -41,6 +41,10 @@ impl crate::backend::BackendStorage for MetalStorage {
         fail!()
     }
 
+    fn const_set(&mut self, _: crate::scalar::Scalar, _: &Layout) -> Result<()> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
     fn to_cpu_storage(&self) -> Result<CpuStorage> {
         Err(Error::NotCompiledWithMetalSupport)
     }
@@ -128,15 +132,27 @@ impl crate::backend::BackendStorage for MetalStorage {
         Err(Error::NotCompiledWithMetalSupport)
     }
 
-    fn scatter_add(
-        &self,
+    fn scatter_set(
+        &mut self,
         _: &Layout,
         _: &Self,
         _: &Layout,
         _: &Self,
         _: &Layout,
         _: usize,
-    ) -> Result<Self> {
+    ) -> Result<()> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
+    fn scatter_add_set(
+        &mut self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: &Self,
+        _: &Layout,
+        _: usize,
+    ) -> Result<()> {
         Err(Error::NotCompiledWithMetalSupport)
     }
 
@@ -218,11 +234,11 @@ impl crate::backend::BackendDevice for MetalDevice {
         Err(Error::NotCompiledWithMetalSupport)
     }
 
-    fn ones_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithMetalSupport)
     }
 
-    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithMetalSupport)
     }
 
@@ -241,4 +257,8 @@ impl crate::backend::BackendDevice for MetalDevice {
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithMetalSupport)
     }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
 }

candle-core/src/error.rs

@@ -1,3 +1,4 @@
+//! Candle-specific Error and Result
 use crate::{DType, DeviceLocation, Layout, MetalError, Shape};
 
 #[derive(Debug, Clone)]
@@ -8,8 +9,14 @@ pub struct MatMulUnexpectedStriding {
     pub msg: &'static str,
 }
 
+impl std::fmt::Debug for Error {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{self}")
+    }
+}
+
 /// Main library error type.
-#[derive(thiserror::Error, Debug)]
+#[derive(thiserror::Error)]
 pub enum Error {
     // === DType Errors ===
     #[error("{msg}, expected: {expected:?}, got: {got:?}")]
@@ -165,6 +172,10 @@ pub enum Error {
     #[error("Metal error {0}")]
     Metal(#[from] MetalError),
 
+    #[cfg(not(target_arch = "wasm32"))]
+    #[error(transparent)]
+    Ug(#[from] ug::Error),
+
     #[error(transparent)]
     TryFromIntError(#[from] core::num::TryFromIntError),
 
@@ -179,6 +190,10 @@ pub enum Error {
     #[error(transparent)]
     ParseInt(#[from] std::num::ParseIntError),
 
+    /// Utf8 parse error.
+    #[error(transparent)]
+    FromUtf8(#[from] std::string::FromUtf8Error),
+
     /// I/O error.
     #[error(transparent)]
     Io(#[from] std::io::Error),
@@ -191,8 +206,14 @@ pub enum Error {
     UnsupportedSafeTensorDtype(safetensors::Dtype),
 
     /// Arbitrary errors wrapping.
-    #[error(transparent)]
-    Wrapped(Box<dyn std::error::Error + Send + Sync>),
+    #[error("{0}")]
+    Wrapped(Box<dyn std::fmt::Display + Send + Sync>),
+
+    #[error("{context}\n{inner}")]
+    Context {
+        inner: Box<Self>,
+        context: Box<dyn std::fmt::Display + Send + Sync>,
+    },
 
     /// Adding path information to an error.
     #[error("path: {path:?} {inner}")]
@@ -210,19 +231,26 @@ pub enum Error {
     /// User generated error message, typically created via `bail!`.
     #[error("{0}")]
     Msg(String),
+
+    #[error("unwrap none")]
+    UnwrapNone,
 }
 
 pub type Result<T> = std::result::Result<T, Error>;
 
 impl Error {
-    pub fn wrap(err: impl std::error::Error + Send + Sync + 'static) -> Self {
+    pub fn wrap(err: impl std::fmt::Display + Send + Sync + 'static) -> Self {
         Self::Wrapped(Box::new(err)).bt()
     }
 
-    pub fn msg(err: impl std::error::Error + Send + Sync + 'static) -> Self {
+    pub fn msg(err: impl std::fmt::Display) -> Self {
         Self::Msg(err.to_string()).bt()
     }
 
+    pub fn debug(err: impl std::fmt::Debug) -> Self {
+        Self::Msg(format!("{err:?}")).bt()
+    }
+
     pub fn bt(self) -> Self {
         let backtrace = std::backtrace::Backtrace::capture();
         match backtrace.status() {
@@ -241,6 +269,13 @@ impl Error {
             path: p.as_ref().to_path_buf(),
         }
     }
+
+    pub fn context(self, c: impl std::fmt::Display + Send + Sync + 'static) -> Self {
+        Self::Context {
+            inner: Box::new(self),
+            context: Box::new(c),
+        }
+    }
 }
 
 #[macro_export]
@@ -263,3 +298,41 @@ pub fn zip<T, U>(r1: Result<T>, r2: Result<U>) -> Result<(T, U)> {
         (_, Err(e)) => Err(e),
     }
 }
+
+// Taken from anyhow.
+pub trait Context<T> {
+    /// Wrap the error value with additional context.
+    fn context<C>(self, context: C) -> Result<T>
+    where
+        C: std::fmt::Display + Send + Sync + 'static;
+
+    /// Wrap the error value with additional context that is evaluated lazily
+    /// only once an error does occur.
+    fn with_context<C, F>(self, f: F) -> Result<T>
+    where
+        C: std::fmt::Display + Send + Sync + 'static,
+        F: FnOnce() -> C;
+}
+
+impl<T> Context<T> for Option<T> {
+    fn context<C>(self, context: C) -> Result<T>
+    where
+        C: std::fmt::Display + Send + Sync + 'static,
+    {
+        match self {
+            Some(v) => Ok(v),
+            None => Err(Error::UnwrapNone.context(context).bt()),
+        }
+    }
+
+    fn with_context<C, F>(self, f: F) -> Result<T>
+    where
+        C: std::fmt::Display + Send + Sync + 'static,
+        F: FnOnce() -> C,
+    {
+        match self {
+            Some(v) => Ok(v),
+            None => Err(Error::UnwrapNone.context(f()).bt()),
+        }
+    }
+}

candle-core/src/indexer.rs

@@ -141,28 +141,117 @@ impl<T> IndexOp<T> for Tensor
 where
     T: Into<TensorIndexer>,
 {
+    ///```rust
+    /// use candle_core::{Tensor, DType, Device, IndexOp};
+    /// let a = Tensor::new(&[
+    ///     [0., 1.],
+    ///     [2., 3.],
+    ///     [4., 5.]
+    /// ], &Device::Cpu)?;
+    ///
+    /// let b = a.i(0)?;
+    /// assert_eq!(b.shape().dims(), &[2]);
+    /// assert_eq!(b.to_vec1::<f64>()?, &[0., 1.]);
+    ///
+    /// let c = a.i(..2)?;
+    /// assert_eq!(c.shape().dims(), &[2, 2]);
+    /// assert_eq!(c.to_vec2::<f64>()?, &[
+    ///     [0., 1.],
+    ///     [2., 3.]
+    /// ]);
+    ///
+    /// let d = a.i(1..)?;
+    /// assert_eq!(d.shape().dims(), &[2, 2]);
+    /// assert_eq!(d.to_vec2::<f64>()?, &[
+    ///     [2., 3.],
+    ///     [4., 5.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     fn i(&self, index: T) -> Result<Tensor, Error> {
         self.index(&[index.into()])
     }
 }
 
+impl<A> IndexOp<(A,)> for Tensor
+where
+    A: Into<TensorIndexer>,
+{
+    ///```rust
+    /// use candle_core::{Tensor, DType, Device, IndexOp};
+    /// let a = Tensor::new(&[
+    ///     [0f32, 1.],
+    ///     [2.  , 3.],
+    ///     [4.  , 5.]
+    /// ], &Device::Cpu)?;
+    ///
+    /// let b = a.i((0,))?;
+    /// assert_eq!(b.shape().dims(), &[2]);
+    /// assert_eq!(b.to_vec1::<f32>()?, &[0., 1.]);
+    ///
+    /// let c = a.i((..2,))?;
+    /// assert_eq!(c.shape().dims(), &[2, 2]);
+    /// assert_eq!(c.to_vec2::<f32>()?, &[
+    ///     [0., 1.],
+    ///     [2., 3.]
+    /// ]);
+    ///
+    /// let d = a.i((1..,))?;
+    /// assert_eq!(d.shape().dims(), &[2, 2]);
+    /// assert_eq!(d.to_vec2::<f32>()?, &[
+    ///     [2., 3.],
+    ///     [4., 5.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    fn i(&self, (a,): (A,)) -> Result<Tensor, Error> {
+        self.index(&[a.into()])
+    }
+}
+#[allow(non_snake_case)]
+impl<A, B> IndexOp<(A, B)> for Tensor
+where
+    A: Into<TensorIndexer>,
+    B: Into<TensorIndexer>,
+{
+    ///```rust
+    /// use candle_core::{Tensor, DType, Device, IndexOp};
+    /// let a = Tensor::new(&[[0f32, 1., 2.], [3., 4., 5.], [6., 7., 8.]], &Device::Cpu)?;
+    ///
+    /// let b = a.i((1, 0))?;
+    /// assert_eq!(b.to_vec0::<f32>()?, 3.);
+    ///
+    /// let c = a.i((..2, 1))?;
+    /// assert_eq!(c.shape().dims(), &[2]);
+    /// assert_eq!(c.to_vec1::<f32>()?, &[1., 4.]);
+    ///
+    /// let d = a.i((2.., ..))?;
+    /// assert_eq!(c.shape().dims(), &[2]);
+    /// assert_eq!(c.to_vec1::<f32>()?, &[1., 4.]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    fn i(&self, (a, b): (A, B)) -> Result<Tensor, Error> {
+        self.index(&[a.into(), b.into()])
+    }
+}
+
 macro_rules! index_op_tuple {
-    ($($t:ident),+) => {
+    ($doc:tt, $($t:ident),+) => {
         #[allow(non_snake_case)]
         impl<$($t),*> IndexOp<($($t,)*)> for Tensor
         where
             $($t: Into<TensorIndexer>,)*
         {
+            #[doc=$doc]
             fn i(&self, ($($t,)*): ($($t,)*)) -> Result<Tensor, Error> {
                 self.index(&[$($t.into(),)*])
             }
         }
     };
 }
-index_op_tuple!(A);
-index_op_tuple!(A, B);
-index_op_tuple!(A, B, C);
-index_op_tuple!(A, B, C, D);
-index_op_tuple!(A, B, C, D, E);
-index_op_tuple!(A, B, C, D, E, F);
-index_op_tuple!(A, B, C, D, E, F, G);
+
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C);
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C, D);
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C, D, E);
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C, D, E, F);
+index_op_tuple!("see [TensorIndex#method.i]", A, B, C, D, E, F, G);

candle-core/src/layout.rs

@@ -1,3 +1,4 @@
+//! Tensor Layouts including contiguous or sparse strides
 use crate::{Error, Result, Shape};
 
 #[derive(Debug, PartialEq, Eq, Clone)]
@@ -35,6 +36,12 @@ impl Layout {
         self.shape.dims()
     }
 
+    /// The dimension size for a specified dimension index.
+    pub fn dim<D: crate::shape::Dim>(&self, dim: D) -> Result<usize> {
+        let dim = dim.to_index(&self.shape, "dim")?;
+        Ok(self.dims()[dim])
+    }
+
     pub fn shape(&self) -> &Shape {
         &self.shape
     }

candle-core/src/lib.rs

@@ -7,8 +7,8 @@
 //!
 //! let a = Tensor::arange(0f32, 6f32, &Device::Cpu)?.reshape((2, 3))?;
 //! let b = Tensor::arange(0f32, 12f32, &Device::Cpu)?.reshape((3, 4))?;
-//!
 //! let c = a.matmul(&b)?;
+//!
 //! # Ok(())}
 //! ```
 //!
@@ -32,6 +32,20 @@
 //! Python can really add overhead in more complex workflows and the [GIL](https://www.backblaze.com/blog/the-python-gil-past-present-and-future/) is a notorious source of headaches.
 //!
 //! Rust is cool, and a lot of the HF ecosystem already has Rust crates [safetensors](https://github.com/huggingface/safetensors) and [tokenizers](https://github.com/huggingface/tokenizers)
+//!
+//! ## Other Crates
+//!
+//! Candle consists of a number of crates. This crate holds core the common data structures but you may wish
+//! to look at the docs for the other crates which can be found here:
+//!
+//! - [candle-core](https://docs.rs/candle-core/). Core Datastructures and DataTypes.
+//! - [candle-nn](https://docs.rs/candle-nn/). Building blocks for Neural Nets.
+//! - [candle-datasets](https://docs.rs/candle-datasets/). Rust access to commonly used Datasets like MNIST.
+//! - [candle-examples](https://docs.rs/candle-examples/). Examples of Candle in Use.
+//! - [candle-onnx](https://docs.rs/candle-onnx/). Loading and using ONNX models.
+//! - [candle-pyo3](https://docs.rs/candle-pyo3/). Access to Candle from Python.
+//! - [candle-transformers](https://docs.rs/candle-transformers/). Candle implemntation of many published transformer models.
+//!
 
 #[cfg(feature = "accelerate")]
 mod accelerate;
@@ -47,7 +61,7 @@ mod custom_op;
 mod device;
 pub mod display;
 mod dtype;
-mod dummy_cuda_backend;
+pub mod dummy_cuda_backend;
 mod dummy_metal_backend;
 pub mod error;
 mod indexer;
@@ -63,7 +77,9 @@ pub mod quantized;
 pub mod safetensors;
 pub mod scalar;
 pub mod shape;
+mod sort;
 mod storage;
+pub mod streaming;
 mod strided_index;
 mod tensor;
 mod tensor_cat;
@@ -74,24 +90,27 @@ mod variable;
 #[cfg(feature = "cudnn")]
 pub use cuda_backend::cudnn;
 
-pub use cpu_backend::CpuStorage;
-pub use custom_op::{CustomOp1, CustomOp2, CustomOp3, InplaceOp1, InplaceOp2, InplaceOp3};
+pub use cpu_backend::{CpuStorage, CpuStorageRef};
+pub use custom_op::{CustomOp1, CustomOp2, CustomOp3, InplaceOp1, InplaceOp2, InplaceOp3, UgIOp1};
 pub use device::{Device, DeviceLocation, NdArray};
 pub use dtype::{DType, DTypeParseError, FloatDType, IntDType, WithDType};
-pub use error::{Error, Result};
-pub use indexer::IndexOp;
+pub use error::{Context, Error, Result};
+pub use indexer::{IndexOp, TensorIndexer};
 pub use layout::Layout;
 pub use shape::{Shape, D};
 pub use storage::Storage;
+pub use streaming::{StreamTensor, StreamingBinOp, StreamingModule};
 pub use strided_index::{StridedBlocks, StridedIndex};
 pub use tensor::{Tensor, TensorId};
 pub use variable::Var;
 
 #[cfg(feature = "cuda")]
-pub use cuda_backend::{CudaDevice, CudaStorage};
+pub use cuda_backend as cuda;
 
 #[cfg(not(feature = "cuda"))]
-pub use dummy_cuda_backend::{CudaDevice, CudaStorage};
+pub use dummy_cuda_backend as cuda;
+
+pub use cuda::{CudaDevice, CudaStorage};
 
 #[cfg(feature = "metal")]
 pub use metal_backend::{MetalDevice, MetalError, MetalStorage};
@@ -121,7 +140,7 @@ impl ToUsize2 for (usize, usize) {
     }
 }
 
-// A simple trait defining a module with forward method using a single argument.
+/// Defining a module with forward method using a single argument.
 pub trait Module {
     fn forward(&self, xs: &Tensor) -> Result<Tensor>;
 }
@@ -141,8 +160,8 @@ impl<M: Module> Module for Option<&M> {
     }
 }
 
-// A trait defining a module with forward method using a single tensor argument and a flag to
-// separate the training and evaluation behaviors.
+/// A single forward method using a single 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>;
 }

candle-core/src/metal_backend/device.rs

@@ -2,9 +2,8 @@ use crate::{DType, Result};
 use candle_metal_kernels::Kernels;
 use metal::{Buffer, CommandBuffer, CommandQueue, MTLResourceOptions, NSUInteger};
 use std::collections::HashMap;
-use std::ffi::c_void;
 use std::path::Path;
-use std::sync::{Arc, Mutex, RwLock, RwLockWriteGuard};
+use std::sync::{Arc, Mutex, RwLock};
 
 use super::MetalError;
 
@@ -22,7 +21,73 @@ impl DeviceId {
 }
 
 type BufferMap = HashMap<(NSUInteger, MTLResourceOptions), Vec<Arc<Buffer>>>;
-type AllocatedBuffers = Arc<RwLock<BufferMap>>;
+pub(crate) struct Commands {
+    /// Single command queue for the entire device.
+    command_queue: CommandQueue,
+    /// One command buffer at a time.
+    /// The scheduler works by allowing multiple
+    /// [ComputeCommandEncoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc)
+    /// on a single command buffer. Using a single command buffer would be fastest on the GPU but
+    /// prevents overlapping of CPU and GPU commands (because command buffer needs to be committed
+    /// to start to work).
+    /// Despite what the documentation says, command buffers are NOT ordered. They are ordered
+    /// for their START time, but there's no guarantee that command buffer1 will finish before
+    /// command buffer2 starts (or there are metal bugs there)
+    command_buffer: CommandBuffer,
+    /// Keeps track of the current amount of compute command encoders on the current
+    /// command buffer
+    /// Arc, RwLock because of the interior mutability.
+    command_buffer_index: usize,
+    /// The maximum amount of [compute command encoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc) per [command buffer](https://developer.apple.com/documentation/metal/mtlcommandbuffer?language=objc)
+    compute_per_buffer: usize,
+}
+
+impl Commands {
+    pub(crate) fn new(command_queue: CommandQueue) -> Result<Self> {
+        let command_buffer = command_queue.new_command_buffer().to_owned();
+        command_buffer.enqueue();
+        let compute_per_buffer = match std::env::var("CANDLE_METAL_COMPUTE_PER_BUFFER") {
+            Ok(val) => val.parse()?,
+            _ => 50,
+        };
+        Ok(Self {
+            command_queue,
+            command_buffer,
+            command_buffer_index: 0,
+            compute_per_buffer,
+        })
+    }
+
+    pub fn command_buffer(&mut self) -> Result<(bool, CommandBuffer)> {
+        let mut command_buffer = self.command_buffer.to_owned();
+        let mut flushed = false;
+        if self.command_buffer_index > self.compute_per_buffer {
+            self.command_buffer.commit();
+            command_buffer = self.command_queue.new_command_buffer().to_owned();
+            self.command_buffer = command_buffer.clone();
+            self.command_buffer_index = 0;
+            flushed = true;
+        }
+        self.command_buffer_index += 1;
+        Ok((flushed, command_buffer))
+    }
+
+    pub fn wait_until_completed(&mut self) -> Result<()> {
+        match self.command_buffer.status() {
+            metal::MTLCommandBufferStatus::Committed
+            | metal::MTLCommandBufferStatus::Scheduled
+            | metal::MTLCommandBufferStatus::Completed => {
+                panic!("Already committed");
+            }
+            _ => {}
+        }
+        self.command_buffer.commit();
+        self.command_buffer.wait_until_completed();
+        self.command_buffer = self.command_queue.new_command_buffer().to_owned();
+
+        Ok(())
+    }
+}
 
 #[derive(Clone)]
 pub struct MetalDevice {
@@ -33,27 +98,8 @@ pub struct MetalDevice {
     /// Raw metal device: <https://developer.apple.com/documentation/metal/mtldevice?language=objc>
     pub(crate) device: metal::Device,
 
-    /// Single command queue for the entire device.
-    pub(crate) command_queue: CommandQueue,
-    /// One command buffer at a time.
-    /// The scheduler works by allowing multiple
-    /// [ComputeCommandEncoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc)
-    /// on a single command buffer. Using a single command buffer would be fastest on the GPU but
-    /// prevents overlapping of CPU and GPU commands (because command buffer needs to be committed
-    /// to start to work).
-    /// Despite what the documentation says, command buffers are NOT ordered. They are ordered
-    /// for their START time, but there's no guarantee that command buffer1 will finish before
-    /// command buffer2 starts (or there are metal bugs there)
-    pub(crate) command_buffer: Arc<RwLock<CommandBuffer>>,
-    /// Keeps track of the current amount of compute command encoders on the current
-    /// command buffer
-    /// Arc, RwLock because of the interior mutability.
-    pub(crate) command_buffer_index: Arc<RwLock<usize>>,
-    /// The maximum amount of [compute command encoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc) per [command buffer](https://developer.apple.com/documentation/metal/mtlcommandbuffer?language=objc)
-    pub(crate) compute_per_buffer: usize,
-    /// Simple keeper struct to keep track of the already compiled kernels so we can reuse them.
-    /// Heavily used by [`candle_metal_kernels`]
-    pub(crate) kernels: Arc<Kernels>,
+    pub(crate) commands: Arc<RwLock<Commands>>,
+
     /// Simple allocator struct.
     /// The buffers are stored in size buckets since ML tends to use similar shapes over and over.
     /// We store the buffers in [`Arc`] because it's much faster than Obj-c internal ref counting
@@ -67,7 +113,11 @@ pub struct MetalDevice {
     ///
     /// Whenever we actually allocate a new buffer, we make a full sweep to clean up unused buffers
     /// (strong_count = 1).
-    pub(crate) buffers: AllocatedBuffers,
+    pub(crate) buffers: Arc<RwLock<BufferMap>>,
+
+    /// Simple keeper struct to keep track of the already compiled kernels so we can reuse them.
+    /// Heavily used by [`candle_metal_kernels`]
+    pub(crate) kernels: Arc<Kernels>,
     /// Seed for random number generation.
     pub(crate) seed: Arc<Mutex<Buffer>>,
 }
@@ -87,6 +137,29 @@ impl std::ops::Deref for MetalDevice {
 }
 
 impl MetalDevice {
+    #[cfg(not(target_arch = "wasm32"))]
+    pub fn compile(
+        &self,
+        func_name: &'static str,
+        kernel: ug::lang::ssa::Kernel,
+    ) -> Result<metal::ComputePipelineState> {
+        let mut buf = vec![];
+        ug_metal::code_gen::gen(&mut buf, func_name, &kernel)?;
+        let metal_code = String::from_utf8(buf)?;
+        let lib = self
+            .device
+            .new_library_with_source(&metal_code, &metal::CompileOptions::new())
+            .map_err(MetalError::from)?;
+        let func = lib
+            .get_function(func_name, None)
+            .map_err(MetalError::from)?;
+        let pl = self
+            .device
+            .new_compute_pipeline_state_with_function(&func)
+            .map_err(MetalError::from)?;
+        Ok(pl)
+    }
+
     pub fn id(&self) -> DeviceId {
         self.id
     }
@@ -95,44 +168,31 @@ impl MetalDevice {
         &self.device
     }
 
-    pub fn command_queue(&self) -> &CommandQueue {
-        &self.command_queue
+    fn drop_unused_buffers(&self) -> Result<()> {
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+        for subbuffers in buffers.values_mut() {
+            let newbuffers = subbuffers
+                .iter()
+                .filter(|s| Arc::strong_count(*s) > 1)
+                .map(Arc::clone)
+                .collect();
+            *subbuffers = newbuffers;
+        }
+        Ok(())
     }
 
     pub fn command_buffer(&self) -> Result<CommandBuffer> {
-        let mut command_buffer_lock = self.command_buffer.try_write().map_err(MetalError::from)?;
-        let mut command_buffer = command_buffer_lock.to_owned();
-        let mut index = self
-            .command_buffer_index
-            .try_write()
-            .map_err(MetalError::from)?;
-        if *index > self.compute_per_buffer {
-            command_buffer.commit();
-            command_buffer = self.command_queue.new_command_buffer().to_owned();
-            *command_buffer_lock = command_buffer.clone();
-            *index = 0;
-
-            self.drop_unused_buffers()?;
+        let mut commands = self.commands.write().map_err(MetalError::from)?;
+        let (flushed, command_buffer) = commands.command_buffer()?;
+        if flushed {
+            self.drop_unused_buffers()?
         }
-        *index += 1;
         Ok(command_buffer)
     }
 
     pub fn wait_until_completed(&self) -> Result<()> {
-        let mut command_buffer = self.command_buffer.try_write().map_err(MetalError::from)?;
-        match command_buffer.status() {
-            metal::MTLCommandBufferStatus::Committed
-            | metal::MTLCommandBufferStatus::Scheduled
-            | metal::MTLCommandBufferStatus::Completed => {
-                panic!("Already committed");
-            }
-            _ => {}
-        }
-        command_buffer.commit();
-        command_buffer.wait_until_completed();
-        *command_buffer = self.command_queue.new_command_buffer().to_owned();
-
-        Ok(())
+        let mut commands = self.commands.write().map_err(MetalError::from)?;
+        commands.wait_until_completed()
     }
 
     pub fn kernels(&self) -> &Kernels {
@@ -175,11 +235,12 @@ impl MetalDevice {
     pub fn new_buffer_with_data<T>(&self, data: &[T]) -> Result<Arc<Buffer>> {
         let size = core::mem::size_of_val(data) as NSUInteger;
         let new_buffer = self.device.new_buffer_with_data(
-            data.as_ptr() as *const c_void,
+            data.as_ptr().cast(),
             size,
             MTLResourceOptions::StorageModeManaged,
         );
-        let mut buffers = self.buffers.try_write().map_err(MetalError::from)?;
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+
         let subbuffers = buffers
             .entry((size, MTLResourceOptions::StorageModeManaged))
             .or_insert(vec![]);
@@ -210,40 +271,6 @@ impl MetalDevice {
         Ok(buffer)
     }
 
-    fn find_available_buffer(
-        &self,
-        size: NSUInteger,
-        option: MTLResourceOptions,
-        buffers: &RwLockWriteGuard<BufferMap>,
-    ) -> Option<Arc<Buffer>> {
-        let mut best_buffer: Option<&Arc<Buffer>> = None;
-        let mut best_buffer_size: NSUInteger = NSUInteger::MAX;
-        for ((buffer_size, buffer_option), subbuffers) in buffers.iter() {
-            if buffer_size >= &size && buffer_size < &best_buffer_size && buffer_option == &option {
-                for sub in subbuffers {
-                    if Arc::strong_count(sub) == 1 {
-                        best_buffer = Some(sub);
-                        best_buffer_size = *buffer_size;
-                    }
-                }
-            }
-        }
-        best_buffer.cloned()
-    }
-
-    fn drop_unused_buffers(&self) -> Result<()> {
-        let mut buffers = self.buffers.try_write().map_err(MetalError::from)?;
-        for subbuffers in buffers.values_mut() {
-            let newbuffers = subbuffers
-                .iter()
-                .filter(|s| Arc::strong_count(*s) > 1)
-                .map(Arc::clone)
-                .collect();
-            *subbuffers = newbuffers;
-        }
-        Ok(())
-    }
-
     /// The critical allocator algorithm
     fn allocate_buffer(
         &self,
@@ -251,8 +278,8 @@ impl MetalDevice {
         option: MTLResourceOptions,
         _name: &str,
     ) -> Result<Arc<Buffer>> {
-        let mut buffers = self.buffers.try_write().map_err(MetalError::from)?;
-        if let Some(b) = self.find_available_buffer(size, option, &buffers) {
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+        if let Some(b) = find_available_buffer(size, option, &buffers) {
             // Cloning also ensures we increment the strong count
             return Ok(b.clone());
         }
@@ -273,7 +300,13 @@ impl MetalDevice {
         let descriptor = metal::CaptureDescriptor::new();
         descriptor.set_destination(metal::MTLCaptureDestination::GpuTraceDocument);
         descriptor.set_capture_device(self);
-        descriptor.set_output_url(path);
+        // The [set_output_url] call requires an absolute path so we convert it if needed.
+        if path.as_ref().is_absolute() {
+            descriptor.set_output_url(path);
+        } else {
+            let path = std::env::current_dir()?.join(path);
+            descriptor.set_output_url(path);
+        }
 
         capture
             .start_capture(&descriptor)
@@ -283,5 +316,25 @@ impl MetalDevice {
 }
 
 fn buf_size(size: NSUInteger) -> NSUInteger {
-    (size - 1).next_power_of_two() as NSUInteger
+    size.saturating_sub(1).next_power_of_two() as NSUInteger
+}
+
+fn find_available_buffer(
+    size: NSUInteger,
+    option: MTLResourceOptions,
+    buffers: &BufferMap,
+) -> Option<Arc<Buffer>> {
+    let mut best_buffer: Option<&Arc<Buffer>> = None;
+    let mut best_buffer_size: NSUInteger = NSUInteger::MAX;
+    for ((buffer_size, buffer_option), subbuffers) in buffers.iter() {
+        if buffer_size >= &size && buffer_size < &best_buffer_size && buffer_option == &option {
+            for sub in subbuffers {
+                if Arc::strong_count(sub) == 1 {
+                    best_buffer = Some(sub);
+                    best_buffer_size = *buffer_size;
+                }
+            }
+        }
+    }
+    best_buffer.cloned()
 }

candle-core/src/npy.rs

@@ -330,7 +330,7 @@ impl Tensor {
         path: P,
     ) -> Result<()> {
         let mut zip = zip::ZipWriter::new(File::create(path.as_ref())?);
-        let options =
+        let options: zip::write::FileOptions<()> =
             zip::write::FileOptions::default().compression_method(zip::CompressionMethod::Stored);
 
         for (name, tensor) in ts.iter() {

candle-core/src/op.rs

@@ -1,3 +1,5 @@
+//! Tensor Opertion Enums and Traits
+//!
 #![allow(clippy::redundant_closure_call)]
 use crate::Tensor;
 use half::{bf16, f16};
@@ -78,6 +80,7 @@ pub enum Op {
     Reduce(Tensor, ReduceOp, Vec<usize>),
     Matmul(Tensor, Tensor),
     Gather(Tensor, Tensor, usize),
+    Scatter(Tensor, Tensor, Tensor, usize),
     ScatterAdd(Tensor, Tensor, Tensor, usize),
     IndexSelect(Tensor, Tensor, usize),
     IndexAdd(Tensor, Tensor, Tensor, usize),

candle-core/src/pickle.rs

@@ -1,7 +1,7 @@
-// Just enough pickle support to be able to read PyTorch checkpoints.
+//! Just enough pickle support to be able to read PyTorch checkpoints.
 // This hardcodes objects that are required for tensor reading, we may want to make this a bit more
 // composable/tensor agnostic at some point.
-use crate::{DType, Error as E, Layout, Result, Tensor};
+use crate::{Context, DType, Error as E, Layout, Result, Tensor};
 use byteorder::{LittleEndian, ReadBytesExt};
 use std::collections::HashMap;
 use std::io::BufRead;
@@ -45,6 +45,7 @@ pub enum OpCode {
     BinFloat = b'G',
     Append = b'a',
     Appends = b'e',
+    Long1 = 0x8a,
 }
 
 // Avoid using FromPrimitive so as not to drag another dependency.
@@ -84,6 +85,7 @@ impl TryFrom<u8> for OpCode {
             b'G' => Ok(Self::BinFloat),
             b'a' => Ok(Self::Append),
             b'e' => Ok(Self::Appends),
+            0x8a => Ok(Self::Long1),
             value => Err(value),
         }
     }
@@ -106,6 +108,7 @@ pub enum Object {
         class_name: String,
     },
     Int(i32),
+    Long(i64),
     Float(f64),
     Unicode(String),
     Bool(bool),
@@ -170,6 +173,14 @@ impl Object {
         }
     }
 
+    pub fn int_or_long(self) -> OResult<i64> {
+        match self {
+            Self::Int(t) => Ok(t as i64),
+            Self::Long(t) => Ok(t),
+            _ => Err(self),
+        }
+    }
+
     pub fn tuple(self) -> OResult<Vec<Self>> {
         match self {
             Self::Tuple(t) => Ok(t),
@@ -537,7 +548,7 @@ impl Stack {
                         crate::bail!("setitems: not an even number of objects")
                     }
                     while let Some(value) = objs.pop() {
-                        let key = objs.pop().unwrap();
+                        let key = objs.pop().context("empty objs")?;
                         d.push((key, value))
                     }
                 } else {
@@ -557,7 +568,7 @@ impl Stack {
                     crate::bail!("setitems: not an even number of objects")
                 }
                 while let Some(value) = objs.pop() {
-                    let key = objs.pop().unwrap();
+                    let key = objs.pop().context("empty objs")?;
                     pydict.push((key, value))
                 }
                 self.push(Object::Dict(pydict))
@@ -590,6 +601,15 @@ impl Stack {
                 let obj = self.new_obj(class, args)?;
                 self.push(obj)
             }
+            OpCode::Long1 => {
+                let n_bytes = r.read_u8()?;
+                let mut v = 0;
+                // Decode the next n bytes in little endian
+                for i in 0..n_bytes {
+                    v |= (r.read_u8()? as i64) << (i * 8);
+                }
+                self.push(Object::Long(v))
+            }
         }
         Ok(false)
     }
@@ -607,10 +627,10 @@ fn rebuild_args(args: Object) -> Result<(Layout, DType, String, usize)> {
     let mut args = args.tuple()?;
     let stride = Vec::<usize>::try_from(args.remove(3))?;
     let size = Vec::<usize>::try_from(args.remove(2))?;
-    let offset = args.remove(1).int()? as usize;
+    let offset = args.remove(1).int_or_long()? as usize;
     let storage = args.remove(0).persistent_load()?;
     let mut storage = storage.tuple()?;
-    let storage_size = storage.remove(4).int()? as usize;
+    let storage_size = storage.remove(4).int_or_long()? as usize;
     let path = storage.remove(2).unicode()?;
     let (_module_name, class_name) = storage.remove(1).class()?;
     let dtype = match class_name.as_str() {
@@ -624,7 +644,11 @@ fn rebuild_args(args: Object) -> Result<(Layout, DType, String, usize)> {
             crate::bail!("unsupported storage type {other}")
         }
     };
-    let layout = Layout::new(crate::Shape::from(size), stride, offset);
+    let layout = Layout::new(
+        crate::Shape::from(size),
+        stride,
+        offset * dtype.size_in_bytes(),
+    );
     Ok((layout, dtype, path, storage_size))
 }
 
@@ -661,7 +685,7 @@ pub fn read_pth_tensor_info<P: AsRef<std::path::Path>>(
         if !file_name.ends_with("data.pkl") {
             continue;
         }
-        let dir_name = std::path::PathBuf::from(file_name.strip_suffix(".pkl").unwrap());
+        let dir_name = std::path::PathBuf::from(file_name.strip_suffix(".pkl").context("no .pkl")?);
         let reader = zip.by_name(file_name)?;
         let mut reader = std::io::BufReader::new(reader);
         let mut stack = Stack::empty();
@@ -792,7 +816,7 @@ impl PthTensors {
 /// # 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.
+///   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>,

candle-core/src/quantized/cuda.rs

@@ -1,13 +1,20 @@
 use super::{GgmlDType, QStorage};
 use crate::quantized::k_quants::GgmlType;
 use crate::{backend::BackendDevice, cuda_backend::WrapErr};
-use crate::{CudaDevice, CudaStorage, Result};
+use crate::{builder_arg as barg, CudaDevice, CudaStorage, Result};
+use half::f16;
 
-use cudarc::driver::{CudaSlice, CudaView, DeviceSlice};
+use cudarc::driver::{CudaSlice, CudaView, PushKernelArg};
+
+#[derive(Clone, Debug)]
+struct PaddedCudaSlice {
+    inner: CudaSlice<u8>,
+    len: usize,
+}
 
 #[derive(Clone, Debug)]
 pub struct QCudaStorage {
-    data: CudaSlice<u8>,
+    data: PaddedCudaSlice,
     dtype: GgmlDType,
     device: CudaDevice,
 }
@@ -29,7 +36,7 @@ pub const CUDA_DEQUANTIZE_BLOCK_SIZE: usize = 256;
 pub const MATRIX_ROW_PADDING: usize = 512;
 
 fn ceil_div(p: usize, q: usize) -> usize {
-    (p + q - 1) / q
+    p.div_ceil(q)
 }
 
 fn pad(p: usize, q: usize) -> usize {
@@ -40,59 +47,59 @@ fn quantize_q8_1(
     src: &CudaView<f32>,
     dst: &mut CudaSlice<u8>,
     elem_count: usize,
+    ky: usize,
     dev: &CudaDevice,
 ) -> Result<()> {
-    use cudarc::driver::LaunchAsync;
-
     let kx = elem_count;
     let kx_padded = pad(kx, MATRIX_ROW_PADDING);
     let num_blocks = ceil_div(kx_padded, CUDA_QUANTIZE_BLOCK_SIZE);
-    let func = dev.get_or_load_func("quantize_q8_1", candle_kernels::QUANTIZED)?;
+    let func = dev.get_or_load_func("quantize_q8_1", &candle_kernels::QUANTIZED)?;
     let cfg = cudarc::driver::LaunchConfig {
-        grid_dim: (num_blocks as u32, 1, 1),
+        grid_dim: (num_blocks as u32, ky as u32, 1),
         block_dim: (CUDA_QUANTIZE_BLOCK_SIZE as u32, 1, 1),
         shared_mem_bytes: 0,
     };
-    let params = (src, dst, kx as i32, kx_padded as i32);
-    unsafe { func.launch(cfg, params) }.w()?;
+    let mut builder = func.builder();
+    builder.arg(src);
+    builder.arg(dst);
+    barg!(builder, kx as i32, kx_padded as i32);
+    unsafe { builder.launch(cfg) }.w()?;
     Ok(())
 }
 
-fn dequantize(
-    data: &CudaSlice<u8>,
+fn dequantize_f32(
+    data: &PaddedCudaSlice,
     dtype: GgmlDType,
     elem_count: usize,
     dev: &CudaDevice,
 ) -> Result<CudaStorage> {
-    use cudarc::driver::LaunchAsync;
-
-    let nb = (elem_count + 255) / 256;
+    let nb = elem_count.div_ceil(256);
     let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
-        GgmlDType::Q4_0 => ("dequantize_block_q4_0", false, 32, nb),
-        GgmlDType::Q4_1 => ("dequantize_block_q4_1", false, 32, nb),
+        GgmlDType::Q4_0 => ("dequantize_block_q4_0_f32", false, 32, nb),
+        GgmlDType::Q4_1 => ("dequantize_block_q4_1_f32", false, 32, nb),
         GgmlDType::Q5_0 => (
-            "dequantize_block_q5_0",
+            "dequantize_block_q5_0_f32",
             false,
             CUDA_DEQUANTIZE_BLOCK_SIZE,
             ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
         ),
         GgmlDType::Q5_1 => (
-            "dequantize_block_q5_1",
+            "dequantize_block_q5_1_f32",
             false,
             CUDA_DEQUANTIZE_BLOCK_SIZE,
             ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
         ),
-        GgmlDType::Q8_0 => ("dequantize_block_q8_0", false, 32, nb),
-        GgmlDType::Q2K => ("dequantize_block_q2_K", true, 64, nb),
-        GgmlDType::Q3K => ("dequantize_block_q3_K", true, 64, nb),
-        GgmlDType::Q4K => ("dequantize_block_q4_K", true, 32, nb),
-        GgmlDType::Q5K => ("dequantize_block_q5_K", true, 64, nb),
-        GgmlDType::Q6K => ("dequantize_block_q6_K", true, 64, nb),
-        GgmlDType::Q8K => ("dequantize_block_q8_K", true, 32, nb),
+        GgmlDType::Q8_0 => ("dequantize_block_q8_0_f32", false, 32, nb),
+        GgmlDType::Q2K => ("dequantize_block_q2_K_f32", true, 64, nb),
+        GgmlDType::Q3K => ("dequantize_block_q3_K_f32", true, 64, nb),
+        GgmlDType::Q4K => ("dequantize_block_q4_K_f32", true, 32, nb),
+        GgmlDType::Q5K => ("dequantize_block_q5_K_f32", true, 64, nb),
+        GgmlDType::Q6K => ("dequantize_block_q6_K_f32", true, 64, nb),
+        GgmlDType::Q8K => ("dequantize_block_q8_K_f32", true, 32, nb),
         _ => crate::bail!("unsupported dtype for dequantize {dtype:?}"),
     };
-    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
-    let dst = unsafe { dev.alloc::<f32>(elem_count).w()? };
+    let func = dev.get_or_load_func(kernel_name, &candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(elem_count)? };
     // See e.g.
     // https://github.com/ggerganov/llama.cpp/blob/cbbd1efa06f8c09f9dff58ff9d9af509cc4c152b/ggml-cuda.cu#L7270
     let cfg = cudarc::driver::LaunchConfig {
@@ -102,30 +109,93 @@ fn dequantize(
     };
 
     if is_k {
-        let params = (data, &dst);
-        unsafe { func.launch(cfg, params) }.w()?;
+        let mut builder = func.builder();
+        builder.arg(&data.inner);
+        builder.arg(&dst);
+        unsafe { builder.launch(cfg) }.w()?;
     } else {
         let nb32 = match dtype {
             GgmlDType::Q5_0 | GgmlDType::Q5_1 => elem_count,
             _ => elem_count / 32,
         };
-        let params = (data, &dst, nb32 as i32);
-        unsafe { func.launch(cfg, params) }.w()?;
+        let mut builder = func.builder();
+        builder.arg(&data.inner);
+        builder.arg(&dst);
+        barg!(builder, nb32 as i32);
+        unsafe { builder.launch(cfg) }.w()?;
+    }
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
+fn dequantize_f16(
+    data: &PaddedCudaSlice,
+    dtype: GgmlDType,
+    elem_count: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    let nb = elem_count.div_ceil(256);
+    let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
+        GgmlDType::Q4_0 => ("dequantize_block_q4_0_f16", false, 32, nb),
+        GgmlDType::Q4_1 => ("dequantize_block_q4_1_f16", false, 32, nb),
+        GgmlDType::Q5_0 => (
+            "dequantize_block_q5_0_f16",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q5_1 => (
+            "dequantize_block_q5_1_f16",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q8_0 => ("dequantize_block_q8_0_f16", false, 32, nb),
+        GgmlDType::Q2K => ("dequantize_block_q2_K_f16", true, 64, nb),
+        GgmlDType::Q3K => ("dequantize_block_q3_K_f16", true, 64, nb),
+        GgmlDType::Q4K => ("dequantize_block_q4_K_f16", true, 32, nb),
+        GgmlDType::Q5K => ("dequantize_block_q5_K_f16", true, 64, nb),
+        GgmlDType::Q6K => ("dequantize_block_q6_K_f16", true, 64, nb),
+        GgmlDType::Q8K => ("dequantize_block_q8_K_f16", true, 32, nb),
+        _ => crate::bail!("unsupported dtype for dequantize {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, &candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f16>(elem_count)? };
+    // See e.g.
+    // https://github.com/ggerganov/llama.cpp/blob/cbbd1efa06f8c09f9dff58ff9d9af509cc4c152b/ggml-cuda.cu#L7270
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (num_blocks as u32, 1, 1),
+        block_dim: (block_dim as u32, 1, 1),
+        shared_mem_bytes: 0,
+    };
+
+    if is_k {
+        let mut builder = func.builder();
+        builder.arg(&data.inner);
+        builder.arg(&dst);
+        unsafe { builder.launch(cfg) }.w()?;
+    } else {
+        let nb32 = match dtype {
+            GgmlDType::Q5_0 | GgmlDType::Q5_1 => elem_count,
+            _ => elem_count / 32,
+        };
+        let mut builder = func.builder();
+        builder.arg(&data.inner);
+        builder.arg(&dst);
+        barg!(builder, nb32 as i32);
+        unsafe { builder.launch(cfg) }.w()?;
     }
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
 }
 
 fn dequantize_mul_mat_vec(
-    data: &CudaSlice<u8>,
+    data: &PaddedCudaSlice,
     y: &CudaView<f32>,
     dtype: GgmlDType,
     ncols: usize,
     nrows: usize,
     dev: &CudaDevice,
 ) -> Result<CudaStorage> {
-    use cudarc::driver::LaunchAsync;
-
-    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    let data_elems = data.len / dtype.type_size() * dtype.block_size();
     if data_elems < ncols * nrows {
         crate::bail!("unexpected data size {}, ncols {ncols} {nrows}", data_elems)
     }
@@ -145,8 +215,8 @@ fn dequantize_mul_mat_vec(
         GgmlDType::Q6K => "dequantize_mul_mat_vec_q6_k",
         _ => crate::bail!("unsupported dtype for quantized matmul {dtype:?}"),
     };
-    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
-    let dst = unsafe { dev.alloc::<f32>(nrows).w()? };
+    let func = dev.get_or_load_func(kernel_name, &candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(nrows)? };
     let block_num_y = ceil_div(nrows, GGML_CUDA_MMV_Y);
     let cfg = cudarc::driver::LaunchConfig {
         grid_dim: (block_num_y as u32, 1, 1),
@@ -154,33 +224,40 @@ fn dequantize_mul_mat_vec(
         shared_mem_bytes: 0,
     };
 
-    let params = (data, y, &dst, ncols as i32, nrows as i32);
-    unsafe { func.launch(cfg, params) }.w()?;
+    let mut builder = func.builder();
+    builder.arg(&data.inner);
+    builder.arg(y);
+    builder.arg(&dst);
+    barg!(builder, ncols as i32, nrows as i32);
+    unsafe { builder.launch(cfg) }.w()?;
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
 }
 
 fn mul_mat_vec_via_q8_1(
-    data: &CudaSlice<u8>,
+    data: &PaddedCudaSlice,
     y: &CudaView<f32>,
     dtype: GgmlDType,
     ncols: usize,
     nrows: usize,
+    b_size: usize,
     dev: &CudaDevice,
 ) -> Result<CudaStorage> {
-    use cudarc::driver::LaunchAsync;
-
-    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    let data_elems = data.len / dtype.type_size() * dtype.block_size();
     if data_elems < ncols * nrows {
         crate::bail!("unexpected data size {}, ncols {ncols} {nrows}", data_elems)
     }
-    if y.len() != ncols {
+    if y.len() != ncols * b_size {
         crate::bail!("unexpected y size {}, ncols {ncols} {nrows}", y.len())
     }
+    if b_size == 0 || b_size > 8 {
+        crate::bail!("only bsize between 1 and 8 are supported, got {b_size}")
+    }
     // Start by quantizing y
     let ncols_padded = pad(ncols, MATRIX_ROW_PADDING);
-    let y_size_in_bytes = ncols_padded * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
-    let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes).w()? };
-    quantize_q8_1(y, &mut y_q8_1, ncols, dev)?;
+    let y_size_in_bytes =
+        b_size * ncols_padded * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
+    let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes)? };
+    quantize_q8_1(y, &mut y_q8_1, ncols, b_size, dev)?;
 
     let kernel_name = match dtype {
         GgmlDType::Q4_0 => "mul_mat_vec_q4_0_q8_1_cuda",
@@ -195,33 +272,118 @@ fn mul_mat_vec_via_q8_1(
         GgmlDType::Q6K => "mul_mat_vec_q6_K_q8_1_cuda",
         _ => crate::bail!("unsupported dtype for quantized matmul {dtype:?}"),
     };
-    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
-    let dst = unsafe { dev.alloc::<f32>(nrows).w()? };
+    let kernel_name = format!("{kernel_name}{b_size}");
+    let func = dev.get_or_load_func(&kernel_name, &candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(nrows * b_size)? };
+    // https://github.com/ggerganov/llama.cpp/blob/facb8b56f8fd3bb10a693bf0943ae9d69d0828ef/ggml-cuda/mmvq.cu#L98
+    let (nblocks, nwarps) = match b_size {
+        1 => (nrows as u32, 4),
+        2..=4 => ((nrows as u32).div_ceil(2), 4),
+        5..=8 => ((nrows as u32).div_ceil(2), 2),
+        _ => crate::bail!("unexpected bsize {b_size}"),
+    };
     let cfg = cudarc::driver::LaunchConfig {
-        grid_dim: (nrows as u32, 1, 1),
+        grid_dim: (nblocks, 1, 1),
+        block_dim: (WARP_SIZE as u32, nwarps, 1),
+        shared_mem_bytes: 0,
+    };
+
+    let mut builder = func.builder();
+    builder.arg(&data.inner);
+    builder.arg(&y_q8_1);
+    builder.arg(&dst);
+    barg!(
+        builder,
+        /* ncols_x */ ncols as i32,
+        /* nrows_x */ nrows as i32,
+        /* nrows_y */ ncols_padded as i32,
+        /* nrows_dst */ nrows as i32
+    );
+    unsafe { builder.launch(cfg) }.w()?;
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
+#[allow(clippy::too_many_arguments)]
+fn mul_mat_via_q8_1(
+    data: &PaddedCudaSlice,
+    y: &CudaView<f32>,
+    dtype: GgmlDType,
+    x_rows: usize,
+    x_cols: usize,
+    y_rows: usize,
+    y_cols: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    let data_elems = data.len / dtype.type_size() * dtype.block_size();
+    if data_elems < x_rows * x_cols {
+        crate::bail!("unexpected lhs size {}, {x_rows} {x_cols}", data_elems)
+    }
+    if y.len() != y_rows * y_cols {
+        crate::bail!("unexpected y size {}, {y_rows} {y_cols}", y.len())
+    }
+    if x_cols != y_rows {
+        crate::bail!("unexpected x/y size {x_rows} {x_cols} {y_rows} {y_cols}")
+    }
+    let k = x_cols;
+    // Start by quantizing y
+    let k_padded = pad(k, MATRIX_ROW_PADDING);
+    let y_size_in_bytes =
+        k_padded * y_cols * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
+    let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes)? };
+    quantize_q8_1(y, &mut y_q8_1, k, y_cols, dev)?;
+
+    let (kernel_name, mmq_x, mmq_y) = match dtype {
+        GgmlDType::Q4_0 => ("mul_mat_q4_0", 64, 128),
+        GgmlDType::Q4_1 => ("mul_mat_q4_1", 64, 128),
+        GgmlDType::Q5_0 => ("mul_mat_q5_0", 128, 64),
+        GgmlDType::Q5_1 => ("mul_mat_q5_1", 128, 64),
+        GgmlDType::Q8_0 => ("mul_mat_q8_0", 128, 64),
+        GgmlDType::Q2K => ("mul_mat_q2_K", 64, 128),
+        GgmlDType::Q3K => ("mul_mat_q3_K", 128, 128),
+        GgmlDType::Q4K => ("mul_mat_q4_K", 64, 128),
+        GgmlDType::Q5K => ("mul_mat_q5_K", 64, 128),
+        GgmlDType::Q6K => ("mul_mat_q6_K", 64, 64),
+        _ => crate::bail!("unsupported dtype for quantized matmul {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, &candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(x_rows * y_cols)? };
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (
+            ceil_div(x_rows, mmq_y) as u32,
+            ceil_div(y_cols, mmq_x) as u32,
+            1,
+        ),
         block_dim: (WARP_SIZE as u32, 4, 1),
         shared_mem_bytes: 0,
     };
 
-    let params = (
-        data,
-        &y_q8_1,
-        &dst,
-        /* ncols_x */ ncols as i32,
-        /* nrows_x */ nrows as i32,
-        /* nrows_y */ ncols as i32,
-        /* nrows_dst */ nrows as i32,
+    let mut builder = func.builder();
+    builder.arg(/* vx */ &data.inner);
+    builder.arg(/* vy */ &y_q8_1);
+    builder.arg(/* dst */ &dst);
+    barg!(
+        builder,
+        /* ncols_x */ x_cols as i32,
+        /* nrows_x */ x_rows as i32,
+        /* ncols_y */ y_cols as i32,
+        /* nrows_y */ k_padded as i32,
+        /* nrows_dst */ x_rows as i32
     );
-    unsafe { func.launch(cfg, params) }.w()?;
+    unsafe { builder.launch(cfg) }.w()?;
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
 }
 
 impl QCudaStorage {
     pub fn zeros(device: &CudaDevice, el_count: usize, dtype: GgmlDType) -> Result<Self> {
         let size_in_bytes = ceil_div(el_count, dtype.block_size()) * dtype.type_size();
-        let data = device.alloc_zeros::<u8>(size_in_bytes).w()?;
+        let padded_size_in_bytes =
+            ceil_div(el_count + MATRIX_ROW_PADDING, dtype.block_size()) * dtype.type_size();
+        let inner = device.alloc_zeros::<u8>(padded_size_in_bytes)?;
         Ok(QCudaStorage {
-            data,
+            data: PaddedCudaSlice {
+                inner,
+                len: size_in_bytes,
+            },
             device: device.clone(),
             dtype,
         })
@@ -257,11 +419,13 @@ impl QCudaStorage {
                 | GgmlDType::Q8K
         );
         if fast_kernel {
-            return dequantize(&self.data, self.dtype, elem_count, self.device());
+            return dequantize_f32(&self.data, self.dtype, elem_count, self.device());
         }
         // Run the dequantization on cpu.
 
-        let buffer = self.device.dtoh_sync_copy(&self.data).w()?;
+        let buffer = self
+            .device
+            .memcpy_dtov(&self.data.inner.slice(..self.data.len))?;
         let mut out = vec![0.0; elem_count];
         let block_len = elem_count / self.dtype.block_size();
         match self.dtype {
@@ -285,12 +449,14 @@ impl QCudaStorage {
             .storage_from_cpu_storage(&crate::CpuStorage::F32(out))
     }
 
+    pub fn dequantize_f16(&self, elem_count: usize) -> Result<CudaStorage> {
+        dequantize_f16(&self.data, self.dtype, elem_count, self.device())
+    }
+
     pub fn quantize(&mut self, src: &CudaStorage) -> Result<()> {
         // Run the quantization on cpu.
         let src = match &src.slice {
-            crate::cuda_backend::CudaStorageSlice::F32(data) => {
-                self.device.dtoh_sync_copy(data).w()?
-            }
+            crate::cuda_backend::CudaStorageSlice::F32(data) => self.device.memcpy_dtov(data)?,
             _ => crate::bail!("only f32 can be quantized"),
         };
         let src_len = src.len();
@@ -298,13 +464,20 @@ impl QCudaStorage {
         let mut qcpu_storage = crate::Device::Cpu.qzeros(src_len, self.dtype)?;
         qcpu_storage.quantize(&src)?;
         let data = qcpu_storage.data()?;
-        let data = self.device.htod_sync_copy(data.as_ref()).w()?;
-        self.data = data;
+        let padded_len =
+            data.len() + MATRIX_ROW_PADDING * self.dtype.type_size() / self.dtype.block_size();
+        let mut inner = unsafe { self.device.alloc::<u8>(padded_len)? };
+        self.device
+            .memcpy_htod(data.as_ref(), &mut inner.slice_mut(..data.len()))?;
+        self.data = PaddedCudaSlice {
+            inner,
+            len: data.len(),
+        };
         Ok(())
     }
 
     pub fn storage_size_in_bytes(&self) -> usize {
-        self.data.len()
+        self.data.len
     }
 
     pub fn fwd(
@@ -313,7 +486,17 @@ impl QCudaStorage {
         storage: &CudaStorage,
         layout: &crate::Layout,
     ) -> Result<(CudaStorage, crate::Shape)> {
-        if matches!(layout.shape().dims(), [1, 1, _] | [1, _]) {
+        let max_bm = if FORCE_DMMV.load(std::sync::atomic::Ordering::Relaxed) {
+            1
+        } else {
+            8
+        };
+        let use_vec_kernel = match layout.shape().dims() {
+            [b, m, _k] => b * m <= max_bm,
+            [b, _k] => *b <= max_bm,
+            _ => false,
+        };
+        if use_vec_kernel {
             self.dequantize_matmul_vec(self_shape, storage, layout)
         } else {
             self.dequantize_matmul(self_shape, storage, layout)
@@ -334,25 +517,31 @@ impl QCudaStorage {
             Some((o1, o2)) => rhs.slice(o1..o2),
             None => Err(crate::Error::RequiresContiguous { op: "dmmv" }.bt())?,
         };
-        let (with_batch, k) = match rhs_l.shape().dims() {
-            [1, 1, k] => (true, k),
-            [1, k] => (false, k),
+        let (b_size, k) = match rhs_l.shape().dims() {
+            [b, m, k] => (b * m, *k),
+            [b, k] => (*b, *k),
             _ => crate::bail!("unexpected rhs shape in dmmv {:?}", rhs_l.shape()),
         };
-        if ncols != *k {
+        if ncols != k {
             crate::bail!("mismatch on matmul dim {self_shape:?} {:?}", rhs_l.shape())
         }
 
         let out = if FORCE_DMMV.load(std::sync::atomic::Ordering::Relaxed) {
             dequantize_mul_mat_vec(&self.data, &rhs, self.dtype, ncols, nrows, self.device())?
         } else {
-            mul_mat_vec_via_q8_1(&self.data, &rhs, self.dtype, ncols, nrows, self.device())?
-        };
-        let out_shape = if with_batch {
-            vec![1, 1, nrows]
-        } else {
-            vec![1, nrows]
+            mul_mat_vec_via_q8_1(
+                &self.data,
+                &rhs,
+                self.dtype,
+                ncols,
+                nrows,
+                b_size,
+                self.device(),
+            )?
         };
+        let mut out_shape = rhs_l.shape().dims().to_vec();
+        out_shape.pop();
+        out_shape.push(nrows);
         Ok((out, out_shape.into()))
     }
 
@@ -373,9 +562,30 @@ impl QCudaStorage {
             crate::bail!("mismatch on matmul dim {self_shape:?} {:?}", layout.shape())
         }
 
-        let data_f32 = self.dequantize(n * k)?;
-        let rhs_l = crate::Layout::new((k, n).into(), vec![1, k], 0).broadcast_as((b, k, n))?;
-        let out = storage.matmul(&data_f32, (b, m, n, k), layout, &rhs_l)?;
+        let out = if FORCE_DMMV.load(std::sync::atomic::Ordering::Relaxed) {
+            let data_f32 = self.dequantize(n * k)?;
+            let rhs_l = crate::Layout::new((k, n).into(), vec![1, k], 0).broadcast_as((b, k, n))?;
+            storage.matmul(&data_f32, (b, m, n, k), layout, &rhs_l)?
+        } else {
+            let storage = storage.as_cuda_slice::<f32>()?;
+            let storage = match layout.contiguous_offsets() {
+                Some((o1, o2)) => storage.slice(o1..o2),
+                None => Err(crate::Error::RequiresContiguous {
+                    op: "quantized-matmul",
+                }
+                .bt())?,
+            };
+            mul_mat_via_q8_1(
+                &self.data,
+                &storage,
+                self.dtype,
+                /* x_rows */ n,
+                /* x_cols */ k,
+                /* y_rows */ k,
+                /* y_cols */ b * m,
+                self.device(),
+            )?
+        };
         let mut out_shape = layout.shape().dims().to_vec();
         out_shape.pop();
         out_shape.push(n);
@@ -390,11 +600,17 @@ pub fn load_quantized<T: super::GgmlType + Send + Sync + 'static>(
     let data = unsafe {
         std::slice::from_raw_parts(data.as_ptr() as *const u8, core::mem::size_of_val(data))
     };
-    let data = device.htod_sync_copy(data).w()?;
+    let dtype = T::DTYPE;
+    let padded_len = data.len() + MATRIX_ROW_PADDING * dtype.type_size() / dtype.block_size();
+    let mut inner = unsafe { device.alloc::<u8>(padded_len)? };
+    device.memcpy_htod(data, &mut inner.slice_mut(..data.len()))?;
     Ok(QStorage::Cuda(QCudaStorage {
-        data,
+        data: PaddedCudaSlice {
+            inner,
+            len: data.len(),
+        },
         device: device.clone(),
-        dtype: T::DTYPE,
+        dtype,
     }))
 }
 
@@ -409,10 +625,10 @@ mod test {
         let el_padded = pad(el, MATRIX_ROW_PADDING);
         let y_size_in_bytes =
             el_padded * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
-        let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes).w()? };
+        let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes)? };
         let vs: Vec<f32> = (0..el).map(|v| v as f32).collect();
-        let y = dev.htod_sync_copy(&vs).w()?;
-        quantize_q8_1(&y.slice(..), &mut y_q8_1, el, &dev)?;
+        let y = dev.memcpy_stod(&vs)?;
+        quantize_q8_1(&y.slice(..), &mut y_q8_1, el, 1, &dev)?;
         Ok(())
     }
 
@@ -421,7 +637,7 @@ mod test {
         let dev = CudaDevice::new(0)?;
         let ncols = 256;
         let vs: Vec<f32> = (0..ncols).map(|v| v as f32).collect();
-        let y = dev.htod_sync_copy(&vs).w()?;
+        let y = dev.memcpy_stod(&vs)?;
         let mut xs = QCudaStorage::zeros(&dev, ncols, GgmlDType::Q4_0)?;
         xs.quantize(&CudaStorage::wrap_cuda_slice(y.clone(), dev.clone()))?;
         let cuda_storage = mul_mat_vec_via_q8_1(
@@ -430,10 +646,11 @@ mod test {
             /* dtype */ GgmlDType::Q4_0,
             /* ncols */ ncols,
             /* nrows */ 1,
+            /* b_size */ 1,
             &dev,
         )?;
         let vs = cuda_storage.as_cuda_slice::<f32>()?;
-        let vs = dev.dtoh_sync_copy(&vs.slice(..)).unwrap();
+        let vs = dev.memcpy_dtov(&vs.slice(..))?;
         assert_eq!(vs.len(), 1);
         // for n = 255, n.(n+1).(2n+1) / 6 = 5559680
         // Q8 means 1/256 precision.
@@ -448,9 +665,73 @@ mod test {
             &dev,
         )?;
         let vs = cuda_storage.as_cuda_slice::<f32>()?;
-        let vs = dev.dtoh_sync_copy(&vs.slice(..)).unwrap();
+        let vs = dev.memcpy_dtov(&vs.slice(..))?;
         assert_eq!(vs.len(), 1);
         assert_eq!(vs[0], 5561851.0);
         Ok(())
     }
+
+    #[test]
+    fn cuda_mm_q8_1() -> Result<()> {
+        let dev = CudaDevice::new(0)?;
+        let ncols = 256;
+        let vs: Vec<f32> = (0..ncols * 4).map(|v| v as f32 / 4.).collect();
+        let y = dev.memcpy_stod(&vs)?;
+        let mut xs = QCudaStorage::zeros(&dev, ncols * 4, GgmlDType::Q4_0)?;
+        xs.quantize(&CudaStorage::wrap_cuda_slice(y.clone(), dev.clone()))?;
+        let cuda_storage = mul_mat_via_q8_1(
+            &xs.data,
+            &y.slice(..),
+            /* dtype */ GgmlDType::Q4_0,
+            /* x_rows */ 4,
+            /* x_cols */ ncols,
+            /* y_rows */ ncols,
+            /* y_cols */ 4,
+            &dev,
+        )?;
+        let vs = cuda_storage.as_cuda_slice::<f32>()?;
+        let vs = dev.memcpy_dtov(&vs.slice(..))?;
+
+        /*
+           x = torch.tensor([float(v) for v in range(1024)]).reshape(4, 256)
+           x @ x.t() / 16
+        tensor([[  347480.0000,   869720.0000,  1391960.0000,  1914200.0000],
+                [  869720.0000,  2440536.0000,  4011352.0000,  5582166.5000],
+                [ 1391960.0000,  4011352.0000,  6630742.0000,  9250132.0000],
+                [ 1914200.0000,  5582166.5000,  9250132.0000, 12918099.0000]])
+                */
+        assert_eq!(vs.len(), 16);
+        assert_eq!(vs[0], 347604.0);
+        assert_eq!(vs[1], 888153.06);
+        assert_eq!(vs[4], 869780.7);
+        assert_eq!(vs[5], 2483145.0);
+        assert_eq!(vs[11], 9407368.0);
+        assert_eq!(vs[14], 9470856.0);
+        assert_eq!(vs[15], 13138824.0);
+        Ok(())
+    }
+
+    // The following test used to fail under compute-sanitizer until #2526.
+    #[test]
+    fn cuda_mm_q8_1_pad() -> Result<()> {
+        let dev = CudaDevice::new(0)?;
+        let (x_rows, ncols, y_cols) = (4, 16, 2048);
+        let vs: Vec<f32> = (0..ncols * y_cols).map(|v| v as f32 / 256.).collect();
+        let y = dev.memcpy_stod(&vs)?;
+        let mut xs = QCudaStorage::zeros(&dev, ncols * x_rows, GgmlDType::Q4_0)?;
+        xs.quantize(&CudaStorage::wrap_cuda_slice(y.clone(), dev.clone()))?;
+        let cuda_storage = mul_mat_via_q8_1(
+            &xs.data,
+            &y.slice(..),
+            /* dtype */ GgmlDType::Q4_0,
+            /* x_rows */ x_rows,
+            /* x_cols */ ncols,
+            /* y_rows */ ncols,
+            /* y_cols */ y_cols,
+            &dev,
+        )?;
+        let vs = cuda_storage.as_cuda_slice::<f32>()?;
+        let _vs = dev.memcpy_dtov(&vs.slice(..))?;
+        Ok(())
+    }
 }

candle-core/src/quantized/dummy_cuda.rs

@@ -24,6 +24,10 @@ impl QCudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    pub fn dequantize_f16(&self, _elem_count: usize) -> Result<CudaStorage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     pub fn quantize(&mut self, _src: &CudaStorage) -> Result<()> {
         Err(Error::NotCompiledWithCudaSupport)
     }

candle-core/src/quantized/ggml_file.rs

@@ -134,7 +134,7 @@ fn from_raw_data<T: super::GgmlType + Send + Sync + 'static>(
     super::QTensor::new(data, dims)
 }
 
-/// Creates a [Tensor] from a raw GGML tensor.
+/// Creates a Tensor from a raw GGML tensor.
 pub fn qtensor_from_ggml(
     ggml_dtype: GgmlDType,
     raw_data: &[u8],

candle-core/src/quantized/gguf_file.rs

@@ -1,9 +1,8 @@
-//! Support for the GGUF file format.
+//! Support for the [GGUF file format](https://github.com/philpax/ggml/blob/gguf-spec/docs/gguf.md).
 //!
-//! Spec: https://github.com/philpax/ggml/blob/gguf-spec/docs/gguf.md
 
 use super::{GgmlDType, QTensor};
-use crate::{Device, Result};
+use crate::{Context, Device, Result};
 use byteorder::{LittleEndian, ReadBytesExt, WriteBytesExt};
 use std::collections::HashMap;
 
@@ -135,7 +134,6 @@ pub enum ValueType {
     // The value is a UTF-8 non-null-terminated string, with length prepended.
     String,
     // The value is an array of other values, with the length and type prepended.
-    ///
     // Arrays can be nested, and the length of the array is the number of elements in the array, not the number of bytes.
     Array,
 }
@@ -218,10 +216,16 @@ impl Value {
         }
     }
 
+    /// This will also automatically upcast any integral types which will not truncate.
     pub fn to_u64(&self) -> Result<u64> {
         match self {
             Self::U64(v) => Ok(*v),
-            v => crate::bail!("not a u64 {v:?}"),
+            // Autoupcast cases here
+            Self::U8(v) => Ok(*v as u64),
+            Self::U16(v) => Ok(*v as u64),
+            Self::U32(v) => Ok(*v as u64),
+            Self::Bool(v) => Ok(*v as u64),
+            v => crate::bail!("not a u64 or upcastable to u64 {v:?}"),
         }
     }
 
@@ -334,7 +338,7 @@ impl Value {
                     if value_type.len() != 1 {
                         crate::bail!("multiple value-types in the same array {value_type:?}")
                     }
-                    value_type.into_iter().next().unwrap()
+                    value_type.into_iter().next().context("empty value_type")?
                 };
                 w.write_u32::<LittleEndian>(value_type.to_u32())?;
                 w.write_u64::<LittleEndian>(v.len() as u64)?;
@@ -453,7 +457,7 @@ impl Content {
             Some(Value::I32(v)) if *v >= 0 => *v as u64,
             _ => DEFAULT_ALIGNMENT,
         };
-        let tensor_data_offset = (position + alignment - 1) / alignment * alignment;
+        let tensor_data_offset = position.div_ceil(alignment) * alignment;
         Ok(Self {
             magic,
             metadata,

candle-core/src/quantized/k_quants.rs

@@ -1850,8 +1850,8 @@ pub fn matmul<T: GgmlType>(
         crate::bail!("unexpected lhs length {} {mkn:?}", lhs.len());
     }
 
-    let k_in_lhs_blocks = (k + T::BLCK_SIZE - 1) / T::BLCK_SIZE;
-    let k_in_rhs_blocks = (k + T::VecDotType::BLCK_SIZE - 1) / T::VecDotType::BLCK_SIZE;
+    let k_in_lhs_blocks = k.div_ceil(T::BLCK_SIZE);
+    let k_in_rhs_blocks = k.div_ceil(T::VecDotType::BLCK_SIZE);
     // TODO: Do not make this copy if the DotType is f32.
     // TODO: Pre-allocate this.
     let mut lhs_b = vec![T::VecDotType::zeros(); m * k_in_lhs_blocks];

candle-core/src/quantized/metal.rs

@@ -152,9 +152,9 @@ impl QMetalStorage {
         // We always use a single batch dimension and stack all the tensors in the batch on the
         // second dimension as the implementation in candle-metal-kernels doesn't handle batch
         // properly.
-        let (b, m) = match dst_shape.len() {
-            3 => (1, dst_shape[0] * dst_shape[1]),
-            2 => (1, dst_shape[0]),
+        let m = match dst_shape.len() {
+            3 => dst_shape[0] * dst_shape[1],
+            2 => dst_shape[0],
             n => crate::bail!("Invalid rank {n} for quantized matmul metal"),
         };
         let last_k = dst_shape.pop().unwrap();
@@ -166,18 +166,23 @@ impl QMetalStorage {
         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)?;
+        // In some cases it would be better to use the mm variant, though it has its drawbacks
+        // around memory alignemnt.
+        for batch_id in 0..m {
+            candle_metal_kernels::call_quantized_matmul_mv_t(
+                device.device(),
+                &command_buffer,
+                device.kernels(),
+                self.dtype.into(),
+                (1, 1, n, k),
+                storage.buffer(),
+                (layout.start_offset() + batch_id * k) * storage.dtype().size_in_bytes(),
+                &self.buffer,
+                batch_id * n * DType::F32.size_in_bytes(),
+                &dst,
+            )
+            .map_err(MetalError::from)?;
+        }
         let dst_storage = crate::MetalStorage::new(dst, device, dst_shape.elem_count(), DType::F32);
         Ok((dst_storage, dst_shape))
     }

candle-core/src/quantized/mod.rs

@@ -1,4 +1,5 @@
-use crate::{CpuStorage, Device, Result, Shape, Storage, Tensor};
+//! Code for GGML and GGUF files
+use crate::{Context, CpuStorage, DType, Device, Result, Shape, Storage, Tensor};
 use k_quants::*;
 use std::borrow::Cow;
 
@@ -360,9 +361,24 @@ impl QTensor {
     pub fn dequantize(&self, device: &Device) -> Result<Tensor> {
         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)
+        crate::tensor::from_storage(storage, self.shape.clone(), none, false).to_device(device)
+    }
+
+    pub fn dequantize_f16(&self, device: &Device) -> Result<Tensor> {
+        // In the CUDA case, we have a specialized kernel as this can be useful for volta
+        // architectures. https://github.com/huggingface/candle/issues/2136
+        match &self.storage {
+            QStorage::Cuda(s) => {
+                let s = s.dequantize_f16(self.shape.elem_count())?;
+                let none = crate::op::BackpropOp::none();
+                crate::tensor::from_storage(Storage::Cuda(s), self.shape.clone(), none, false)
+                    .to_device(device)
+            }
+            _ => {
+                let s = self.dequantize(device)?.to_dtype(crate::DType::F16)?;
+                Ok(s)
+            }
+        }
     }
 
     pub fn storage_size_in_bytes(&self) -> usize {
@@ -378,6 +394,7 @@ impl QTensor {
 pub enum QMatMul {
     QTensor(std::sync::Arc<QTensor>),
     Tensor(Tensor),
+    TensorF16(Tensor),
 }
 
 thread_local! {
@@ -391,6 +408,17 @@ thread_local! {
     }
 }
 
+thread_local! {
+    static DEQUANTIZE_ALL_F16: bool = {
+        match std::env::var("CANDLE_DEQUANTIZE_ALL_F16") {
+            Ok(s) => {
+                !s.is_empty() && s != "0"
+            },
+            Err(_) => false,
+        }
+    }
+}
+
 impl QMatMul {
     pub fn from_arc(qtensor: std::sync::Arc<QTensor>) -> Result<Self> {
         let dequantize = match qtensor.dtype() {
@@ -400,6 +428,9 @@ impl QMatMul {
         let t = if dequantize {
             let tensor = qtensor.dequantize(&qtensor.device())?;
             Self::Tensor(tensor)
+        } else if DEQUANTIZE_ALL_F16.with(|b| *b) {
+            let tensor = qtensor.dequantize_f16(&qtensor.device())?;
+            Self::TensorF16(tensor)
         } else {
             Self::QTensor(qtensor)
         };
@@ -409,6 +440,25 @@ impl QMatMul {
     pub fn from_qtensor(qtensor: QTensor) -> Result<Self> {
         Self::from_arc(std::sync::Arc::new(qtensor))
     }
+
+    pub fn dequantize_f16(&self) -> Result<Tensor> {
+        match self {
+            Self::QTensor(t) => t.dequantize_f16(&t.device()),
+            Self::Tensor(t) => t.to_dtype(DType::F16),
+            Self::TensorF16(t) => Ok(t.clone()),
+        }
+    }
+
+    pub fn forward_via_f16(&self, xs: &Tensor) -> Result<Tensor> {
+        let w = self.dequantize_f16()?;
+        let in_dtype = xs.dtype();
+        let w = match *xs.dims() {
+            [b1, b2, _, _] => w.broadcast_left((b1, b2))?.t()?,
+            [bsize, _, _] => w.broadcast_left(bsize)?.t()?,
+            _ => w.t()?,
+        };
+        xs.to_dtype(DType::F16)?.matmul(&w)?.to_dtype(in_dtype)
+    }
 }
 
 impl crate::CustomOp1 for QTensor {
@@ -431,7 +481,7 @@ impl crate::CustomOp1 for QTensor {
             crate::bail!("input tensor has only one dimension {layout:?}")
         }
         let mut dst_shape = src_shape.dims().to_vec();
-        let last_k = dst_shape.pop().unwrap();
+        let last_k = dst_shape.pop().context("empty dst_shape")?;
         if last_k != k {
             crate::bail!("input tensor {layout:?} incompatible with {:?}", self.shape)
         }
@@ -486,6 +536,15 @@ impl crate::Module for QMatMul {
                 };
                 xs.matmul(&w)
             }
+            Self::TensorF16(w) => {
+                let in_dtype = xs.dtype();
+                let w = match *xs.dims() {
+                    [b1, b2, _, _] => w.broadcast_left((b1, b2))?.t()?,
+                    [bsize, _, _] => w.broadcast_left(bsize)?.t()?,
+                    _ => w.t()?,
+                };
+                xs.to_dtype(DType::F16)?.matmul(&w)?.to_dtype(in_dtype)
+            }
         }
     }
 }

candle-core/src/safetensors.rs

@@ -1,3 +1,14 @@
+//! Module to load `safetensor` files into CPU/GPU memory.
+//!
+//! There are multiple ways to load tensors from safetensor files:
+//! - `load` function for loading directly into memory and returning a HashMap of tensors
+//! - `MmapedSafetensors` for memory mapping files and avoiding full allocation
+//! - `SliceSafetensors` for working with in-memory buffers
+//! - `BufferedSafetensors` for owning a buffer of data
+//!
+//! Tensors can also be serialized to safetensor format using the `save` function or
+//! `Tensor::save_safetensors` method.
+//!
 use crate::{DType, Device, Error, Result, Tensor, WithDType};
 use safetensors::tensor as st;
 use safetensors::tensor::SafeTensors;
@@ -171,7 +182,7 @@ pub trait Load {
     fn load(&self, device: &Device) -> Result<Tensor>;
 }
 
-impl<'a> Load for st::TensorView<'a> {
+impl Load for st::TensorView<'_> {
     fn load(&self, device: &Device) -> Result<Tensor> {
         convert(self, device)
     }
@@ -349,6 +360,30 @@ impl MmapedSafetensors {
     }
 }
 
+pub struct SliceSafetensors<'a> {
+    safetensors: SafeTensors<'a>,
+}
+
+impl<'a> SliceSafetensors<'a> {
+    /// Creates a wrapper around a binary buffer and deserialize the safetensors header.
+    pub fn new(buffer: &'a [u8]) -> Result<Self> {
+        let safetensors = safetensors::SafeTensors::deserialize(buffer)?;
+        Ok(Self { safetensors })
+    }
+
+    pub fn load(&self, name: &str, dev: &Device) -> Result<Tensor> {
+        self.safetensors.tensor(name)?.load(dev)
+    }
+
+    pub fn tensors(&self) -> Vec<(String, st::TensorView<'_>)> {
+        self.safetensors.tensors()
+    }
+
+    pub fn get(&self, name: &str) -> Result<st::TensorView<'_>> {
+        Ok(self.safetensors.tensor(name)?)
+    }
+}
+
 pub struct BufferedSafetensors {
     safetensors: yoke::Yoke<SafeTensors_<'static>, Vec<u8>>,
 }

candle-core/src/scalar.rs

@@ -1,4 +1,74 @@
-use crate::{Result, Tensor, WithDType};
+//! TensorScalar Enum and Trait
+//!
+use crate::{DType, Result, Tensor, WithDType};
+use half::{bf16, f16};
+
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub enum Scalar {
+    U8(u8),
+    U32(u32),
+    I64(i64),
+    BF16(bf16),
+    F16(f16),
+    F32(f32),
+    F64(f64),
+}
+
+impl<T: WithDType> From<T> for Scalar {
+    fn from(value: T) -> Self {
+        value.to_scalar()
+    }
+}
+
+impl Scalar {
+    pub fn zero(dtype: DType) -> Self {
+        match dtype {
+            DType::U8 => Scalar::U8(0),
+            DType::U32 => Scalar::U32(0),
+            DType::I64 => Scalar::I64(0),
+            DType::BF16 => Scalar::BF16(bf16::ZERO),
+            DType::F16 => Scalar::F16(f16::ZERO),
+            DType::F32 => Scalar::F32(0.0),
+            DType::F64 => Scalar::F64(0.0),
+        }
+    }
+
+    pub fn one(dtype: DType) -> Self {
+        match dtype {
+            DType::U8 => Scalar::U8(1),
+            DType::U32 => Scalar::U32(1),
+            DType::I64 => Scalar::I64(1),
+            DType::BF16 => Scalar::BF16(bf16::ONE),
+            DType::F16 => Scalar::F16(f16::ONE),
+            DType::F32 => Scalar::F32(1.0),
+            DType::F64 => Scalar::F64(1.0),
+        }
+    }
+
+    pub fn dtype(&self) -> DType {
+        match self {
+            Scalar::U8(_) => DType::U8,
+            Scalar::U32(_) => DType::U32,
+            Scalar::I64(_) => DType::I64,
+            Scalar::BF16(_) => DType::BF16,
+            Scalar::F16(_) => DType::F16,
+            Scalar::F32(_) => DType::F32,
+            Scalar::F64(_) => DType::F64,
+        }
+    }
+
+    pub fn to_f64(&self) -> f64 {
+        match self {
+            Scalar::U8(v) => *v as f64,
+            Scalar::U32(v) => *v as f64,
+            Scalar::I64(v) => *v as f64,
+            Scalar::BF16(v) => v.to_f64(),
+            Scalar::F16(v) => v.to_f64(),
+            Scalar::F32(v) => *v as f64,
+            Scalar::F64(v) => *v,
+        }
+    }
+}
 
 pub enum TensorScalar {
     Tensor(Tensor),

candle-core/src/shape.rs

@@ -43,43 +43,22 @@ impl From<usize> for Shape {
     }
 }
 
-impl From<(usize,)> for Shape {
-    fn from(d1: (usize,)) -> Self {
-        Self(vec![d1.0])
+macro_rules! impl_from_tuple {
+    ($tuple:ty, $($index:tt),+) => {
+        impl From<$tuple> for Shape {
+            fn from(d: $tuple) -> Self {
+                Self(vec![$(d.$index,)+])
+            }
+        }
     }
 }
 
-impl From<(usize, usize)> for Shape {
-    fn from(d12: (usize, usize)) -> Self {
-        Self(vec![d12.0, d12.1])
-    }
-}
-
-impl From<(usize, usize, usize)> for Shape {
-    fn from(d123: (usize, usize, usize)) -> Self {
-        Self(vec![d123.0, d123.1, d123.2])
-    }
-}
-
-impl From<(usize, usize, usize, usize)> for Shape {
-    fn from(d1234: (usize, usize, usize, usize)) -> Self {
-        Self(vec![d1234.0, d1234.1, d1234.2, d1234.3])
-    }
-}
-
-impl From<(usize, usize, usize, usize, usize)> for Shape {
-    fn from(d12345: (usize, usize, usize, usize, usize)) -> Self {
-        Self(vec![d12345.0, d12345.1, d12345.2, d12345.3, d12345.4])
-    }
-}
-
-impl From<(usize, usize, usize, usize, usize, usize)> for Shape {
-    fn from(d123456: (usize, usize, usize, usize, usize, usize)) -> Self {
-        Self(vec![
-            d123456.0, d123456.1, d123456.2, d123456.3, d123456.4, d123456.5,
-        ])
-    }
-}
+impl_from_tuple!((usize,), 0);
+impl_from_tuple!((usize, usize), 0, 1);
+impl_from_tuple!((usize, usize, usize), 0, 1, 2);
+impl_from_tuple!((usize, usize, usize, usize), 0, 1, 2, 3);
+impl_from_tuple!((usize, usize, usize, usize, usize), 0, 1, 2, 3, 4);
+impl_from_tuple!((usize, usize, usize, usize, usize, usize), 0, 1, 2, 3, 4, 5);
 
 impl From<Vec<usize>> for Shape {
     fn from(dims: Vec<usize>) -> Self {
@@ -142,6 +121,12 @@ impl Shape {
         &self.0
     }
 
+    /// The dimension size for a specified dimension index.
+    pub fn dim<D: Dim>(&self, dim: D) -> Result<usize> {
+        let dim = dim.to_index(self, "dim")?;
+        Ok(self.dims()[dim])
+    }
+
     /// The total number of elements, this is the product of all dimension sizes.
     pub fn elem_count(&self) -> usize {
         self.0.iter().product()
@@ -304,6 +289,7 @@ impl Dim for usize {
 pub enum D {
     Minus1,
     Minus2,
+    Minus(usize),
 }
 
 impl D {
@@ -311,6 +297,7 @@ impl D {
         let dim = match self {
             Self::Minus1 => -1,
             Self::Minus2 => -2,
+            Self::Minus(u) => -(*u as i32),
         };
         Error::DimOutOfRange {
             shape: shape.clone(),
@@ -327,6 +314,7 @@ impl Dim for D {
         match self {
             Self::Minus1 if rank >= 1 => Ok(rank - 1),
             Self::Minus2 if rank >= 2 => Ok(rank - 2),
+            Self::Minus(u) if *u > 0 && rank >= *u => Ok(rank - *u),
             _ => Err(self.out_of_range(shape, op)),
         }
     }
@@ -336,6 +324,7 @@ impl Dim for D {
         match self {
             Self::Minus1 => Ok(rank),
             Self::Minus2 if rank >= 1 => Ok(rank - 1),
+            Self::Minus(u) if *u > 0 && rank + 1 >= *u => Ok(rank + 1 - *u),
             _ => Err(self.out_of_range(shape, op)),
         }
     }
@@ -626,4 +615,20 @@ mod tests {
         let shape = Shape::from((299, 792, 458));
         assert_eq!(shape.stride_contiguous(), [458 * 792, 458, 1]);
     }
+
+    #[test]
+    fn test_from_tuple() {
+        let shape = Shape::from((2,));
+        assert_eq!(shape.dims(), &[2]);
+        let shape = Shape::from((2, 3));
+        assert_eq!(shape.dims(), &[2, 3]);
+        let shape = Shape::from((2, 3, 4));
+        assert_eq!(shape.dims(), &[2, 3, 4]);
+        let shape = Shape::from((2, 3, 4, 5));
+        assert_eq!(shape.dims(), &[2, 3, 4, 5]);
+        let shape = Shape::from((2, 3, 4, 5, 6));
+        assert_eq!(shape.dims(), &[2, 3, 4, 5, 6]);
+        let shape = Shape::from((2, 3, 4, 5, 6, 7));
+        assert_eq!(shape.dims(), &[2, 3, 4, 5, 6, 7]);
+    }
 }

candle-core/src/sort.rs

@@ -0,0 +1,250 @@
+use crate::{Result, Tensor};
+use rayon::prelude::*;
+
+#[derive(Debug, Clone, Copy)]
+struct ArgSort {
+    asc: bool,
+    last_dim: usize,
+}
+
+impl ArgSort {
+    fn asort<T: crate::WithDType>(&self, vs: &[T], layout: &crate::Layout) -> Vec<u32> {
+        #[allow(clippy::uninit_vec)]
+        // Safety: indexes are set later in the parallelized section.
+        let mut sort_indexes = unsafe {
+            let el_count = layout.shape().elem_count();
+            let mut v = Vec::with_capacity(el_count);
+            v.set_len(el_count);
+            v
+        };
+        if self.asc {
+            sort_indexes
+                .par_chunks_exact_mut(self.last_dim)
+                .zip(vs.par_chunks_exact(self.last_dim))
+                .for_each(|(indexes, vs)| {
+                    indexes
+                        .iter_mut()
+                        .enumerate()
+                        .for_each(|(i, v)| *v = i as u32);
+                    indexes.sort_by(|&i, &j| {
+                        vs[i as usize]
+                            .partial_cmp(&vs[j as usize])
+                            .unwrap_or(std::cmp::Ordering::Greater)
+                    })
+                });
+        } else {
+            sort_indexes
+                .par_chunks_exact_mut(self.last_dim)
+                .zip(vs.par_chunks_exact(self.last_dim))
+                .for_each(|(indexes, vs)| {
+                    indexes
+                        .iter_mut()
+                        .enumerate()
+                        .for_each(|(i, v)| *v = i as u32);
+                    indexes.sort_by(|&j, &i| {
+                        vs[i as usize]
+                            .partial_cmp(&vs[j as usize])
+                            .unwrap_or(std::cmp::Ordering::Greater)
+                    })
+                });
+        }
+        sort_indexes
+    }
+}
+
+#[cfg(feature = "cuda")]
+mod cuda {
+    use super::*;
+    use crate::cuda_backend::cudarc::driver::{
+        CudaSlice, DeviceRepr, LaunchConfig, ValidAsZeroBits,
+    };
+    use crate::cuda_backend::{kernel_name, kernels, CudaStorageSlice as S, WrapErr};
+    use crate::{CudaDevice, WithDType};
+
+    impl crate::cuda_backend::Map1Any for ArgSort {
+        fn f<T: DeviceRepr + WithDType + ValidAsZeroBits, W: Fn(CudaSlice<T>) -> S>(
+            &self,
+            src: &CudaSlice<T>,
+            dev: &CudaDevice,
+            layout: &crate::Layout,
+            _wrap: W,
+        ) -> Result<S> {
+            use cudarc::driver::PushKernelArg;
+
+            let slice = match layout.contiguous_offsets() {
+                None => crate::bail!("input has to be contiguous"),
+                Some((o1, o2)) => src.slice(o1..o2),
+            };
+            let elem_count = layout.shape().elem_count();
+            let dst = unsafe { dev.alloc::<u32>(elem_count)? };
+            let func = if self.asc {
+                dev.get_or_load_func(&kernel_name::<T>("asort_asc"), &kernels::SORT)?
+            } else {
+                dev.get_or_load_func(&kernel_name::<T>("asort_desc"), &kernels::SORT)?
+            };
+            let ncols = self.last_dim;
+            let nrows = elem_count / ncols;
+            let ncols_pad = next_power_of_2(ncols);
+            let cfg = LaunchConfig {
+                grid_dim: (1, nrows as u32, 1),
+                block_dim: (ncols_pad as u32, 1, 1),
+                shared_mem_bytes: (ncols_pad * std::mem::size_of::<u32>()) as u32,
+            };
+            let stream = dev.cuda_stream();
+            let mut builder = stream.launch_builder(&func);
+            let ncols = ncols as i32;
+            let ncols_pad = ncols_pad as i32;
+            builder.arg(&slice).arg(&dst).arg(&ncols).arg(&ncols_pad);
+            unsafe { builder.launch(cfg) }.w()?;
+            Ok(S::U32(dst))
+        }
+    }
+}
+
+impl crate::CustomOp1 for ArgSort {
+    fn name(&self) -> &'static str {
+        "argsort"
+    }
+
+    fn cpu_fwd(
+        &self,
+        storage: &crate::CpuStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::CpuStorage, crate::Shape)> {
+        let sort_indexes = match storage {
+            crate::CpuStorage::U8(vs) => self.asort(vs, layout),
+            crate::CpuStorage::U32(vs) => self.asort(vs, layout),
+            crate::CpuStorage::I64(vs) => self.asort(vs, layout),
+            crate::CpuStorage::BF16(vs) => self.asort(vs, layout),
+            crate::CpuStorage::F16(vs) => self.asort(vs, layout),
+            crate::CpuStorage::F32(vs) => self.asort(vs, layout),
+            crate::CpuStorage::F64(vs) => self.asort(vs, layout),
+        };
+        let sort_indexes = crate::CpuStorage::U32(sort_indexes);
+        Ok((sort_indexes, layout.shape().into()))
+    }
+
+    #[cfg(feature = "cuda")]
+    fn cuda_fwd(
+        &self,
+        storage: &crate::CudaStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::CudaStorage, crate::Shape)> {
+        use crate::backend::BackendStorage;
+        use crate::cuda_backend::Map1Any;
+        let dev = storage.device();
+        let slice = self.map(&storage.slice, dev, layout)?;
+        let dst = crate::cuda_backend::CudaStorage {
+            slice,
+            device: dev.clone(),
+        };
+        Ok((dst, layout.shape().clone()))
+    }
+
+    #[cfg(feature = "metal")]
+    fn metal_fwd(
+        &self,
+        storage: &crate::MetalStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::MetalStorage, crate::Shape)> {
+        use crate::backend::BackendStorage;
+        use crate::DType;
+
+        let name = {
+            if self.asc {
+                match storage.dtype() {
+                    DType::BF16 => "asort_asc_bf16",
+                    DType::F16 => "asort_asc_f16",
+                    DType::F32 => "asort_asc_f32",
+                    DType::F64 => "asort_asc_f64",
+                    DType::U8 => "asort_asc_u8",
+                    DType::U32 => "asort_asc_u32",
+                    DType::I64 => "asort_asc_i64",
+                }
+            } else {
+                match storage.dtype() {
+                    DType::BF16 => "asort_desc_bf16",
+                    DType::F16 => "asort_desc_f16",
+                    DType::F32 => "asort_desc_f32",
+                    DType::F64 => "asort_desc_f64",
+                    DType::U8 => "asort_desc_u8",
+                    DType::U32 => "asort_desc_u32",
+                    DType::I64 => "asort_desc_i64",
+                }
+            }
+        };
+        let device = storage.device();
+        let kernels = device.kernels();
+        let command_buffer = device.command_buffer()?;
+        let el = layout.shape().elem_count();
+        let ncols = self.last_dim;
+        let nrows = el / ncols;
+        let src = crate::metal_backend::buffer_o(storage.buffer(), layout, storage.dtype());
+        let dst = device.new_buffer(el, DType::U32, "asort")?;
+        let mut ncols_pad = 1;
+        while ncols_pad < ncols {
+            ncols_pad *= 2;
+        }
+        candle_metal_kernels::call_arg_sort(
+            device.metal_device(),
+            &command_buffer,
+            kernels,
+            name,
+            nrows,
+            ncols,
+            ncols_pad,
+            src,
+            &dst,
+        )
+        .map_err(crate::Error::wrap)?;
+        let dst = crate::MetalStorage::new(dst, device.clone(), el, DType::U32);
+        Ok((dst, layout.shape().clone()))
+    }
+}
+
+#[allow(unused)]
+fn next_power_of_2(x: usize) -> usize {
+    let mut n = 1;
+    while n < x {
+        n *= 2
+    }
+    n
+}
+
+impl Tensor {
+    /// Returns the indices that sort the tensor along the last dimension.
+    ///
+    /// If `asc` is `true`, sorting is in ascending order. Otherwise sorting is performed in
+    /// descending order. The sort is unstable so there is no guarantees on the final order when it
+    /// comes to ties.
+    pub fn arg_sort_last_dim(&self, asc: bool) -> Result<Tensor> {
+        if !self.is_contiguous() {
+            return Err(crate::Error::RequiresContiguous {
+                op: "arg_sort_last_dim",
+            });
+        }
+        let last_dim = match self.dims().last() {
+            None => crate::bail!("empty last-dim in arg-sort"),
+            Some(last_dim) => *last_dim,
+        };
+        // No need for a backward pass for arg sort.
+        self.apply_op1_no_bwd(&ArgSort { asc, last_dim })
+    }
+
+    /// Sorts the tensor along the last dimension, returns the sorted tensor together with the
+    /// sorted indexes.
+    ///
+    /// If `asc` is `true`, sorting is in ascending order. Otherwise sorting is performed in
+    /// descending order. The sort is unstable so there is no guarantees on the final order when it
+    /// comes to ties.
+    pub fn sort_last_dim(&self, asc: bool) -> Result<(Tensor, Tensor)> {
+        if !self.is_contiguous() {
+            return Err(crate::Error::RequiresContiguous {
+                op: "sort_last_dim",
+            });
+        }
+        let asort = self.arg_sort_last_dim(asc)?;
+        let sorted = self.gather(&asort, crate::D::Minus1)?;
+        Ok((sorted, asort))
+    }
+}

candle-core/src/storage.rs

@@ -1,5 +1,6 @@
 use crate::backend::BackendStorage;
 use crate::op::{self, CmpOp, ReduceOp};
+use crate::scalar::Scalar;
 use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, MetalStorage, Result, Shape};
 use crate::{CustomOp1, CustomOp2, CustomOp3, InplaceOp1, InplaceOp2, InplaceOp3};
 
@@ -73,6 +74,14 @@ impl Storage {
         }
     }
 
+    pub(crate) fn const_set(&mut self, v: Scalar, l: &Layout) -> Result<()> {
+        match self {
+            Storage::Cpu(storage) => storage.const_set(v, l),
+            Storage::Cuda(storage) => storage.const_set(v, l),
+            Storage::Metal(storage) => storage.const_set(v, l),
+        }
+    }
+
     pub(crate) fn affine(&self, layout: &Layout, mul: f64, add: f64) -> Result<Self> {
         match self {
             Storage::Cpu(storage) => {
@@ -619,32 +628,56 @@ impl Storage {
         }
     }
 
-    pub(crate) fn scatter_add(
-        &self,
+    pub(crate) fn scatter_set(
+        &mut self,
         l: &Layout,
         indexes: &Self,
         indexes_l: &Layout,
         source: &Self,
         source_l: &Layout,
         d: usize,
-    ) -> Result<Self> {
+    ) -> Result<()> {
+        self.same_device(indexes, "scatter-set")?;
+        self.same_device(source, "scatter-set")?;
+        match (self, indexes, source) {
+            (Self::Cpu(s), Self::Cpu(indexes), Self::Cpu(source)) => {
+                s.scatter_set(l, indexes, indexes_l, source, source_l, d)?;
+            }
+            (Self::Cuda(s), Self::Cuda(indexes), Self::Cuda(source)) => {
+                s.scatter_set(l, indexes, indexes_l, source, source_l, d)?;
+            }
+            (Self::Metal(s), Self::Metal(indexes), Self::Metal(source)) => {
+                s.scatter_set(l, indexes, indexes_l, source, source_l, d)?;
+            }
+            _ => unreachable!(),
+        }
+        Ok(())
+    }
+
+    pub(crate) fn scatter_add(
+        &mut self,
+        l: &Layout,
+        indexes: &Self,
+        indexes_l: &Layout,
+        source: &Self,
+        source_l: &Layout,
+        d: usize,
+    ) -> Result<()> {
         self.same_device(indexes, "scatter-add")?;
         self.same_device(source, "scatter-add")?;
         match (self, indexes, source) {
             (Self::Cpu(s), Self::Cpu(indexes), Self::Cpu(source)) => {
-                let storage = s.scatter_add(l, indexes, indexes_l, source, source_l, d)?;
-                Ok(Self::Cpu(storage))
+                s.scatter_add_set(l, indexes, indexes_l, source, source_l, d)?;
             }
             (Self::Cuda(s), Self::Cuda(indexes), Self::Cuda(source)) => {
-                let storage = s.scatter_add(l, indexes, indexes_l, source, source_l, d)?;
-                Ok(Self::Cuda(storage))
+                s.scatter_add_set(l, indexes, indexes_l, source, source_l, d)?;
             }
             (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))
+                s.scatter_add_set(l, indexes, indexes_l, source, source_l, d)?;
             }
             _ => unreachable!(),
         }
+        Ok(())
     }
 
     pub(crate) fn index_add(

candle-core/src/streaming.rs

@@ -0,0 +1,208 @@
+//! StreamTensror useful for streaming ops.
+//!
+use crate::{Result, Shape, Tensor};
+
+pub trait Dim: crate::shape::Dim + Copy {}
+impl<T: crate::shape::Dim + Copy> Dim for T {}
+
+/// A stream tensor is used in streaming module. It can either contain an actual tensor or be
+/// empty.
+#[derive(Clone)]
+pub struct StreamTensor(Option<Tensor>);
+
+impl std::fmt::Debug for StreamTensor {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match &self.0 {
+            Some(t) => write!(f, "{:?}", t.shape()),
+            None => write!(f, "Empty"),
+        }
+    }
+}
+
+impl std::convert::From<Option<Tensor>> for StreamTensor {
+    fn from(value: Option<Tensor>) -> Self {
+        Self(value)
+    }
+}
+
+impl std::convert::From<Tensor> for StreamTensor {
+    fn from(value: Tensor) -> Self {
+        Self(Some(value))
+    }
+}
+
+impl std::convert::From<()> for StreamTensor {
+    fn from(_value: ()) -> Self {
+        Self(None)
+    }
+}
+
+impl StreamTensor {
+    pub fn empty() -> Self {
+        Self(None)
+    }
+
+    pub fn from_tensor(tensor: Tensor) -> Self {
+        Self(Some(tensor))
+    }
+
+    pub fn shape(&self) -> Option<&Shape> {
+        self.0.as_ref().map(|t| t.shape())
+    }
+
+    pub fn cat2<D: Dim>(&self, rhs: &Self, dim: D) -> Result<Self> {
+        let xs = match (&self.0, &rhs.0) {
+            (Some(lhs), Some(rhs)) => {
+                let xs = Tensor::cat(&[lhs, rhs], dim)?;
+                Some(xs)
+            }
+            (Some(xs), None) | (None, Some(xs)) => Some(xs.clone()),
+            (None, None) => None,
+        };
+        Ok(Self(xs))
+    }
+
+    pub fn seq_len<D: Dim>(&self, dim: D) -> Result<usize> {
+        match &self.0 {
+            None => Ok(0),
+            Some(v) => v.dim(dim),
+        }
+    }
+
+    pub fn reset(&mut self) {
+        self.0 = None
+    }
+
+    pub fn narrow<D: Dim>(&self, dim: D, offset: usize, len: usize) -> Result<StreamTensor> {
+        let t = match &self.0 {
+            None => None,
+            Some(t) => {
+                let seq_len = t.dim(dim)?;
+                if seq_len <= offset {
+                    None
+                } else {
+                    let t = t.narrow(dim, offset, usize::min(len, seq_len - offset))?;
+                    Some(t)
+                }
+            }
+        };
+        Ok(Self(t))
+    }
+
+    /// Splits the Streaming Tensor on the time axis `dim` with the first `lhs_len` elements
+    /// returned in the first output and the remaining in the second output.
+    pub fn split<D: Dim>(&self, dim: D, lhs_len: usize) -> Result<(Self, Self)> {
+        match &self.0 {
+            None => Ok((Self::empty(), Self::empty())),
+            Some(t) => {
+                let seq_len = t.dim(dim)?;
+                let lhs_len = usize::min(seq_len, lhs_len);
+                if lhs_len == 0 {
+                    Ok((Self::empty(), t.clone().into()))
+                } else {
+                    let lhs = Self::from_tensor(t.narrow(dim, 0, lhs_len)?);
+                    let rhs_len = seq_len - lhs_len;
+                    let rhs = if rhs_len == 0 {
+                        Self::empty()
+                    } else {
+                        Self::from_tensor(t.narrow(dim, lhs_len, rhs_len)?)
+                    };
+                    Ok((lhs, rhs))
+                }
+            }
+        }
+    }
+
+    pub fn as_option(&self) -> Option<&Tensor> {
+        self.0.as_ref()
+    }
+
+    pub fn apply<M: crate::Module>(&self, m: &M) -> Result<Self> {
+        match &self.0 {
+            None => Ok(Self::empty()),
+            Some(t) => Ok(Self::from_tensor(t.apply(m)?)),
+        }
+    }
+}
+
+/// Streaming modules take as input a stream tensor and return a stream tensor. They may perform
+/// some internal buffering so that enough data has been received for the module to be able to
+/// perform some operations.
+pub trait StreamingModule {
+    // TODO: Should we also have a flush method?
+    fn step(&mut self, xs: &StreamTensor) -> Result<StreamTensor>;
+    fn reset_state(&mut self);
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum BinOp {
+    Add,
+    Mul,
+    Sub,
+    Div,
+}
+
+#[derive(Debug, Clone)]
+pub struct StreamingBinOp {
+    prev_lhs: StreamTensor,
+    prev_rhs: StreamTensor,
+    pub op: BinOp,
+    pub dim: crate::D,
+}
+
+impl StreamingBinOp {
+    pub fn new(op: BinOp, dim: crate::D) -> Self {
+        Self {
+            prev_lhs: StreamTensor::empty(),
+            prev_rhs: StreamTensor::empty(),
+            op,
+            dim,
+        }
+    }
+
+    pub fn reset_state(&mut self) {
+        self.prev_lhs.reset();
+        self.prev_rhs.reset();
+    }
+
+    pub fn forward(&self, lhs: &Tensor, rhs: &Tensor) -> Result<Tensor> {
+        match self.op {
+            BinOp::Add => Tensor::add(lhs, rhs),
+            BinOp::Mul => Tensor::mul(lhs, rhs),
+            BinOp::Sub => Tensor::sub(lhs, rhs),
+            BinOp::Div => Tensor::div(lhs, rhs),
+        }
+    }
+
+    pub fn step(&mut self, lhs: &StreamTensor, rhs: &StreamTensor) -> Result<StreamTensor> {
+        let lhs = StreamTensor::cat2(&self.prev_lhs, lhs, self.dim)?;
+        let rhs = StreamTensor::cat2(&self.prev_rhs, rhs, self.dim)?;
+        let lhs_len = lhs.seq_len(self.dim)?;
+        let rhs_len = rhs.seq_len(self.dim)?;
+        let common_len = usize::min(lhs_len, rhs_len);
+        let (lhs, prev_lhs) = lhs.split(self.dim, common_len)?;
+        let (rhs, prev_rhs) = rhs.split(self.dim, common_len)?;
+        let ys = match (lhs.0, rhs.0) {
+            (Some(lhs), Some(rhs)) => {
+                let ys = self.forward(&lhs, &rhs)?;
+                StreamTensor::from_tensor(ys)
+            }
+            (None, None) => StreamTensor::empty(),
+            (lhs, rhs) => crate::bail!("INTERNAL ERROR inconsistent lhs and rhs {lhs:?} {rhs:?}"),
+        };
+        self.prev_lhs = prev_lhs;
+        self.prev_rhs = prev_rhs;
+        Ok(ys)
+    }
+}
+
+/// Simple wrapper that doesn't do any buffering.
+pub struct Map<T: crate::Module>(T);
+
+impl<T: crate::Module> StreamingModule for Map<T> {
+    fn reset_state(&mut self) {}
+
+    fn step(&mut self, xs: &StreamTensor) -> Result<StreamTensor> {
+        xs.apply(&self.0)
+    }
+}

candle-core/src/strided_index.rs

@@ -32,14 +32,11 @@ impl<'a> StridedIndex<'a> {
     }
 }
 
-impl<'a> Iterator for StridedIndex<'a> {
+impl Iterator for StridedIndex<'_> {
     type Item = usize;
 
     fn next(&mut self) -> Option<Self::Item> {
-        let storage_index = match self.next_storage_index {
-            None => return None,
-            Some(storage_index) => storage_index,
-        };
+        let storage_index = self.next_storage_index?;
         let mut updated = false;
         let mut next_storage_index = storage_index;
         for ((multi_i, max_i), stride_i) in self

candle-core/src/tensor.rs

@@ -3,7 +3,7 @@
 use crate::backend::{BackendDevice, BackendStorage};
 use crate::op::{BackpropOp, BinaryOp, CmpOp, Op, ReduceOp, UnaryOp};
 use crate::scalar::TensorOrScalar;
-use crate::shape::{Dim, Dims};
+use crate::shape::{Dim, Dims, ShapeWithOneHole};
 use crate::{bail, storage::Storage, DType, Device, Error, Layout, Result, Shape};
 use std::sync::{Arc, RwLock};
 
@@ -79,6 +79,9 @@ macro_rules! unary_op {
     ($fn_name:ident, $op_name:ident) => {
         pub fn $fn_name(&self) -> Result<Self> {
             let shape = self.shape();
+            if shape.elem_count() == 0 {
+                return Ok(self.clone());
+            }
             let storage = self
                 .storage()
                 .unary_impl::<crate::op::$op_name>(self.layout())?;
@@ -92,6 +95,9 @@ macro_rules! binary_op {
     ($fn_name:ident, $op_name:ident) => {
         pub fn $fn_name(&self, rhs: &Self) -> Result<Self> {
             let shape = self.same_shape_binary_op(rhs, stringify!($fn_name))?;
+            if shape.elem_count() == 0 {
+                return Ok(self.clone());
+            }
             let storage = self.storage().binary_impl::<crate::op::$op_name>(
                 &*rhs.storage(),
                 self.layout(),
@@ -114,6 +120,9 @@ macro_rules! binary_op_scalar {
                     .broadcast_as(self.shape())?,
             };
             let shape = self.same_shape_binary_op(&rhs, stringify!($fn_name))?;
+            if self.elem_count() == 0 {
+                return Ok(self.clone());
+            }
             let storage = self.storage().binary_impl::<crate::op::$op_name>(
                 &*rhs.storage(),
                 self.layout(),
@@ -176,7 +185,9 @@ impl Tensor {
     ) -> Result<Self> {
         let none = BackpropOp::none();
         let shape = shape.into();
-        let storage = device.ones(&shape, dtype)?;
+        let mut storage = unsafe { device.alloc_uninit(&shape, dtype)? };
+        let layout = Layout::contiguous(shape.clone());
+        storage.const_set(crate::scalar::Scalar::one(dtype), &layout)?;
         Ok(from_storage(storage, shape, none, is_variable))
     }
 
@@ -193,6 +204,18 @@ impl Tensor {
         Self::ones_impl(shape, dtype, device, false)
     }
 
+    pub fn const_set(&self, value: crate::scalar::Scalar) -> Result<()> {
+        self.storage_mut().const_set(value, self.layout())
+    }
+
+    pub fn zero_set(&self) -> Result<()> {
+        self.const_set(crate::scalar::Scalar::zero(self.dtype()))
+    }
+
+    pub fn one_set(&self) -> Result<()> {
+        self.const_set(crate::scalar::Scalar::one(self.dtype()))
+    }
+
     /// Creates a new tensor filled with ones with same shape, dtype, and device as the other tensor.
     ///
     /// ```rust
@@ -233,7 +256,7 @@ impl Tensor {
         Self::zeros_impl(shape, dtype, device, false)
     }
 
-    /// Creates a new tensor filled with ones with same shape, dtype, and device as the other
+    /// Creates a new tensor filled with zeros with same shape, dtype, and device as the other
     /// tensor.
     ///
     /// ```rust
@@ -359,17 +382,37 @@ 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.
+    /// Returns a new tensor with all the elements having the same specified value.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::full(3.5, (2, 4), &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec2::<f64>()?, &[
+    ///     [3.5, 3.5, 3.5, 3.5],
+    ///     [3.5, 3.5, 3.5, 3.5],
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
     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)
+        let none = BackpropOp::none();
+        let shape = shape.into();
+        let mut storage = unsafe { device.alloc_uninit(&shape, D::DTYPE)? };
+        let layout = Layout::contiguous(shape.clone());
+        storage.const_set(value.to_scalar(), &layout)?;
+        Ok(from_storage(storage, shape, none, false))
     }
 
     /// Creates a new 1D tensor from an iterator.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::from_iter( [1.0, 2.0, 3.0, 4.0].into_iter(), &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec1::<f64>()?, &[1.0, 2.0, 3.0, 4.0]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     pub fn from_iter<D: crate::WithDType>(
         iter: impl IntoIterator<Item = D>,
         device: &Device,
@@ -381,12 +424,26 @@ impl Tensor {
 
     /// Creates a new 1D tensor with values from the interval `[start, end)` taken with a common
     /// difference `1` from `start`.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::arange(2., 5., &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec1::<f64>()?, &[2., 3., 4.]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     pub fn arange<D: crate::WithDType>(start: D, end: D, device: &Device) -> Result<Self> {
         Self::arange_step(start, end, D::one(), device)
     }
 
     /// Creates a new 1D tensor with values from the interval `[start, end)` taken with a common
     /// difference `step` from `start`.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::arange_step(2.0, 4.0, 0.5, &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec1::<f64>()?, &[2.0, 2.5, 3.0, 3.5]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     pub fn arange_step<D: crate::WithDType>(
         start: D,
         end: D,
@@ -413,17 +470,13 @@ impl Tensor {
         Self::from_vec_impl(data, len, device, false)
     }
 
-    pub(crate) fn from_vec_impl<S: Into<Shape>, D: crate::WithDType>(
+    pub(crate) fn from_vec_impl<S: ShapeWithOneHole, D: crate::WithDType>(
         data: Vec<D>,
         shape: S,
         device: &Device,
         is_variable: bool,
     ) -> Result<Self> {
-        let shape = shape.into();
-        let buffer_size = data.len();
-        if buffer_size != shape.elem_count() {
-            return Err(Error::ShapeMismatch { buffer_size, shape }.bt());
-        }
+        let shape = shape.into_shape(data.len())?;
         let storage = device.storage_owned(data)?;
         let none = BackpropOp::none();
         Ok(from_storage(storage, shape, none, is_variable))
@@ -432,7 +485,17 @@ impl Tensor {
     /// Creates a new tensor initialized with values from the input vector. The number of elements
     /// in this vector must be the same as the number of elements defined by the shape.
     /// If the device is cpu, no data copy is made.
-    pub fn from_vec<S: Into<Shape>, D: crate::WithDType>(
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::from_vec(vec!{1., 2., 3., 4., 5., 6.}, (2, 3), &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec2::<f64>()?, &[
+    ///     [1., 2., 3.],
+    ///     [4., 5., 6.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn from_vec<S: ShapeWithOneHole, D: crate::WithDType>(
         data: Vec<D>,
         shape: S,
         device: &Device,
@@ -442,12 +505,26 @@ impl Tensor {
 
     /// Creates a new tensor initialized with values from the input slice. The number of elements
     /// in this vector must be the same as the number of elements defined by the shape.
-    pub fn from_slice<S: Into<Shape>, D: crate::WithDType>(
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let values = vec![1., 2., 3., 4., 5., 6., 7., 8.];
+    /// let a = Tensor::from_slice(&values[1..7], (2, 3), &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec2::<f64>()?, &[
+    ///     [2., 3., 4.],
+    ///     [5., 6., 7.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn from_slice<S: ShapeWithOneHole, D: crate::WithDType>(
         array: &[D],
         shape: S,
         device: &Device,
     ) -> Result<Self> {
-        Self::new_impl(array, shape.into(), device, false)
+        let shape = shape.into_shape(array.len())?;
+        let storage = device.storage_from_slice(array)?;
+        let none = BackpropOp::none();
+        Ok(from_storage(storage, shape, none, false))
     }
 
     pub(crate) fn same_shape_binary_op(&self, rhs: &Self, op: &'static str) -> Result<&Shape> {
@@ -573,9 +650,9 @@ impl Tensor {
     ///
     /// * `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.
+    ///   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
     ///
@@ -646,6 +723,9 @@ impl Tensor {
     /// # Ok::<(), candle_core::Error>(())
     /// ```
     pub fn affine(&self, mul: f64, add: f64) -> Result<Self> {
+        if self.elem_count() == 0 {
+            return Ok(self.clone());
+        }
         let storage = self.storage().affine(self.layout(), mul, add)?;
         let op = BackpropOp::new1(self, |arg| Op::Affine { arg, mul, add });
         Ok(from_storage(storage, self.shape(), op, false))
@@ -653,6 +733,9 @@ impl Tensor {
 
     /// Applies the Exponential Linear Unit (ELU) function on each element of the input tensor.
     pub fn elu(&self, alpha: f64) -> Result<Self> {
+        if self.elem_count() == 0 {
+            return Ok(self.clone());
+        }
         let storage = self.storage().elu(self.layout(), alpha)?;
         let op = BackpropOp::new1(self, |t| Op::Elu(t, alpha));
         Ok(from_storage(storage, self.shape(), op, false))
@@ -660,6 +743,9 @@ impl Tensor {
 
     /// Raise the tensor to some float exponent `e`.
     pub fn powf(&self, e: f64) -> Result<Self> {
+        if self.elem_count() == 0 {
+            return Ok(self.clone());
+        }
         let storage = self.storage().powf(self.layout(), e)?;
         let op = BackpropOp::new1(self, |t| Op::Powf(t, e));
         Ok(from_storage(storage, self.shape(), op, false))
@@ -706,6 +792,30 @@ impl Tensor {
 
     /// Returns a new tensor that is a narrowed version of the input, the dimension `dim`
     /// ranges from `start` to `start + len`.
+    /// ```
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::new(&[
+    ///     [0f32, 1., 2.],
+    ///     [3.  , 4., 5.],
+    ///     [6.  , 7., 8.]
+    /// ], &Device::Cpu)?;
+    ///
+    /// let b = a.narrow(0, 1, 2)?;
+    /// assert_eq!(b.shape().dims(), &[2, 3]);
+    /// assert_eq!(b.to_vec2::<f32>()?, &[
+    ///     [3., 4., 5.],
+    ///     [6., 7., 8.]
+    /// ]);
+    ///
+    /// let c = a.narrow(1, 1, 1)?;
+    /// assert_eq!(c.shape().dims(), &[3, 1]);
+    /// assert_eq!(c.to_vec2::<f32>()?, &[
+    ///     [1.],
+    ///     [4.],
+    ///     [7.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     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")?;
@@ -1154,6 +1264,9 @@ impl Tensor {
         let n = b_dims[dim - 1];
 
         let c_shape = Shape::from(&a_dims[..dim - 2]).extend(&[m, n]);
+        if c_shape.elem_count() == 0 || k == 0 {
+            return Tensor::zeros(c_shape, self.dtype(), self.device());
+        }
         let batching: usize = a_dims[..dim - 2].iter().product();
         let batching_b: usize = b_dims[..dim - 2].iter().product();
         if k != k2 || batching != batching_b {
@@ -1245,8 +1358,7 @@ impl Tensor {
         self.index_select(ids, 0)
     }
 
-    pub fn scatter_add<D: Dim>(&self, indexes: &Self, source: &Self, dim: D) -> Result<Self> {
-        let dim = dim.to_index(self.shape(), "scatter-add")?;
+    fn scatter_checks(&self, indexes: &Self, source: &Self, dim: usize) -> Result<()> {
         let source_dims = source.dims();
         let self_dims = self.dims();
         let mismatch = if source_dims.len() != self_dims.len() {
@@ -1263,7 +1375,7 @@ impl Tensor {
         };
         if mismatch {
             Err(Error::ShapeMismatchBinaryOp {
-                op: "scatter-add (self, src)",
+                op: "scatter (self, src)",
                 lhs: self.shape().clone(),
                 rhs: source.shape().clone(),
             }
@@ -1271,13 +1383,44 @@ impl Tensor {
         }
         if indexes.dims() != source.dims() {
             Err(Error::ShapeMismatchBinaryOp {
-                op: "scatter-add (indexes, src)",
+                op: "scatter (indexes, src)",
                 lhs: indexes.shape().clone(),
                 rhs: source.shape().clone(),
             }
             .bt())?
         }
-        let storage = self.storage().scatter_add(
+        Ok(())
+    }
+
+    pub fn scatter<D: Dim>(&self, indexes: &Self, source: &Self, dim: D) -> Result<Self> {
+        let dim = dim.to_index(self.shape(), "scatter")?;
+        self.scatter_checks(indexes, source, dim)?;
+        let shape = self.shape();
+        let mut storage = unsafe { self.device().alloc_uninit(shape, self.dtype())? };
+        self.storage()
+            .copy_strided_src(&mut storage, 0, self.layout())?;
+        let layout = Layout::contiguous(shape);
+        storage.scatter_set(
+            &layout,
+            &indexes.storage(),
+            indexes.layout(),
+            &source.storage(),
+            source.layout(),
+            dim,
+        )?;
+        let op = BackpropOp::new3(self, indexes, source, |t1, t2, t3| {
+            Op::Scatter(t1, t2, t3, dim)
+        });
+        Ok(from_storage(storage, self.shape(), op, false))
+    }
+
+    pub fn scatter_set<D: Dim>(&self, indexes: &Self, source: &Self, dim: D) -> Result<()> {
+        if self.same_storage(source) {
+            crate::bail!("cannot use slice_set when self and src share their storage")
+        }
+        let dim = dim.to_index(self.shape(), "scatter-set")?;
+        self.scatter_checks(indexes, source, dim)?;
+        self.storage_mut().scatter_set(
             self.layout(),
             &indexes.storage(),
             indexes.layout(),
@@ -1285,12 +1428,48 @@ impl Tensor {
             source.layout(),
             dim,
         )?;
+        Ok(())
+    }
+
+    pub fn scatter_add<D: Dim>(&self, indexes: &Self, source: &Self, dim: D) -> Result<Self> {
+        let dim = dim.to_index(self.shape(), "scatter-add")?;
+        self.scatter_checks(indexes, source, dim)?;
+        let shape = self.shape();
+        let mut storage = unsafe { self.device().alloc_uninit(shape, self.dtype())? };
+        self.storage()
+            .copy_strided_src(&mut storage, 0, self.layout())?;
+        let layout = Layout::contiguous(shape);
+        storage.scatter_add(
+            &layout,
+            &indexes.storage(),
+            indexes.layout(),
+            &source.storage(),
+            source.layout(),
+            dim,
+        )?;
         let op = BackpropOp::new3(self, indexes, source, |t1, t2, t3| {
             Op::ScatterAdd(t1, t2, t3, dim)
         });
         Ok(from_storage(storage, self.shape(), op, false))
     }
 
+    pub fn scatter_add_set<D: Dim>(&self, indexes: &Self, source: &Self, dim: D) -> Result<()> {
+        if self.same_storage(source) {
+            crate::bail!("cannot use slice_set when self and src share their storage")
+        }
+        let dim = dim.to_index(self.shape(), "scatter-add-set")?;
+        self.scatter_checks(indexes, source, dim)?;
+        self.storage_mut().scatter_add(
+            self.layout(),
+            &indexes.storage(),
+            indexes.layout(),
+            &source.storage(),
+            source.layout(),
+            dim,
+        )?;
+        Ok(())
+    }
+
     /// 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")?;
@@ -1416,14 +1595,15 @@ impl Tensor {
     /// # Arguments
     ///
     /// * `self` - The input tensor.
-    /// * `indexes` - The indices of elements to gather, this should have the same shape as `self`
-    ///   but can have a different number of elements on the target dimension.
+    /// * `indexes` - The indices of elements to gather, this should have same number of dimensions as `self`
+    ///   and indexes.dims()[d] <= self.dims()[d] for all dimensions d != dim
     /// * `dim` - the target dimension.
     ///
     /// The resulting tensor has the same shape as `indexes` and use values from `self` indexed on
     /// dimension `dim` by the values in `indexes`.
     pub fn gather<D: Dim>(&self, indexes: &Self, dim: D) -> Result<Self> {
         let dim = dim.to_index(self.shape(), "gather")?;
+
         let self_dims = self.dims();
         let indexes_dims = indexes.dims();
         let mismatch = if indexes_dims.len() != self_dims.len() {
@@ -1431,7 +1611,7 @@ impl Tensor {
         } else {
             let mut mismatch = false;
             for (i, (&d1, &d2)) in self_dims.iter().zip(indexes_dims.iter()).enumerate() {
-                if i != dim && d1 != d2 {
+                if i != dim && d1 < d2 {
                     mismatch = true;
                     break;
                 }
@@ -1655,6 +1835,42 @@ impl Tensor {
         &self.op
     }
 
+    /// Computes the max of all the elements in this tensor and returns a tensor holding this
+    /// scalar with zero dimensions.
+    ///
+    /// ```rust
+    /// use candle_core::{Tensor, Device};
+    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
+    /// let tensor = tensor.max_all()?;
+    /// assert_eq!(tensor.to_scalar::<f32>()?, 5.);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn max_all(&self) -> Result<Tensor> {
+        if self.rank() == 0 {
+            Ok(self.clone())
+        } else {
+            self.flatten_all()?.max(0)
+        }
+    }
+
+    /// Computes the min of all the elements in this tensor and returns a tensor holding this
+    /// scalar with zero dimensions.
+    ///
+    /// ```rust
+    /// use candle_core::{Tensor, Device};
+    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
+    /// let tensor = tensor.min_all()?;
+    /// assert_eq!(tensor.to_scalar::<f32>()?, 0.);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn min_all(&self) -> Result<Tensor> {
+        if self.rank() == 0 {
+            Ok(self.clone())
+        } else {
+            self.flatten_all()?.min(0)
+        }
+    }
+
     /// Computes the sum of all the elements in this tensor and returns a tensor holding this
     /// scalar with zero dimensions.
     ///
@@ -1921,7 +2137,11 @@ impl Tensor {
                 }
                 (Storage::Cpu(storage), Device::Cpu) => Storage::Cpu(storage.clone()),
                 _ => {
-                    bail!("not implemented yet")
+                    bail!(
+                        "not implemented yet, self.device: {:?}, device: {:?}",
+                        self.device(),
+                        device
+                    )
                 }
             };
             let op = BackpropOp::new1(self, Op::ToDevice);
@@ -2052,7 +2272,7 @@ impl Tensor {
     ///
     /// # Ok::<(), candle_core::Error>(())
     /// ```
-    pub fn reshape<S: crate::shape::ShapeWithOneHole>(&self, s: S) -> Result<Tensor> {
+    pub fn reshape<S: ShapeWithOneHole>(&self, s: S) -> Result<Tensor> {
         let shape = s.into_shape(self.elem_count())?;
         if shape.elem_count() != self.elem_count() {
             return Err(Error::ShapeMismatchBinaryOp {
@@ -2411,9 +2631,19 @@ impl Tensor {
 
     /// 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()
+        let sum_dims = sum_dims.to_indexes(self.shape(), "log-sum-exp")?;
+        if sum_dims.is_empty() {
+            return Ok(self.clone());
+        }
+        let max = sum_dims[1..]
+            .iter()
+            .try_fold(self.max_keepdim(sum_dims[0])?, |max, &dim| {
+                max.max_keepdim(dim)
+            })?;
+        let exp = self.broadcast_sub(&max)?.exp()?;
+        let sum = exp.sum(sum_dims.clone())?;
+
+        sum.log()? + max.squeeze_dims(&sum_dims)
     }
 
     /// Pointwise pow operation.
@@ -2425,6 +2655,28 @@ impl Tensor {
     pub fn broadcast_pow(&self, rhs: &Tensor) -> Result<Self> {
         rhs.broadcast_mul(&self.log()?)?.exp()
     }
+
+    /// Returns a new tensor with the order of elements reversed along the specified dimensions.
+    /// This function makes a copy of the tensor’s data.
+    ///
+    /// ```rust
+    /// # use candle_core::{Tensor, Device};
+    /// let t = Tensor::arange(0., 6., &Device::Cpu)?.reshape((2, 3))?;
+    /// assert_eq!(t.to_vec2::<f64>()?, &[[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]]);
+    /// let t_flipped = t.flip(&[0])?;
+    /// assert_eq!(t_flipped.to_vec2::<f64>()?, &[[3.0, 4.0, 5.0], [0.0, 1.0, 2.0]]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn flip(&self, dims: &[usize]) -> Result<Tensor> {
+        let mut result = self.clone();
+        for &dim in dims.iter() {
+            let size = result.dim(dim)?;
+            let indices: Vec<i64> = (0..size).rev().map(|x| x as i64).collect();
+            let indices_tensor = Tensor::from_vec(indices, (size,), result.device())?;
+            result = result.index_select(&indices_tensor, dim)?;
+        }
+        Ok(result)
+    }
 }
 
 macro_rules! bin_trait {

candle-core/src/tensor_cat.rs

@@ -1,4 +1,4 @@
-use crate::{shape::Dim, Error, Result, Shape, Tensor};
+use crate::{shape::Dim, Context, Error, Result, Shape, Tensor};
 
 impl Tensor {
     /// Concatenates two or more tensors along a particular dimension.
@@ -134,7 +134,7 @@ impl Tensor {
                     .bt())?
                 }
             }
-            let next_offset = offsets.last().unwrap() + arg.elem_count();
+            let next_offset = offsets.last().context("empty offsets")? + arg.elem_count();
             offsets.push(next_offset);
         }
         let shape = Shape::from(cat_dims);
@@ -235,4 +235,69 @@ impl Tensor {
         }
         Ok(crate::tensor::from_storage(storage, shape, op, false))
     }
+
+    /// Set the values on `self` using values from `src`. The copy starts at the specified
+    /// `offset` for the target dimension `dim` on `self`.
+    /// `self` and `src` must have the same shape except on dimension `dim` where the `self` size
+    /// has to be greater than or equal to `offset` plus the `src` size.
+    ///
+    /// Note that this modifies `self` in place and as such is not compatible with
+    /// back-propagation.  
+    pub fn slice_set<D: Dim>(&self, src: &Self, dim: D, offset: usize) -> Result<()> {
+        let dim = dim.to_index(self.shape(), "slice-set")?;
+        if !self.is_contiguous() || !src.is_contiguous() {
+            Err(Error::RequiresContiguous { op: "slice-set" }.bt())?
+        }
+        if self.same_storage(src) {
+            crate::bail!("cannot use slice_set when self and src share their storage")
+        }
+        if self.dtype() != src.dtype() {
+            Err(Error::DTypeMismatchBinaryOp {
+                lhs: self.dtype(),
+                rhs: src.dtype(),
+                op: "slice-set",
+            }
+            .bt())?
+        }
+        if self.device().location() != src.device().location() {
+            Err(Error::DeviceMismatchBinaryOp {
+                lhs: self.device().location(),
+                rhs: src.device().location(),
+                op: "slice-set",
+            }
+            .bt())?
+        }
+        if self.rank() != src.rank() {
+            Err(Error::UnexpectedNumberOfDims {
+                expected: self.rank(),
+                got: src.rank(),
+                shape: self.shape().clone(),
+            }
+            .bt())?
+        }
+        for (dim_idx, (v1, v2)) in self.dims().iter().zip(src.dims().iter()).enumerate() {
+            if dim_idx == dim && *v2 + offset > *v1 {
+                crate::bail!("shape mismatch on target dim, dst: {v1}, src: {v2} + {offset}")
+            }
+            if dim_idx != dim && v1 != v2 {
+                crate::bail!("shape mismatch on dim {dim_idx}, {v1} <> {v2}")
+            }
+        }
+        let block_size: usize = src.dims().iter().skip(1 + dim).product();
+        let d1: usize = src.dims().iter().take(dim).product();
+        let d2 = block_size * src.dims()[dim];
+        let dst_o = self.layout().start_offset() + offset * block_size;
+        let src_o = src.layout().start_offset();
+        src.storage().copy2d(
+            &mut self.storage_mut(),
+            d1,
+            d2,
+            /* src_s */ d2,
+            /* dst_s */ block_size * self.dims()[dim],
+            src_o,
+            dst_o,
+        )?;
+
+        Ok(())
+    }
 }

candle-core/src/test_utils.rs

@@ -24,6 +24,15 @@ macro_rules! test_device {
     };
 }
 
+pub fn assert_tensor_eq(t1: &Tensor, t2: &Tensor) -> Result<()> {
+    assert_eq!(t1.shape(), t2.shape());
+    // Default U8 may not be large enough to hold the sum (`t.sum_all` defaults to the dtype of `t`)
+    let eq_tensor = t1.eq(t2)?.to_dtype(crate::DType::U32)?;
+    let all_equal = eq_tensor.sum_all()?;
+    assert_eq!(all_equal.to_scalar::<u32>()?, eq_tensor.elem_count() as u32);
+    Ok(())
+}
+
 pub fn to_vec0_round(t: &Tensor, digits: i32) -> Result<f32> {
     let b = 10f32.powi(digits);
     let t = t.to_vec0::<f32>()?;

candle-core/src/utils.rs

@@ -1,3 +1,4 @@
+//! Useful functions for checking features.
 use std::str::FromStr;
 
 pub fn get_num_threads() -> usize {

candle-core/src/variable.rs

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

candle-core/tests/conv_tests.rs

@@ -53,6 +53,20 @@ 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 = {
+        let t = Tensor::cat(&[&t.zeros_like()?, &t, &t.zeros_like()?], 0)?;
+        t.conv1d(&w, /*padding*/ 1, 1, 1, 1)?
+    };
+    assert_eq!(res.dims(), [3, 2, 5]);
+    // Same as pytorch default padding: use zeros.
+    assert_eq!(
+        test_utils::to_vec1_round(&res.i(0)?.flatten_all()?, 4)?,
+        [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(&res.i(1)?.flatten_all()?, 4)?,
+        [2.4509, 2.6357, -1.3336, 4.1393, 0.5657, 1.8091, -1.1784, 3.5675, 0.5069, 3.3352]
+    );
 
     let w = w.transpose(0, 1)?;
     // The CPU kernels applied in the contiguous and non contiguous cases are different.
@@ -163,6 +177,22 @@ fn conv2d(dev: &Device) -> Result<()> {
             10.389, 3.6023, -4.2808, 0.2672, 5.3646, -5.2023, -2.1955, -9.4075
         ]
     );
+    let res = {
+        let t = Tensor::cat(&[&t.zeros_like()?, &t, &t.zeros_like()?], 0)?;
+        t.conv2d(&w, 0, 1, 1, 1)?
+    };
+    assert_eq!(res.dims(), [3, 2, 3, 3]);
+    assert_eq!(
+        test_utils::to_vec1_round(&res.i(0)?.flatten_all()?, 4)?,
+        [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(&res.i(1)?.flatten_all()?, 4)?,
+        [
+            -4.2812, 2.0923, 5.2187, 7.5184, 0.752, -14.9426, 10.0087, 4.391, 0.2918, 1.6715,
+            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)?;
 
@@ -730,6 +760,103 @@ fn conv2d_grad(dev: &Device) -> Result<()> {
             ]
         ]
     );
+
+    // Test the same, but then with the following properties, t & w are unmodified.
+    let padding = 1;
+    let outpadding = 1;
+    let dilation = 1;
+    let stride = 2;
+
+    let res = t.conv_transpose2d(&w, padding, outpadding, stride, dilation)?;
+    let loss = res.sqr()?.sum_all()?;
+    assert_eq!(test_utils::to_vec0_round(&loss, 0)?, 3627.0); // torch gives 3626.8560
+
+    let grads = loss.backward()?;
+
+    let grad_t = grads.get(&t).unwrap();
+    let grad_w = grads.get(&w).unwrap();
+    assert_eq!(grad_t.dims(), [1, 4, 7, 5]);
+    assert_eq!(grad_w.dims(), [4, 2, 3, 5]);
+
+    #[rustfmt::skip]
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_t.i(0)?, 1)?,
+        [
+            [
+                [  13.2,  -40.7,   -9.7,  -47.3,  -82.7],
+                [ -98.2,    9.7,   57.7,   -6.2,  180.7],
+                [ 100.2,   24.1,    3.7, -100.5,  -48.1],
+                [  -0.3,   13.5,   -2.9,   80.0,  -49.8],
+                [  47.2,  -25.6,  -74.4,   61.2,  -18.4],
+                [   4.6,  -69.5,   27.9,   66.5,  -88.1],
+                 // 4th column on next row; torch is 4.2
+                [ -12.0,   79.2,  -40.0,    4.1,  -97.1],
+            ],
+            [
+                [ -42.2,  -36.5,  -51.1,    7.5,   32.3],
+                [  74.1,  -44.6,  -68.8,   19.5,    7.7],
+                [ 137.1,   54.2,  153.8,  -58.0,   45.5],
+                [  24.4,  -56.8,    9.7,  -41.0,  -14.5],
+                [  -3.7,   72.6,    8.3,  134.8,   40.5],
+                [  43.2,  -56.9,  -47.5,  -89.4,  -95.4],
+                [  68.2,  108.1,  -80.0,   57.0, -121.1]
+            ],
+            [
+                [  31.1,  -11.4,  -34.8,   33.1,  -44.2],
+                [  29.4,  -31.6,  -40.2,   13.7,   13.1],
+                [  -0.8,  -83.8,   -7.8,  -17.3,   78.2],
+                [  12.0, -118.7,  137.5,  -76.7,   50.8],
+                [ -28.7, -114.2,   -3.7,  -96.3,  -13.8],
+                [ -31.8,   28.5,  -14.3,    4.6,   13.4],
+                [  28.0,   -0.2,  -38.9,  -29.7,  -59.0]
+            ],
+            [
+                [ -16.8,   38.5,   15.5,   26.6,   48.9],
+                [  14.5,   49.6,  -24.8,   65.6,   61.7],
+                [  22.1,  -64.7,   -4.3,  -51.0,   36.3],
+                [  31.0,  -88.9,   47.1, -123.5,   -3.8],
+                [ -14.8,  -39.8,  128.2, -110.3,   42.6],
+                // 1st column on next row; torch is -7.2
+                [  -7.1,   95.3,  -21.3,  -58.7,  -13.9], 
+                [  26.9,   21.3,   16.1,   70.3,   32.1]
+            ]
+        ]
+    );
+
+    #[rustfmt::skip]
+    assert_eq!(
+        test_utils::to_vec1_round(&grad_w.flatten_all()?, 1)?,
+        [
+            // 2nd value; torch gets -3.2, 3rd value; torch gets 221.8
+           -2.460e+01, -3.100e+00,  2.219e+02,  7.400e+00,  5.620e+01,
+            7.420e+01,  7.830e+01,  8.900e+00,  1.050e+01,  2.810e+01,
+            5.100e+00, -1.046e+02, -1.572e+02,  8.710e+01, -9.840e+01,
+           -4.230e+01, -1.898e+02,  1.860e+01, -3.570e+01,  9.810e+01,
+            4.680e+01,  1.182e+02,  4.020e+01, -1.900e+00,  1.508e+02,
+            1.094e+02,  1.018e+02, -4.620e+01,  1.591e+02, -2.320e+01,
+            // 5th value; torch gets 7.1
+           -8.450e+01, -4.600e+00,  6.330e+01,  1.123e+02, -7.000e+00,
+            1.101e+02, -6.620e+01,  2.090e+01, -5.120e+01,  8.990e+01,
+            9.050e+01, -6.990e+01,  6.800e+01, -9.250e+01,  1.380e+02,
+            4.720e+01,  4.710e+01,  6.210e+01,  8.870e+01,  2.098e+02,
+            3.870e+01, -1.390e+01,  6.270e+01,  1.484e+02, -9.920e+01,
+           -4.200e+01, -1.505e+02, -1.480e+01, -2.620e+01,  8.220e+01,
+           -3.350e+01, -2.260e+01, -1.198e+02, -5.080e+01,  1.259e+02,
+            5.600e+01,  9.270e+01,  1.209e+02,  6.590e+01, -8.330e+01,
+            7.000e+00, -2.600e+01, -1.133e+02,  3.870e+01,  4.020e+01,
+           -6.300e+00, -8.710e+01, -5.150e+01, -8.510e+01,  2.000e-01,
+            3.640e+01, -6.100e+00,  6.590e+01, -2.700e+00,  6.550e+01,
+            // 4th value; torch gets 3.8
+            5.300e+00, -6.760e+01, -4.270e+01, -3.900e+00,  2.880e+01,
+            5.260e+01,  6.170e+01, -1.203e+02, -1.610e+01,  7.740e+01,
+           -1.008e+02, -1.070e+01, -9.900e+00,  3.300e+00, -2.620e+01,
+           -4.440e+01,  2.580e+01, -6.920e+01, -4.220e+01,  1.108e+02,
+            1.240e+01, -3.440e+01, -2.800e+00,  7.880e+01, -6.690e+01,
+            1.480e+01,  2.310e+01, -4.260e+01, -1.500e+00, -4.760e+01,
+            5.350e+01, -2.260e+01,  8.000e-01, -3.840e+01, -2.500e+00
+        ]
+    );
+
     Ok(())
 }
 

candle-core/tests/custom_op_tests.rs

@@ -143,3 +143,39 @@ fn inplace_op1() -> Result<()> {
     );
     Ok(())
 }
+
+#[cfg(any(feature = "cuda", feature = "metal"))]
+#[allow(clippy::approx_constant)]
+#[test]
+fn ug_op() -> Result<()> {
+    let kernel = {
+        use ug::lang::op;
+
+        let layout = ug::Layout::from_shape(&[12]);
+        let ptr = op::Arg::ptr(ug::DType::F32);
+        let src = op::load(ptr.id(), layout.clone(), ug::DType::F32)?;
+        let src = op::unary(op::UnaryOp::Exp, src)?;
+        let st = op::store(ptr.id(), layout, src)?;
+        let kernel = op::Kernel::new("exp".to_string(), vec![ptr], vec![st]);
+        let opts: ug::lower_op::Opts = Default::default();
+        kernel.lower(&opts)?
+    };
+    let device = if candle_core::utils::cuda_is_available() {
+        Device::new_cuda(0)?
+    } else if candle_core::utils::metal_is_available() {
+        Device::new_metal(0)?
+    } else {
+        candle_core::bail!("metal/cuda is mandatory for this test")
+    };
+    let op = candle_core::UgIOp1::new("test", kernel, &device)?;
+    let t = Tensor::arange(0u32, 12u32, &device)?.to_dtype(DType::F32)?;
+    t.inplace_op1(&op)?;
+    assert_eq!(
+        to_vec1_round(&t, 2)?,
+        &[
+            1.0, 2.72, 7.39, 20.09, 54.6, 148.41, 403.43, 1096.63, 2980.96, 8103.08, 22026.47,
+            59874.13
+        ]
+    );
+    Ok(())
+}

candle-core/tests/grad_tests.rs

@@ -1,6 +1,6 @@
 #![allow(clippy::approx_constant)]
 use anyhow::{Context, Result};
-use candle_core::{test_device, test_utils, Device, Shape, Tensor, Var};
+use candle_core::{test_device, test_utils, DType, Device, Shape, Tensor, Var};
 
 fn simple_grad(device: &Device) -> Result<()> {
     let x = Var::new(&[3f32, 1., 4.], device)?;
@@ -505,6 +505,36 @@ fn binary_grad(device: &Device) -> Result<()> {
     Ok(())
 }
 
+#[test]
+fn test_flip_backprop() -> Result<()> {
+    let device = &Device::Cpu;
+
+    // Create a tensor (leaf node) that requires gradients
+    let x = Var::ones((2, 2), DType::F64, device)?;
+    let weights = Tensor::arange(1.0, 5.0, device)?.reshape((2, 2))?;
+
+    let y = x.matmul(&weights)?;
+    let expected_y = Tensor::from_vec(vec![4.0, 6.0, 4.0, 6.0], (2, 2), device)?;
+    candle_core::test_utils::assert_tensor_eq(&y, &expected_y)?;
+
+    let z = y.flip(&[1])?;
+    let expected_z = Tensor::from_vec(vec![6.0, 4.0, 6.0, 4.0], (2, 2), device)?;
+    candle_core::test_utils::assert_tensor_eq(&z, &expected_z)?;
+
+    let loss = z.sum_all()?;
+
+    let grad_store = loss.backward()?;
+    let grad_x = grad_store.get_id(x.id()).unwrap();
+
+    let flipped_weights = weights.flip(&[1])?;
+    let dloss_dy = Tensor::ones((2, 2), DType::F64, device)?;
+    // dloss/dx = dloss/dy @ dy/dx = ones @ weight.flip.T
+    let expected_grad = dloss_dy.matmul(&flipped_weights.t()?)?;
+    candle_core::test_utils::assert_tensor_eq(grad_x, &expected_grad)?;
+
+    Ok(())
+}
+
 test_device!(
     simple_grad,
     simple_grad_cpu,

candle-core/tests/matmul_tests.rs

@@ -49,6 +49,20 @@ fn matmul(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn matmul_bf16(device: &Device) -> Result<()> {
+    if !device.supports_bf16() {
+        return Ok(());
+    }
+    let data = vec![1.0f32, 2.0, 3.0, 4.0];
+    let a = Tensor::from_slice(&data, (2, 2), device)?.to_dtype(DType::BF16)?;
+    let data = vec![1.0f32, 2.0, 3.0, 4.0];
+    let b = Tensor::from_slice(&data, (2, 2), device)?.to_dtype(DType::BF16)?;
+
+    let c = a.matmul(&b)?.to_dtype(DType::F32)?;
+    assert_eq!(c.to_vec2::<f32>()?, &[[7.0f32, 10.0], [15.0, 22.0]]);
+    Ok(())
+}
+
 fn broadcast_matmul(device: &Device) -> Result<()> {
     let lhs = Tensor::randn(0f32, 1f32, (3, 1, 4, 5), device)?;
     let rhs = Tensor::randn(0f32, 1f32, (6, 5, 2), device)?;
@@ -96,6 +110,12 @@ fn mm_layout(device: &Device) -> Result<()> {
 }
 
 test_device!(matmul, matmul_cpu, matmul_gpu, matmul_metal);
+test_device!(
+    matmul_bf16,
+    matmul_bf16_cpu,
+    matmul_bf16_gpu,
+    matmul_bf16_metal
+);
 test_device!(
     broadcast_matmul,
     broadcast_matmul_cpu,

candle-core/tests/quantized_tests.rs

@@ -3,7 +3,7 @@ use candle_core::{
     quantized::{self, GgmlDType},
     test_device,
     test_utils::to_vec2_round,
-    Device, Module, Result, Tensor,
+    DType, Device, IndexOp, Module, Result, Tensor,
 };
 use quantized::{k_quants, GgmlType};
 use rand::prelude::*;
@@ -47,18 +47,14 @@ fn test_matmul(
 }
 
 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), device)?;
+    let lhs_s = (0..(m * k)).map(|v| v as f32).collect::<Vec<_>>();
+    let lhs = Tensor::from_slice(&lhs_s, (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<_>>();
     k_quants::BlockQ4_0::from_float(&rhs, &mut rhs_t)?;
-    k_quants::matmul((m, k, n), &lhs, &rhs_t, &mut dst)?;
+    k_quants::matmul((m, k, n), &lhs_s, &rhs_t, &mut dst)?;
     assert_eq!(
         dst.iter().map(|x| x.round()).collect::<Vec<_>>(),
         &[
@@ -67,7 +63,7 @@ fn quantized_matmul(device: &Device) -> Result<()> {
         ]
     );
     let tensor_rhs = Tensor::from_slice(&rhs, (n, k), device)?.t()?;
-    let mm = tensor_lhs.matmul(&tensor_rhs)?;
+    let mm = lhs.matmul(&tensor_rhs)?;
     assert_eq!(
         mm.to_vec2::<f32>()?,
         &[
@@ -79,7 +75,7 @@ fn quantized_matmul(device: &Device) -> Result<()> {
 
     let qtensor = quantized::QTensor::quantize(&tensor_rhs.t()?, GgmlDType::Q4_0)?;
     let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
-    let res = matmul.forward(&tensor_lhs)?;
+    let res = matmul.forward(&lhs)?;
     match device {
         Device::Metal(_) => assert_eq!(
             to_vec2_round(&res, 0)?,
@@ -89,7 +85,15 @@ fn quantized_matmul(device: &Device) -> Result<()> {
                 [341970.0, 994574.0, 1656181.0, 2302182.0]
             ]
         ),
-        _ => assert_eq!(
+        Device::Cuda(_) => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [84866.0, 214045.0, 344676.0, 473707.0],
+                [213425.0, 604313.0, 1000431.0, 1387960.0],
+                [342030.0, 994630.0, 1656248.0, 2302250.0]
+            ]
+        ),
+        Device::Cpu => assert_eq!(
             to_vec2_round(&res, 0)?,
             &[
                 [85120.0, 214562.0, 345455.0, 474748.0],
@@ -98,22 +102,16 @@ fn quantized_matmul(device: &Device) -> Result<()> {
             ]
         ),
     }
-
     test_matmul(device, (1, 3, 4, 256), GgmlDType::Q4_0)?;
-
     Ok(())
 }
 
 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))
+    let lhs_s = (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), device)?;
+    let lhs = Tensor::from_slice(&lhs_s, (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)
@@ -121,7 +119,7 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
         .collect::<Vec<_>>();
     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)?;
+    k_quants::matmul((m, k, n), &lhs_s, &rhs_t, &mut dst)?;
     assert_eq!(
         dst.iter().map(|x| x.round()).collect::<Vec<_>>(),
         &[
@@ -129,7 +127,7 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
             -196472.0, 63012.0, 324585.0, 587902.0
         ]
     );
-    let mm = tensor_lhs.matmul(&tensor_rhs)?;
+    let mm = lhs.matmul(&tensor_rhs)?;
     assert_eq!(
         to_vec2_round(&mm, 0)?,
         &[
@@ -141,7 +139,7 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
 
     let qtensor = quantized::QTensor::quantize(&tensor_rhs.t()?, GgmlDType::Q4_0)?;
     let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
-    let res = matmul.forward(&tensor_lhs)?;
+    let res = matmul.forward(&lhs)?;
     match device {
         Device::Metal(_) => assert_eq!(
             to_vec2_round(&res, 0)?,
@@ -151,7 +149,15 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
                 [-196102.0, 63022.0, 324233.0, 587191.0]
             ]
         ),
-        _ => assert_eq!(
+        Device::Cuda(_) => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [243740.0, -19762.0, -285476.0, -550498.0],
+                [23774.0, 21645.0, 19395.0, 18364.0],
+                [-196045.0, 63030.0, 324120.0, 587079.0]
+            ]
+        ),
+        Device::Cpu => assert_eq!(
             to_vec2_round(&res, 0)?,
             &[
                 [243524.0, -19596.0, -285051.0, -549815.0],
@@ -160,22 +166,58 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
             ]
         ),
     }
-
+    let lhs2 = Tensor::stack(&[&lhs, &lhs], 0)?;
+    let res2 = matmul.forward(&lhs2)?;
+    let res2 = res2.i(1)?;
+    let diff = (res - res2)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    if device.is_cuda() {
+        assert!(diff < 0.1);
+    } else {
+        assert_eq!(diff, 0.);
+    }
     Ok(())
 }
 
-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 qmm_batch(dev: &Device) -> Result<()> {
+    let (lhs, rhs, _mm) = get_random_tensors(2, 256, 6, dev)?;
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q2K)?;
+    let rhs = quantized::QMatMul::from_qtensor(rhs)?;
+    let mm = rhs.forward(&lhs)?;
+    assert_eq!(mm.shape().dims(), [2, 6]);
+    let lhs2 = Tensor::cat(&[&lhs, &lhs], 0)?;
+    let mm2 = rhs.forward(&lhs2)?;
+    assert_eq!(mm2.shape().dims(), [4, 6]);
+    let diff2 = (mm2.i(2..)? - &mm)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff2, 0.0);
+    let lhs3 = Tensor::cat(&[&lhs2, &lhs], 0)?;
+    let mm3 = rhs.forward(&lhs3)?;
+    assert_eq!(mm3.shape().dims(), [6, 6]);
+    let diff3 = (mm3.i(2..4)? - &mm)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff3, 0.0);
+    let diff3 = (mm3.i(4..)? - &mm)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff3, 0.0);
+    let lhs4 = Tensor::cat(&[&lhs3, &lhs3], 0)?;
+    let mm4 = rhs.forward(&lhs4)?;
+    assert_eq!(mm4.shape().dims(), [12, 6]);
+    let diff4 = (mm4.i(..6)? - &mm3)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    if dev.is_cuda() {
+        // We use a different kernel for sizes from 1 to 8 on cuda which explains
+        // the difference here.
+        assert!(0. < diff4 && diff4 < 1e-4)
+    } else {
+        assert_eq!(diff4, 0.0)
+    };
+    let diff4 = (mm4.i(6..)? - &mm4.i(..6)?)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff4, 0.0);
+    Ok(())
+}
+
+test_device!(quantized_matmul, qmm_cpu, qmm_cuda, qmm_metal);
+test_device!(quantized_matmul_neg, qmm_n_cpu, qmm_n_cuda, qmm_n_metal);
+test_device!(qmm_batch, qmm_b_cpu, qmm_b_cuda, qmm_b_metal);
 
 fn quantize_q4_0(device: &Device) -> Result<()> {
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
@@ -183,6 +225,13 @@ fn quantize_q4_0(device: &Device) -> Result<()> {
     let src = Tensor::from_slice(&src, (32 * 4,), device)?;
     let quant = quantized::QTensor::quantize(&src, GgmlDType::Q4_0)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     assert_eq!(
         dst.to_vec1::<f32>()?,
         &[
@@ -209,6 +258,13 @@ fn quantize_q4_1(device: &Device) -> Result<()> {
     let src = Tensor::from_slice(&src, (32 * 4,), device)?;
     let quant = quantized::QTensor::quantize(&src, GgmlDType::Q4_1)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     assert_eq!(
         round_vector(&dst.to_vec1::<f32>()?),
         &[
@@ -235,6 +291,13 @@ fn quantize_q5_0(device: &Device) -> Result<()> {
     let src = Tensor::from_slice(&src, (32 * 4,), device)?;
     let quant = quantized::QTensor::quantize(&src, GgmlDType::Q5_0)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     assert_eq!(
         round_vector(&dst.to_vec1::<f32>()?),
         &[
@@ -261,6 +324,13 @@ fn quantize_q5_1(device: &Device) -> Result<()> {
     let src = Tensor::from_slice(&src, (32 * 4,), device)?;
     let quant = quantized::QTensor::quantize(&src, GgmlDType::Q5_1)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     assert_eq!(
         round_vector(&dst.to_vec1::<f32>()?),
         &[
@@ -345,6 +415,13 @@ fn ggml_quantization_error_test(dtype: GgmlDType, device: &Device, max_error: f3
     let src = Tensor::from_slice(&src, (GGML_TEST_SIZE,), device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     let error = calculate_rmse(&src.to_vec1::<f32>()?, &dst.to_vec1::<f32>()?);
     if error > max_error {
         bail!(
@@ -362,6 +439,13 @@ fn quantize_q2k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -381,6 +465,13 @@ fn quantize_q2k(device: &Device) -> Result<()> {
     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 dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -395,6 +486,13 @@ fn quantize_q3k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -414,6 +512,13 @@ fn quantize_q3k(device: &Device) -> Result<()> {
     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 dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -428,6 +533,13 @@ fn quantize_q4k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -447,6 +559,13 @@ fn quantize_q4k(device: &Device) -> Result<()> {
     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 dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -461,6 +580,13 @@ fn quantize_q5k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -480,6 +606,13 @@ fn quantize_q5k(device: &Device) -> Result<()> {
     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 dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -494,6 +627,13 @@ fn quantize_q6k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -513,6 +653,13 @@ fn quantize_q6k(device: &Device) -> Result<()> {
     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 dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -527,6 +674,13 @@ fn quantize_q8k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -546,6 +700,13 @@ fn quantize_q8k(device: &Device) -> Result<()> {
     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 dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -719,10 +880,10 @@ fn get_random_tensors(
     let mut rng = StdRng::seed_from_u64(314159265358979);
 
     let lhs = (0..m * k)
-        .map(|_| rng.gen::<f32>() - 0.5)
+        .map(|_| rng.random::<f32>() - 0.5)
         .collect::<Vec<_>>();
     let rhs = (0..n * k)
-        .map(|_| rng.gen::<f32>() - 0.5)
+        .map(|_| rng.random::<f32>() - 0.5)
         .collect::<Vec<_>>();
 
     let lhs = Tensor::from_vec(lhs, (m, k), device)?;

candle-core/tests/serialization_tests.rs

@@ -1,5 +1,31 @@
 use candle_core::{DType, Result, Tensor};
 
+struct TmpFile(std::path::PathBuf);
+
+impl TmpFile {
+    fn create(base: &str) -> TmpFile {
+        let filename = std::env::temp_dir().join(format!(
+            "candle-{}-{}-{:?}",
+            base,
+            std::process::id(),
+            std::thread::current().id(),
+        ));
+        TmpFile(filename)
+    }
+}
+
+impl std::convert::AsRef<std::path::Path> for TmpFile {
+    fn as_ref(&self) -> &std::path::Path {
+        self.0.as_path()
+    }
+}
+
+impl Drop for TmpFile {
+    fn drop(&mut self) {
+        std::fs::remove_file(&self.0).unwrap()
+    }
+}
+
 #[test]
 fn npy() -> Result<()> {
     let npy = Tensor::read_npy("tests/test.npy")?;
@@ -22,3 +48,24 @@ fn npz() -> Result<()> {
     );
     Ok(())
 }
+
+#[test]
+fn safetensors() -> Result<()> {
+    use candle_core::safetensors::Load;
+
+    let tmp_file = TmpFile::create("st");
+    let t = Tensor::arange(0f32, 24f32, &candle_core::Device::Cpu)?;
+    t.save_safetensors("t", &tmp_file)?;
+    // Load from file.
+    let st = candle_core::safetensors::load(&tmp_file, &candle_core::Device::Cpu)?;
+    let t2 = st.get("t").unwrap();
+    let diff = (&t - t2)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff, 0f32);
+    // Load from bytes.
+    let bytes = std::fs::read(tmp_file)?;
+    let st = candle_core::safetensors::SliceSafetensors::new(&bytes)?;
+    let t2 = st.get("t").unwrap().load(&candle_core::Device::Cpu);
+    let diff = (&t - t2)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff, 0f32);
+    Ok(())
+}

candle-core/tests/tensor_tests.rs

@@ -25,14 +25,66 @@ fn ones(device: &Device) -> Result<()> {
         Tensor::ones((2, 3), DType::F32, device)?.to_vec2::<f32>()?,
         [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
     );
+    if !device.is_metal() {
+        assert_eq!(
+            Tensor::ones((2, 3), DType::F64, device)?.to_vec2::<f64>()?,
+            [[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]],
+        Tensor::ones((2, 3), DType::F16, device)?.to_vec2::<half::f16>()?,
+        [
+            [
+                half::f16::from_f32(1.0),
+                half::f16::from_f32(1.0),
+                half::f16::from_f32(1.0)
+            ],
+            [
+                half::f16::from_f32(1.0),
+                half::f16::from_f32(1.0),
+                half::f16::from_f32(1.0)
+            ]
+        ],
+    );
+    assert_eq!(
+        Tensor::ones((2, 3), DType::BF16, device)?.to_vec2::<half::bf16>()?,
+        [
+            [
+                half::bf16::from_f32(1.0),
+                half::bf16::from_f32(1.0),
+                half::bf16::from_f32(1.0)
+            ],
+            [
+                half::bf16::from_f32(1.0),
+                half::bf16::from_f32(1.0),
+                half::bf16::from_f32(1.0)
+            ]
+        ],
     );
     Ok(())
 }
 
 fn full(device: &Device) -> Result<()> {
+    let tensor = Tensor::zeros((3, 4), DType::U32, device)?;
+    tensor.const_set(42u32.into())?;
+    assert_eq!(
+        tensor.to_vec2::<u32>()?,
+        [[42, 42, 42, 42], [42, 42, 42, 42], [42, 42, 42, 42]]
+    );
+    tensor.i((.., 2))?.const_set(1337u32.into())?;
+    assert_eq!(
+        tensor.to_vec2::<u32>()?,
+        [[42, 42, 1337, 42], [42, 42, 1337, 42], [42, 42, 1337, 42]]
+    );
+    tensor.i((2, ..))?.const_set(1u32.into())?;
+    assert_eq!(
+        tensor.to_vec2::<u32>()?,
+        [[42, 42, 1337, 42], [42, 42, 1337, 42], [1, 1, 1, 1]]
+    );
+    Ok(())
+}
+
+fn const_set(device: &Device) -> Result<()> {
     assert_eq!(
         Tensor::full(42u32, (2, 3), device)?.to_vec2::<u32>()?,
         [[42, 42, 42], [42, 42, 42]],
@@ -96,6 +148,40 @@ fn clamp(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn asort(device: &Device) -> Result<()> {
+    let data = &[[3f32, 1., 4., 1.1, 5.], [2.1, 1., 7., 8., 2.]];
+    let tensor = Tensor::new(data, device)?;
+    let indexes = tensor.arg_sort_last_dim(true)?;
+    assert_eq!(
+        indexes.to_vec2::<u32>()?,
+        [[1, 3, 0, 2, 4], [1, 4, 0, 2, 3]],
+    );
+    let indexes = tensor.arg_sort_last_dim(false)?;
+    assert_eq!(
+        indexes.to_vec2::<u32>()?,
+        [[4, 2, 0, 3, 1], [3, 2, 0, 4, 1]],
+    );
+    let (sorted, indexes) = tensor.sort_last_dim(true)?;
+    assert_eq!(
+        indexes.to_vec2::<u32>()?,
+        [[1, 3, 0, 2, 4], [1, 4, 0, 2, 3]],
+    );
+    assert_eq!(
+        sorted.to_vec2::<f32>()?,
+        [[1.0, 1.1, 3.0, 4.0, 5.0], [1.0, 2.0, 2.1, 7.0, 8.0]]
+    );
+    let (sorted, indexes) = tensor.sort_last_dim(false)?;
+    assert_eq!(
+        indexes.to_vec2::<u32>()?,
+        [[4, 2, 0, 3, 1], [3, 2, 0, 4, 1]],
+    );
+    assert_eq!(
+        sorted.to_vec2::<f32>()?,
+        [[5.0, 4.0, 3.0, 1.1, 1.0], [8.0, 7.0, 2.1, 2.0, 1.0]]
+    );
+    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)?;
@@ -159,6 +245,19 @@ fn unary_op(device: &Device) -> Result<()> {
         tensor.sign()?.to_vec1::<f32>()?,
         [-1., -1., -1., 0., 0., 1., 1., 1., 1.]
     );
+    let tensor = Tensor::new(&[-1.0f32, 0., -2., 3.], device)?;
+    let y = tensor.elu(2.)?;
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [-1.2642, 0.0000, -1.7293, 3.0000]
+    );
+    // This test failed on metal prior to the following PR:
+    // https://github.com/huggingface/candle/pull/2490
+    let y = tensor.reshape((2, 2))?.t()?.elu(2.)?.flatten_all()?;
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [-1.2642, -1.7293, 0.0000, 3.0000]
+    );
     Ok(())
 }
 
@@ -631,6 +730,32 @@ fn broadcast(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn slice_set(device: &Device) -> Result<()> {
+    let (b, h, max_t, d) = (2, 4, 7, 3);
+    let cache = Tensor::zeros((b, h, max_t, d), DType::F32, device)?;
+    let tensor = Tensor::randn(0f32, 1f32, (b, h, 4, d), device)?;
+    cache.slice_set(&tensor, 2, 0)?;
+    let cache_t = cache.narrow(2, 0, 4)?;
+    let diff = (cache_t - &tensor)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
+    cache.slice_set(&tensor, 2, 1)?;
+    let cache_t = cache.narrow(2, 1, 4)?;
+    let diff = (cache_t - &tensor)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
+    let ones = Tensor::ones((b, h, 1, d), DType::F32, device)?;
+    cache.slice_set(&ones, 2, 6)?;
+    let diff = cache.narrow(2, 5, 1)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
+    let diff = (cache.narrow(2, 6, 1)? - 1.)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
+    // This used to create a deadlock rather than returning an actual error.
+    assert!(cache.slice_set(&cache, 0, 0).is_err());
+    Ok(())
+}
+
 fn cat(device: &Device) -> Result<()> {
     // 1D
     let t1 = Tensor::new(&[3f32, 1., 4.], device)?;
@@ -723,6 +848,31 @@ fn embeddings(device: &Device) -> Result<()> {
     Ok(())
 }
 
+#[test]
+fn index_select_fail() -> Result<()> {
+    // Check that an error is properly reported on out of bounds.
+    let ids = Tensor::new(&[4u32, 2u32, 1u32], &Device::Cpu)?;
+    let t = Tensor::new(&[[0f32, 1f32], [2f32, 3f32], [4f32, 5f32]], &Device::Cpu)?;
+    let hs = t.index_select(&ids, 0);
+    assert!(hs.is_err());
+    Ok(())
+}
+
+// The test below triggers an unwinding panic as there is a panic within the
+// #[cfg(feature = "cuda")]
+// #[test]
+// #[should_panic]
+// fn index_select_fail_gpu() {
+//     // Check that a panic happens for out of bounds in cuda
+//     if let Ok(device) = Device::new_cuda(0) {
+//         if let Ok(ids) = Tensor::new(&[4u32, 2u32, 1u32], &device) {
+//             if let Ok(t) = Tensor::new(&[[0f32, 1f32], [2f32, 3f32], [4f32, 5f32]], &device) {
+//                 let _ = t.index_select(&ids, 0);
+//             }
+//         }
+//     }
+// }
+
 fn cmp(device: &Device) -> Result<()> {
     let t1 = Tensor::new(&[[0f32, 1f32], [2f32, 3f32], [4f32, 5f32]], device)?;
     let t2 = Tensor::new(&[[1f32, 0f32], [3f32, 3f32], [4f32, 7f32]], device)?;
@@ -877,7 +1027,7 @@ fn slice_scatter(device: &Device) -> Result<()> {
     Ok(())
 }
 
-fn scatter_add(device: &Device) -> Result<()> {
+fn scatter(device: &Device) -> Result<()> {
     let t = Tensor::arange(0f32, 12f32, device)?.reshape((4, 3))?;
     assert_eq!(
         t.to_vec2::<f32>()?,
@@ -901,6 +1051,17 @@ fn scatter_add(device: &Device) -> Result<()> {
         ]
     );
 
+    let hs = init.scatter(&ids, &t, 1)?;
+    assert_eq!(
+        hs.to_vec2::<f32>()?,
+        &[
+            [0.0, 1.0, 2.0, 1.0, 1.0],
+            [5.0, 1.0, 1.0, 3.0, 4.0],
+            [1.0, 8.0, 1.0, 7.0, 1.0],
+            [10.0, 1.0, 9.0, 1.0, 11.0]
+        ]
+    );
+
     let init = Tensor::ones((6, 3), DType::F32, device)?;
     let hs = init.scatter_add(&ids, &t, 0)?;
     assert_eq!(
@@ -914,6 +1075,30 @@ fn scatter_add(device: &Device) -> Result<()> {
             [1.0, 1.0, 1.0]
         ]
     );
+    let hs = init.scatter(&ids, &t, 0)?;
+    assert_eq!(
+        hs.to_vec2::<f32>()?,
+        &[
+            [0.0, 10.0, 5.0],
+            [1.0, 1.0, 8.0],
+            [9.0, 1.0, 2.0],
+            [6.0, 7.0, 1.0],
+            [1.0, 4.0, 11.0],
+            [1.0, 1.0, 1.0]
+        ]
+    );
+    init.scatter_set(&ids, &t, 0)?;
+    assert_eq!(
+        init.to_vec2::<f32>()?,
+        &[
+            [0.0, 10.0, 5.0],
+            [1.0, 1.0, 8.0],
+            [9.0, 1.0, 2.0],
+            [6.0, 7.0, 1.0],
+            [1.0, 4.0, 11.0],
+            [1.0, 1.0, 1.0]
+        ]
+    );
     Ok(())
 }
 
@@ -946,6 +1131,280 @@ fn gather(device: &Device) -> Result<()> {
     let ids = Tensor::new(&[[0u32, 2u32, 0u32], [0u32, 1u32, 1u32]], device)?;
     let hs = t.gather(&ids, 0)?;
     assert_eq!(hs.to_vec2::<f32>()?, &[[0.0, 7.0, 2.0], [0.0, 4.0, 5.0]]);
+
+    // Random data
+
+    // Dim: 0
+    let t = Tensor::new(
+        &[
+            [
+                [108_f32, -47., 16., -56., -83., -130., 210.],
+                [253., 95., 151., 228., -210., -123., -127.],
+                [-9., -217., 2., -78., 163., 245., -204.],
+                [-246., 79., -238., 88., -226., -184., 171.],
+                [8., -48., -153., 234., -34., 166., -153.],
+                [124., 0., -10., -61., -242., -15., -238.],
+            ],
+            [
+                [12., -64., -199., 244., -240., 156., -128.],
+                [173., -57., 4., -198., 233., -110., 238.],
+                [95., 82., 0., 240., 53., -211., 209.],
+                [-122., 167., -212., 227., -144., 61., 118.],
+                [-63., -146., 200., 244., 168., -167., 116.],
+                [-125., -147., 110., -253., -178., -250., -18.],
+            ],
+            [
+                [57., 86., -50., 56., 92., 205., -78.],
+                [-137., -156., -18., 248., -61., -239., 14.],
+                [-248., -30., -50., -70., -251., 250., -83.],
+                [-221., 67., 72., 59., -24., -154., 232.],
+                [-144., -23., -74., 5., 93., 171., 205.],
+                [46., -77., -38., -226., 246., 161., -17.],
+            ],
+            [
+                [-153., -231., -236., 161., 126., 2., -22.],
+                [-229., -41., 209., 164., 234., 160., 57.],
+                [223., 254., -186., -162., -46., -160., -102.],
+                [65., 30., 213., -253., 59., 224., -154.],
+                [-82., -203., -177., 17., 31., -256., -246.],
+                [176., -135., -65., 54., -56., 210., 76.],
+            ],
+            [
+                [-10., -245., 168., 124., -14., -33., -178.],
+                [25., -43., -39., 132., -89., 169., 179.],
+                [187., -215., 32., -133., 87., -7., -168.],
+                [-224., -215., -5., -230., -58., -162., 128.],
+                [158., -137., -122., -100., -202., -83., 136.],
+                [30., -185., -144., 250., 209., -40., 127.],
+            ],
+            [
+                [-196., 108., -245., 122., 146., -228., 62.],
+                [-1., -66., 160., 137., 13., -172., -21.],
+                [244., 199., -164., 28., 119., -175., 198.],
+                [-62., 253., -162., 195., -95., -230., -211.],
+                [123., -72., -26., -107., -139., 64., 245.],
+                [11., -126., -182., 108., -12., 184., -127.],
+            ],
+            [
+                [-159., 126., 176., 161., 73., -111., -138.],
+                [-187., 214., -217., -33., -223., -201., -212.],
+                [-61., -120., -166., -172., -95., 53., 196.],
+                [-33., 86., 134., -152., 154., -53., 74.],
+                [186., -28., -154., -174., 141., -109., 217.],
+                [82., 35., 252., 145., 181., 74., -87.],
+            ],
+        ],
+        device,
+    )?;
+
+    let ids = Tensor::new(
+        &[
+            [
+                [6_u32, 6, 4, 3, 4, 4, 6],
+                [3, 3, 2, 4, 4, 4, 6],
+                [3, 3, 0, 2, 4, 6, 4],
+                [2, 5, 1, 2, 6, 6, 1],
+                [2, 1, 6, 5, 3, 2, 3],
+                [6, 1, 0, 1, 0, 2, 6],
+            ],
+            [
+                [4, 6, 4, 3, 3, 3, 2],
+                [4, 3, 2, 4, 4, 4, 6],
+                [2, 3, 0, 2, 4, 6, 4],
+                [6, 5, 1, 2, 6, 6, 1],
+                [4, 1, 6, 5, 3, 2, 3],
+                [1, 1, 0, 1, 0, 2, 6],
+            ],
+            [
+                [3, 6, 4, 3, 3, 3, 2],
+                [2, 3, 2, 4, 4, 4, 6],
+                [4, 3, 0, 2, 4, 6, 4],
+                [0, 5, 1, 2, 6, 6, 1],
+                [6, 1, 6, 5, 3, 2, 3],
+                [4, 1, 0, 1, 0, 2, 6],
+            ],
+            [
+                [0, 6, 4, 3, 3, 3, 2],
+                [5, 3, 2, 4, 4, 4, 6],
+                [0, 3, 0, 2, 4, 6, 4],
+                [3, 5, 1, 2, 6, 6, 1],
+                [0, 1, 6, 5, 3, 2, 3],
+                [3, 1, 0, 1, 0, 2, 6],
+            ],
+        ],
+        device,
+    )?;
+
+    let hs = t.gather(&ids, 0)?;
+    assert_eq!(
+        hs.to_vec3::<f32>()?,
+        &[
+            [
+                [-159_f32, 126., 168., 161., -14., -33., -138.],
+                [-229., -41., -18., 132., -89., 169., -212.],
+                [223., 254., 2., -70., 87., 53., -168.],
+                [-221., 253., -212., 59., 154., -53., 118.],
+                [-144., -146., -154., -107., 31., 171., -246.],
+                [82., -147., -10., -253., -242., 161., -87.]
+            ],
+            [
+                [-10., 126., 168., 161., 126., 2., -78.],
+                [25., -41., -18., 132., -89., 169., -212.],
+                [-248., 254., 2., -70., 87., 53., -168.],
+                [-33., 253., -212., 59., 154., -53., 118.],
+                [158., -146., -154., -107., 31., 171., -246.],
+                [-125., -147., -10., -253., -242., 161., -87.]
+            ],
+            [
+                [-153., 126., 168., 161., 126., 2., -78.],
+                [-137., -41., -18., 132., -89., 169., -212.],
+                [187., 254., 2., -70., 87., 53., -168.],
+                [-246., 253., -212., 59., 154., -53., 118.],
+                [186., -146., -154., -107., 31., 171., -246.],
+                [30., -147., -10., -253., -242., 161., -87.]
+            ],
+            [
+                [108., 126., 168., 161., 126., 2., -78.],
+                [-1., -41., -18., 132., -89., 169., -212.],
+                [-9., 254., 2., -70., 87., 53., -168.],
+                [65., 253., -212., 59., 154., -53., 118.],
+                [8., -146., -154., -107., 31., 171., -246.],
+                [176., -147., -10., -253., -242., 161., -87.]
+            ]
+        ]
+    );
+
+    // Dim: 1
+    let t = Tensor::new(
+        &[
+            [
+                [-117_f32, -175., 69., -163.],
+                [200., 242., -21., -67.],
+                [179., 150., -126., -75.],
+                [-118., 38., -138., -13.],
+                [-221., 136., -185., 180.],
+                [58., 182., -204., -149.],
+            ],
+            [
+                [3., -148., -58., -154.],
+                [-43., 45., -108., 4.],
+                [-69., -249., -71., -21.],
+                [80., 110., -152., -235.],
+                [-88., 7., 92., -250.],
+                [-186., 207., -242., 98.],
+            ],
+            [
+                [238., 19., 64., -242.],
+                [-150., -97., 218., 58.],
+                [111., -233., 204., -212.],
+                [-242., -232., 83., 42.],
+                [153., 62., -251., 219.],
+                [-117., 36., -119., 10.],
+            ],
+            [
+                [215., 159., -169., -27.],
+                [-83., 101., -88., 169.],
+                [-205., 93., 225., -64.],
+                [-162., 240., 214., 23.],
+                [-112., 6., 21., 245.],
+                [-38., 113., 93., 215.],
+            ],
+            [
+                [91., -188., -148., 101.],
+                [74., 203., -35., 55.],
+                [-116., -130., -153., -96.],
+                [58., 22., -45., -194.],
+                [-221., -134., 73., 159.],
+                [-203., -254., 31., 235.],
+            ],
+            [
+                [105., -53., 61., 186.],
+                [-195., 234., 75., -1.],
+                [51., 139., 160., -108.],
+                [-173., -167., 161., 19.],
+                [83., -246., 156., -222.],
+                [109., 39., -149., 137.],
+            ],
+        ],
+        device,
+    )?;
+
+    let ids = Tensor::new(
+        &[
+            [[4_u32, 4, 4, 2]],
+            [[0, 4, 4, 3]],
+            [[1, 5, 3, 4]],
+            [[0, 3, 3, 2]],
+            [[1, 1, 5, 2]],
+            [[1, 4, 5, 4]],
+        ],
+        device,
+    )?;
+
+    let hs = t.gather(&ids, 1)?;
+    assert_eq!(
+        hs.to_vec3::<f32>()?,
+        &[
+            [[-221., 136., -185., -75.]],
+            [[3., 7., 92., -235.]],
+            [[-150., 36., 83., 219.]],
+            [[215., 240., 214., -64.]],
+            [[74., 203., 31., -96.]],
+            [[-195., -246., -149., -222.]]
+        ]
+    );
+
+    // Dim: 2
+    let t = Tensor::new(
+        &[
+            [[-162_f32, 202.], [-126., -39.], [35., -65.], [1., 80.]],
+            [[37., 248.], [-191., 89.], [117., -40.], [-217., 220.]],
+        ],
+        device,
+    )?;
+
+    let ids = Tensor::new(&[[[1_u32], [0], [1], [1]], [[0], [1], [0], [1]]], device)?;
+
+    let hs = t.gather(&ids, 2)?;
+    assert_eq!(
+        hs.to_vec3::<f32>()?,
+        &[
+            [[202.], [-126.], [-65.], [80.]],
+            [[37.], [89.], [117.], [220.]]
+        ]
+    );
+
+    let t = Tensor::new(
+        &[
+            [[-21_f32, -197.], [194., 122.]],
+            [[255., -106.], [-191., 250.]],
+            [[33., -117.], [43., 10.]],
+            [[-130., 238.], [-217., -92.]],
+        ],
+        device,
+    )?;
+
+    let ids = Tensor::new(
+        &[
+            [[0_u32, 1], [1, 0]],
+            [[1, 0], [0, 1]],
+            [[0, 1], [0, 1]],
+            [[1, 0], [1, 0]],
+        ],
+        device,
+    )?;
+
+    let hs = t.gather(&ids, 2)?;
+    assert_eq!(
+        hs.to_vec3::<f32>()?,
+        &[
+            [[-21., -197.], [122., 194.]],
+            [[-106., 255.], [-191., 250.]],
+            [[33., -117.], [43., 10.]],
+            [[238., -130.], [-92., -217.]]
+        ]
+    );
+
     Ok(())
 }
 
@@ -1083,14 +1542,37 @@ fn randn(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn zero_dim(device: &Device) -> Result<()> {
+    let t = Tensor::zeros((4, 0, 1), DType::F32, device)?;
+    assert_eq!(t.dims3()?, (4, 0, 1));
+    let t2 = Tensor::zeros((4, 3, 1), DType::F32, device)?;
+    let t_cat = Tensor::cat(&[&t, &t2], 1)?;
+    assert_eq!(t_cat.dims3()?, (4, 3, 1));
+    let t_cat = Tensor::cat(&[&t, &t], 1)?;
+    assert_eq!(t_cat.dims3()?, (4, 0, 1));
+    let t_unary = t.sqrt()?;
+    assert_eq!(t_unary.dims3()?, (4, 0, 1));
+    let t_plus = (&t + 1.)?;
+    assert_eq!(t_plus.dims3()?, (4, 0, 1));
+    let t_mm = t2.matmul(&t.t()?)?;
+    assert_eq!(t_mm.dims3()?, (4, 3, 0));
+    let t_mm = t.matmul(&t2.t()?)?;
+    assert_eq!(t_mm.dims3()?, (4, 0, 3));
+    let t_mm = t.t()?.matmul(&t)?;
+    assert_eq!(t_mm.dims3()?, (4, 1, 1));
+    Ok(())
+}
+
 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!(const_set, cs_cpu, cs_gpu, cs_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!(slice_set, ss_cpu, ss_gpu, ss_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);
@@ -1116,12 +1598,7 @@ test_device!(
 );
 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!(scatter, scatter_cpu, scatter_gpu, scatter_metal);
 test_device!(
     slice_scatter,
     slice_scatter_cpu,
@@ -1130,7 +1607,9 @@ test_device!(
 );
 test_device!(randn, randn_cpu, randn_gpu, randn_metal);
 test_device!(clamp, clamp_cpu, clamp_gpu, clamp_metal);
+test_device!(asort, asort_cpu, asort_gpu, asort_metal);
 test_device!(var, var_cpu, var_gpu, var_metal);
+test_device!(zero_dim, zero_dim_cpu, zero_dim_gpu, zero_dim_metal);
 
 // There was originally a bug on the CPU implementation for randn
 // https://github.com/huggingface/candle/issues/381
@@ -1244,11 +1723,29 @@ fn assert_close(a: &Tensor, b: &Tensor, epsilon: f64) -> Result<()> {
 
 #[test]
 fn log_sum_exp() -> Result<()> {
-    let input = Tensor::new(&[[1f64, 2., 3.], [4., 5., 6.]], &Device::Cpu)?;
+    let input = Tensor::new(
+        &[
+            [[1f64, 2., 3.], [4., 5., 6.]],
+            [[-1000.0, -999.0, -1001.0], [1000.0, 999.0, 1001.0]],
+        ],
+        &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)?;
+    let expected = Tensor::new(&[[3.4076, 6.4076], [-998.5924, 1001.4076]], &Device::Cpu)?;
+    assert_eq!(output.dims(), expected.dims());
+    assert_close(&output.flatten_all()?, &expected.flatten_all()?, 0.00001)?;
+
+    assert_eq!(
+        input.log_sum_exp((0, 1))?.to_vec1::<f64>()?,
+        [1000.0, 999.0, 1001.0]
+    );
+    assert_eq!(
+        input.log_sum_exp(())?.to_vec3::<f64>()?,
+        input.to_vec3::<f64>()?
+    );
+
     Ok(())
 }
 
@@ -1263,3 +1760,77 @@ fn pow() -> Result<()> {
     );
     Ok(())
 }
+
+#[test]
+fn test_flip_1d() -> Result<()> {
+    // 1D: [0, 1, 2, 3, 4]
+    let t = Tensor::arange(0.0, 5.0, &Device::Cpu)?.reshape((5,))?;
+    let flipped = t.flip(&[0])?;
+    // Expected: [4, 3, 2, 1, 0]
+    let expected = Tensor::from_vec(vec![4.0, 3.0, 2.0, 1.0, 0.0], (5,), &Device::Cpu)?;
+    candle_core::test_utils::assert_tensor_eq(&flipped, &expected)?;
+    Ok(())
+}
+
+#[test]
+fn test_flip_2d() -> Result<()> {
+    // 2D:
+    // [[0, 1, 2],
+    //  [3, 4, 5]]
+    let t = Tensor::arange(0.0, 6.0, &Device::Cpu)?.reshape((2, 3))?;
+    let flipped = t.flip(&[0, 1])?;
+    // Expected:
+    // [[5, 4, 3],
+    //  [2, 1, 0]]
+    let expected = Tensor::from_vec(vec![5.0, 4.0, 3.0, 2.0, 1.0, 0.0], (2, 3), &Device::Cpu)?;
+    candle_core::test_utils::assert_tensor_eq(&flipped, &expected)?;
+    Ok(())
+}
+
+#[test]
+fn test_flip_3d_channels() -> Result<()> {
+    // 3D:
+    // [[[0,1,2],
+    //   [3,4,5]],
+    //
+    //  [[6,7,8],
+    //   [9,10,11]]]
+    let t = Tensor::arange(0.0, 12.0, &Device::Cpu)?.reshape((2, 2, 3))?;
+    let flipped = t.flip(&[2])?;
+    // Expected:
+    // [[[2,1,0],
+    //   [5,4,3]],
+    //
+    //  [[8,7,6],
+    //   [11,10,9]]]
+    let expected = Tensor::from_vec(
+        vec![2.0, 1.0, 0.0, 5.0, 4.0, 3.0, 8.0, 7.0, 6.0, 11.0, 10.0, 9.0],
+        (2, 2, 3),
+        &Device::Cpu,
+    )?;
+    candle_core::test_utils::assert_tensor_eq(&flipped, &expected)?;
+    Ok(())
+}
+
+#[test]
+fn tensor_new() -> Result<()> {
+    let t1 = Tensor::new(vec![1f32, 2.0, 3.0], &Device::Cpu)?;
+    assert_eq!(t1.to_vec1::<f32>()?, [1.0, 2.0, 3.0]);
+    let t2 = Tensor::new(vec![vec![1f32, 2., 3.], vec![4., 5., 6.]], &Device::Cpu)?;
+    assert_eq!(t2.to_vec2::<f32>()?, [[1., 2., 3.], [4., 5., 6.]]);
+    let t3 = Tensor::new(
+        vec![
+            vec![vec![1f32, 2., 3.], vec![4., 5., 6.]],
+            vec![vec![3f32, 1., 4.], vec![1., 5., 9.]],
+        ],
+        &Device::Cpu,
+    )?;
+    assert_eq!(
+        t3.to_vec3::<f32>()?,
+        [
+            [[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]],
+            [[3.0, 1.0, 4.0], [1.0, 5.0, 9.0]]
+        ]
+    );
+    Ok(())
+}

candle-datasets/src/batcher.rs

@@ -78,7 +78,7 @@ impl<I: Iterator<Item = Tensor>> Iterator for Batcher<Iter1<I>> {
             match self.inner.inner.next() {
                 Some(item) => items.push(item),
                 None => {
-                    if self.return_last_incomplete_batch {
+                    if self.return_last_incomplete_batch && !items.is_empty() {
                         break;
                     }
                     return None;
@@ -102,7 +102,7 @@ impl<I: Iterator<Item = (Tensor, Tensor)>> Iterator for Batcher<Iter2<I>> {
                     ys.push(y)
                 }
                 None => {
-                    if self.return_last_incomplete_batch {
+                    if self.return_last_incomplete_batch && !xs.is_empty() && !ys.is_empty() {
                         break;
                     }
                     return None;
@@ -127,7 +127,7 @@ impl<I: Iterator<Item = Result<Tensor>>> Iterator for Batcher<IterResult1<I>> {
             match self.inner.inner.next() {
                 Some(item) => items.push(item),
                 None => {
-                    if self.return_last_incomplete_batch {
+                    if self.return_last_incomplete_batch && !items.is_empty() {
                         break;
                     }
                     return None;
@@ -154,7 +154,7 @@ impl<I: Iterator<Item = Result<(Tensor, Tensor)>>> Iterator for Batcher<IterResu
                 }
                 Some(Err(err)) => errs.push(err),
                 None => {
-                    if self.return_last_incomplete_batch {
+                    if self.return_last_incomplete_batch && !xs.is_empty() && !ys.is_empty() {
                         break;
                     }
                     return None;

candle-datasets/src/nlp/tinystories.rs

@@ -60,8 +60,8 @@ pub struct DatasetRandomIter<'a> {
 
 impl<'a> DatasetRandomIter<'a> {
     pub fn new(ds: &'a Dataset, valid: bool, seq_len: usize, device: Device) -> Self {
+        use rand::rng;
         use rand::seq::SliceRandom;
-        use rand::thread_rng;
 
         let all_tokens = if valid {
             &ds.valid_tokens
@@ -69,13 +69,13 @@ impl<'a> DatasetRandomIter<'a> {
             &ds.train_tokens
         };
         let mut tokens = all_tokens.iter().collect::<Vec<_>>();
-        tokens.shuffle(&mut thread_rng());
+        tokens.shuffle(&mut rng());
         let current_tokens = tokens.pop().unwrap();
         let seq_len_in_bytes = seq_len * 2;
         let mut indexes_in_bytes = (0..current_tokens.len() - seq_len_in_bytes)
             .step_by(seq_len_in_bytes)
             .collect::<Vec<_>>();
-        indexes_in_bytes.shuffle(&mut thread_rng());
+        indexes_in_bytes.shuffle(&mut rng());
         Self {
             all_tokens,
             tokens,
@@ -87,26 +87,26 @@ impl<'a> DatasetRandomIter<'a> {
     }
 }
 
-impl<'a> Iterator for DatasetRandomIter<'a> {
+impl Iterator for DatasetRandomIter<'_> {
     type Item = Result<(Tensor, Tensor)>;
 
     fn next(&mut self) -> Option<Self::Item> {
         use byteorder::{LittleEndian, ReadBytesExt};
+        use rand::rng;
         use rand::seq::SliceRandom;
-        use rand::thread_rng;
 
         let seq_len = self.seq_len;
         if self.indexes_in_bytes.is_empty() {
             if self.tokens.is_empty() {
                 self.tokens = self.all_tokens.iter().collect();
-                self.tokens.shuffle(&mut thread_rng());
+                self.tokens.shuffle(&mut rng());
             }
             self.current_tokens = self.tokens.pop().unwrap();
             let seq_len_in_bytes = self.seq_len * 2;
             self.indexes_in_bytes = (0..self.current_tokens.len() - seq_len_in_bytes)
                 .step_by(seq_len_in_bytes)
                 .collect::<Vec<_>>();
-            self.indexes_in_bytes.shuffle(&mut thread_rng());
+            self.indexes_in_bytes.shuffle(&mut rng());
         }
         let start_idx = self.indexes_in_bytes.pop().unwrap();
         let bytes = &self.current_tokens[start_idx..start_idx + 2 * (seq_len + 1)];

candle-datasets/src/vision/cifar.rs

@@ -72,6 +72,8 @@ fn load_parquet(parquet: SerializedFileReader<std::fs::File>) -> Result<(Tensor,
             if let parquet::record::Field::Group(subrow) = field {
                 for (_name, field) in subrow.get_column_iter() {
                     if let parquet::record::Field::Bytes(value) = field {
+                        // image-rs crate convention is to load in (width, height, channels) order
+                        // See: https://docs.rs/image/latest/image/trait.ImageDecoder.html#tymethod.dimensions
                         let image = image::load_from_memory(value.data()).unwrap();
                         buffer_images.extend(image.to_rgb8().as_raw());
                     }
@@ -81,8 +83,10 @@ fn load_parquet(parquet: SerializedFileReader<std::fs::File>) -> Result<(Tensor,
             }
         }
     }
-    let images = (Tensor::from_vec(buffer_images, (samples, 3, 32, 32), &Device::Cpu)?
-        .to_dtype(DType::U8)?
+    // Reorder image-rs convention (width, height, channels) to candle/pytorch convolution convention (channels, height, width)
+    let images = (Tensor::from_vec(buffer_images, (samples, 32, 32, 3), &Device::Cpu)?
+        .to_dtype(DType::F32)?
+        .permute((0, 3, 2, 1))?
         / 255.)?;
     let labels = Tensor::from_vec(buffer_labels, (samples,), &Device::Cpu)?;
     Ok((images, labels))

candle-datasets/src/vision/mnist.rs

@@ -89,7 +89,7 @@ fn load_parquet(parquet: SerializedFileReader<std::fs::File>) -> Result<(Tensor,
 
 pub fn load() -> Result<crate::vision::Dataset> {
     let api = Api::new().map_err(|e| Error::Msg(format!("Api error: {e}")))?;
-    let dataset_id = "mnist".to_string();
+    let dataset_id = "ylecun/mnist".to_string();
     let repo = Repo::with_revision(
         dataset_id,
         RepoType::Dataset,

candle-examples/Cargo.toml

@@ -25,7 +25,9 @@ hf-hub = { workspace = true, features = ["tokio"] }
 image = { workspace = true }
 intel-mkl-src = { workspace = true, optional = true }
 num-traits = { workspace = true }
-pyo3 = { version = "0.21.0", features = ["auto-initialize"], optional = true }
+palette = { version = "0.7.6", optional = true }
+enterpolation = { version = "0.2.1", optional = true}
+pyo3 = { version = "0.22.0", features = ["auto-initialize", "abi3-py311"], optional = true }
 rayon = { workspace = true }
 rubato = { version = "0.15.0", optional = true }
 safetensors = { workspace = true }
@@ -33,7 +35,8 @@ serde = { workspace = true }
 serde_json = { workspace = true }
 symphonia = { version = "0.5.3", features = ["all"], optional = true }
 tokenizers = { workspace = true, features = ["onig"] }
-cpal= { version = "0.15.2", optional = true }
+cpal = { version = "0.15.2", optional = true }
+pdf2image = { version = "0.1.2" , optional = true}
 
 [dev-dependencies]
 anyhow = { workspace = true }
@@ -47,7 +50,7 @@ tracing = { workspace = true }
 tracing-chrome = { workspace = true }
 tracing-subscriber = { workspace = true }
 # Necessary to disambiguate with tokio in wasm examples which are 1.28.1
-tokio = "1.29.1"
+tokio = "1.43.0"
 
 [build-dependencies]
 anyhow = { workspace = true }
@@ -57,14 +60,17 @@ bindgen_cuda = { version = "0.1.1", optional = true }
 default = []
 accelerate = ["dep:accelerate-src", "candle/accelerate", "candle-nn/accelerate", "candle-transformers/accelerate"]
 cuda = ["candle/cuda", "candle-nn/cuda", "candle-transformers/cuda", "dep:bindgen_cuda"]
-cudnn = ["candle/cudnn"]
+cudnn = ["candle/cudnn", "candle-nn/cudnn", "candle-transformers/cudnn"]
 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"]
+microphone = ["cpal", "rubato"]
 encodec = ["cpal", "symphonia", "rubato"]
+mimi = ["cpal", "symphonia", "rubato"]
+snac = ["cpal", "symphonia", "rubato"]
+depth_anything_v2 = ["palette", "enterpolation"]
 
 [[example]]
 name = "llama_multiprocess"
@@ -98,6 +104,26 @@ required-features = ["candle-datasets"]
 name = "llama2-c"
 required-features = ["candle-datasets"]
 
+[[example]]
+name = "mimi"
+required-features = ["mimi"]
+
+[[example]]
+name = "snac"
+required-features = ["snac"]
+
 [[example]]
 name = "encodec"
 required-features = ["encodec"]
+
+[[example]]
+name = "depth_anything_v2"
+required-features = ["depth_anything_v2"]
+
+[[example]]
+name = "silero-vad"
+required-features = ["onnx"]
+
+[[example]]
+name = "colpali"
+required-features = ["pdf2image"]

candle-examples/examples/based/README.md

@@ -0,0 +1,20 @@
+# candle-based
+
+Experimental, not instruction-tuned small LLM from the Hazy Research group, combining local and linear attention layers.
+
+[Blogpost](https://hazyresearch.stanford.edu/blog/2024-03-03-based)
+
+[Simple linear attention language models balance the recall-throughput tradeoff](https://arxiv.org/abs/2402.18668)
+
+## Running an example
+
+```bash
+$ cargo run --example based --release -- --prompt "Flying monkeys are" --which 1b-50b --sample-len 100
+
+Flying monkeys are a common sight in the wild, but they are also a threat to humans.
+
+The new study, published today (July 31) in the journal Science Advances, shows that the monkeys are using their brains to solve the problem of how to get around the problem.
+
+"We found that the monkeys were using a strategy called 'cognitive mapping' - they would use their brains to map out the route ahead," says lead author Dr. David J. Smith from the University of California
+
+```

candle-examples/examples/based/main.rs

@@ -0,0 +1,275 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::{Error as E, Result};
+use clap::{Parser, ValueEnum};
+
+use candle_transformers::models::based::Model;
+
+use candle::{DType, Device, Tensor};
+use candle_examples::token_output_stream::TokenOutputStream;
+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: TokenOutputStream,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+}
+
+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,
+        device: &Device,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            tokenizer: TokenOutputStream::new(tokenizer),
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            device: device.clone(),
+        }
+    }
+
+    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
+        use std::io::Write;
+        self.tokenizer.clear();
+        let mut tokens = self
+            .tokenizer
+            .tokenizer()
+            .encode(prompt, true)
+            .map_err(E::msg)?
+            .get_ids()
+            .to_vec();
+        for &t in tokens.iter() {
+            if let Some(t) = self.tokenizer.next_token(t)? {
+                print!("{t}")
+            }
+        }
+        std::io::stdout().flush()?;
+
+        let mut generated_tokens = 0usize;
+        let eos_token = match self.tokenizer.get_token("<|endoftext|>") {
+            Some(token) => token,
+            None => anyhow::bail!("cannot find the <|endoftext|> token"),
+        };
+        let start_gen = std::time::Instant::now();
+        for index in 0..sample_len {
+            let context_size = if index > 0 { 1 } else { tokens.len() };
+            let start_pos = tokens.len().saturating_sub(context_size);
+            let ctxt = &tokens[start_pos..];
+            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+            let logits = self.model.forward(&input, start_pos)?;
+            let logits = logits.squeeze(0)?.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;
+            }
+            if let Some(t) = self.tokenizer.next_token(next_token)? {
+                print!("{t}");
+                std::io::stdout().flush()?;
+            }
+        }
+        let dt = start_gen.elapsed();
+        if let Some(rest) = self.tokenizer.decode_rest().map_err(E::msg)? {
+            print!("{rest}");
+        }
+        std::io::stdout().flush()?;
+        println!(
+            "\n{generated_tokens} tokens generated ({:.2} token/s)",
+            generated_tokens as f64 / dt.as_secs_f64(),
+        );
+        Ok(())
+    }
+}
+
+#[derive(Clone, Debug, Copy, PartialEq, Eq, ValueEnum)]
+enum Which {
+    #[value(name = "360m")]
+    W360m,
+    #[value(name = "1b")]
+    W1b,
+    #[value(name = "1b-50b")]
+    W1b50b,
+}
+
+#[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,
+
+    #[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 = 10000)]
+    sample_len: usize,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long, default_value = "refs/pr/1")]
+    revision: String,
+
+    #[arg(long)]
+    config_file: Option<String>,
+
+    #[arg(long)]
+    tokenizer_file: Option<String>,
+
+    #[arg(long)]
+    weight_files: 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,
+
+    #[arg(long, default_value = "360m")]
+    which: Which,
+}
+
+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,
+        None => match args.which {
+            Which::W360m => "hazyresearch/based-360m".to_string(),
+            Which::W1b => "hazyresearch/based-1b".to_string(),
+            Which::W1b50b => "hazyresearch/based-1b-50b".to_string(),
+        },
+    };
+    let repo = api.repo(Repo::with_revision(
+        model_id,
+        RepoType::Model,
+        args.revision,
+    ));
+    let config_file = match args.config_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("config.json")?,
+    };
+    let filenames = match args.weight_files {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        None => vec![repo.get("model.safetensors")?],
+    };
+
+    let repo = api.model("openai-community/gpt2".to_string());
+    let tokenizer_file = match args.tokenizer_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("tokenizer.json")?,
+    };
+
+    println!("retrieved the files in {:?}", start.elapsed());
+    let tokenizer = Tokenizer::from_file(tokenizer_file).map_err(E::msg)?;
+
+    let start = std::time::Instant::now();
+    let config = serde_json::from_reader(std::fs::File::open(config_file)?)?;
+    let device = candle_examples::device(args.cpu)?;
+    let dtype = if device.is_cuda() {
+        DType::BF16
+    } else {
+        DType::F32
+    };
+
+    let mut vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
+    if args.which == Which::W1b50b {
+        vb = vb.pp("model");
+    };
+
+    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,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}

candle-examples/examples/beit/README.md

@@ -0,0 +1,20 @@
+# candle-beit
+
+[Beit](https://arxiv.org/abs/2106.08254) is a computer vision model.
+In this example, it is used as an ImageNet classifier: the model returns the
+probability for the image to belong to each of the 1000 ImageNet categories.
+
+## Running some example
+
+```bash
+cargo run --example beit --release -- --image candle-examples/examples/yolo-v8/assets/bike.jpg
+
+> mountain bike, all-terrain bike, off-roader: 56.16%
+> bicycle-built-for-two, tandem bicycle, tandem: 3.08%
+> maillot                 : 2.23%
+> alp                     : 0.88%
+> crash helmet            : 0.85%
+
+```
+
+![Leading group, Giro d'Italia 2021](../yolo-v8/assets/bike.jpg)

candle-examples/examples/beit/main.rs

@@ -0,0 +1,79 @@
+//! BEiT: BERT Pre-Training of Image Transformers
+//! https://github.com/microsoft/unilm/tree/master/beit
+
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use clap::Parser;
+
+use candle::{DType, Device, IndexOp, Result, Tensor, D};
+use candle_nn::{Module, VarBuilder};
+use candle_transformers::models::beit;
+
+/// Loads an image from disk using the image crate, this returns a tensor with shape
+/// (3, 384, 384). Beit special normalization is applied.
+pub fn load_image384_beit_norm<P: AsRef<std::path::Path>>(p: P) -> Result<Tensor> {
+    let img = image::ImageReader::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(&[0.5f32, 0.5, 0.5], &Device::Cpu)?.reshape((3, 1, 1))?;
+    let std = Tensor::new(&[0.5f32, 0.5, 0.5], &Device::Cpu)?.reshape((3, 1, 1))?;
+    (data.to_dtype(candle::DType::F32)? / 255.)?
+        .broadcast_sub(&mean)?
+        .broadcast_div(&std)
+}
+
+#[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 = load_image384_beit_norm(args.image)?.to_device(&device)?;
+    println!("loaded image {image:?}");
+
+    let model_file = match args.model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model("vincent-espitalier/candle-beit".into());
+            api.get("beit_base_patch16_384.in22k_ft_in22k_in1k.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+    let model = beit::vit_base(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/bert/main.rs

@@ -126,7 +126,7 @@ fn main() -> Result<()> {
         println!("Loaded and encoded {:?}", start.elapsed());
         for idx in 0..args.n {
             let start = std::time::Instant::now();
-            let ys = model.forward(&token_ids, &token_type_ids)?;
+            let ys = model.forward(&token_ids, &token_type_ids, None)?;
             if idx == 0 {
                 println!("{ys}");
             }
@@ -163,11 +163,19 @@ fn main() -> Result<()> {
                 Ok(Tensor::new(tokens.as_slice(), device)?)
             })
             .collect::<Result<Vec<_>>>()?;
+        let attention_mask = tokens
+            .iter()
+            .map(|tokens| {
+                let tokens = tokens.get_attention_mask().to_vec();
+                Ok(Tensor::new(tokens.as_slice(), device)?)
+            })
+            .collect::<Result<Vec<_>>>()?;
 
         let token_ids = Tensor::stack(&token_ids, 0)?;
+        let attention_mask = Tensor::stack(&attention_mask, 0)?;
         let token_type_ids = token_ids.zeros_like()?;
         println!("running inference on batch {:?}", token_ids.shape());
-        let embeddings = model.forward(&token_ids, &token_type_ids)?;
+        let embeddings = model.forward(&token_ids, &token_type_ids, Some(&attention_mask))?;
         println!("generated embeddings {:?}", embeddings.shape());
         // Apply some avg-pooling by taking the mean embedding value for all tokens (including padding)
         let (_n_sentence, n_tokens, _hidden_size) = embeddings.dims3()?;

candle-examples/examples/blip/main.rs

@@ -55,7 +55,7 @@ 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)?
+    let img = image::ImageReader::open(p)?
         .decode()
         .map_err(candle::Error::wrap)?
         .resize_to_fill(384, 384, image::imageops::FilterType::Triangle);

candle-examples/examples/chatglm/README.md

@@ -0,0 +1,13 @@
+# candle-chatglm
+
+Uses `THUDM/chatglm3-6b` to generate chinese text. Will not generate text for english (usually).
+ 
+## Text Generation
+
+```bash
+cargo run --example chatglm --release  -- --prompt "部署门槛较低等众多优秀特 "
+
+> 部署门槛较低等众多优秀特 点，使得其成为了一款备受欢迎的AI助手。
+> 
+> 作为一款人工智能助手，ChatGLM3-6B
+```

candle-examples/examples/chinese_clip/README.md

@@ -0,0 +1,42 @@
+# candle-chinese-clip
+
+Contrastive Language-Image Pre-Training (CLIP) is an architecture trained on
+pairs of images with related texts. This one is trained using in chinese instead of english.
+
+## Running on cpu
+
+```bash
+$ cargo run --example chinese_clip --release -- --images "candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg","candle-examples/examples/yolo-v8/assets/bike.jpg" --cpu --sequences "一场自行车比赛","两只猫的照片","一个机器人拿着蜡烛"
+
+> Results for image: candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg
+>
+> 2025-03-25T19:22:01.325177Z  INFO chinese_clip: Probability: 0.0000% Text: 一场自行车比赛 
+> 2025-03-25T19:22:01.325179Z  INFO chinese_clip: Probability: 0.0000% Text: 两只猫的照片 
+> 2025-03-25T19:22:01.325181Z  INFO chinese_clip: Probability: 100.0000% Text: 一个机器人拿着蜡烛 
+> 2025-03-25T19:22:01.325183Z  INFO chinese_clip: 
+> 
+> Results for image: candle-examples/examples/yolo-v8/assets/bike.jpg
+> 
+> 2025-03-25T19:22:01.325184Z  INFO chinese_clip: Probability: 100.0000% Text: 一场自行车比赛 
+> 2025-03-25T19:22:01.325186Z  INFO chinese_clip: Probability: 0.0000% Text: 两只猫的照片 
+> 2025-03-25T19:22:01.325187Z  INFO chinese_clip: Probability: 0.0000% Text: 一个机器人拿着蜡烛 
+```
+
+## Running on metal
+
+```bash 
+$ cargo run --features metal --example chinese_clip --release -- --images "candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg","candle-examples/examples/yolo-v8/assets/bike.jpg" --cpu --sequences "一场自行车比赛","两只猫的照片","一个机器人拿着蜡烛"
+
+> Results for image: candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg
+>
+> 2025-03-25T19:22:01.325177Z  INFO chinese_clip: Probability: 0.0000% Text: 一场自行车比赛 
+> 2025-03-25T19:22:01.325179Z  INFO chinese_clip: Probability: 0.0000% Text: 两只猫的照片 
+> 2025-03-25T19:22:01.325181Z  INFO chinese_clip: Probability: 100.0000% Text: 一个机器人拿着蜡烛 
+> 2025-03-25T19:22:01.325183Z  INFO chinese_clip: 
+> 
+> Results for image: candle-examples/examples/yolo-v8/assets/bike.jpg
+> 
+> 2025-03-25T19:22:01.325184Z  INFO chinese_clip: Probability: 100.0000% Text: 一场自行车比赛 
+> 2025-03-25T19:22:01.325186Z  INFO chinese_clip: Probability: 0.0000% Text: 两只猫的照片 
+> 2025-03-25T19:22:01.325187Z  INFO chinese_clip: Probability: 0.0000% Text: 一个机器人拿着蜡烛 
+```

candle-examples/examples/chinese_clip/main.rs

@@ -0,0 +1,224 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use candle::{DType, Device, Tensor};
+use candle_nn as nn;
+use candle_transformers::models::chinese_clip::{ChineseClipConfig, ChineseClipModel};
+use clap::Parser;
+use tokenizers::Tokenizer;
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    tokenizer: Option<String>,
+
+    #[arg(long, use_value_delimiter = true)]
+    images: Option<Vec<String>>,
+
+    #[arg(long)]
+    cpu: bool,
+
+    #[arg(long, use_value_delimiter = true)]
+    sequences: Option<Vec<String>>,
+}
+
+fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    tracing_subscriber::fmt::init();
+
+    let device = candle_examples::device(args.cpu)?;
+    let var = load_weights(args.model, &device)?;
+    let clip_model = ChineseClipModel::new(var, &ChineseClipConfig::clip_vit_base_patch16())?;
+    tracing::info!("Transformer loaded. ");
+
+    let (pixel_values, vec_imgs) = load_images(args.images, &device)?;
+    tracing::info!("Images loaded. ");
+
+    let tokenizer = load_tokenizer()?;
+    let (input_ids, type_ids, attention_mask, text_sequences) =
+        tokenize_sequences(args.sequences, &tokenizer, &device)?;
+
+    tracing::info!("Computing ... ");
+    let (_logits_per_text, logits_per_image) = clip_model.forward(
+        &pixel_values,
+        &input_ids,
+        Some(&type_ids),
+        Some(&attention_mask),
+    )?;
+    let softmax_image = nn::ops::softmax(&logits_per_image, 1)?;
+
+    let softmax_image_vec = softmax_image.flatten_all()?.to_vec1::<f32>()?;
+
+    let probability_vec = softmax_image_vec
+        .iter()
+        .map(|v| v * 100.0)
+        .collect::<Vec<f32>>();
+
+    let probability_per_image = probability_vec.len() / vec_imgs.len();
+
+    for (i, img) in vec_imgs.iter().enumerate() {
+        let start = i * probability_per_image;
+        let end = start + probability_per_image;
+        let prob = &probability_vec[start..end];
+        tracing::info!("\n\nResults for image: {}\n", img);
+
+        for (i, p) in prob.iter().enumerate() {
+            tracing::info!("Probability: {:.4}% Text: {} ", p, text_sequences[i]);
+        }
+    }
+
+    Ok(())
+}
+
+pub fn load_weights(model: Option<String>, device: &Device) -> anyhow::Result<nn::VarBuilder> {
+    let model_file = match model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let repo = hf_hub::Repo::with_revision(
+                "OFA-Sys/chinese-clip-vit-base-patch16".to_string(),
+                hf_hub::RepoType::Model,
+                "refs/pr/3".to_string(),
+            );
+            let api = api.repo(repo);
+            api.get("model.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+
+    Ok(unsafe { nn::VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, device)? })
+}
+
+pub fn load_tokenizer() -> anyhow::Result<Tokenizer> {
+    let tokenizer_file = {
+        let api = hf_hub::api::sync::Api::new()?;
+        let repo = hf_hub::Repo::with_revision(
+            "OFA-Sys/chinese-clip-vit-base-patch16".to_string(),
+            hf_hub::RepoType::Model,
+            "refs/pr/3".to_string(),
+        );
+        let api = api.repo(repo);
+        api.get("tokenizer.json")?
+    };
+
+    Tokenizer::from_file(tokenizer_file).map_err(anyhow::Error::msg)
+}
+
+pub fn tokenize_sequences(
+    sequences: Option<Vec<String>>,
+    tokenizer: &Tokenizer,
+    device: &Device,
+) -> anyhow::Result<(Tensor, Tensor, Tensor, Vec<String>)> {
+    let vec_seq = match sequences {
+        Some(seq) => seq,
+        None => vec![
+            "自行车比赛".to_string(),
+            "两只猫咪".to_string(),
+            "拿着蜡烛的机器人".to_string(),
+        ],
+    };
+
+    let mut input_ids = vec![];
+    let mut type_ids = vec![];
+    let mut attention_mask = vec![];
+    let mut max_len = 0;
+
+    for seq in vec_seq.clone() {
+        let encoding = tokenizer.encode(seq, true).map_err(anyhow::Error::msg)?;
+        input_ids.push(encoding.get_ids().to_vec());
+        type_ids.push(encoding.get_type_ids().to_vec());
+        attention_mask.push(encoding.get_attention_mask().to_vec());
+        if encoding.get_ids().len() > max_len {
+            max_len = encoding.get_ids().len();
+        }
+    }
+
+    let pad_id = *tokenizer
+        .get_vocab(true)
+        .get("[PAD]")
+        .ok_or(anyhow::Error::msg("No pad token"))?;
+
+    let input_ids: Vec<Vec<u32>> = input_ids
+        .iter_mut()
+        .map(|item| {
+            item.extend(vec![pad_id; max_len - item.len()]);
+            item.to_vec()
+        })
+        .collect();
+
+    let type_ids: Vec<Vec<u32>> = type_ids
+        .iter_mut()
+        .map(|item| {
+            item.extend(vec![0; max_len - item.len()]);
+            item.to_vec()
+        })
+        .collect();
+
+    let attention_mask: Vec<Vec<u32>> = attention_mask
+        .iter_mut()
+        .map(|item| {
+            item.extend(vec![0; max_len - item.len()]);
+            item.to_vec()
+        })
+        .collect();
+
+    let input_ids = Tensor::new(input_ids, device)?;
+    let type_ids = Tensor::new(type_ids, device)?;
+    let attention_mask = Tensor::new(attention_mask, device)?;
+
+    Ok((input_ids, type_ids, attention_mask, vec_seq))
+}
+
+pub fn load_images(
+    images: Option<Vec<String>>,
+    device: &Device,
+) -> anyhow::Result<(Tensor, Vec<String>)> {
+    let vec_imgs = match images {
+        Some(imgs) => imgs,
+        None => vec![
+            "candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg".to_string(),
+            "candle-examples/examples/yolo-v8/assets/bike.jpg".to_string(),
+        ],
+    };
+
+    let mut images = vec![];
+
+    for path in vec_imgs.iter() {
+        let tensor = load_image(path, 224, device)?;
+        images.push(tensor);
+    }
+
+    let images = Tensor::stack(&images, 0)?.to_device(device)?;
+    Ok((images, vec_imgs))
+}
+
+fn load_image<T: AsRef<std::path::Path>>(
+    path: T,
+    image_size: usize,
+    device: &Device,
+) -> anyhow::Result<Tensor> {
+    let img = image::ImageReader::open(path)?.decode()?;
+    let (height, width) = (image_size, image_size);
+    let img = img.resize_to_fill(
+        width as u32,
+        height as u32,
+        image::imageops::FilterType::Triangle,
+    );
+
+    let img = img.to_rgb8().into_raw();
+    let img = Tensor::from_vec(img, (height, width, 3), device)?.permute((2, 0, 1))?;
+    let mean = Tensor::new(&[0.48145466f32, 0.4578275, 0.40821073], device)?.reshape((3, 1, 1))?;
+    let std =
+        Tensor::new(&[0.26862954f32, 0.261_302_6, 0.275_777_1], device)?.reshape((3, 1, 1))?;
+    let img = (img.to_dtype(DType::F32)? / 255.)?
+        .broadcast_sub(&mean)?
+        .broadcast_div(&std)?;
+
+    Ok(img)
+}

candle-examples/examples/clip/README.md

@@ -1,4 +1,4 @@
-Contrastive Language-Image Pre-Training
+# candle-clip
 
 Contrastive Language-Image Pre-Training (CLIP) is an architecture trained on
 pairs of images with related texts.

candle-examples/examples/clip/main.rs

@@ -12,7 +12,6 @@ use candle_nn::{ops::softmax, VarBuilder};
 use candle_transformers::models::clip;
 
 use tokenizers::Tokenizer;
-use tracing::info;
 
 #[derive(Parser)]
 struct Args {
@@ -33,22 +32,19 @@ struct Args {
 }
 
 fn load_image<T: AsRef<std::path::Path>>(path: T, image_size: usize) -> anyhow::Result<Tensor> {
-    let img = image::io::Reader::open(path)?.decode()?;
+    let img = image::ImageReader::open(path)?.decode()?;
     let (height, width) = (image_size, image_size);
     let img = img.resize_to_fill(
         width as u32,
         height as u32,
         image::imageops::FilterType::Triangle,
     );
-
     let img = img.to_rgb8();
-
     let img = img.into_raw();
     let img = Tensor::from_vec(img, (height, width, 3), &Device::Cpu)?
         .permute((2, 0, 1))?
         .to_dtype(DType::F32)?
         .affine(2. / 255., -1.)?;
-    // .unsqueeze(0)?;
     Ok(img)
 }
 
@@ -57,24 +53,16 @@ fn load_images<T: AsRef<std::path::Path>>(
     image_size: usize,
 ) -> anyhow::Result<Tensor> {
     let mut images = vec![];
-
     for path in paths {
         let tensor = load_image(path, image_size)?;
         images.push(tensor);
     }
-
     let images = Tensor::stack(&images, 0)?;
-
     Ok(images)
 }
 
 pub fn main() -> anyhow::Result<()> {
-    // std::env::set_var("RUST_BACKTRACE", "full");
-
     let args = Args::parse();
-
-    tracing_subscriber::fmt::init();
-
     let model_file = match args.model {
         None => {
             let api = hf_hub::api::sync::Api::new()?;
@@ -89,13 +77,9 @@ pub fn main() -> anyhow::Result<()> {
         }
         Some(model) => model.into(),
     };
-
     let tokenizer = get_tokenizer(args.tokenizer)?;
-
     let config = clip::ClipConfig::vit_base_patch32();
-
     let device = candle_examples::device(args.cpu)?;
-
     let vec_imgs = match args.images {
         Some(imgs) => imgs,
         None => vec![
@@ -103,43 +87,29 @@ pub fn main() -> anyhow::Result<()> {
             "candle-examples/examples/yolo-v8/assets/bike.jpg".to_string(),
         ],
     };
-
-    // let image = load_image(args.image, config.image_size)?.to_device(&device)?;
     let images = load_images(&vec_imgs, config.image_size)?.to_device(&device)?;
-
     let vb =
         unsafe { VarBuilder::from_mmaped_safetensors(&[model_file.clone()], DType::F32, &device)? };
-
     let model = clip::ClipModel::new(vb, &config)?;
-
     let (input_ids, vec_seq) = tokenize_sequences(args.sequences, &tokenizer, &device)?;
-
     let (_logits_per_text, logits_per_image) = model.forward(&images, &input_ids)?;
-
     let softmax_image = softmax(&logits_per_image, 1)?;
-
     let softmax_image_vec = softmax_image.flatten_all()?.to_vec1::<f32>()?;
-
-    info!("softmax_image_vec: {:?}", softmax_image_vec);
-
+    println!("softmax_image_vec: {:?}", softmax_image_vec);
     let probability_vec = softmax_image_vec
         .iter()
         .map(|v| v * 100.0)
         .collect::<Vec<f32>>();
-
     let probability_per_image = probability_vec.len() / vec_imgs.len();
-
     for (i, img) in vec_imgs.iter().enumerate() {
         let start = i * probability_per_image;
         let end = start + probability_per_image;
         let prob = &probability_vec[start..end];
-        info!("\n\nResults for image: {}\n", img);
-
+        println!("\n\nResults for image: {}\n", img);
         for (i, p) in prob.iter().enumerate() {
-            info!("Probability: {:.4}% Text: {} ", p, vec_seq[i]);
+            println!("Probability: {:.4}% Text: {} ", p, vec_seq[i]);
         }
     }
-
     Ok(())
 }
 
@@ -156,7 +126,6 @@ pub fn get_tokenizer(tokenizer: Option<String>) -> anyhow::Result<Tokenizer> {
         }
         Some(file) => file.into(),
     };
-
     Tokenizer::from_file(tokenizer).map_err(E::msg)
 }
 
@@ -169,7 +138,6 @@ pub fn tokenize_sequences(
         .get_vocab(true)
         .get("<|endoftext|>")
         .ok_or(E::msg("No pad token"))?;
-
     let vec_seq = match sequences {
         Some(seq) => seq,
         None => vec![
@@ -178,16 +146,12 @@ pub fn tokenize_sequences(
             "a robot holding a candle".to_string(),
         ],
     };
-
     let mut tokens = vec![];
-
     for seq in vec_seq.clone() {
         let encoding = tokenizer.encode(seq, true).map_err(E::msg)?;
         tokens.push(encoding.get_ids().to_vec());
     }
-
     let max_len = tokens.iter().map(|v| v.len()).max().unwrap_or(0);
-
     // Pad the sequences to have the same length
     for token_vec in tokens.iter_mut() {
         let len_diff = max_len - token_vec.len();
@@ -195,8 +159,6 @@ pub fn tokenize_sequences(
             token_vec.extend(vec![pad_id; len_diff]);
         }
     }
-
     let input_ids = Tensor::new(tokens, device)?;
-
     Ok((input_ids, vec_seq))
 }

candle-examples/examples/codegeex4-9b/README.org

@@ -0,0 +1,96 @@
+* candle-codegeex4_9b
+THUDM/CodeGeeX4 is a versatile model for all AI software development scenarios, including code completion, code interpreter, web search, function calling, repository-level Q&A and much more.
+
+- [[https://github.com/THUDM/CodeGeeX4][Github]]
+- [[https://codegeex.cn/][HomePage]]
+- [[https://huggingface.co/THUDM/codegeex4-all-9b][huggingface]]  
+
+** Running with ~cuda~
+
+#+begin_src shell
+  cargo run --example codegeex4-9b --release --features cuda   -- --prompt "please write a insertion sort in rust" --sample-len 300
+#+end_src
+
+** Running with ~cpu~
+#+begin_src shell
+  cargo run --example codegeex4-9b --release -- --cpu   --prompt "please write a insertion sort in rust" --sample-len 300
+#+end_src
+
+** Output_Example
+*** Input
+#+begin_src shell
+  cargo run  --release --features cuda -- --prompt 'please write a FFT in rust' --sample-len 500 --cache /root/autodl-tmp 
+#+end_src
+
+*** Output
+#+begin_src shell
+  avx: false, neon: false, simd128: false, f16c: false
+  temp: 0.95 repeat-penalty: 1.10 repeat-last-n: 64
+  cache path /root/autodl-tmp
+  Prompt: [please write a FFT in rust]
+  Using Seed 11511762269791786684
+  DType is BF16
+  transofrmer layers create
+  模型加载完毕 4
+  starting the inference loop
+
+   开始生成
+  samplelen 500
+
+  500 tokens generated (34.60 token/s)
+  Result:
+
+  Sure, I can help you with that. Here's an example of a Fast Fourier Transform (FFT) implementation in Rust:
+
+  ```rust
+  use num_complex::Complex;
+
+  fn fft(input: &[Complex<f64> > ] ) -> Vec<Complex<f64> > > {
+      let n = input.len();
+    
+      if n == 1 {
+	  return vec![input[0]]];
+      }
+    
+      let mut even = vec![];
+      let mut odd = vec![];
+    
+      for i in 0..n {
+
+	      if i % 2 == 0 {
+	      even.push(input[i]);
+	  } else {
+	      odd.push(input[i]);
+	  }
+      }
+    
+      let even_fft = fft(&even);
+      let odd_fft = fft(&odd);
+    
+      let mut output = vec![];
+    
+      for k in 0..n/2 {
+	  let t = Complex::new(0.0, -2.0 * std::f64::consts::PI * (k as f64) / (n as f64))) ).exp();
+        
+	  output.push(even_fft[k] + odd_fft[k] * t]);
+	  output.push(even_fft[k] - odd_fft[k] * t]);
+      }
+    
+      return output;
+  }
+  ```
+
+  This implementation uses the Cooley-Tukey algorithm to perform the FFT. The function takes an array of complex numbers and returns an array of complex numbers which is the result of the FFT.
+#+end_src
+
+
+*  Citation
+#+begin_src
+  @inproceedings{zheng2023codegeex,
+  title={CodeGeeX: A Pre-Trained Model for Code Generation with Multilingual Benchmarking on HumanEval-X},
+  author={Qinkai Zheng and Xiao Xia and Xu Zou and Yuxiao Dong and Shan Wang and Yufei Xue and Zihan Wang and Lei Shen and Andi Wang and Yang Li and Teng Su and Zhilin Yang and Jie Tang},
+  booktitle={Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
+  pages={5673--5684},
+  year={2023}
+}
+#+end_src