Add RandomNormal ONNX operator (#2200)

* Use flash-attn in gemma.

* Fix flash-attn for head dim 256.

.github/dependabot.yml

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

.github/workflows/ci_cuda.yaml

@@ -5,47 +5,15 @@ on:
   pull_request:
 
 jobs:
-  start-runner:
-    name: Start self-hosted EC2 runner
-    runs-on: ubuntu-latest
-    env:
-      AWS_REGION: us-east-1
-      EC2_AMI_ID: ami-03cfed9ea28f4b002
-      EC2_INSTANCE_TYPE: g5.xlarge
-      EC2_SUBNET_ID: subnet-931b34f5,subnet-ecb993cd,subnet-943dc2d8,subnet-45371f1a,subnet-ee93e0df,subnet-fddc3dfc
-      EC2_SECURITY_GROUP: sg-030175c435ac141d6
-    outputs:
-      label: ${{ steps.start-ec2-runner.outputs.label }}
-      ec2-instance-id: ${{ steps.start-ec2-runner.outputs.ec2-instance-id }}
-    steps:
-      - name: Configure AWS credentials
-        uses: aws-actions/configure-aws-credentials@v1
-        with:
-          aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
-          aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
-          aws-region: ${{ env.AWS_REGION }}
-      - name: Start EC2 runner
-        id: start-ec2-runner
-        uses: philschmid/philschmid-ec2-github-runner@main
-        with:
-          mode: start
-          github-token: ${{ secrets.GH_PERSONAL_ACCESS_TOKEN }}
-          ec2-image-id: ${{ env.EC2_AMI_ID }}
-          ec2-instance-type: ${{ env.EC2_INSTANCE_TYPE }}
-          subnet-id: ${{ env.EC2_SUBNET_ID }}
-          security-group-id: ${{ env.EC2_SECURITY_GROUP }}
-          aws-resource-tags: > # optional, requires additional permissions
-            [
-              {"Key": "Name", "Value": "ec2-tgi-github-runner"},
-              {"Key": "GitHubRepository", "Value": "${{ github.repository }}"}
-            ]
-
   test-cuda:
     concurrency:
       group: ${{ github.workflow }}-${{ github.job }}-${{ github.head_ref || github.run_id }}
       cancel-in-progress: true
-    needs: start-runner # required to start the main job when the runner is ready
-    runs-on: ${{ needs.start-runner.outputs.label }} # run the job on the newly created runner
+    runs-on: [single-gpu, nvidia-gpu, t4, ci]
+    container:
+      image: nvidia/cuda:12.3.1-devel-ubuntu22.04
+      options: --gpus 0 
+    if: ${{ github.event.pull_request.head.repo.full_name == github.event.pull_request.base.repo.full_name }}
     permissions:
       contents: write
       packages: write
@@ -56,32 +24,10 @@ jobs:
     steps:
       - name: Checkout repository
         uses: actions/checkout@v3
+      - name: Install dependencies
+        run: apt-get update && apt install curl build-essential libssl-dev protobuf-compiler pkg-config -y
       - name: Install Rust Stable
-        run: curl https://sh.rustup.rs -sSf | sh -s -- -y
+        uses: actions-rust-lang/setup-rust-toolchain@v1
       - uses: Swatinem/rust-cache@v2
-      - run: apt-get update -y && apt-get install libssl-dev protobuf-compiler -y
       - name: Test (cuda)
-        run: PATH=$PATH:/usr/local/cuda-11.8/bin/ /root/.cargo/bin/cargo test --features cuda
-  stop-runner:
-    name: Stop self-hosted EC2 runner
-    needs:
-      - start-runner
-      - test-cuda
-    runs-on: ubuntu-latest
-    env:
-      AWS_REGION: us-east-1
-    if: ${{ always() }} # required to stop the runner even if the error happened in the previous jobs
-    steps:
-      - name: Configure AWS credentials
-        uses: aws-actions/configure-aws-credentials@v1
-        with:
-          aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
-          aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
-          aws-region: ${{ env.AWS_REGION }}
-      - name: Stop EC2 runner
-        uses: philschmid/philschmid-ec2-github-runner@main
-        with:
-          mode: stop
-          github-token: ${{ secrets.GH_PERSONAL_ACCESS_TOKEN }}
-          label: ${{ needs.start-runner.outputs.label }}
-          ec2-instance-id: ${{ needs.start-runner.outputs.ec2-instance-id }}
+        run: cargo test --features cuda

CHANGELOG.md

@@ -63,7 +63,7 @@ This documents the main changes to the `candle` crate.
   [760](https://github.com/huggingface/candle/pull/760).
 - Add the Segment-Anything Model (SAM) as an example
   [773](https://github.com/huggingface/candle/pull/773).
-- TinyViT backbone for the segemnt anything example
+- TinyViT backbone for the segment anything example
   [787](https://github.com/huggingface/candle/pull/787).
 - Shape with holes support
   [770](https://github.com/huggingface/candle/pull/770).

Cargo.toml

@@ -9,16 +9,18 @@ members = [
     "candle-transformers",
     "candle-wasm-examples/*",
     "candle-wasm-tests",
+    "tensor-tools",
 ]
 exclude = [
    "candle-flash-attn",
    "candle-kernels",
+   "candle-metal-kernels",
    "candle-onnx",
 ]
 resolver = "2"
 
 [workspace.package]
-version = "0.3.0"
+version = "0.5.1"
 edition = "2021"
 description = "Minimalist ML framework."
 repository = "https://github.com/huggingface/candle"
@@ -27,38 +29,50 @@ categories = ["science"]
 license = "MIT OR Apache-2.0"
 
 [workspace.dependencies]
+ab_glyph = "0.2.23"
 accelerate-src = { version = "0.3.2" }
 anyhow = { version = "1", features = ["backtrace"] }
 byteorder = "1.4.3"
+candle = { path = "./candle-core", package = "candle-core", version = "0.5.1" }
+candle-datasets = { path = "./candle-datasets", version = "0.5.1" }
+candle-flash-attn = { path = "./candle-flash-attn", version = "0.5.1" }
+candle-kernels = { path = "./candle-kernels", version = "0.5.1" }
+candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.5.1" }
+candle-nn = { path = "./candle-nn", version = "0.5.1" }
+candle-onnx = { path = "./candle-onnx", version = "0.5.1" }
+candle-transformers = { path = "./candle-transformers", version = "0.5.1" }
 clap = { version = "4.2.4", features = ["derive"] }
-cudarc = { version = "0.9.14", features = ["f16"] }
-gemm = { version = "0.16.6", features = ["wasm-simd128-enable"] }
+criterion = { version = "0.5.1", default-features=false }
+cudarc = { version = "0.11.1", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16", "cuda-version-from-build-system", "dynamic-linking"], default-features=false }
+fancy-regex = "0.13.0"
+gemm = { version = "0.17.0", features = ["wasm-simd128-enable"] }
 hf-hub = "0.3.0"
 half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
-image = { version = "0.24.7", default-features = false, features = ["jpeg", "png"] }
-imageproc = { version = "0.23.0", default-features = false }
+hound = "3.5.1"
+image = { version = "0.25.0", default-features = false, features = ["jpeg", "png"] }
+imageproc = { version = "0.24.0", default-features = false }
 intel-mkl-src = { version = "0.8.1", features = ["mkl-static-lp64-iomp"] }
 libc = { version = "0.2.147" }
 log = "0.4"
-memmap2 = { version = "0.7.1", features = ["stable_deref_trait"] }
+memmap2 = { version = "0.9.3", features = ["stable_deref_trait"] }
 num_cpus = "1.15.0"
 num-traits = "0.2.15"
-parquet = { version = "45.0.0" }
+parquet = { version = "51.0.0" }
 rand = "0.8.5"
 rand_distr = "0.4.3"
 rayon = "1.7.0"
-rusttype = { version = "0.9", default-features = false }
-safetensors = "0.3.1"
+safetensors = "0.4.1"
 serde = { version = "1.0.171", features = ["derive"] }
+serde_plain = "1.0.2"
 serde_json = "1.0.99"
 thiserror = "1"
-tokenizers = { version = "0.13.4", default-features = false }
+tokenizers = { version = "0.19.1", default-features = false }
 tracing = "0.1.37"
 tracing-chrome = "0.7.1"
 tracing-subscriber = "0.3.7"
-wav = "1.0.0"
 yoke = { version = "0.7.2", features = ["derive"] }
-zip = { version = "0.6.6", default-features = false }
+zip = { version = "1.1.1", default-features = false }
+metal = { version = "0.27.0", features = ["mps"]}
 
 [profile.release-with-debug]
 inherits = "release"

README.md

@@ -54,21 +54,38 @@ These online demos run entirely in your browser:
 - [whisper](https://huggingface.co/spaces/lmz/candle-whisper): speech recognition.
 - [LLaMA2](https://huggingface.co/spaces/lmz/candle-llama2): text generation.
 - [T5](https://huggingface.co/spaces/radames/Candle-T5-Generation-Wasm): text generation.
-- [Phi-v1.5](https://huggingface.co/spaces/radames/Candle-Phi-1.5-Wasm): text generation.
+- [Phi-1.5, and Phi-2](https://huggingface.co/spaces/radames/Candle-Phi-1.5-Wasm): text generation.
 - [Segment Anything Model](https://huggingface.co/spaces/radames/candle-segment-anything-wasm): Image segmentation.
 - [BLIP](https://huggingface.co/spaces/radames/Candle-BLIP-Image-Captioning): image captioning.
 
 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.
+- [LLaMA v1, v2, and v3](./candle-examples/examples/llama/): general LLM, includes
+  the SOLAR-10.7B variant.
 - [Falcon](./candle-examples/examples/falcon/): general LLM.
-- [Phi-v1 and Phi-v1.5](./candle-examples/examples/phi/): a 1.3b general LLM with performance on par with LLaMA-v2 7b.
+- [Gemma](./candle-examples/examples/gemma/): 2b and 7b 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.
+  pre-trained on 1T tokens of English and code datasets. Also supports
+  StableLM-2, a 1.6b LLM trained on 2T tokens, as well as the code variants.
+- [Mamba](./candle-examples/examples/mamba/): an inference only
+  implementation of the Mamba state space model.
 - [Mistral7b-v0.1](./candle-examples/examples/mistral/): a 7b general LLM with
-  performance larger than all publicly available 13b models as of 2023-09-28.
-- [StarCoder](./candle-examples/examples/bigcode/): LLM specialized to code generation.
+  better performance than all publicly available 13b models as of 2023-09-28.
+- [Mixtral8x7b-v0.1](./candle-examples/examples/mixtral/): a sparse mixture of
+  experts 8x7b general LLM with better performance than a Llama 2 70B model with
+  much faster inference.
+- [StarCoder](./candle-examples/examples/bigcode/) and
+  [StarCoder2](./candle-examples/examples/starcoder2/): LLM specialized to code generation.
+- [Qwen1.5](./candle-examples/examples/qwen/): Bilingual (English/Chinese) LLMs.
+- [RWKV v5 and v6](./candle-examples/examples/rwkv/): An RNN with transformer level LLM
+  performance.
 - [Replit-code-v1.5](./candle-examples/examples/replit-code/): a 3.3b LLM specialized for code completion.
+- [Yi-6B / Yi-34B](./candle-examples/examples/yi/): two bilingual
+  (English/Chinese) general LLMs with 6b and 34b parameters.
 - [Quantized LLaMA](./candle-examples/examples/quantized/): quantized version of
   the LLaMA model using the same quantization techniques as
   [llama.cpp](https://github.com/ggerganov/llama.cpp).
@@ -76,7 +93,7 @@ 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/quantized/assets/aoc.gif" width="600">
   
 - [Stable Diffusion](./candle-examples/examples/stable-diffusion/): text to
-  image generative model, support for the 1.5, 2.1, and SDXL 1.0 versions.
+  image generative model, support for the 1.5, 2.1, SDXL 1.0 and Turbo versions.
 
 <img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg" width="200">
 
@@ -95,16 +112,29 @@ 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 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.
 - [T5](./candle-examples/examples/t5), [Bert](./candle-examples/examples/bert/),
   [JinaBert](./candle-examples/examples/jina-bert/) : useful for sentence embeddings.
 - [DINOv2](./candle-examples/examples/dinov2/): computer vision model trained
   using self-supervision (can be used for imagenet classification, depth
   evaluation, segmentation).
+- [VGG](./candle-examples/examples/vgg/),
+  [RepVGG](./candle-examples/examples/repvgg): computer vision models.
 - [BLIP](./candle-examples/examples/blip/): image to text model, can be used to
   generate captions for an image.
+- [CLIP](./candle-examples/examples/clip/): multi-model vision and language
+  model.
+- [TrOCR](./candle-examples/examples/trocr/): a transformer OCR model, with
+  dedicated submodels for hand-writing and printed recognition.
 - [Marian-MT](./candle-examples/examples/marian-mt/): neural machine translation
   model, generates the translated text from the input text.
+- [Moondream](./candle-examples/examples/moondream/): tiny computer-vision model 
+  that can answer real-world questions about images.
 
 Run them using commands like:
 ```
@@ -120,7 +150,7 @@ There are also some wasm examples for whisper and
 [whisper](https://huggingface.co/spaces/lmz/candle-whisper),
 [llama2](https://huggingface.co/spaces/lmz/candle-llama2),
 [T5](https://huggingface.co/spaces/radames/Candle-T5-Generation-Wasm),
-[Phi-v1.5](https://huggingface.co/spaces/radames/Candle-Phi-1.5-Wasm),
+[Phi-1.5, and Phi-2](https://huggingface.co/spaces/radames/Candle-Phi-1.5-Wasm),
 [Segment Anything Model](https://huggingface.co/spaces/radames/candle-segment-anything-wasm).
 
 For LLaMA2, run the following command to retrieve the weight files and start a
@@ -137,17 +167,22 @@ And then head over to
 <!--- ANCHOR: useful_libraries --->
 
 ## Useful External Resources
-- [`candle-tutorial`](https://github.com/ToluClassics/candle-tutorial): a
+- [`candle-tutorial`](https://github.com/ToluClassics/candle-tutorial): A
   very detailed tutorial showing how to convert a PyTorch model to Candle.
-- [`optimisers`](https://github.com/KGrewal1/optimisers): a collection of optimisers
+- [`candle-lora`](https://github.com/EricLBuehler/candle-lora): Efficient and
+  ergonomic LoRA implementation for Candle. `candle-lora` has      
+  out-of-the-box LoRA support for many models from Candle, which can be found
+  [here](https://github.com/EricLBuehler/candle-lora/tree/master/candle-lora-transformers/examples).
+- [`optimisers`](https://github.com/KGrewal1/optimisers): A collection of optimisers
   including SGD with momentum, AdaGrad, AdaDelta, AdaMax, NAdam, RAdam, and RMSprop.
-- [`candle-lora`](https://github.com/EricLBuehler/candle-lora): a LoRA implementation
-  that conforms to the official `peft` implementation.
 - [`candle-vllm`](https://github.com/EricLBuehler/candle-vllm): Efficient platform for inference and
   serving local LLMs including an OpenAI compatible API server.
-- [`candle-ext`](https://github.com/mokeyish/candle-ext): an extension library to Candle that provides PyTorch functions not currently available in Candle.
-- [`kalosm`](https://github.com/floneum/floneum/tree/master/kalosm): A multi-modal meta-framework in Rust for interfacing with local pre-trained models with support for controlled generation, custom samplers, in-memory vector databases, audio transcription, and more.
+- [`candle-ext`](https://github.com/mokeyish/candle-ext): An extension library to Candle that provides PyTorch functions not currently available in Candle.
+- [`candle-coursera-ml`](https://github.com/vishpat/candle-coursera-ml): Implementation of ML algorithms from Coursera's [Machine Learning Specialization](https://www.coursera.org/specializations/machine-learning-introduction) course.
+- [`kalosm`](https://github.com/floneum/floneum/tree/master/interfaces/kalosm): A multi-modal meta-framework in Rust for interfacing with local pre-trained models with support for controlled generation, custom samplers, in-memory vector databases, audio transcription, and more.
 - [`candle-sampling`](https://github.com/EricLBuehler/candle-sampling): Sampling techniques for Candle.
+- [`gpt-from-scratch-rs`](https://github.com/jeroenvlek/gpt-from-scratch-rs): A port of Andrej Karpathy's _Let's build GPT_ tutorial on YouTube showcasing the Candle API on a toy problem.
+- [`candle-einops`](https://github.com/tomsanbear/candle-einops): A pure rust implementation of the python [einops](https://github.com/arogozhnikov/einops) library.
 
 If you have an addition to this list, please submit a pull request.
 
@@ -166,27 +201,45 @@ 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.
+        - LLaMA v1, v2, and v3 with variants such as SOLAR-10.7B.
         - Falcon.
-        - StarCoder.
-        - Phi v1.5.
+        - StarCoder, StarCoder2.
+        - Phi 1, 1.5, 2, and 3.
+        - Mamba, Minimal Mamba
+        - Gemma 2b and 7b.
         - Mistral 7b v0.1.
-        - StableLM-3B-4E1T.
+        - Mixtral 8x7b v0.1.
+        - StableLM-3B-4E1T, StableLM-2-1.6B, Stable-Code-3B.
         - Replit-code-v1.5-3B.
         - Bert.
+        - Yi-6B and Yi-34B.
+        - Qwen1.5, Qwen1.5 MoE.
+        - RWKV v5 and v6.
+    - Quantized LLMs.
+        - Llama 7b, 13b, 70b, as well as the chat and code variants.
+        - Mistral 7b, and 7b instruct.
+        - Mixtral 8x7b.
+        - Zephyr 7b a and b (Mistral-7b based).
+        - OpenChat 3.5 (Mistral-7b based).
     - Text to text.
-        - T5 and its variants: FlanT5, MADLAD400 (translation), CoEdit (Grammar correction).
+        - T5 and its variants: FlanT5, UL2, MADLAD400 (translation), CoEdit (Grammar correction).
         - Marian MT (Machine Translation).
-    - Whisper (multi-lingual support).
     - Text to image.
         - Stable Diffusion v1.5, v2.1, XL v1.0.
         - Wurstchen v2.
     - Image to text.
         - BLIP.
+        - TrOCR.
+    - Audio.
+        - Whisper, multi-lingual speech-to-text.
+        - EnCodec, audio compression model.
+        - MetaVoice-1B, text-to-speech model.
     - Computer Vision Models.
-        - DINOv2, ConvMixer, EfficientNet, ResNet, ViT.
+        - DINOv2, ConvMixer, EfficientNet, ResNet, ViT, VGG, RepVGG, ConvNeXT,
+          ConvNeXTv2, MobileOne, EfficientVit (MSRA).
         - yolo-v3, yolo-v8.
         - Segment-Anything Model (SAM).
+        - SegFormer.
 - File formats: load models from safetensors, npz, ggml, or PyTorch files.
 - Serverless (on CPU), small and fast deployments.
 - Quantization support using the llama.cpp quantized types.
@@ -323,9 +376,9 @@ 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
@@ -355,3 +408,10 @@ 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/Cargo.toml

@@ -11,11 +11,11 @@ readme = "README.md"
 
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
-candle = { path = "../candle-core", version = "0.3.0", package = "candle-core" }
-candle-datasets = { path = "../candle-datasets", version = "0.3.0" }
-candle-nn = { path = "../candle-nn", version = "0.3.0" }
-candle-transformers = { path = "../candle-transformers", version = "0.3.0" }
-candle-flash-attn = { path = "../candle-flash-attn", version = "0.3.0", optional = true }
+candle = { workspace = true }
+candle-datasets = { workspace = true }
+candle-nn = { workspace = true }
+candle-transformers = { workspace = true }
+candle-flash-attn = { workspace = true, optional = true }
 safetensors = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
@@ -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/lib.rs

@@ -28,6 +28,7 @@ let weights = candle::safetensors::load(weights_filename, &Device::Cpu).unwrap()
     #[rustfmt::skip]
     #[test]
     fn book_hub_2() {
+        {
 // ANCHOR: book_hub_2
 use candle::Device;
 use hf_hub::api::sync::Api;
@@ -45,9 +46,10 @@ let weights = candle::safetensors::load_buffer(&mmap[..], &Device::Cpu).unwrap()
         assert_eq!(weights.len(), 206);
     }
 
-    #[rustfmt::skip]
-    #[test]
-    fn book_hub_3() {
+    // #[rustfmt::skip]
+    // #[test]
+    // fn book_hub_3() {
+    {
 // ANCHOR: book_hub_3
 use candle::{DType, Device, Tensor};
 use hf_hub::api::sync::Api;
@@ -79,7 +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;
@@ -102,6 +104,7 @@ let tp_tensor = Tensor::from_raw_buffer(&raw, dtype, &tp_shape, &Device::Cpu).un
         assert_eq!(view.shape(), &[768, 768]);
         assert_eq!(tp_tensor.dims(), &[192, 768]);
     }
+}
 
     #[rustfmt::skip]
     #[test]

candle-core/Cargo.toml

@@ -12,7 +12,9 @@ readme = "README.md"
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
 byteorder = { workspace = true }
-candle-kernels = { path = "../candle-kernels", version = "0.3.0", optional = true }
+candle-kernels = { workspace = true, optional = true }
+candle-metal-kernels = { workspace = true, optional = true }
+metal = { workspace = true, optional = true}
 cudarc = { workspace = true, optional = true }
 gemm = { workspace = true }
 half = { workspace = true }
@@ -32,6 +34,8 @@ zip = { workspace = true }
 [dev-dependencies]
 anyhow = { workspace = true }
 clap = { workspace = true }
+criterion = { workspace = true }
+
 
 [features]
 default = []
@@ -39,3 +43,8 @@ cuda = ["cudarc", "dep:candle-kernels"]
 cudnn = ["cuda", "cudarc/cudnn"]
 mkl = ["dep:libc", "dep:intel-mkl-src"]
 accelerate = ["dep:libc", "dep:accelerate-src"]
+metal = ["dep:metal", "dep:candle-metal-kernels"]
+
+[[bench]]
+name = "bench_main"
+harness = false

candle-core/benches/bench_main.rs

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

candle-core/benches/benchmarks/affine.rs

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

candle-core/benches/benchmarks/conv_transpose2d.rs

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

candle-core/benches/benchmarks/matmul.rs

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

candle-core/benches/benchmarks/mod.rs

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

candle-core/benches/benchmarks/qmatmul.rs

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

candle-core/benches/benchmarks/random.rs

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

candle-core/benches/benchmarks/unary.rs

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

candle-core/benches/benchmarks/where_cond.rs

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

candle-core/examples/basics.rs

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

candle-core/examples/cuda_basics.rs

@@ -9,21 +9,22 @@ use candle_core::{Device, Tensor};
 
 fn main() -> Result<()> {
     let device = Device::new_cuda(0)?;
-    let in_t = Tensor::rand(-1f32, 1f32, (1, 3, 12, 7), &device)?;
-    let k_t = Tensor::rand(-1f32, 1f32, (6, 3, 1, 1), &device)?;
-    let out_t = in_t.conv2d(&k_t, 0, 1, 1, 1)?;
-    println!("{out_t}");
-    let in_t = in_t.to_device(&Device::Cpu)?;
-    let k_t = k_t.to_device(&Device::Cpu)?;
-    let out_t2 = in_t.conv2d(&k_t, 0, 1, 1, 1)?;
-    let diff = (out_t.to_device(&Device::Cpu)? - out_t2)?
-        .sqr()?
-        .sum_all()?;
-    println!("{diff}");
-
-    let t = Tensor::randn(0f32, 1f32, (2, 4, 96, 96), &device)?;
-    let w = Tensor::randn(0f32, 1f32, (320, 4, 3, 3), &device)?;
-    let res = t.conv2d(&w, 1, 1, 1, 1)?;
-    println!("{res:?}");
+    let x = Tensor::randn(0f32, 1.0, (8 * 4096, 8 * 4096), &device)?;
+    candle_core::cuda::set_gemm_reduced_precision_f32(false);
+    let _x1 = x.matmul(&x)?;
+    drop(_x1);
+    let start_time = std::time::Instant::now();
+    let _x1 = x.matmul(&x)?;
+    device.synchronize()?;
+    println!("fp32: {:?}", start_time.elapsed());
+    drop(_x1);
+    candle_core::cuda::set_gemm_reduced_precision_f32(true);
+    let _x1 = x.matmul(&x)?;
+    drop(_x1);
+    let start_time = std::time::Instant::now();
+    let _x1 = x.matmul(&x)?;
+    device.synchronize()?;
+    println!("tf32: {:?}", start_time.elapsed());
+    drop(_x1);
     Ok(())
 }

candle-core/src/accelerate.rs

@@ -380,6 +380,16 @@ pub fn vd_tanh_inplace(y: &mut [f64]) {
     unsafe { ffi::vvtanh(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
 }
 
+#[inline]
+pub fn vs_exp_inplace(y: &mut [f32]) {
+    unsafe { ffi::vvexpf(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
+}
+
+#[inline]
+pub fn vd_exp_inplace(y: &mut [f64]) {
+    unsafe { ffi::vvexp(y.as_mut_ptr(), y.as_ptr(), &(y.len() as i32)) }
+}
+
 #[inline]
 pub fn vs_gelu(vs: &[f32], ys: &mut [f32]) {
     for (&v, y) in vs.iter().zip(ys.iter_mut()) {
@@ -402,6 +412,28 @@ pub fn vd_gelu(vs: &[f64], ys: &mut [f64]) {
     }
 }
 
+#[inline]
+pub fn vs_silu(vs: &[f32], ys: &mut [f32]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vs_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
+#[inline]
+pub fn vd_silu(vs: &[f64], ys: &mut [f64]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vd_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
 macro_rules! binary_op {
     ($fn_name:ident, $ty:ty, $accelerate_name:ident) => {
         #[inline]

candle-core/src/backend.rs

@@ -98,6 +98,19 @@ pub trait BackendStorage: Sized {
     ) -> Result<Self>;
 
     fn copy_strided_src(&self, _: &mut Self, _: usize, _: &Layout) -> Result<()>;
+
+    #[allow(clippy::too_many_arguments)]
+    // Similar to cudaMemcpy2D, though values are in elements and not in bytes.
+    fn copy2d(
+        &self,
+        _: &mut Self,
+        _d1: usize,
+        _d2: usize,
+        _src_stride1: usize,
+        _dst_stride1: usize,
+        _src_offset: usize,
+        _dst_offset: usize,
+    ) -> Result<()>;
 }
 
 pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
@@ -114,11 +127,24 @@ pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
 
     fn ones_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;
 
+    /// # Safety
+    /// This function is unsafe as it doesn't initialize the underlying data store.
+    /// The caller should ensure that the data is properly initialized as early as possible
+    /// after this call.
+    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;
+
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage>;
+
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage>;
 
+    fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage>;
+
     fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;
 
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;
 
     fn set_seed(&self, _: u64) -> Result<()>;
+
+    /// Synchronize should block until all the operations on the device are completed.
+    fn synchronize(&self) -> Result<()>;
 }

candle-core/src/backprop.rs

@@ -1,3 +1,4 @@
+/// Methods for backpropagation of gradients.
 use crate::op::{BinaryOp, Op, ReduceOp, UnaryOp};
 use crate::{Error, Result, Tensor, TensorId};
 use std::collections::HashMap;
@@ -111,10 +112,11 @@ impl Tensor {
                     }
                     Op::Unary(_node, UnaryOp::Ceil)
                     | Op::Unary(_node, UnaryOp::Floor)
-                    | Op::Unary(_node, UnaryOp::Round) => nodes,
+                    | Op::Unary(_node, UnaryOp::Round)
+                    | Op::Unary(_node, UnaryOp::Sign) => nodes,
                     Op::Reshape(node)
-                    | Op::UpsampleNearest1D(node)
-                    | Op::UpsampleNearest2D(node)
+                    | Op::UpsampleNearest1D { arg: node, .. }
+                    | Op::UpsampleNearest2D { arg: node, .. }
                     | Op::AvgPool2D { arg: node, .. }
                     | Op::MaxPool2D { arg: node, .. }
                     | Op::Copy(node)
@@ -175,7 +177,7 @@ impl Tensor {
             // the backprop graph of the backprop itself. This would be an issue for second order
             // derivatives but these are out of scope at the moment.
             let do_not_detach = CANDLE_GRAD_DO_NOT_DETACH.with(|b| *b);
-            let grad = if do_not_detach { grad } else { grad.detach()? };
+            let grad = if do_not_detach { grad } else { grad.detach() };
             if let Some(op) = node.op() {
                 match op {
                     Op::Binary(lhs, rhs, BinaryOp::Add) => {
@@ -250,6 +252,7 @@ impl Tensor {
                             out_padding,
                             *stride,
                             *dilation,
+                            /* groups */ 1,
                         )?;
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&grad_arg)?;
@@ -309,9 +312,32 @@ impl Tensor {
                     Op::ConvTranspose1D { .. } => Err(Error::BackwardNotSupported {
                         op: "conv-transpose1d",
                     })?,
-                    Op::ConvTranspose2D { .. } => Err(Error::BackwardNotSupported {
-                        op: "conv-transpose2d",
-                    })?,
+                    Op::ConvTranspose2D {
+                        arg,
+                        kernel,
+                        padding,
+                        stride,
+                        dilation,
+                        output_padding: _output_padding,
+                    } => {
+                        let grad_arg = grad.conv2d(kernel, *padding, *dilation, *stride, 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)?
+                            .transpose(0, 1)?;
+                        let sum_grad = grads.or_insert(kernel)?;
+                        let (_, _, k0, k1) = kernel.dims4()?;
+                        let (_, _, g_k0, g_k1) = grad_kernel.dims4()?;
+                        let grad_kernel = if g_k0 != k0 || g_k1 != k1 {
+                            grad_kernel.narrow(2, 0, k0)?.narrow(3, 0, k1)?
+                        } else {
+                            grad_kernel
+                        };
+                        *sum_grad = sum_grad.add(&grad_kernel)?;
+                    }
                     Op::AvgPool2D {
                         arg,
                         kernel_size,
@@ -347,12 +373,39 @@ impl Tensor {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&grad_arg)?;
                     }
-                    Op::UpsampleNearest1D { .. } => Err(Error::BackwardNotSupported {
-                        op: "upsample-nearest1d",
-                    })?,
-                    Op::UpsampleNearest2D { .. } => Err(Error::BackwardNotSupported {
-                        op: "upsample-nearest2d",
-                    })?,
+                    Op::UpsampleNearest1D { arg, target_size } => {
+                        let (_n, c, size) = arg.dims3()?;
+                        if target_size % size != 0 {
+                            crate::bail!("backward not supported for non integer upscaling factors")
+                        }
+                        let scale = target_size / size;
+
+                        let kernel = Tensor::ones((c, 1, scale), arg.dtype(), arg.device())?;
+                        let conv_sum = grad.conv1d(&kernel, 0, scale, 1, c)?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = conv_sum;
+                    }
+                    Op::UpsampleNearest2D {
+                        arg,
+                        target_h,
+                        target_w,
+                    } => {
+                        let (_n, c, h, w) = arg.dims4()?;
+                        if target_h % h != 0 || target_w % w != 0 {
+                            crate::bail!("backward not supported for non integer upscaling factors")
+                        }
+                        let scale_h = target_h / h;
+                        let scale_w = target_w / w;
+
+                        if scale_h != scale_w {
+                            crate::bail!("backward not supported for non uniform upscaling factors")
+                        };
+                        let kernel =
+                            Tensor::ones((c, 1, scale_h, scale_w), arg.dtype(), arg.device())?;
+                        let conv_sum = grad.conv2d(&kernel, 0, scale_h, 1, c)?;
+                        let sum_grad = grads.or_insert(arg)?;
+                        *sum_grad = conv_sum;
+                    }
                     Op::SliceScatter0(lhs, rhs, start_rhs) => {
                         let rhs_sum_grad = grads.or_insert(rhs)?;
                         let rhs_grad = grad.narrow(0, *start_rhs, rhs.dim(0)?)?;
@@ -436,7 +489,6 @@ impl Tensor {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&grad)?;
                     }
-                    Op::Cmp(_args, _) => {}
                     Op::Reduce(arg, ReduceOp::Max, reduced_dims) => {
                         let node = broadcast_back(arg, node, reduced_dims)?;
                         let grad = broadcast_back(arg, &grad, reduced_dims)?;
@@ -526,20 +578,18 @@ impl Tensor {
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&arg_grad)?
                     }
-                    Op::Reduce(_, ReduceOp::ArgMin, _) => {}
-                    Op::Reduce(_, ReduceOp::ArgMax, _) => {}
+                    Op::Unary(_, UnaryOp::Floor)
+                    | Op::Unary(_, UnaryOp::Round)
+                    | Op::Reduce(_, ReduceOp::ArgMin, _)
+                    | Op::Reduce(_, ReduceOp::ArgMax, _)
+                    | Op::Unary(_, UnaryOp::Sign)
+                    | Op::Cmp(_, _) => {}
                     Op::Reshape(arg) => {
                         let arg_grad = grad.reshape(arg.dims())?;
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&arg_grad)?
                     }
                     Op::Unary(_, UnaryOp::Ceil) => Err(Error::BackwardNotSupported { op: "ceil" })?,
-                    Op::Unary(_, UnaryOp::Floor) => {
-                        Err(Error::BackwardNotSupported { op: "floor" })?
-                    }
-                    Op::Unary(_, UnaryOp::Round) => {
-                        Err(Error::BackwardNotSupported { op: "round" })?
-                    }
                     Op::Unary(arg, UnaryOp::Gelu) => {
                         let sum_grad = grads.or_insert(arg)?;
                         let cube = arg.powf(3.)?;
@@ -571,6 +621,13 @@ impl Tensor {
                         let relu_grad = arg.ge(&arg.zeros_like()?)?.to_dtype(arg.dtype())?;
                         *sum_grad = sum_grad.add(&(&grad * relu_grad)?)?
                     }
+                    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 = (arg.neg()?.exp()? + 1.)?.recip()?;
+                        let silu_grad = (&sigmoid_arg * (1. + (arg * (1. - &sigmoid_arg)?)?)?)?;
+                        *sum_grad = sum_grad.add(&(&grad * silu_grad)?)?
+                    }
                     Op::Elu(arg, alpha) => {
                         // d/dx elu(x) = 1 for x > 0, alpha * e^x for x <= 0
                         let sum_grad = grads.or_insert(arg)?;
@@ -655,30 +712,38 @@ impl Tensor {
     }
 }
 
+/// A store for gradients, associating a tensor id to the corresponding gradient tensor, used for back propagation.
 #[derive(Debug)]
 pub struct GradStore(HashMap<TensorId, Tensor>);
 
 impl GradStore {
+    /// Create a new gradient store
     fn new() -> Self {
         GradStore(HashMap::new())
     }
 
+    /// Get the gradient tensor corresponding to the given tensor id
     pub fn get_id(&self, id: TensorId) -> Option<&Tensor> {
         self.0.get(&id)
     }
 
+    /// Get the gradient tensor associated with the given tensor
     pub fn get(&self, tensor: &Tensor) -> Option<&Tensor> {
         self.0.get(&tensor.id())
     }
 
+    /// Remove the gradient tensor associated with the given tensor, returning it if it exists
     pub fn remove(&mut self, tensor: &Tensor) -> Option<Tensor> {
         self.0.remove(&tensor.id())
     }
 
+    /// Insert a gradient tensor associated with the given tensor, returning the previous gradient tensor if it existed
     pub fn insert(&mut self, tensor: &Tensor, grad: Tensor) -> Option<Tensor> {
         self.0.insert(tensor.id(), grad)
     }
 
+    /// Get the gradient tensor associated with the given tensor, or, if it does not exist,
+    /// insert a tensor of zeroes, with the same shape and type as the given tensors and return it
     fn or_insert(&mut self, tensor: &Tensor) -> Result<&mut Tensor> {
         use std::collections::hash_map::Entry;
         let grad = match self.0.entry(tensor.id()) {

candle-core/src/conv.rs

@@ -187,36 +187,16 @@ impl Tensor {
         }
     }
 
-    /// Applies a 1D transposed convolution over the input tensor.
-    pub fn conv_transpose1d(
+    fn conv_transpose1d_single_group(
         &self,
         kernel: &Self,
-        padding: usize,
-        output_padding: usize,
-        stride: usize,
-        dilation: usize,
+        params: &ParamsConvTranspose1D,
     ) -> Result<Self> {
-        let (b_size, c_in, l_in) = self.dims3()?;
-        let (c_in_k, c_out, k_size) = kernel.dims3()?;
-        if c_in != c_in_k {
-            crate::bail!("in_channel mismatch between input ({c_in}) and kernel ({c_in_k})")
-        }
-        let params = ParamsConvTranspose1D {
-            b_size,
-            l_in,
-            k_size,
-            c_out,
-            c_in,
-            padding,
-            output_padding,
-            stride,
-            dilation,
-        };
         let storage = self.storage().conv_transpose1d(
             self.layout(),
             &kernel.storage(),
             kernel.layout(),
-            &params,
+            params,
         )?;
         let op = BackpropOp::new2(self, kernel, |arg, kernel| Op::ConvTranspose1D {
             arg,
@@ -230,6 +210,49 @@ impl Tensor {
         Ok(crate::tensor::from_storage(storage, out_dims, op, false))
     }
 
+    /// Applies a 1D transposed convolution over the input tensor.
+    pub fn conv_transpose1d(
+        &self,
+        kernel: &Self,
+        padding: usize,
+        output_padding: usize,
+        stride: usize,
+        dilation: usize,
+        groups: usize,
+    ) -> Result<Self> {
+        let (c_in_k, c_out, k_size) = kernel.dims3()?;
+        let (b_size, c_in, l_in) = self.dims3()?;
+        if c_in != c_in_k {
+            crate::bail!("in_channel mismatch between input ({c_in}) and kernel ({c_in_k})")
+        }
+        if c_in % groups != 0 {
+            crate::bail!("in_channel {c_in} is not divisible by the number of groups")
+        }
+        let params = ParamsConvTranspose1D {
+            b_size,
+            l_in,
+            k_size,
+            c_out,
+            c_in: c_in / groups,
+            padding,
+            output_padding,
+            stride,
+            dilation,
+        };
+        if groups == 1 {
+            self.conv_transpose1d_single_group(kernel, &params)
+        } else {
+            let blocks = self.chunk(groups, 1)?;
+            let kernel = kernel.chunk(groups, 0)?;
+            let blocks = blocks
+                .iter()
+                .zip(&kernel)
+                .map(|(block, kernel)| block.conv_transpose1d_single_group(kernel, &params))
+                .collect::<Result<Vec<_>>>()?;
+            Tensor::cat(&blocks, 1)
+        }
+    }
+
     fn conv2d_single_group(&self, kernel: &Self, params: &ParamsConv2D) -> Result<Self> {
         let storage =
             self.storage()

candle-core/src/cpu/mod.rs

@@ -1,6 +1,7 @@
 pub mod erf;
 pub mod kernels;
 
+#[allow(unused)]
 trait Cpu<const ARR: usize> {
     type Unit;
     type Array;
@@ -18,6 +19,7 @@ trait Cpu<const ARR: usize> {
     unsafe fn vec_store(mem_addr: *mut f32, a: Self::Unit);
 }
 
+#[allow(unused)]
 trait CpuF16<const ARR: usize> {
     type Unit;
     type Array;

candle-core/src/cpu_backend/mod.rs

@@ -4,7 +4,13 @@ use crate::{DType, Error, IntDType, Layout, Result, Shape, WithDType};
 use half::{bf16, f16};
 use rayon::prelude::*;
 
+mod utils;
+pub use utils::{
+    binary_map, binary_map_vec, unary_map, unary_map_vec, Map1, Map1Any, Map2, Map2U8,
+};
+
 const USE_IM2COL_CONV1D: bool = true;
+const USE_IM2COL_CONV1D_TR: bool = true;
 const USE_IM2COL_CONV2D: bool = true;
 
 // TODO: Maybe we should not implement [Clone] here and instead have an explicit allocator +
@@ -20,105 +26,20 @@ 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;
 
-pub trait Map1 {
-    fn f<T: WithDType>(&self, vs: &[T], layout: &Layout) -> Result<Vec<T>>;
-
-    fn map(&self, vs: &CpuStorage, layout: &Layout) -> Result<CpuStorage> {
-        match vs {
-            CpuStorage::U8(vs) => Ok(CpuStorage::U8(self.f(vs, layout)?)),
-            CpuStorage::U32(vs) => Ok(CpuStorage::U32(self.f(vs, layout)?)),
-            CpuStorage::I64(vs) => Ok(CpuStorage::I64(self.f(vs, layout)?)),
-            CpuStorage::BF16(vs) => Ok(CpuStorage::BF16(self.f(vs, layout)?)),
-            CpuStorage::F16(vs) => Ok(CpuStorage::F16(self.f(vs, layout)?)),
-            CpuStorage::F32(vs) => Ok(CpuStorage::F32(self.f(vs, layout)?)),
-            CpuStorage::F64(vs) => Ok(CpuStorage::F64(self.f(vs, layout)?)),
-        }
-    }
-}
-
-pub trait Map1Any {
-    fn f<T: WithDType, W: Fn(Vec<T>) -> CpuStorage>(
-        &self,
-        vs: &[T],
-        layout: &Layout,
-        wrap: W,
-    ) -> Result<CpuStorage>;
-
-    fn map(&self, vs: &CpuStorage, layout: &Layout) -> Result<CpuStorage> {
-        match vs {
-            CpuStorage::U8(vs) => Ok(self.f(vs, layout, CpuStorage::U8)?),
-            CpuStorage::U32(vs) => Ok(self.f(vs, layout, CpuStorage::U32)?),
-            CpuStorage::I64(vs) => Ok(self.f(vs, layout, CpuStorage::I64)?),
-            CpuStorage::BF16(vs) => Ok(self.f(vs, layout, CpuStorage::BF16)?),
-            CpuStorage::F16(vs) => Ok(self.f(vs, layout, CpuStorage::F16)?),
-            CpuStorage::F32(vs) => Ok(self.f(vs, layout, CpuStorage::F32)?),
-            CpuStorage::F64(vs) => Ok(self.f(vs, layout, CpuStorage::F64)?),
-        }
-    }
-}
-
-type C = CpuStorage;
-pub trait Map2 {
-    const OP: &'static str;
-    fn f<T: WithDType>(&self, v1: &[T], l1: &Layout, v2: &[T], l2: &Layout) -> Result<Vec<T>>;
-
-    fn map(
-        &self,
-        v1: &CpuStorage,
-        l1: &Layout,
-        v2: &CpuStorage,
-        l2: &Layout,
-    ) -> Result<CpuStorage> {
-        match (v1, v2) {
-            (C::U8(v1), C::U8(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
-            (C::U32(v1), C::U32(v2)) => Ok(C::U32(self.f(v1, l1, v2, l2)?)),
-            (C::I64(v1), C::I64(v2)) => Ok(C::I64(self.f(v1, l1, v2, l2)?)),
-            (C::BF16(v1), C::BF16(v2)) => Ok(C::BF16(self.f(v1, l1, v2, l2)?)),
-            (C::F16(v1), C::F16(v2)) => Ok(C::F16(self.f(v1, l1, v2, l2)?)),
-            (C::F32(v1), C::F32(v2)) => Ok(C::F32(self.f(v1, l1, v2, l2)?)),
-            (C::F64(v1), C::F64(v2)) => Ok(C::F64(self.f(v1, l1, v2, l2)?)),
-            _ => Err(Error::DTypeMismatchBinaryOp {
-                lhs: v1.dtype(),
-                rhs: v2.dtype(),
-                op: Self::OP,
-            }
-            .bt()),
-        }
-    }
-}
-
-pub trait Map2U8 {
-    const OP: &'static str;
-    fn f<T: WithDType>(&self, v1: &[T], l1: &Layout, v2: &[T], l2: &Layout) -> Result<Vec<u8>>;
-
-    fn map(
-        &self,
-        v1: &CpuStorage,
-        l1: &Layout,
-        v2: &CpuStorage,
-        l2: &Layout,
-    ) -> Result<CpuStorage> {
-        match (v1, v2) {
-            (C::U8(v1), C::U8(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
-            (C::U32(v1), C::U32(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
-            (C::I64(v1), C::I64(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
-            (C::BF16(v1), C::BF16(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
-            (C::F16(v1), C::F16(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
-            (C::F32(v1), C::F32(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
-            (C::F64(v1), C::F64(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
-            _ => Err(Error::DTypeMismatchBinaryOp {
-                lhs: v1.dtype(),
-                rhs: v2.dtype(),
-                op: Self::OP,
-            }
-            .bt()),
-        }
-    }
-}
-
 struct Cmp(CmpOp);
 impl Map2U8 for Cmp {
     const OP: &'static str = "cmp";
@@ -365,275 +286,6 @@ impl<'a> Map1 for ReduceSum<'a> {
     }
 }
 
-pub fn unary_map<T: Copy, U: Copy, F: FnMut(T) -> U>(
-    vs: &[T],
-    layout: &Layout,
-    mut f: F,
-) -> Vec<U> {
-    match layout.strided_blocks() {
-        crate::StridedBlocks::SingleBlock { start_offset, len } => vs
-            [start_offset..start_offset + len]
-            .iter()
-            .map(|&v| f(v))
-            .collect(),
-        crate::StridedBlocks::MultipleBlocks {
-            block_start_index,
-            block_len,
-        } => {
-            let mut result = Vec::with_capacity(layout.shape().elem_count());
-            // Specialize the case where block_len is one to avoid the second loop.
-            if block_len == 1 {
-                for index in block_start_index {
-                    let v = unsafe { vs.get_unchecked(index) };
-                    result.push(f(*v))
-                }
-            } else {
-                for index in block_start_index {
-                    for offset in 0..block_len {
-                        let v = unsafe { vs.get_unchecked(index + offset) };
-                        result.push(f(*v))
-                    }
-                }
-            }
-            result
-        }
-    }
-}
-
-pub fn unary_map_vec<T: Copy, U: Copy, F: FnMut(T) -> U, FV: FnMut(&[T], &mut [U])>(
-    vs: &[T],
-    layout: &Layout,
-    mut f: F,
-    mut f_vec: FV,
-) -> Vec<U> {
-    match layout.strided_blocks() {
-        crate::StridedBlocks::SingleBlock { start_offset, len } => {
-            let mut ys: Vec<U> = Vec::with_capacity(len);
-            let ys_to_set = ys.spare_capacity_mut();
-            let ys_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(ys_to_set) };
-            f_vec(&vs[start_offset..start_offset + len], ys_to_set);
-            // SAFETY: values are all set by f_vec.
-            unsafe { ys.set_len(len) };
-            ys
-        }
-        crate::StridedBlocks::MultipleBlocks {
-            block_start_index,
-            block_len,
-        } => {
-            let el_count = layout.shape().elem_count();
-            // Specialize the case where block_len is one to avoid the second loop.
-            if block_len == 1 {
-                let mut result = Vec::with_capacity(el_count);
-                for index in block_start_index {
-                    let v = unsafe { vs.get_unchecked(index) };
-                    result.push(f(*v))
-                }
-                result
-            } else {
-                let mut ys: Vec<U> = Vec::with_capacity(el_count);
-                let ys_to_set = ys.spare_capacity_mut();
-                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(ys_to_set) };
-                let mut dst_index = 0;
-                for src_index in block_start_index {
-                    let vs = &vs[src_index..src_index + block_len];
-                    let ys = &mut ys_to_set[dst_index..dst_index + block_len];
-                    f_vec(vs, ys);
-                    dst_index += block_len;
-                }
-                // SAFETY: values are all set by f_vec.
-                unsafe { ys.set_len(el_count) };
-                ys
-            }
-        }
-    }
-}
-
-// This function maps over two strided index sequences.
-pub fn binary_map<T: Copy, U: Copy, F: FnMut(T, T) -> U>(
-    lhs_l: &Layout,
-    rhs_l: &Layout,
-    lhs: &[T],
-    rhs: &[T],
-    mut f: F,
-) -> Vec<U> {
-    match (lhs_l.contiguous_offsets(), rhs_l.contiguous_offsets()) {
-        (Some((o_l1, o_l2)), Some((o_r1, o_r2))) => lhs[o_l1..o_l2]
-            .iter()
-            .zip(rhs[o_r1..o_r2].iter())
-            .map(|(&l, &r)| f(l, r))
-            .collect(),
-        (Some((o_l1, o_l2)), None) => {
-            // TODO: Maybe we want to avoid going through the layout twice.
-            match rhs_l.offsets_b() {
-                Some(ob) => {
-                    let mut i_in_block = 0;
-                    let mut i_right_broadcast = 0;
-                    lhs[o_l1..o_l2]
-                        .iter()
-                        .map(|&l| {
-                            let r = unsafe { rhs.get_unchecked(i_in_block + ob.start) };
-                            i_right_broadcast += 1;
-                            if i_right_broadcast >= ob.right_broadcast {
-                                i_in_block += 1;
-                                i_right_broadcast = 0;
-                            }
-                            if i_in_block >= ob.len {
-                                i_in_block = 0
-                            }
-                            f(l, *r)
-                        })
-                        .collect()
-                }
-                None => lhs_l
-                    .strided_index()
-                    .zip(rhs_l.strided_index())
-                    .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
-                    .collect(),
-            }
-        }
-        (None, Some((o_r1, o_r2))) => {
-            // TODO: Maybe we want to avoid going through the layout twice.
-            match lhs_l.offsets_b() {
-                Some(ob) => {
-                    let mut i_in_block = 0;
-                    let mut i_right_broadcast = 0;
-                    rhs[o_r1..o_r2]
-                        .iter()
-                        .map(|&r| {
-                            let l = unsafe { lhs.get_unchecked(i_in_block + ob.start) };
-                            i_right_broadcast += 1;
-                            if i_right_broadcast >= ob.right_broadcast {
-                                i_in_block += 1;
-                                i_right_broadcast = 0;
-                            }
-                            if i_in_block >= ob.len {
-                                i_in_block = 0
-                            }
-                            f(*l, r)
-                        })
-                        .collect()
-                }
-                None => lhs_l
-                    .strided_index()
-                    .zip(rhs_l.strided_index())
-                    .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
-                    .collect(),
-            }
-        }
-        _ => lhs_l
-            .strided_index()
-            .zip(rhs_l.strided_index())
-            .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
-            .collect(),
-    }
-}
-
-// Similar to binary_map but with vectorized variants.
-pub fn binary_map_vec<T: Copy, F: FnMut(T, T) -> T, FV: FnMut(&[T], &[T], &mut [T])>(
-    lhs_l: &Layout,
-    rhs_l: &Layout,
-    lhs: &[T],
-    rhs: &[T],
-    mut f: F,
-    mut f_vec: FV,
-) -> Vec<T> {
-    let el_count = lhs_l.shape().elem_count();
-    match (lhs_l.contiguous_offsets(), rhs_l.contiguous_offsets()) {
-        (Some((o_l1, o_l2)), Some((o_r1, o_r2))) => {
-            let mut ys: Vec<T> = Vec::with_capacity(el_count);
-            let ys_to_set = ys.spare_capacity_mut();
-            let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
-            f_vec(&lhs[o_l1..o_l2], &rhs[o_r1..o_r2], ys_to_set);
-            // SAFETY: values are all set by f_vec.
-            unsafe { ys.set_len(el_count) };
-            ys
-        }
-        (Some((o_l1, o_l2)), None) => match rhs_l.offsets_b() {
-            Some(ob) if ob.right_broadcast == 1 => {
-                let rhs = &rhs[ob.start..ob.start + ob.len];
-                let mut ys: Vec<T> = Vec::with_capacity(el_count);
-                let ys_to_set = ys.spare_capacity_mut();
-                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
-                let mut dst_i = 0;
-                for src_i in (o_l1..o_l2).step_by(ob.len) {
-                    f_vec(
-                        &lhs[src_i..src_i + ob.len],
-                        rhs,
-                        &mut ys_to_set[dst_i..dst_i + ob.len],
-                    );
-                    dst_i += ob.len;
-                }
-                // SAFETY: values are all set by f_vec.
-                unsafe { ys.set_len(el_count) };
-                ys
-            }
-            Some(ob) => {
-                let rhs = &rhs[ob.start..ob.start + ob.len];
-                let mut ys = lhs[o_l1..o_l2].to_vec();
-                for idx_l in 0..ob.left_broadcast {
-                    let start = idx_l * ob.len * ob.right_broadcast;
-                    for (i, &r) in rhs.iter().enumerate() {
-                        let start = start + i * ob.right_broadcast;
-                        for v in ys[start..start + ob.right_broadcast].iter_mut() {
-                            *v = f(*v, r)
-                        }
-                    }
-                }
-                ys
-            }
-            None => lhs_l
-                .strided_index()
-                .zip(rhs_l.strided_index())
-                .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
-                .collect(),
-        },
-        (None, Some((o_r1, o_r2))) => match lhs_l.offsets_b() {
-            Some(ob) if ob.right_broadcast == 1 => {
-                let lhs = &lhs[ob.start..ob.start + ob.len];
-                let mut ys: Vec<T> = Vec::with_capacity(el_count);
-                let ys_to_set = ys.spare_capacity_mut();
-                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
-                let mut dst_i = 0;
-                for src_i in (o_r1..o_r2).step_by(ob.len) {
-                    f_vec(
-                        lhs,
-                        &rhs[src_i..src_i + ob.len],
-                        &mut ys_to_set[dst_i..dst_i + ob.len],
-                    );
-                    dst_i += ob.len;
-                }
-                // SAFETY: values are all set by f_vec.
-                unsafe { ys.set_len(el_count) };
-                ys
-            }
-            Some(ob) => {
-                let lhs = &lhs[ob.start..ob.start + ob.len];
-                let mut ys = rhs[o_r1..o_r2].to_vec();
-                for idx_l in 0..ob.left_broadcast {
-                    let start = idx_l * ob.len * ob.right_broadcast;
-                    for (i, &l) in lhs.iter().enumerate() {
-                        let start = start + i * ob.right_broadcast;
-                        for v in ys[start..start + ob.right_broadcast].iter_mut() {
-                            *v = f(l, *v)
-                        }
-                    }
-                }
-                ys
-            }
-            None => lhs_l
-                .strided_index()
-                .zip(rhs_l.strided_index())
-                .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
-                .collect(),
-        },
-        _ => lhs_l
-            .strided_index()
-            .zip(rhs_l.strided_index())
-            .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
-            .collect(),
-    }
-}
-
 struct Affine(f64, f64);
 
 impl Map1 for Affine {
@@ -1022,6 +674,26 @@ impl<'a, I: IntDType> Map2 for IndexAdd<'a, I> {
     }
 }
 
+#[allow(clippy::too_many_arguments)]
+fn copy2d_<T: Copy>(
+    src: &[T],
+    dst: &mut [T],
+    d1: usize,
+    d2: usize,
+    src_stride1: usize,
+    dst_stride1: usize,
+    src_offset: usize,
+    dst_offset: usize,
+) {
+    for i1 in 0..d1 {
+        let dst_idx = i1 * dst_stride1 + dst_offset;
+        let src_idx = i1 * src_stride1 + src_offset;
+        let dst = &mut dst[dst_idx..dst_idx + d2];
+        let src = &src[src_idx..src_idx + d2];
+        dst.copy_from_slice(src)
+    }
+}
+
 fn copy_strided_src_<T: Copy>(src: &[T], dst: &mut [T], dst_offset: usize, src_l: &Layout) {
     match src_l.strided_blocks() {
         crate::StridedBlocks::SingleBlock { start_offset, len } => {
@@ -1256,6 +928,34 @@ impl Map1 for Im2Col {
     }
 }
 
+struct Col2Im1D {
+    stride: usize,
+}
+
+impl Map1 for Col2Im1D {
+    fn f<T: WithDType>(&self, col: &[T], l: &Layout) -> Result<Vec<T>> {
+        let (b_size, l_in, c_out, k_size) = l.shape().dims4()?;
+        let stride = self.stride;
+        let l_out = (l_in - 1) * stride + k_size;
+        let mut im = vec![T::zero(); b_size * c_out * l_out];
+        let (dst_s0, dst_s1) = (c_out * l_out, l_out);
+        let (src_s0, src_s1, src_s2) = (c_out * k_size * l_in, c_out * k_size, k_size);
+        for l_in_i in 0..l_in {
+            for k_i in 0..k_size {
+                let l_out_i = l_in_i * stride + k_i;
+                for b_i in 0..b_size {
+                    for c_i in 0..c_out {
+                        let dst_idx = b_i * dst_s0 + c_i * dst_s1 + l_out_i;
+                        let src_idx = b_i * src_s0 + l_in_i * src_s1 + c_i * src_s2 + k_i;
+                        im[dst_idx] += col[src_idx]
+                    }
+                }
+            }
+        }
+        Ok(im)
+    }
+}
+
 struct ConvTranspose1D<'a>(&'a crate::conv::ParamsConvTranspose1D);
 
 impl<'a> Map2 for ConvTranspose1D<'a> {
@@ -1263,6 +963,7 @@ impl<'a> Map2 for ConvTranspose1D<'a> {
     fn f<T: WithDType>(&self, inp: &[T], inp_l: &Layout, k: &[T], k_l: &Layout) -> Result<Vec<T>> {
         let p = self.0;
         let inp = &inp[inp_l.start_offset()..];
+        let k = &k[k_l.start_offset()..];
         let (inp_s0, inp_s1, inp_s2) = crate::shape::dims3(inp_l.stride())?;
         let (k_s0, k_s1, k_s2) = crate::shape::dims3(k_l.stride())?;
         let l_out = p.l_out();
@@ -1514,6 +1215,30 @@ impl MatMul {
         }))
         .bt()
     }
+
+    fn ab_skip(&self, lhs_l: &Layout, rhs_l: &Layout) -> Result<(usize, usize)> {
+        let lhs_stride = lhs_l.stride();
+        let rhs_stride = rhs_l.stride();
+        let rank = lhs_stride.len();
+        let (_b, m, n, k) = self.0;
+        let a_skip: usize = match lhs_stride[..rank - 2] {
+            [s1, stride] if s1 == stride * lhs_l.dims()[1] => stride,
+            [_, stride] if lhs_l.dims()[0] == 1 => stride,
+            [stride, _] if lhs_l.dims()[1] == 1 => stride,
+            [stride] => stride,
+            [] => m * k,
+            _ => Err(self.striding_error(lhs_l, rhs_l, "non-contiguous lhs"))?,
+        };
+        let b_skip: usize = match rhs_stride[..rank - 2] {
+            [s1, stride] if s1 == stride * rhs_l.dims()[1] => stride,
+            [_, stride] if rhs_l.dims()[0] == 1 => stride,
+            [stride, _] if rhs_l.dims()[1] == 1 => stride,
+            [stride] => stride,
+            [] => n * k,
+            _ => Err(self.striding_error(lhs_l, rhs_l, "non-contiguous rhs"))?,
+        };
+        Ok((a_skip, b_skip))
+    }
 }
 
 impl Map2 for MatMul {
@@ -1547,18 +1272,7 @@ impl Map2 for MatMul {
         let rhs_cs = rhs_stride[rank - 1];
         let rhs_rs = rhs_stride[rank - 2];
 
-        let a_skip: usize = match lhs_stride[..rank - 2] {
-            [s1, stride] if s1 == stride * lhs_l.dims()[1] => stride,
-            [stride] => stride,
-            [] => m * k,
-            _ => Err(self.striding_error(lhs_l, rhs_l, "non-contiguous lhs"))?,
-        };
-        let b_skip: usize = match rhs_stride[..rank - 2] {
-            [s1, stride] if s1 == stride * rhs_l.dims()[1] => stride,
-            [stride] => stride,
-            [] => n * k,
-            _ => Err(self.striding_error(lhs_l, rhs_l, "non-contiguous rhs"))?,
-        };
+        let (a_skip, b_skip) = self.ab_skip(lhs_l, rhs_l)?;
         let c_skip: usize = m * n;
 
         let dst_shape: Shape = (m, n).into();
@@ -1618,20 +1332,8 @@ impl Map2 for MatMul {
 
         let lhs_stride = lhs_l.stride();
         let rhs_stride = rhs_l.stride();
-        let rank = lhs_stride.len();
 
-        let a_skip: usize = match lhs_stride[..rank - 2] {
-            [s1, stride] if s1 == stride * lhs_l.dims()[1] => stride,
-            [stride] => stride,
-            [] => m * k,
-            _ => Err(self.striding_error(lhs_l, rhs_l, "non-contiguous lhs"))?,
-        };
-        let b_skip: usize = match rhs_stride[..rank - 2] {
-            [s1, stride] if s1 == stride * rhs_l.dims()[1] => stride,
-            [stride] => stride,
-            [] => n * k,
-            _ => Err(self.striding_error(lhs_l, rhs_l, "non-contiguous rhs"))?,
-        };
+        let (a_skip, b_skip) = self.ab_skip(lhs_l, rhs_l)?;
         let c_skip: usize = m * n;
 
         let rhs_m1 = rhs_stride[rhs_stride.len() - 1];
@@ -1639,7 +1341,7 @@ impl Map2 for MatMul {
         let lhs_m1 = lhs_stride[lhs_stride.len() - 1];
         let lhs_m2 = lhs_stride[lhs_stride.len() - 2];
 
-        let (lda, transa) = if rhs_m1 == 1 && rhs_m2 == n {
+        let (lda, transa) = if (rhs_m1 == 1 || n == 1) && (rhs_m2 == n || k == 1) {
             (n as i32, b'N')
         } else if rhs_m1 == k && rhs_m2 == 1 {
             (k as i32, b'T')
@@ -1647,7 +1349,7 @@ impl Map2 for MatMul {
             Err(self.striding_error(lhs_l, rhs_l, "non-contiguous rhs"))?
         };
         // The b tensor has dims batching, m, k (lhs)
-        let (ldb, transb) = if lhs_m1 == 1 && lhs_m2 == k {
+        let (ldb, transb) = if (lhs_m1 == 1 || k == 1) && (lhs_m2 == k || m == 1) {
             (k as i32, b'N')
         } else if lhs_m1 == m && lhs_m2 == 1 {
             (m as i32, b'T')
@@ -1721,20 +1423,8 @@ impl Map2 for MatMul {
 
         let lhs_stride = lhs_l.stride();
         let rhs_stride = rhs_l.stride();
-        let rank = lhs_stride.len();
 
-        let a_skip: usize = match lhs_stride[..rank - 2] {
-            [s1, stride] if s1 == stride * lhs_l.dims()[1] => stride,
-            [stride] => stride,
-            [] => m * k,
-            _ => Err(self.striding_error(lhs_l, rhs_l, "non-contiguous lhs"))?,
-        };
-        let b_skip: usize = match rhs_stride[..rank - 2] {
-            [s1, stride] if s1 == stride * rhs_l.dims()[1] => stride,
-            [stride] => stride,
-            [] => n * k,
-            _ => Err(self.striding_error(lhs_l, rhs_l, "non-contiguous rhs"))?,
-        };
+        let (a_skip, b_skip) = self.ab_skip(lhs_l, rhs_l)?;
         let c_skip: usize = m * n;
 
         let rhs_m1 = rhs_stride[rhs_stride.len() - 1];
@@ -1742,7 +1432,7 @@ impl Map2 for MatMul {
         let lhs_m1 = lhs_stride[lhs_stride.len() - 1];
         let lhs_m2 = lhs_stride[lhs_stride.len() - 2];
 
-        let (lda, transa) = if rhs_m1 == 1 && rhs_m2 == n {
+        let (lda, transa) = if (rhs_m1 == 1 || n == 1) && (rhs_m2 == n || k == 1) {
             (n as i32, b'N')
         } else if rhs_m1 == k && rhs_m2 == 1 {
             (k as i32, b'T')
@@ -1750,7 +1440,7 @@ impl Map2 for MatMul {
             Err(self.striding_error(lhs_l, rhs_l, "non-contiguous rhs"))?
         };
         // The b tensor has dims batching, m, k (lhs)
-        let (ldb, transb) = if lhs_m1 == 1 && lhs_m2 == k {
+        let (ldb, transb) = if (lhs_m1 == 1 || k == 1) && (lhs_m2 == k || m == 1) {
             (k as i32, b'N')
         } else if lhs_m1 == m && lhs_m2 == 1 {
             (m as i32, b'T')
@@ -2422,6 +2112,48 @@ impl BackendStorage for CpuStorage {
         }
     }
 
+    fn copy2d(
+        &self,
+        dst: &mut Self,
+        d1: usize,
+        d2: usize,
+        src_s: usize,
+        dst_s: usize,
+        src_o: usize,
+        dst_o: usize,
+    ) -> Result<()> {
+        match (self, dst) {
+            (Self::U8(src), Self::U8(dst)) => copy2d_(src, dst, d1, d2, src_s, dst_s, src_o, dst_o),
+            (Self::U32(src), Self::U32(dst)) => {
+                copy2d_(src, dst, d1, d2, src_s, dst_s, src_o, dst_o)
+            }
+            (Self::I64(src), Self::I64(dst)) => {
+                copy2d_(src, dst, d1, d2, src_s, dst_s, src_o, dst_o)
+            }
+            (Self::BF16(src), Self::BF16(dst)) => {
+                copy2d_(src, dst, d1, d2, src_s, dst_s, src_o, dst_o)
+            }
+            (Self::F16(src), Self::F16(dst)) => {
+                copy2d_(src, dst, d1, d2, src_s, dst_s, src_o, dst_o)
+            }
+            (Self::F32(src), Self::F32(dst)) => {
+                copy2d_(src, dst, d1, d2, src_s, dst_s, src_o, dst_o)
+            }
+            (Self::F64(src), Self::F64(dst)) => {
+                copy2d_(src, dst, d1, d2, src_s, dst_s, src_o, dst_o)
+            }
+            (_, dst) => {
+                return Err(Error::DTypeMismatchBinaryOp {
+                    lhs: self.dtype(),
+                    rhs: dst.dtype(),
+                    op: "copy2d",
+                }
+                .bt());
+            }
+        }
+        Ok(())
+    }
+
     fn copy_strided_src(&self, dst: &mut Self, dst_offset: usize, src_l: &Layout) -> Result<()> {
         match (self, dst) {
             (Self::U8(src), Self::U8(dst)) => copy_strided_src_(src, dst, dst_offset, src_l),
@@ -2490,7 +2222,10 @@ impl BackendStorage for CpuStorage {
             col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
         } else {
             // Make the kernel contiguous if not already the case.
-            let mut kernel_c = self.device().zeros_impl(kernel_l.shape(), kernel.dtype())?;
+            let mut kernel_c = unsafe {
+                self.device()
+                    .alloc_uninit(kernel_l.shape(), kernel.dtype())?
+            };
             kernel.copy_strided_src(&mut kernel_c, 0, kernel_l)?;
             let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
                 .transpose(1, 2)?
@@ -2498,7 +2233,7 @@ impl BackendStorage for CpuStorage {
             col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
         };
         let res_l = Layout::contiguous((b, l_out, params.c_out)).transpose(1, 2)?;
-        let mut res_t = self.device().zeros_impl(res_l.shape(), res.dtype())?;
+        let mut res_t = unsafe { self.device().alloc_uninit(res_l.shape(), res.dtype())? };
         res.copy_strided_src(&mut res_t, 0, &res_l)?;
         Ok(res_t)
     }
@@ -2510,7 +2245,52 @@ impl BackendStorage for CpuStorage {
         kernel_l: &Layout,
         params: &crate::conv::ParamsConvTranspose1D,
     ) -> Result<Self> {
-        ConvTranspose1D(params).map(self, l, kernel, kernel_l)
+        let can_use_col2im = kernel_l.is_contiguous()
+            && params.dilation == 1
+            && params.padding == 0
+            && params.output_padding == 0;
+        if USE_IM2COL_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() {
+                crate::bail!(
+                    "convtr1d: the second argument (kernel) has to be contiguous {kernel_l:?}"
+                )
+            }
+            if c_in != c_in2 {
+                crate::bail!(
+                    "convtr1d: shape mismatch on c_in {:?} {:?}",
+                    l.shape(),
+                    kernel_l.shape()
+                )
+            }
+            let col = {
+                // This merges the last two dimensions of the kernel together.
+                let kernel_l_mm = Layout::new(
+                    (b_size, c_in, k_size * c_out).into(),
+                    vec![0, k_size * c_out, 1],
+                    kernel_l.start_offset(),
+                );
+                self.matmul(
+                    kernel,
+                    (
+                        b_size,
+                        /* m */ l_in,
+                        /* n */ c_out * k_size,
+                        /* k */ c_in,
+                    ),
+                    &l.transpose(1, 2)?,
+                    &kernel_l_mm,
+                )?
+            };
+            let col_l = Layout::contiguous((b_size, l_in, c_out, k_size));
+            Col2Im1D {
+                stride: params.stride,
+            }
+            .map(&col, &col_l)
+        } else {
+            ConvTranspose1D(params).map(self, l, kernel, kernel_l)
+        }
     }
 
     fn conv2d(
@@ -2544,7 +2324,10 @@ impl BackendStorage for CpuStorage {
             col.matmul(kernel, (b, m, n, k), &col_l, &kernel_l)?
         } else {
             // Make the kernel contiguous if not already the case.
-            let mut kernel_c = self.device().zeros_impl(kernel_l.shape(), kernel.dtype())?;
+            let mut kernel_c = unsafe {
+                self.device()
+                    .alloc_uninit(kernel_l.shape(), kernel.dtype())?
+            };
             kernel.copy_strided_src(&mut kernel_c, 0, kernel_l)?;
             let kernel_l = Layout::contiguous_with_offset((1, n, k), kernel_l.start_offset())
                 .transpose(1, 2)?
@@ -2554,7 +2337,7 @@ impl BackendStorage for CpuStorage {
         let res_l = Layout::contiguous((b, h_out, w_out, params.c_out))
             .transpose(1, 2)?
             .transpose(1, 3)?;
-        let mut res_t = self.device().zeros_impl(res_l.shape(), res.dtype())?;
+        let mut res_t = unsafe { self.device().alloc_uninit(res_l.shape(), res.dtype())? };
         res.copy_strided_src(&mut res_t, 0, &res_l)?;
         Ok(res_t)
     }
@@ -2574,7 +2357,7 @@ impl BackendStorage for CpuStorage {
             Self::U8(ids) => IndexSelect { ids, ids_l, dim }.map(self, l),
             Self::U32(ids) => IndexSelect { ids, ids_l, dim }.map(self, l),
             Self::I64(ids) => IndexSelect { ids, ids_l, dim }.map(self, l),
-            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "index-select")),
+            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "index-select").bt()),
         }
     }
 
@@ -2583,7 +2366,7 @@ impl BackendStorage for CpuStorage {
             Self::U8(ids) => Gather { ids, ids_l, dim }.map(self, l),
             Self::U32(ids) => Gather { ids, ids_l, dim }.map(self, l),
             Self::I64(ids) => Gather { ids, ids_l, dim }.map(self, l),
-            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "gather")),
+            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "gather").bt()),
         }
     }
 
@@ -2600,7 +2383,7 @@ impl BackendStorage for CpuStorage {
             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),
-            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "scatter-add")),
+            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "scatter-add").bt()),
         }
     }
 
@@ -2673,10 +2456,18 @@ 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())
     }
 
+    fn storage_from_cpu_storage_owned(&self, s: CpuStorage) -> Result<Self::Storage> {
+        Ok(s)
+    }
+
     fn new(_: usize) -> Result<Self> {
         Ok(Self)
     }
@@ -2778,6 +2569,53 @@ impl BackendDevice for CpuDevice {
         }
     }
 
+    #[allow(clippy::uninit_vec)]
+    unsafe fn alloc_uninit(&self, shape: &Shape, dtype: DType) -> Result<CpuStorage> {
+        let elem_count = shape.elem_count();
+        // The code below is highly unsafe but hopefully not directly unsound as we only consider
+        // types that are Copy, not Drop, and for which all bit patterns are proper values.
+        // It's still pretty risky, see the following for more details:
+        // https://github.com/rust-lang/rust-clippy/issues/4483
+        let storage = match dtype {
+            DType::U8 => {
+                let mut v = Vec::with_capacity(elem_count);
+                v.set_len(elem_count);
+                CpuStorage::U8(v)
+            }
+            DType::U32 => {
+                let mut v = Vec::with_capacity(elem_count);
+                v.set_len(elem_count);
+                CpuStorage::U32(v)
+            }
+            DType::I64 => {
+                let mut v = Vec::with_capacity(elem_count);
+                v.set_len(elem_count);
+                CpuStorage::I64(v)
+            }
+            DType::BF16 => {
+                let mut v = Vec::with_capacity(elem_count);
+                v.set_len(elem_count);
+                CpuStorage::BF16(v)
+            }
+            DType::F16 => {
+                let mut v = Vec::with_capacity(elem_count);
+                v.set_len(elem_count);
+                CpuStorage::F16(v)
+            }
+            DType::F32 => {
+                let mut v = Vec::with_capacity(elem_count);
+                v.set_len(elem_count);
+                CpuStorage::F32(v)
+            }
+            DType::F64 => {
+                let mut v = Vec::with_capacity(elem_count);
+                v.set_len(elem_count);
+                CpuStorage::F64(v)
+            }
+        };
+        Ok(storage)
+    }
+
     fn ones_impl(&self, shape: &Shape, dtype: DType) -> Result<CpuStorage> {
         let elem_count = shape.elem_count();
         let storage = match dtype {
@@ -2805,6 +2643,10 @@ impl BackendDevice for CpuDevice {
         };
         Ok(storage)
     }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
 }
 
 #[macro_export]

candle-core/src/cpu_backend/utils.rs

@@ -0,0 +1,350 @@
+/// Helper functions to write CPU kernels.
+use crate::backend::BackendStorage;
+use crate::{Error, Layout, Result, WithDType};
+
+type C = super::CpuStorage;
+pub trait Map1 {
+    fn f<T: WithDType>(&self, vs: &[T], layout: &Layout) -> Result<Vec<T>>;
+
+    fn map(&self, vs: &C, layout: &Layout) -> Result<C> {
+        match vs {
+            C::U8(vs) => Ok(C::U8(self.f(vs, layout)?)),
+            C::U32(vs) => Ok(C::U32(self.f(vs, layout)?)),
+            C::I64(vs) => Ok(C::I64(self.f(vs, layout)?)),
+            C::BF16(vs) => Ok(C::BF16(self.f(vs, layout)?)),
+            C::F16(vs) => Ok(C::F16(self.f(vs, layout)?)),
+            C::F32(vs) => Ok(C::F32(self.f(vs, layout)?)),
+            C::F64(vs) => Ok(C::F64(self.f(vs, layout)?)),
+        }
+    }
+}
+
+pub trait Map1Any {
+    fn f<T: WithDType, W: Fn(Vec<T>) -> C>(&self, vs: &[T], layout: &Layout, wrap: W) -> Result<C>;
+
+    fn map(&self, vs: &C, layout: &Layout) -> Result<C> {
+        match vs {
+            C::U8(vs) => Ok(self.f(vs, layout, C::U8)?),
+            C::U32(vs) => Ok(self.f(vs, layout, C::U32)?),
+            C::I64(vs) => Ok(self.f(vs, layout, C::I64)?),
+            C::BF16(vs) => Ok(self.f(vs, layout, C::BF16)?),
+            C::F16(vs) => Ok(self.f(vs, layout, C::F16)?),
+            C::F32(vs) => Ok(self.f(vs, layout, C::F32)?),
+            C::F64(vs) => Ok(self.f(vs, layout, C::F64)?),
+        }
+    }
+}
+
+pub trait Map2 {
+    const OP: &'static str;
+    fn f<T: WithDType>(&self, v1: &[T], l1: &Layout, v2: &[T], l2: &Layout) -> Result<Vec<T>>;
+
+    fn map(&self, v1: &C, l1: &Layout, v2: &C, l2: &Layout) -> Result<C> {
+        match (v1, v2) {
+            (C::U8(v1), C::U8(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::U32(v1), C::U32(v2)) => Ok(C::U32(self.f(v1, l1, v2, l2)?)),
+            (C::I64(v1), C::I64(v2)) => Ok(C::I64(self.f(v1, l1, v2, l2)?)),
+            (C::BF16(v1), C::BF16(v2)) => Ok(C::BF16(self.f(v1, l1, v2, l2)?)),
+            (C::F16(v1), C::F16(v2)) => Ok(C::F16(self.f(v1, l1, v2, l2)?)),
+            (C::F32(v1), C::F32(v2)) => Ok(C::F32(self.f(v1, l1, v2, l2)?)),
+            (C::F64(v1), C::F64(v2)) => Ok(C::F64(self.f(v1, l1, v2, l2)?)),
+            _ => Err(Error::DTypeMismatchBinaryOp {
+                lhs: v1.dtype(),
+                rhs: v2.dtype(),
+                op: Self::OP,
+            }
+            .bt()),
+        }
+    }
+}
+
+pub trait Map2U8 {
+    const OP: &'static str;
+    fn f<T: WithDType>(&self, v1: &[T], l1: &Layout, v2: &[T], l2: &Layout) -> Result<Vec<u8>>;
+
+    fn map(&self, v1: &C, l1: &Layout, v2: &C, l2: &Layout) -> Result<C> {
+        match (v1, v2) {
+            (C::U8(v1), C::U8(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::U32(v1), C::U32(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::I64(v1), C::I64(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::BF16(v1), C::BF16(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::F16(v1), C::F16(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::F32(v1), C::F32(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            (C::F64(v1), C::F64(v2)) => Ok(C::U8(self.f(v1, l1, v2, l2)?)),
+            _ => Err(Error::DTypeMismatchBinaryOp {
+                lhs: v1.dtype(),
+                rhs: v2.dtype(),
+                op: Self::OP,
+            }
+            .bt()),
+        }
+    }
+}
+
+pub fn binary_map<T: Copy, U: Copy, F: FnMut(T, T) -> U>(
+    lhs_l: &Layout,
+    rhs_l: &Layout,
+    lhs: &[T],
+    rhs: &[T],
+    mut f: F,
+) -> Vec<U> {
+    match (lhs_l.contiguous_offsets(), rhs_l.contiguous_offsets()) {
+        (Some((o_l1, o_l2)), Some((o_r1, o_r2))) => lhs[o_l1..o_l2]
+            .iter()
+            .zip(rhs[o_r1..o_r2].iter())
+            .map(|(&l, &r)| f(l, r))
+            .collect(),
+        (Some((o_l1, o_l2)), None) => {
+            // TODO: Maybe we want to avoid going through the layout twice.
+            match rhs_l.offsets_b() {
+                Some(ob) => {
+                    let mut i_in_block = 0;
+                    let mut i_right_broadcast = 0;
+                    lhs[o_l1..o_l2]
+                        .iter()
+                        .map(|&l| {
+                            let r = unsafe { rhs.get_unchecked(i_in_block + ob.start) };
+                            i_right_broadcast += 1;
+                            if i_right_broadcast >= ob.right_broadcast {
+                                i_in_block += 1;
+                                i_right_broadcast = 0;
+                            }
+                            if i_in_block >= ob.len {
+                                i_in_block = 0
+                            }
+                            f(l, *r)
+                        })
+                        .collect()
+                }
+                None => lhs_l
+                    .strided_index()
+                    .zip(rhs_l.strided_index())
+                    .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+                    .collect(),
+            }
+        }
+        (None, Some((o_r1, o_r2))) => {
+            // TODO: Maybe we want to avoid going through the layout twice.
+            match lhs_l.offsets_b() {
+                Some(ob) => {
+                    let mut i_in_block = 0;
+                    let mut i_right_broadcast = 0;
+                    rhs[o_r1..o_r2]
+                        .iter()
+                        .map(|&r| {
+                            let l = unsafe { lhs.get_unchecked(i_in_block + ob.start) };
+                            i_right_broadcast += 1;
+                            if i_right_broadcast >= ob.right_broadcast {
+                                i_in_block += 1;
+                                i_right_broadcast = 0;
+                            }
+                            if i_in_block >= ob.len {
+                                i_in_block = 0
+                            }
+                            f(*l, r)
+                        })
+                        .collect()
+                }
+                None => lhs_l
+                    .strided_index()
+                    .zip(rhs_l.strided_index())
+                    .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+                    .collect(),
+            }
+        }
+        _ => lhs_l
+            .strided_index()
+            .zip(rhs_l.strided_index())
+            .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+            .collect(),
+    }
+}
+
+// Similar to binary_map but with vectorized variants.
+pub fn binary_map_vec<T: Copy, F: FnMut(T, T) -> T, FV: FnMut(&[T], &[T], &mut [T])>(
+    lhs_l: &Layout,
+    rhs_l: &Layout,
+    lhs: &[T],
+    rhs: &[T],
+    mut f: F,
+    mut f_vec: FV,
+) -> Vec<T> {
+    let el_count = lhs_l.shape().elem_count();
+    match (lhs_l.contiguous_offsets(), rhs_l.contiguous_offsets()) {
+        (Some((o_l1, o_l2)), Some((o_r1, o_r2))) => {
+            let mut ys: Vec<T> = Vec::with_capacity(el_count);
+            let ys_to_set = ys.spare_capacity_mut();
+            let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
+            f_vec(&lhs[o_l1..o_l2], &rhs[o_r1..o_r2], ys_to_set);
+            // SAFETY: values are all set by f_vec.
+            unsafe { ys.set_len(el_count) };
+            ys
+        }
+        (Some((o_l1, o_l2)), None) => match rhs_l.offsets_b() {
+            Some(ob) if ob.right_broadcast == 1 => {
+                let rhs = &rhs[ob.start..ob.start + ob.len];
+                let mut ys: Vec<T> = Vec::with_capacity(el_count);
+                let ys_to_set = ys.spare_capacity_mut();
+                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
+                let mut dst_i = 0;
+                for src_i in (o_l1..o_l2).step_by(ob.len) {
+                    f_vec(
+                        &lhs[src_i..src_i + ob.len],
+                        rhs,
+                        &mut ys_to_set[dst_i..dst_i + ob.len],
+                    );
+                    dst_i += ob.len;
+                }
+                // SAFETY: values are all set by f_vec.
+                unsafe { ys.set_len(el_count) };
+                ys
+            }
+            Some(ob) => {
+                let rhs = &rhs[ob.start..ob.start + ob.len];
+                let mut ys = lhs[o_l1..o_l2].to_vec();
+                for idx_l in 0..ob.left_broadcast {
+                    let start = idx_l * ob.len * ob.right_broadcast;
+                    for (i, &r) in rhs.iter().enumerate() {
+                        let start = start + i * ob.right_broadcast;
+                        for v in ys[start..start + ob.right_broadcast].iter_mut() {
+                            *v = f(*v, r)
+                        }
+                    }
+                }
+                ys
+            }
+            None => lhs_l
+                .strided_index()
+                .zip(rhs_l.strided_index())
+                .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+                .collect(),
+        },
+        (None, Some((o_r1, o_r2))) => match lhs_l.offsets_b() {
+            Some(ob) if ob.right_broadcast == 1 => {
+                let lhs = &lhs[ob.start..ob.start + ob.len];
+                let mut ys: Vec<T> = Vec::with_capacity(el_count);
+                let ys_to_set = ys.spare_capacity_mut();
+                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
+                let mut dst_i = 0;
+                for src_i in (o_r1..o_r2).step_by(ob.len) {
+                    f_vec(
+                        lhs,
+                        &rhs[src_i..src_i + ob.len],
+                        &mut ys_to_set[dst_i..dst_i + ob.len],
+                    );
+                    dst_i += ob.len;
+                }
+                // SAFETY: values are all set by f_vec.
+                unsafe { ys.set_len(el_count) };
+                ys
+            }
+            Some(ob) => {
+                let lhs = &lhs[ob.start..ob.start + ob.len];
+                let mut ys = rhs[o_r1..o_r2].to_vec();
+                for idx_l in 0..ob.left_broadcast {
+                    let start = idx_l * ob.len * ob.right_broadcast;
+                    for (i, &l) in lhs.iter().enumerate() {
+                        let start = start + i * ob.right_broadcast;
+                        for v in ys[start..start + ob.right_broadcast].iter_mut() {
+                            *v = f(l, *v)
+                        }
+                    }
+                }
+                ys
+            }
+            None => lhs_l
+                .strided_index()
+                .zip(rhs_l.strided_index())
+                .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+                .collect(),
+        },
+        _ => lhs_l
+            .strided_index()
+            .zip(rhs_l.strided_index())
+            .map(|(lhs_i, rhs_i)| f(lhs[lhs_i], rhs[rhs_i]))
+            .collect(),
+    }
+}
+
+pub fn unary_map<T: Copy, U: Copy, F: FnMut(T) -> U>(
+    vs: &[T],
+    layout: &Layout,
+    mut f: F,
+) -> Vec<U> {
+    match layout.strided_blocks() {
+        crate::StridedBlocks::SingleBlock { start_offset, len } => vs
+            [start_offset..start_offset + len]
+            .iter()
+            .map(|&v| f(v))
+            .collect(),
+        crate::StridedBlocks::MultipleBlocks {
+            block_start_index,
+            block_len,
+        } => {
+            let mut result = Vec::with_capacity(layout.shape().elem_count());
+            // Specialize the case where block_len is one to avoid the second loop.
+            if block_len == 1 {
+                for index in block_start_index {
+                    let v = unsafe { vs.get_unchecked(index) };
+                    result.push(f(*v))
+                }
+            } else {
+                for index in block_start_index {
+                    for offset in 0..block_len {
+                        let v = unsafe { vs.get_unchecked(index + offset) };
+                        result.push(f(*v))
+                    }
+                }
+            }
+            result
+        }
+    }
+}
+
+pub fn unary_map_vec<T: Copy, U: Copy, F: FnMut(T) -> U, FV: FnMut(&[T], &mut [U])>(
+    vs: &[T],
+    layout: &Layout,
+    mut f: F,
+    mut f_vec: FV,
+) -> Vec<U> {
+    match layout.strided_blocks() {
+        crate::StridedBlocks::SingleBlock { start_offset, len } => {
+            let mut ys: Vec<U> = Vec::with_capacity(len);
+            let ys_to_set = ys.spare_capacity_mut();
+            let ys_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(ys_to_set) };
+            f_vec(&vs[start_offset..start_offset + len], ys_to_set);
+            // SAFETY: values are all set by f_vec.
+            unsafe { ys.set_len(len) };
+            ys
+        }
+        crate::StridedBlocks::MultipleBlocks {
+            block_start_index,
+            block_len,
+        } => {
+            let el_count = layout.shape().elem_count();
+            // Specialize the case where block_len is one to avoid the second loop.
+            if block_len == 1 {
+                let mut result = Vec::with_capacity(el_count);
+                for index in block_start_index {
+                    let v = unsafe { vs.get_unchecked(index) };
+                    result.push(f(*v))
+                }
+                result
+            } else {
+                let mut ys: Vec<U> = Vec::with_capacity(el_count);
+                let ys_to_set = ys.spare_capacity_mut();
+                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(ys_to_set) };
+                let mut dst_index = 0;
+                for src_index in block_start_index {
+                    let vs = &vs[src_index..src_index + block_len];
+                    let ys = &mut ys_to_set[dst_index..dst_index + block_len];
+                    f_vec(vs, ys);
+                    dst_index += block_len;
+                }
+                // SAFETY: values are all set by f_vec.
+                unsafe { ys.set_len(el_count) };
+                ys
+            }
+        }
+    }
+}

candle-core/src/cuda_backend/device.rs

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

candle-core/src/cuda_backend/error.rs

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

candle-core/src/cuda_backend/utils.rs

@@ -0,0 +1,134 @@
+/// Helper functions to plug cuda kernels in candle.
+use crate::{Layout, Result, Shape, WithDType};
+pub use cudarc;
+use cudarc::driver::{CudaSlice, DeviceRepr, ValidAsZeroBits};
+
+use super::{CudaDevice, CudaError, WrapErr};
+
+pub type S = super::CudaStorageSlice;
+
+pub trait Map1 {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        src: &CudaSlice<T>,
+        dev: &CudaDevice,
+        layout: &Layout,
+    ) -> Result<CudaSlice<T>>;
+
+    fn map(&self, s: &S, d: &CudaDevice, l: &Layout) -> Result<S> {
+        let out = match s {
+            S::U8(s) => S::U8(self.f(s, d, l)?),
+            S::U32(s) => S::U32(self.f(s, d, l)?),
+            S::I64(s) => S::I64(self.f(s, d, l)?),
+            S::BF16(s) => S::BF16(self.f(s, d, l)?),
+            S::F16(s) => S::F16(self.f(s, d, l)?),
+            S::F32(s) => S::F32(self.f(s, d, l)?),
+            S::F64(s) => S::F64(self.f(s, d, l)?),
+        };
+        Ok(out)
+    }
+}
+
+pub trait Map2 {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        src1: &CudaSlice<T>,
+        layout1: &Layout,
+        src2: &CudaSlice<T>,
+        layout2: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<CudaSlice<T>>;
+
+    fn map(&self, s1: &S, l1: &Layout, s2: &S, l2: &Layout, d: &CudaDevice) -> Result<S> {
+        let out = match (s1, s2) {
+            (S::U8(s1), S::U8(s2)) => S::U8(self.f(s1, l1, s2, l2, d)?),
+            (S::U32(s1), S::U32(s2)) => S::U32(self.f(s1, l1, s2, l2, d)?),
+            (S::I64(s1), S::I64(s2)) => S::I64(self.f(s1, l1, s2, l2, d)?),
+            (S::BF16(s1), S::BF16(s2)) => S::BF16(self.f(s1, l1, s2, l2, d)?),
+            (S::F16(s1), S::F16(s2)) => S::F16(self.f(s1, l1, s2, l2, d)?),
+            (S::F32(s1), S::F32(s2)) => S::F32(self.f(s1, l1, s2, l2, d)?),
+            (S::F64(s1), S::F64(s2)) => S::F64(self.f(s1, l1, s2, l2, d)?),
+            _ => Err(CudaError::InternalError("dtype mismatch in binary op"))?,
+        };
+        Ok(out)
+    }
+}
+
+pub trait Map2InPlace {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        dst: &mut CudaSlice<T>,
+        dst_shape: &Shape,
+        src: &CudaSlice<T>,
+        src_l: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<()>;
+
+    fn map(
+        &self,
+        dst: &mut S,
+        dst_s: &Shape,
+        src: &S,
+        src_l: &Layout,
+        d: &CudaDevice,
+    ) -> Result<()> {
+        match (dst, src) {
+            (S::U8(dst), S::U8(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::U32(dst), S::U32(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::I64(dst), S::I64(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::BF16(dst), S::BF16(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::F16(dst), S::F16(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::F32(dst), S::F32(src)) => self.f(dst, dst_s, src, src_l, d),
+            (S::F64(dst), S::F64(src)) => self.f(dst, dst_s, src, src_l, d),
+            _ => Err(CudaError::InternalError("dtype mismatch in binary op"))?,
+        }
+    }
+}
+
+pub trait Map1Any {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits, W: Fn(CudaSlice<T>) -> S>(
+        &self,
+        src: &CudaSlice<T>,
+        dev: &CudaDevice,
+        layout: &Layout,
+        wrap: W,
+    ) -> Result<S>;
+
+    fn map(&self, s: &S, d: &CudaDevice, l: &Layout) -> Result<S> {
+        let out = match s {
+            S::U8(s) => self.f(s, d, l, S::U8)?,
+            S::U32(s) => self.f(s, d, l, S::U32)?,
+            S::I64(s) => self.f(s, d, l, S::I64)?,
+            S::BF16(s) => self.f(s, d, l, S::BF16)?,
+            S::F16(s) => self.f(s, d, l, S::F16)?,
+            S::F32(s) => self.f(s, d, l, S::F32)?,
+            S::F64(s) => self.f(s, d, l, S::F64)?,
+        };
+        Ok(out)
+    }
+}
+
+pub trait Map2Any {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        src1: &CudaSlice<T>,
+        layout1: &Layout,
+        src2: &CudaSlice<T>,
+        layout2: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<S>;
+
+    fn map(&self, s1: &S, l1: &Layout, s2: &S, l2: &Layout, d: &CudaDevice) -> Result<S> {
+        let out = match (s1, s2) {
+            (S::U8(s1), S::U8(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::U32(s1), S::U32(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::I64(s1), S::I64(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::BF16(s1), S::BF16(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::F16(s1), S::F16(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::F32(s1), S::F32(s2)) => self.f(s1, l1, s2, l2, d)?,
+            (S::F64(s1), S::F64(s2)) => self.f(s1, l1, s2, l2, d)?,
+            _ => Err(CudaError::InternalError("dtype mismatch in binary op")).w()?,
+        };
+        Ok(out)
+    }
+}

candle-core/src/custom_op.rs

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

candle-core/src/device.rs

@@ -8,12 +8,14 @@ use crate::{CpuStorage, DType, Result, Shape, Storage, WithDType};
 pub enum DeviceLocation {
     Cpu,
     Cuda { gpu_id: usize },
+    Metal { gpu_id: usize },
 }
 
 #[derive(Debug, Clone)]
 pub enum Device {
     Cpu,
     Cuda(crate::CudaDevice),
+    Metal(crate::MetalDevice),
 }
 
 pub trait NdArray {
@@ -128,10 +130,15 @@ impl Device {
         Ok(Self::Cuda(crate::CudaDevice::new(ordinal)?))
     }
 
+    pub fn new_metal(ordinal: usize) -> Result<Self> {
+        Ok(Self::Metal(crate::MetalDevice::new(ordinal)?))
+    }
+
     pub fn set_seed(&self, seed: u64) -> Result<()> {
         match self {
-            Self::Cpu => crate::cpu_backend::CpuDevice.set_seed(seed),
+            Self::Cpu => CpuDevice.set_seed(seed),
             Self::Cuda(c) => c.set_seed(seed),
+            Self::Metal(m) => m.set_seed(seed),
         }
     }
 
@@ -139,6 +146,7 @@ impl Device {
         match (self, rhs) {
             (Self::Cpu, Self::Cpu) => true,
             (Self::Cuda(lhs), Self::Cuda(rhs)) => lhs.same_device(rhs),
+            (Self::Metal(lhs), Self::Metal(rhs)) => lhs.same_device(rhs),
             _ => false,
         }
     }
@@ -147,21 +155,20 @@ impl Device {
         match self {
             Self::Cpu => DeviceLocation::Cpu,
             Self::Cuda(device) => device.location(),
+            Device::Metal(device) => device.location(),
         }
     }
 
     pub fn is_cpu(&self) -> bool {
-        match self {
-            Self::Cpu => true,
-            Self::Cuda(_) => false,
-        }
+        matches!(self, Self::Cpu)
     }
 
     pub fn is_cuda(&self) -> bool {
-        match self {
-            Self::Cpu => false,
-            Self::Cuda(_) => true,
-        }
+        matches!(self, Self::Cuda(_))
+    }
+
+    pub fn is_metal(&self) -> bool {
+        matches!(self, Self::Metal(_))
     }
 
     pub fn cuda_if_available(ordinal: usize) -> Result<Self> {
@@ -194,6 +201,10 @@ impl Device {
                     Ok(Storage::Cuda(storage))
                 }
             }
+            Device::Metal(device) => {
+                let storage = device.rand_uniform(shape, dtype, lo, up)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 
@@ -228,6 +239,10 @@ impl Device {
                     Ok(Storage::Cuda(storage))
                 }
             }
+            Device::Metal(device) => {
+                let storage = device.rand_normal(shape, dtype, mean, std)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 
@@ -250,6 +265,10 @@ impl Device {
                 let storage = device.ones_impl(shape, dtype)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = device.ones_impl(shape, dtype)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 
@@ -263,6 +282,41 @@ impl Device {
                 let storage = device.zeros_impl(shape, dtype)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = device.zeros_impl(shape, dtype)?;
+                Ok(Storage::Metal(storage))
+            }
+        }
+    }
+
+    pub(crate) unsafe fn alloc_uninit(&self, shape: &Shape, dtype: DType) -> Result<Storage> {
+        match self {
+            Device::Cpu => {
+                let storage = CpuDevice.alloc_uninit(shape, dtype)?;
+                Ok(Storage::Cpu(storage))
+            }
+            Device::Cuda(device) => {
+                let storage = device.alloc_uninit(shape, dtype)?;
+                Ok(Storage::Cuda(storage))
+            }
+            Device::Metal(device) => {
+                let storage = device.alloc_uninit(shape, dtype)?;
+                Ok(Storage::Metal(storage))
+            }
+        }
+    }
+
+    pub(crate) fn storage_from_slice<D: WithDType>(&self, data: &[D]) -> Result<Storage> {
+        match self {
+            Device::Cpu => Ok(Storage::Cpu(data.to_cpu_storage())),
+            Device::Cuda(device) => {
+                let storage = device.storage_from_slice(data)?;
+                Ok(Storage::Cuda(storage))
+            }
+            Device::Metal(device) => {
+                let storage = device.storage_from_slice(data)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 
@@ -271,9 +325,14 @@ impl Device {
             Device::Cpu => Ok(Storage::Cpu(array.to_cpu_storage())),
             Device::Cuda(device) => {
                 let storage = array.to_cpu_storage();
-                let storage = device.storage_from_cpu_storage(&storage)?;
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = array.to_cpu_storage();
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
+                Ok(Storage::Metal(storage))
+            }
         }
     }
 
@@ -282,9 +341,22 @@ impl Device {
             Device::Cpu => Ok(Storage::Cpu(S::to_cpu_storage_owned(data))),
             Device::Cuda(device) => {
                 let storage = S::to_cpu_storage_owned(data);
-                let storage = device.storage_from_cpu_storage(&storage)?;
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
                 Ok(Storage::Cuda(storage))
             }
+            Device::Metal(device) => {
+                let storage = S::to_cpu_storage_owned(data);
+                let storage = device.storage_from_cpu_storage_owned(storage)?;
+                Ok(Storage::Metal(storage))
+            }
+        }
+    }
+
+    pub fn synchronize(&self) -> Result<()> {
+        match self {
+            Self::Cpu => Ok(()),
+            Self::Cuda(d) => d.synchronize(),
+            Self::Metal(d) => d.synchronize(),
         }
     }
 }

candle-core/src/display.rs

@@ -14,6 +14,9 @@ impl Tensor {
             crate::DeviceLocation::Cuda { gpu_id } => {
                 format!(", cuda:{}", gpu_id)
             }
+            crate::DeviceLocation::Metal { gpu_id } => {
+                format!(", metal:{}", gpu_id)
+            }
         };
 
         write!(f, "Tensor[")?;
@@ -62,12 +65,13 @@ impl std::fmt::Debug for Tensor {
 }
 
 /// Options for Tensor pretty printing
+#[derive(Debug, Clone)]
 pub struct PrinterOptions {
-    precision: usize,
-    threshold: usize,
-    edge_items: usize,
-    line_width: usize,
-    sci_mode: Option<bool>,
+    pub precision: usize,
+    pub threshold: usize,
+    pub edge_items: usize,
+    pub line_width: usize,
+    pub sci_mode: Option<bool>,
 }
 
 static PRINT_OPTS: std::sync::Mutex<PrinterOptions> =
@@ -86,6 +90,10 @@ impl PrinterOptions {
     }
 }
 
+pub fn print_options() -> &'static std::sync::Mutex<PrinterOptions> {
+    &PRINT_OPTS
+}
+
 pub fn set_print_options(options: PrinterOptions) {
     *PRINT_OPTS.lock().unwrap() = options
 }
@@ -114,6 +122,26 @@ pub fn set_print_options_full() {
     }
 }
 
+pub fn set_line_width(line_width: usize) {
+    PRINT_OPTS.lock().unwrap().line_width = line_width
+}
+
+pub fn set_precision(precision: usize) {
+    PRINT_OPTS.lock().unwrap().precision = precision
+}
+
+pub fn set_edge_items(edge_items: usize) {
+    PRINT_OPTS.lock().unwrap().edge_items = edge_items
+}
+
+pub fn set_threshold(threshold: usize) {
+    PRINT_OPTS.lock().unwrap().threshold = threshold
+}
+
+pub fn set_sci_mode(sci_mode: Option<bool>) {
+    PRINT_OPTS.lock().unwrap().sci_mode = sci_mode
+}
+
 struct FmtSize {
     current_size: usize,
 }
@@ -476,6 +504,9 @@ impl std::fmt::Display for Tensor {
             crate::DeviceLocation::Cuda { gpu_id } => {
                 format!(", cuda:{}", gpu_id)
             }
+            crate::DeviceLocation::Metal { gpu_id } => {
+                format!(", metal:{}", gpu_id)
+            }
         };
 
         write!(

candle-core/src/dtype.rs

@@ -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)]
@@ -23,7 +23,15 @@ pub enum DType {
 }
 
 #[derive(Debug, PartialEq, Eq)]
-pub struct DTypeParseError;
+pub struct DTypeParseError(String);
+
+impl std::fmt::Display for DTypeParseError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "cannot parse '{}' as a dtype", self.0)
+    }
+}
+
+impl std::error::Error for DTypeParseError {}
 
 impl std::str::FromStr for DType {
     type Err = DTypeParseError;
@@ -36,7 +44,7 @@ impl std::str::FromStr for DType {
             "f16" => Ok(Self::F16),
             "f32" => Ok(Self::F32),
             "f64" => Ok(Self::F64),
-            _ => Err(DTypeParseError),
+            _ => Err(DTypeParseError(s.to_string())),
         }
     }
 }
@@ -92,12 +100,14 @@ pub trait WithDType:
     + 'static
     + Send
     + Sync
+    + std::any::Any
     + crate::cpu::kernels::VecOps
 {
     const DTYPE: DType;
 
     fn from_f64(v: f64) -> Self;
     fn to_f64(self) -> f64;
+    fn cpu_storage_ref(data: &[Self]) -> CpuStorageRef<'_>;
     fn to_cpu_storage_owned(data: Vec<Self>) -> CpuStorage;
 
     fn to_cpu_storage(data: &[Self]) -> CpuStorage {
@@ -121,6 +131,10 @@ macro_rules! with_dtype {
                 $to_f64(self)
             }
 
+            fn cpu_storage_ref(data: &[Self]) -> CpuStorageRef<'_> {
+                CpuStorageRef::$dtype(data)
+            }
+
             fn to_cpu_storage_owned(data: Vec<Self>) -> CpuStorage {
                 CpuStorage::$dtype(data)
             }

candle-core/src/dummy_cuda_backend.rs

@@ -154,6 +154,19 @@ impl crate::backend::BackendStorage for CudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn copy2d(
+        &self,
+        _: &mut Self,
+        _: usize,
+        _: usize,
+        _: usize,
+        _: usize,
+        _: usize,
+        _: usize,
+    ) -> Result<()> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn avg_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self> {
         Err(Error::NotCompiledWithCudaSupport)
     }
@@ -197,10 +210,22 @@ impl crate::backend::BackendDevice for CudaDevice {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
@@ -208,4 +233,38 @@ impl crate::backend::BackendDevice for CudaDevice {
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
 }
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn gemm_reduced_precision_f16() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn set_gemm_reduced_precision_f16(_: bool) {}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn gemm_reduced_precision_bf16() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn set_gemm_reduced_precision_bf16(_: bool) {}
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn gemm_reduced_precision_f32() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn set_gemm_reduced_precision_f32(_b: bool) {}

candle-core/src/dummy_metal_backend.rs

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

candle-core/src/error.rs

@@ -1,4 +1,4 @@
-use crate::{DType, DeviceLocation, Layout, Shape};
+use crate::{DType, DeviceLocation, Layout, MetalError, Shape};
 
 #[derive(Debug, Clone)]
 pub struct MatMulUnexpectedStriding {
@@ -152,6 +152,9 @@ pub enum Error {
     #[error("the candle crate has not been built with cuda support")]
     NotCompiledWithCudaSupport,
 
+    #[error("the candle crate has not been built with metal support")]
+    NotCompiledWithMetalSupport,
+
     #[error("cannot find tensor {path}")]
     CannotFindTensor { path: String },
 
@@ -159,6 +162,9 @@ pub enum Error {
     #[error(transparent)]
     Cuda(Box<dyn std::error::Error + Send + Sync>),
 
+    #[error("Metal error {0}")]
+    Metal(#[from] MetalError),
+
     #[error(transparent)]
     TryFromIntError(#[from] core::num::TryFromIntError),
 
@@ -213,10 +219,14 @@ impl Error {
         Self::Wrapped(Box::new(err)).bt()
     }
 
-    pub fn msg(err: impl std::error::Error + Send + Sync + 'static) -> Self {
+    pub fn msg(err: impl std::error::Error) -> Self {
         Self::Msg(err.to_string()).bt()
     }
 
+    pub fn debug(err: impl std::fmt::Debug) -> Self {
+        Self::Msg(format!("{err:?}")).bt()
+    }
+
     pub fn bt(self) -> Self {
         let backtrace = std::backtrace::Backtrace::capture();
         match backtrace.status() {

candle-core/src/indexer.rs

@@ -64,7 +64,7 @@ impl Tensor {
 #[derive(Debug)]
 /// Generic structure used to index a slice of the tensor
 pub enum TensorIndexer {
-    /// This selects the elemnts for which an index has some specific value.
+    /// This selects the elements for which an index has some specific value.
     Select(usize),
     /// This is a regular slice, purely indexing a chunk of the tensor
     Narrow(Bound<usize>, Bound<usize>),
@@ -104,37 +104,31 @@ impl From<&Tensor> for TensorIndexer {
     }
 }
 
-macro_rules! impl_from_range {
-    ($range_type:ty) => {
-        impl From<$range_type> for TensorIndexer {
-            fn from(range: $range_type) -> Self {
-                use std::ops::Bound::*;
+trait RB: RangeBounds<usize> {}
+impl RB for Range<usize> {}
+impl RB for RangeFrom<usize> {}
+impl RB for RangeFull {}
+impl RB for RangeInclusive<usize> {}
+impl RB for RangeTo<usize> {}
+impl RB for RangeToInclusive<usize> {}
 
-                let start = match range.start_bound() {
-                    Included(idx) => Included(*idx),
-                    Excluded(idx) => Excluded(*idx),
-                    Unbounded => Unbounded,
-                };
-
-                let end = match range.end_bound() {
-                    Included(idx) => Included(*idx),
-                    Excluded(idx) => Excluded(*idx),
-                    Unbounded => Unbounded,
-                };
-
-                TensorIndexer::Narrow(start, end)
-            }
-        }
-    };
+impl<T: RB> From<T> for TensorIndexer {
+    fn from(range: T) -> Self {
+        use std::ops::Bound::*;
+        let start = match range.start_bound() {
+            Included(idx) => Included(*idx),
+            Excluded(idx) => Excluded(*idx),
+            Unbounded => Unbounded,
+        };
+        let end = match range.end_bound() {
+            Included(idx) => Included(*idx),
+            Excluded(idx) => Excluded(*idx),
+            Unbounded => Unbounded,
+        };
+        TensorIndexer::Narrow(start, end)
+    }
 }
 
-impl_from_range!(Range<usize>);
-impl_from_range!(RangeFrom<usize>);
-impl_from_range!(RangeFull);
-impl_from_range!(RangeInclusive<usize>);
-impl_from_range!(RangeTo<usize>);
-impl_from_range!(RangeToInclusive<usize>);
-
 /// Trait used to implement multiple signatures for ease of use of the slicing
 /// of a tensor
 pub trait IndexOp<T> {

candle-core/src/layout.rs

@@ -70,7 +70,7 @@ impl Layout {
         self.shape.is_fortran_contiguous(&self.stride)
     }
 
-    pub(crate) fn narrow(&self, dim: usize, start: usize, len: usize) -> Result<Self> {
+    pub fn narrow(&self, dim: usize, start: usize, len: usize) -> Result<Self> {
         let dims = self.shape().dims();
         if dim >= dims.len() {
             Err(Error::DimOutOfRange {
@@ -99,7 +99,7 @@ impl Layout {
         })
     }
 
-    pub(crate) fn transpose(&self, dim1: usize, dim2: usize) -> Result<Self> {
+    pub fn transpose(&self, dim1: usize, dim2: usize) -> Result<Self> {
         let rank = self.shape.rank();
         if rank <= dim1 || rank <= dim2 {
             Err(Error::UnexpectedNumberOfDims {
@@ -120,7 +120,7 @@ impl Layout {
         })
     }
 
-    pub(crate) fn permute(&self, idxs: &[usize]) -> Result<Self> {
+    pub fn permute(&self, idxs: &[usize]) -> Result<Self> {
         let is_permutation =
             idxs.len() == self.shape.rank() && (0..idxs.len()).all(|i| idxs.contains(&i));
         if !is_permutation {

candle-core/src/lib.rs

@@ -14,7 +14,7 @@
 //!
 //! ## Features
 //!
-//! - Simple syntax (looks and like PyTorch)
+//! - Simple syntax (looks and feels like PyTorch)
 //! - CPU and Cuda backends (and M1 support)
 //! - Enable serverless (CPU) small and fast deployments
 //! - Model training
@@ -37,44 +37,51 @@
 mod accelerate;
 pub mod backend;
 pub mod backprop;
-mod conv;
+pub mod conv;
 mod convert;
 pub mod cpu;
 pub mod cpu_backend;
 #[cfg(feature = "cuda")]
 pub mod cuda_backend;
-#[cfg(feature = "cudnn")]
-pub mod cudnn;
+mod custom_op;
 mod device;
 pub mod display;
 mod dtype;
-mod dummy_cuda_backend;
+pub mod dummy_cuda_backend;
+mod dummy_metal_backend;
 pub mod error;
 mod indexer;
 pub mod layout;
+#[cfg(feature = "metal")]
+pub mod metal_backend;
 #[cfg(feature = "mkl")]
 mod mkl;
 pub mod npy;
-mod op;
+pub mod op;
 pub mod pickle;
 pub mod quantized;
 pub mod safetensors;
 pub mod scalar;
 pub mod shape;
+mod sort;
 mod storage;
 mod strided_index;
 mod tensor;
+mod tensor_cat;
 pub mod test_utils;
 pub mod utils;
 mod variable;
 
-pub use cpu_backend::CpuStorage;
-pub use device::{Device, DeviceLocation};
-pub use dtype::{DType, FloatDType, IntDType, WithDType};
+#[cfg(feature = "cudnn")]
+pub use cuda_backend::cudnn;
+
+pub use cpu_backend::{CpuStorage, CpuStorageRef};
+pub use custom_op::{CustomOp1, CustomOp2, CustomOp3, InplaceOp1, InplaceOp2, InplaceOp3};
+pub use device::{Device, DeviceLocation, NdArray};
+pub use dtype::{DType, DTypeParseError, FloatDType, IntDType, WithDType};
 pub use error::{Error, Result};
 pub use indexer::IndexOp;
 pub use layout::Layout;
-pub use op::{CustomOp1, CustomOp2, CustomOp3};
 pub use shape::{Shape, D};
 pub use storage::Storage;
 pub use strided_index::{StridedBlocks, StridedIndex};
@@ -82,10 +89,18 @@ pub use tensor::{Tensor, TensorId};
 pub use variable::Var;
 
 #[cfg(feature = "cuda")]
-pub use cuda_backend::{CudaDevice, CudaStorage};
+pub use cuda_backend as cuda;
 
 #[cfg(not(feature = "cuda"))]
-pub use dummy_cuda_backend::{CudaDevice, CudaStorage};
+pub use dummy_cuda_backend as cuda;
+
+pub use cuda::{CudaDevice, CudaStorage};
+
+#[cfg(feature = "metal")]
+pub use metal_backend::{MetalDevice, MetalError, MetalStorage};
+
+#[cfg(not(feature = "metal"))]
+pub use dummy_metal_backend::{MetalDevice, MetalError, MetalStorage};
 
 #[cfg(feature = "mkl")]
 extern crate intel_mkl_src;
@@ -114,18 +129,21 @@ pub trait Module {
     fn forward(&self, xs: &Tensor) -> Result<Tensor>;
 }
 
-impl Module for quantized::QMatMul {
-    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
-        self.forward(xs)
-    }
-}
-
 impl<T: Fn(&Tensor) -> Result<Tensor>> Module for T {
     fn forward(&self, xs: &Tensor) -> Result<Tensor> {
         self(xs)
     }
 }
 
+impl<M: Module> Module for Option<&M> {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        match self {
+            None => Ok(xs.clone()),
+            Some(m) => m.forward(xs),
+        }
+    }
+}
+
 // A trait defining a module with forward method using a single tensor argument and a flag to
 // separate the training and evaluation behaviors.
 pub trait ModuleT {

candle-core/src/metal_backend/device.rs

@@ -0,0 +1,287 @@
+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 super::MetalError;
+
+/// Unique identifier for cuda devices.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
+pub struct DeviceId(usize);
+
+impl DeviceId {
+    pub(crate) fn new() -> Self {
+        // https://users.rust-lang.org/t/idiomatic-rust-way-to-generate-unique-id/33805
+        use std::sync::atomic;
+        static COUNTER: atomic::AtomicUsize = atomic::AtomicUsize::new(1);
+        Self(COUNTER.fetch_add(1, atomic::Ordering::Relaxed))
+    }
+}
+
+type BufferMap = HashMap<(NSUInteger, MTLResourceOptions), Vec<Arc<Buffer>>>;
+type AllocatedBuffers = Arc<RwLock<BufferMap>>;
+
+#[derive(Clone)]
+pub struct MetalDevice {
+    /// Unique identifier, the registryID is not sufficient as it identifies the GPU rather than
+    /// the device itself.
+    pub(crate) id: DeviceId,
+
+    /// Raw metal device: <https://developer.apple.com/documentation/metal/mtldevice?language=objc>
+    pub(crate) device: metal::Device,
+
+    /// 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>,
+    /// Simple allocator struct.
+    /// The buffers are stored in size buckets since ML tends to use similar shapes over and over.
+    /// We store the buffers in [`Arc`] because it's much faster than Obj-c internal ref counting
+    /// (could be linked to FFI communication overhead).
+    ///
+    /// Whenever a buffer has a strong_count==1, we can reuse it, it means it was dropped in the
+    /// graph calculation, and only we the allocator kept a reference to it, therefore it's free
+    /// to be reused. However, in order for this to work, we need to guarantee the order of
+    /// operation, so that this buffer is not being used by another kernel at the same time.
+    /// Arc is the CPU reference count, it doesn't mean anything on the GPU side of things.
+    ///
+    /// Whenever we actually allocate a new buffer, we make a full sweep to clean up unused buffers
+    /// (strong_count = 1).
+    pub(crate) buffers: AllocatedBuffers,
+    /// Seed for random number generation.
+    pub(crate) seed: Arc<Mutex<Buffer>>,
+}
+
+impl std::fmt::Debug for MetalDevice {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "MetalDevice({:?})", self.id)
+    }
+}
+
+impl std::ops::Deref for MetalDevice {
+    type Target = metal::DeviceRef;
+
+    fn deref(&self) -> &Self::Target {
+        &self.device
+    }
+}
+
+impl MetalDevice {
+    pub fn id(&self) -> DeviceId {
+        self.id
+    }
+
+    pub fn metal_device(&self) -> &metal::Device {
+        &self.device
+    }
+
+    pub fn command_queue(&self) -> &CommandQueue {
+        &self.command_queue
+    }
+
+    pub fn command_buffer(&self) -> Result<CommandBuffer> {
+        let mut command_buffer_lock = self.command_buffer.write().map_err(MetalError::from)?;
+        let mut command_buffer = command_buffer_lock.to_owned();
+        let mut index = self
+            .command_buffer_index
+            .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()?;
+        }
+        *index += 1;
+        Ok(command_buffer)
+    }
+
+    pub fn wait_until_completed(&self) -> Result<()> {
+        let mut command_buffer = self.command_buffer.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(())
+    }
+
+    pub fn kernels(&self) -> &Kernels {
+        &self.kernels
+    }
+
+    pub fn device(&self) -> &metal::Device {
+        &self.device
+    }
+
+    /// Creates a new buffer (not necessarily zeroed).
+    /// The buffer is [MTLPrivate](https://developer.apple.com/documentation/metal/mtlstoragemode)
+    /// This means the buffer data cannot be read on the CPU directly.
+    ///
+    /// [`name`] is only used to keep track of the resource origin in case of bugs
+    pub fn new_buffer(
+        &self,
+        element_count: usize,
+        dtype: DType,
+        name: &str,
+    ) -> Result<Arc<Buffer>> {
+        let size = (element_count * dtype.size_in_bytes()) as NSUInteger;
+        self.allocate_buffer(size, MTLResourceOptions::StorageModePrivate, name)
+    }
+
+    /// Creates a new buffer (not necessarily zeroed).
+    /// The buffer is [MTLManaged](https://developer.apple.com/documentation/metal/mtlstoragemode)
+    /// This means the buffer can be read on the CPU but will require manual
+    /// synchronization when the CPU memory is modified
+    /// Used as a bridge to gather data back from the GPU
+    pub fn new_buffer_managed(&self, size: NSUInteger) -> Result<Arc<Buffer>> {
+        self.allocate_buffer(size, MTLResourceOptions::StorageModeManaged, "managed")
+    }
+
+    /// Creates a new buffer from data.
+    /// The buffer is [MTLManaged](https://developer.apple.com/documentation/metal/mtlstoragemode)
+    ///
+    /// Does not require synchronization, as [newBufferWithBytes](https://developer.apple.com/documentation/metal/mtldevice/1433429-newbufferwithbytes)
+    /// allocates the buffer and copies over the existing data before returning the MTLBuffer.
+    pub fn new_buffer_with_data<T>(&self, data: &[T]) -> Result<Arc<Buffer>> {
+        let size = core::mem::size_of_val(data) as NSUInteger;
+        let new_buffer = self.device.new_buffer_with_data(
+            data.as_ptr() as *const c_void,
+            size,
+            MTLResourceOptions::StorageModeManaged,
+        );
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+        let subbuffers = buffers
+            .entry((size, MTLResourceOptions::StorageModeManaged))
+            .or_insert(vec![]);
+
+        let new_buffer = Arc::new(new_buffer);
+        subbuffers.push(new_buffer.clone());
+        Ok(new_buffer)
+    }
+
+    pub fn allocate_zeros(&self, size_in_bytes: usize) -> Result<Arc<Buffer>> {
+        let buffer = self.allocate_buffer(
+            size_in_bytes as NSUInteger,
+            MTLResourceOptions::StorageModePrivate,
+            "allocate_zeros",
+        )?;
+        let command_buffer = self.command_buffer()?;
+        command_buffer.set_label("zeros");
+        let blit = command_buffer.new_blit_command_encoder();
+        blit.fill_buffer(
+            &buffer,
+            metal::NSRange {
+                location: 0,
+                length: buffer.length(),
+            },
+            0,
+        );
+        blit.end_encoding();
+        Ok(buffer)
+    }
+
+    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.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,
+        size: NSUInteger,
+        option: MTLResourceOptions,
+        _name: &str,
+    ) -> Result<Arc<Buffer>> {
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+        if let Some(b) = self.find_available_buffer(size, option, &buffers) {
+            // Cloning also ensures we increment the strong count
+            return Ok(b.clone());
+        }
+
+        let size = buf_size(size);
+        let subbuffers = buffers.entry((size, option)).or_insert(vec![]);
+
+        let new_buffer = self.device.new_buffer(size as NSUInteger, option);
+        let new_buffer = Arc::new(new_buffer);
+        subbuffers.push(new_buffer.clone());
+
+        Ok(new_buffer)
+    }
+
+    /// Create a metal GPU capture trace on [`path`].
+    pub fn capture<P: AsRef<Path>>(&self, path: P) -> Result<()> {
+        let capture = metal::CaptureManager::shared();
+        let descriptor = metal::CaptureDescriptor::new();
+        descriptor.set_destination(metal::MTLCaptureDestination::GpuTraceDocument);
+        descriptor.set_capture_device(self);
+        descriptor.set_output_url(path);
+
+        capture
+            .start_capture(&descriptor)
+            .map_err(MetalError::from)?;
+        Ok(())
+    }
+}
+
+fn buf_size(size: NSUInteger) -> NSUInteger {
+    size.saturating_sub(1).next_power_of_two() as NSUInteger
+}

candle-core/src/mkl.rs

@@ -333,6 +333,16 @@ pub fn vd_tanh_inplace(y: &mut [f64]) {
     unsafe { ffi::vdTanh(y.len() as i32, y.as_ptr(), y.as_mut_ptr()) }
 }
 
+#[inline]
+pub fn vs_exp_inplace(y: &mut [f32]) {
+    unsafe { ffi::vsExp(y.len() as i32, y.as_ptr(), y.as_mut_ptr()) }
+}
+
+#[inline]
+pub fn vd_exp_inplace(y: &mut [f64]) {
+    unsafe { ffi::vdExp(y.len() as i32, y.as_ptr(), y.as_mut_ptr()) }
+}
+
 #[inline]
 pub fn vs_gelu(vs: &[f32], ys: &mut [f32]) {
     for (&v, y) in vs.iter().zip(ys.iter_mut()) {
@@ -355,6 +365,28 @@ pub fn vd_gelu(vs: &[f64], ys: &mut [f64]) {
     }
 }
 
+#[inline]
+pub fn vs_silu(vs: &[f32], ys: &mut [f32]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vs_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
+#[inline]
+pub fn vd_silu(vs: &[f64], ys: &mut [f64]) {
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = -v
+    }
+    vd_exp_inplace(ys);
+    for (&v, y) in vs.iter().zip(ys.iter_mut()) {
+        *y = v / (1.0 + *y)
+    }
+}
+
 macro_rules! binary_op {
     ($fn_name:ident, $ty:ty, $mkl_name:ident) => {
         #[inline]

candle-core/src/npy.rs

@@ -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

@@ -1,5 +1,5 @@
 #![allow(clippy::redundant_closure_call)]
-use crate::{CpuStorage, CudaStorage, Layout, Result, Shape, Tensor};
+use crate::Tensor;
 use half::{bf16, f16};
 use num_traits::float::Float;
 
@@ -61,10 +61,12 @@ pub enum UnaryOp {
     GeluErf,
     Erf,
     Relu,
+    Silu,
     Tanh,
     Floor,
     Ceil,
     Round,
+    Sign,
 }
 
 #[derive(Clone)]
@@ -131,8 +133,15 @@ pub enum Op {
         stride: (usize, usize),
     },
 
-    UpsampleNearest1D(Tensor),
-    UpsampleNearest2D(Tensor),
+    UpsampleNearest1D {
+        arg: Tensor,
+        target_size: usize,
+    },
+    UpsampleNearest2D {
+        arg: Tensor,
+        target_h: usize,
+        target_w: usize,
+    },
 
     Cat(Vec<Tensor>, usize),
 
@@ -153,126 +162,23 @@ pub enum Op {
     Permute(Tensor, Vec<usize>),
     Elu(Tensor, f64),
     Powf(Tensor, f64),
-    CustomOp1(Tensor, std::sync::Arc<Box<dyn CustomOp1 + Send + Sync>>),
+    CustomOp1(
+        Tensor,
+        std::sync::Arc<Box<dyn crate::CustomOp1 + Send + Sync>>,
+    ),
     CustomOp2(
         Tensor,
         Tensor,
-        std::sync::Arc<Box<dyn CustomOp2 + Send + Sync>>,
+        std::sync::Arc<Box<dyn crate::CustomOp2 + Send + Sync>>,
     ),
     CustomOp3(
         Tensor,
         Tensor,
         Tensor,
-        std::sync::Arc<Box<dyn CustomOp3 + Send + Sync>>,
+        std::sync::Arc<Box<dyn crate::CustomOp3 + Send + Sync>>,
     ),
 }
 
-/// Unary ops that can be defined in user-land.
-pub trait CustomOp1 {
-    // Box<dyn> does not support const yet, so use a function to get the name.
-    fn name(&self) -> &'static str;
-
-    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cpu_fwd(&self, storage: &CpuStorage, layout: &Layout) -> Result<(CpuStorage, Shape)>;
-
-    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cuda_fwd(&self, _storage: &CudaStorage, _layout: &Layout) -> Result<(CudaStorage, Shape)> {
-        Err(crate::Error::Cuda(
-            format!("no cuda implementation for {}", self.name()).into(),
-        ))
-    }
-
-    /// This function takes as argument the argument `arg` used in the forward pass, the result
-    /// produced by the forward operation `res` and the gradient of the result `grad_res`.
-    /// The function should return the gradient of the argument.
-    fn bwd(&self, _arg: &Tensor, _res: &Tensor, _grad_res: &Tensor) -> Result<Option<Tensor>> {
-        Err(crate::Error::BackwardNotSupported { op: self.name() })
-    }
-}
-
-pub trait CustomOp2 {
-    fn name(&self) -> &'static str;
-
-    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cpu_fwd(
-        &self,
-        s1: &CpuStorage,
-        l1: &Layout,
-        s2: &CpuStorage,
-        l2: &Layout,
-    ) -> Result<(CpuStorage, Shape)>;
-
-    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cuda_fwd(
-        &self,
-        _: &CudaStorage,
-        _: &Layout,
-        _: &CudaStorage,
-        _: &Layout,
-    ) -> Result<(CudaStorage, Shape)> {
-        Err(crate::Error::Cuda(
-            format!("no cuda implementation for {}", self.name()).into(),
-        ))
-    }
-
-    fn bwd(
-        &self,
-        _arg1: &Tensor,
-        _arg2: &Tensor,
-        _res: &Tensor,
-        _grad_res: &Tensor,
-    ) -> Result<(Option<Tensor>, Option<Tensor>)> {
-        Err(crate::Error::BackwardNotSupported { op: self.name() })
-    }
-}
-
-pub trait CustomOp3 {
-    fn name(&self) -> &'static str;
-
-    /// The forward pass, as run on a cpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cpu_fwd(
-        &self,
-        s1: &CpuStorage,
-        l1: &Layout,
-        s2: &CpuStorage,
-        l2: &Layout,
-        s3: &CpuStorage,
-        l3: &Layout,
-    ) -> Result<(CpuStorage, Shape)>;
-
-    /// The forward pass, as run on a gpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn cuda_fwd(
-        &self,
-        _: &CudaStorage,
-        _: &Layout,
-        _: &CudaStorage,
-        _: &Layout,
-        _: &CudaStorage,
-        _: &Layout,
-    ) -> Result<(CudaStorage, Shape)> {
-        Err(crate::Error::Cuda(
-            format!("no cuda implementation for {}", self.name()).into(),
-        ))
-    }
-
-    fn bwd(
-        &self,
-        _arg1: &Tensor,
-        _arg2: &Tensor,
-        _arg3: &Tensor,
-        _res: &Tensor,
-        _grad_res: &Tensor,
-    ) -> Result<(Option<Tensor>, Option<Tensor>, Option<Tensor>)> {
-        Err(crate::Error::BackwardNotSupported { op: self.name() })
-    }
-}
-
 pub trait UnaryOpT {
     const NAME: &'static str;
     const KERNEL: &'static str;
@@ -344,10 +250,12 @@ pub(crate) struct Gelu;
 pub(crate) struct GeluErf;
 pub(crate) struct Erf;
 pub(crate) struct Relu;
+pub(crate) struct Silu;
 pub(crate) struct Tanh;
 pub(crate) struct Floor;
 pub(crate) struct Ceil;
 pub(crate) struct Round;
+pub(crate) struct Sign;
 
 macro_rules! bin_op {
     ($op:ident, $name: literal, $e: expr, $f32_vec: ident, $f64_vec: ident) => {
@@ -551,7 +459,15 @@ unary_op!(Recip, "recip", v, v.recip());
 unary_op!(Sqr, "sqr", v, v * v, vs_sqr, vd_sqr);
 unary_op!(Sqrt, "sqrt", v, v.sqrt(), vs_sqrt, vd_sqrt);
 
-/// `gelu` operation
+// Hardcode the value for sqrt(2/pi)
+// https://github.com/huggingface/candle/issues/1982
+#[allow(clippy::excessive_precision)]
+const SQRT_TWO_OVER_PI_F32: f32 = 0.79788456080286535587989211986876373;
+#[allow(clippy::excessive_precision)]
+const SQRT_TWO_OVER_PI_F64: f64 = 0.79788456080286535587989211986876373;
+
+/// Tanh based approximation of the `gelu` operation
+/// GeluErf is the more precise one.
 /// <https://en.wikipedia.org/wiki/Activation_function#Comparison_of_activation_functions>
 impl UnaryOpT for Gelu {
     const NAME: &'static str = "gelu";
@@ -562,7 +478,7 @@ impl UnaryOpT for Gelu {
             * v
             * (bf16::ONE
                 + bf16::tanh(
-                    (bf16::from_f32_const(2.0) / bf16::PI).sqrt()
+                    bf16::from_f32_const(SQRT_TWO_OVER_PI_F32)
                         * v
                         * (bf16::ONE + bf16::from_f32_const(0.044715) * v * v),
                 ))
@@ -573,22 +489,18 @@ impl UnaryOpT for Gelu {
             * v
             * (f16::ONE
                 + f16::tanh(
-                    (f16::from_f32_const(2.0) / f16::PI).sqrt()
+                    f16::from_f32_const(SQRT_TWO_OVER_PI_F32)
                         * v
                         * (f16::ONE + f16::from_f32_const(0.044715) * v * v),
                 ))
     }
     #[inline(always)]
     fn f32(v: f32) -> f32 {
-        0.5 * v
-            * (1.0
-                + f32::tanh((2.0f32 / std::f32::consts::PI).sqrt() * v * (1.0 + 0.044715 * v * v)))
+        0.5 * v * (1.0 + f32::tanh(SQRT_TWO_OVER_PI_F32 * v * (1.0 + 0.044715 * v * v)))
     }
     #[inline(always)]
     fn f64(v: f64) -> f64 {
-        0.5 * v
-            * (1.0
-                + f64::tanh((2.0f64 / std::f64::consts::PI).sqrt() * v * (1.0 + 0.044715 * v * v)))
+        0.5 * v * (1.0 + f64::tanh(SQRT_TWO_OVER_PI_F64 * v * (1.0 + 0.044715 * v * v)))
     }
     #[inline(always)]
     fn u8(_: u8) -> u8 {
@@ -677,6 +589,77 @@ impl UnaryOpT for Erf {
     }
 }
 
+/// Silu operation
+impl UnaryOpT for Silu {
+    const NAME: &'static str = "silu";
+    const V: Self = Silu;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        v / (bf16::ONE + (-v).exp())
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        v / (f16::ONE + (-v).exp())
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        v / (1.0 + (-v).exp())
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        v / (1.0 + (-v).exp())
+    }
+    #[inline(always)]
+    fn u8(_: u8) -> u8 {
+        0
+    }
+    #[inline(always)]
+    fn u32(_: u32) -> u32 {
+        0
+    }
+    #[inline(always)]
+    fn i64(_: i64) -> i64 {
+        0
+    }
+    const KERNEL: &'static str = "usilu";
+
+    #[cfg(feature = "mkl")]
+    const F32_VEC: bool = true;
+
+    #[cfg(feature = "mkl")]
+    #[inline(always)]
+    fn f32_vec(xs: &[f32], ys: &mut [f32]) {
+        crate::mkl::vs_silu(xs, ys)
+    }
+
+    #[cfg(feature = "mkl")]
+    const F64_VEC: bool = true;
+
+    #[cfg(feature = "mkl")]
+    #[inline(always)]
+    fn f64_vec(xs: &[f64], ys: &mut [f64]) {
+        crate::mkl::vd_silu(xs, ys)
+    }
+
+    #[cfg(feature = "accelerate")]
+    const F32_VEC: bool = true;
+
+    #[cfg(feature = "accelerate")]
+    #[inline(always)]
+    fn f32_vec(xs: &[f32], ys: &mut [f32]) {
+        crate::accelerate::vs_silu(xs, ys)
+    }
+
+    #[cfg(feature = "accelerate")]
+    const F64_VEC: bool = true;
+
+    #[cfg(feature = "accelerate")]
+    #[inline(always)]
+    fn f64_vec(xs: &[f64], ys: &mut [f64]) {
+        crate::accelerate::vd_silu(xs, ys)
+    }
+}
+
 impl UnaryOpT for Abs {
     const NAME: &'static str = "abs";
     const KERNEL: &'static str = "uabs";
@@ -944,3 +927,37 @@ impl std::ops::Deref for BackpropOp {
         &self.0
     }
 }
+
+impl UnaryOpT for Sign {
+    const NAME: &'static str = "sign";
+    const KERNEL: &'static str = "usign";
+    const V: Self = Sign;
+    #[inline(always)]
+    fn bf16(v: bf16) -> bf16 {
+        bf16::from((v > bf16::ZERO) as i8) - bf16::from((v < bf16::ZERO) as i8)
+    }
+    #[inline(always)]
+    fn f16(v: f16) -> f16 {
+        f16::from((v > f16::ZERO) as i8) - f16::from((v < f16::ZERO) as i8)
+    }
+    #[inline(always)]
+    fn f32(v: f32) -> f32 {
+        f32::from(v > 0.) - f32::from(v < 0.)
+    }
+    #[inline(always)]
+    fn f64(v: f64) -> f64 {
+        f64::from(v > 0.) - f64::from(v < 0.)
+    }
+    #[inline(always)]
+    fn u8(v: u8) -> u8 {
+        u8::min(1, v)
+    }
+    #[inline(always)]
+    fn u32(v: u32) -> u32 {
+        u32::min(1, v)
+    }
+    #[inline(always)]
+    fn i64(v: i64) -> i64 {
+        (v > 0) as i64 - (v < 0) as i64
+    }
+}

candle-core/src/pickle.rs

@@ -42,7 +42,7 @@ pub enum OpCode {
     Stop = b'.',
     NewObj = 0x81,
     EmptyList = b']',
-    BinFloat = b'g',
+    BinFloat = b'G',
     Append = b'a',
     Appends = b'e',
 }
@@ -217,6 +217,13 @@ impl Object {
                 let args = args.remove(1);
                 (callable, args)
             }
+            Object::Class {
+                module_name,
+                class_name,
+            } if module_name == "torch._utils" && class_name == "_rebuild_parameter" => {
+                let mut args = args.tuple()?;
+                args.remove(0).reduce()?
+            }
             _ => (callable, args),
         };
         match callable {
@@ -227,13 +234,11 @@ impl Object {
             _ => return Ok(None),
         };
         let (layout, dtype, file_path, storage_size) = rebuild_args(args)?;
-        let mut path = dir_name.to_path_buf();
-        path.push(file_path);
         Ok(Some(TensorInfo {
             name,
             dtype,
             layout,
-            path: path.to_string_lossy().into_owned(),
+            path: format!("{}/{}", dir_name.to_string_lossy(), file_path),
             storage_size,
         }))
     }
@@ -345,8 +350,10 @@ impl Stack {
                 module_name,
                 class_name,
             } => {
-                if module_name == "collections" && class_name == "OrderedDict" {
-                    // TODO: have a separate ordered dict.
+                if module_name == "collections"
+                    && (class_name == "OrderedDict" || class_name == "defaultdict")
+                {
+                    // TODO: have a separate ordered dict and a separate default dict.
                     Some(Object::Dict(vec![]))
                 } else {
                     None
@@ -455,7 +462,10 @@ impl Stack {
                 self.push(Object::Int(arg))
             }
             OpCode::BinFloat => {
-                let arg = r.read_f64::<LittleEndian>()?;
+                // Somehow floats are encoded using BigEndian whereas int types use LittleEndian.
+                // https://github.com/python/cpython/blob/0c80da4c14d904a367968955544dd6ae58c8101c/Lib/pickletools.py#L855
+                // https://github.com/pytorch/pytorch/blob/372d078f361e726bb4ac0884ac334b04c58179ef/torch/_weights_only_unpickler.py#L243
+                let arg = r.read_f64::<byteorder::BigEndian>()?;
                 self.push(Object::Float(arg))
             }
             OpCode::BinUnicode => {
@@ -627,9 +637,16 @@ pub struct TensorInfo {
     pub storage_size: usize,
 }
 
+/// Read the tensor info from a .pth file.
+///
+/// # Arguments
+/// * `file` - The path to the .pth file.
+/// * `verbose` - Whether to print debug information.
+/// * `key` - Optional key to retrieve `state_dict` from the pth file.
 pub fn read_pth_tensor_info<P: AsRef<std::path::Path>>(
     file: P,
     verbose: bool,
+    key: Option<&str>,
 ) -> Result<Vec<TensorInfo>> {
     let file = std::fs::File::open(file)?;
     let zip_reader = std::io::BufReader::new(file);
@@ -651,8 +668,9 @@ pub fn read_pth_tensor_info<P: AsRef<std::path::Path>>(
         stack.read_loop(&mut reader)?;
         let obj = stack.finalize()?;
         if VERBOSE || verbose {
-            println!("{obj:?}");
+            println!("{obj:#?}");
         }
+
         let obj = match obj {
             Object::Build { callable, args } => match *callable {
                 Object::Reduce { callable, args: _ } => match *callable {
@@ -666,6 +684,24 @@ pub fn read_pth_tensor_info<P: AsRef<std::path::Path>>(
             },
             obj => obj,
         };
+
+        // If key is provided, then we need to extract the state_dict from the object.
+        let obj = if let Some(key) = key {
+            if let Object::Dict(key_values) = obj {
+                key_values
+                    .into_iter()
+                    .find(|(k, _)| *k == Object::Unicode(key.to_owned()))
+                    .map(|(_, v)| v)
+                    .ok_or_else(|| E::Msg(format!("key {key} not found")))?
+            } else {
+                obj
+            }
+        } else {
+            obj
+        };
+
+        // If the object is a dict, then we can extract the tensor info from it.
+        // NOTE: We are assuming that the `obj` is state_dict by this stage.
         if let Object::Dict(key_values) = obj {
             for (name, value) in key_values.into_iter() {
                 match value.into_tensor_info(name, &dir_name) {
@@ -688,8 +724,8 @@ pub struct PthTensors {
 }
 
 impl PthTensors {
-    pub fn new<P: AsRef<std::path::Path>>(path: P) -> Result<Self> {
-        let tensor_infos = read_pth_tensor_info(path.as_ref(), false)?;
+    pub fn new<P: AsRef<std::path::Path>>(path: P, key: Option<&str>) -> Result<Self> {
+        let tensor_infos = read_pth_tensor_info(path.as_ref(), false, key)?;
         let tensor_infos = tensor_infos
             .into_iter()
             .map(|ti| (ti.name.to_string(), ti))
@@ -703,6 +739,7 @@ impl PthTensors {
     }
 
     pub fn get(&self, name: &str) -> Result<Option<Tensor>> {
+        use std::io::Read;
         let tensor_info = match self.tensor_infos.get(name) {
             None => return Ok(None),
             Some(tensor_info) => tensor_info,
@@ -711,27 +748,56 @@ impl PthTensors {
         let zip_reader = std::io::BufReader::new(std::fs::File::open(&self.path)?);
         let mut zip = zip::ZipArchive::new(zip_reader)?;
         let mut reader = zip.by_name(&tensor_info.path)?;
+        let is_fortran_contiguous = tensor_info.layout.is_fortran_contiguous();
+        let rank = tensor_info.layout.shape().rank();
 
-        // Reading the data is a bit tricky as it can be strided, use an offset, etc.
-        // For now only support the basic case.
-        if tensor_info.layout.start_offset() != 0 || !tensor_info.layout.is_contiguous() {
+        // Reading the data is a bit tricky as it can be strided, for now only support the basic
+        // case and when the tensor is fortran contiguous.
+        if !tensor_info.layout.is_contiguous() && !is_fortran_contiguous {
             crate::bail!(
                 "cannot retrieve non-contiguous tensors {:?}",
                 tensor_info.layout
             )
         }
+        let start_offset = tensor_info.layout.start_offset();
+        if start_offset > 0 {
+            std::io::copy(
+                &mut reader.by_ref().take(start_offset as u64),
+                &mut std::io::sink(),
+            )?;
+        }
         let tensor = Tensor::from_reader(
             tensor_info.layout.shape().clone(),
             tensor_info.dtype,
             &mut reader,
         )?;
-        Ok(Some(tensor))
+
+        if rank > 1 && is_fortran_contiguous {
+            // Reverse the shape, e.g. Shape(2, 3, 4) -> Shape(4, 3, 2)
+            let shape_reversed: Vec<_> = tensor_info.layout.dims().iter().rev().cloned().collect();
+            let tensor = tensor.reshape(shape_reversed)?;
+
+            // Permute (transpose) the dimensions, e.g. Shape(4, 3, 2) -> Shape(2, 3, 4)
+            let dim_indeces_reversed: Vec<_> = (0..rank).rev().collect();
+            let tensor = tensor.permute(dim_indeces_reversed)?;
+            Ok(Some(tensor))
+        } else {
+            Ok(Some(tensor))
+        }
     }
 }
 
-/// Read all the tensors from a PyTorch pth file.
-pub fn read_all<P: AsRef<std::path::Path>>(path: P) -> Result<Vec<(String, Tensor)>> {
-    let pth = PthTensors::new(path)?;
+/// Read all the tensors from a PyTorch pth file with a given key.
+///
+/// # Arguments
+/// * `path` - Path to the pth file.
+/// * `key` - Optional key to retrieve `state_dict` from the pth file. Sometimes the pth file
+///           contains multiple objects and the state_dict is the one we are interested in.
+pub fn read_all_with_key<P: AsRef<std::path::Path>>(
+    path: P,
+    key: Option<&str>,
+) -> Result<Vec<(String, Tensor)>> {
+    let pth = PthTensors::new(path, key)?;
     let tensor_names = pth.tensor_infos.keys();
     let mut tensors = Vec::with_capacity(tensor_names.len());
     for name in tensor_names {
@@ -741,3 +807,11 @@ pub fn read_all<P: AsRef<std::path::Path>>(path: P) -> Result<Vec<(String, Tenso
     }
     Ok(tensors)
 }
+
+/// Read all the tensors from a PyTorch pth file.
+///
+/// # Arguments
+/// * `path` - Path to the pth file.
+pub fn read_all<P: AsRef<std::path::Path>>(path: P) -> Result<Vec<(String, Tensor)>> {
+    read_all_with_key(path, None)
+}

candle-core/src/quantized/avx.rs

@@ -353,7 +353,7 @@ pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Res
                 q3 = q3.add(32);
 
                 // Prepare low and high bits
-                // We hardcode the shifts here to avoid loading them into a seperate register
+                // We hardcode the shifts here to avoid loading them into a separate register
                 let q3l_0 = _mm256_and_si256(q3bits, m3);
                 let q3h_0 = if j == 0 {
                     _mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, 0)), 0)
@@ -586,7 +586,7 @@ pub(crate) fn vec_dot_q5k_q8k(n: usize, xs: &[BlockQ5K], ys: &[BlockQ8K]) -> Res
                 let q5bits = _mm256_loadu_si256(q5 as *const __m256i);
                 q5 = q5.add(32);
 
-                //Similar to q3k we hardcode the shifts here to avoid loading them into a seperate register
+                //Similar to q3k we hardcode the shifts here to avoid loading them into a separate register
                 let q5l_0 = _mm256_and_si256(q5bits, m4);
                 let q5l_0_shift_input = _mm256_and_si256(hbits, hmask);
                 let q5l_0_right_shift = match j {

candle-core/src/quantized/cuda.rs

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

candle-core/src/quantized/dummy_cuda.rs

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

candle-core/src/quantized/dummy_metal.rs

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

candle-core/src/quantized/ggml_file.rs

@@ -1,7 +1,7 @@
 //! Support for the GGML file format.
 
-use super::{k_quants, GgmlDType};
-use crate::Result;
+use super::{k_quants, GgmlDType, QStorage};
+use crate::{Device, Result};
 use byteorder::{LittleEndian, ReadBytesExt};
 use std::collections::HashMap;
 
@@ -121,11 +121,17 @@ fn from_raw_data<T: super::GgmlType + Send + Sync + 'static>(
     raw_data: &[u8],
     size_in_bytes: usize,
     dims: Vec<usize>,
+    device: &Device,
 ) -> Result<super::QTensor> {
     let raw_data_ptr = raw_data.as_ptr();
     let n_blocks = size_in_bytes / std::mem::size_of::<T>();
     let data = unsafe { std::slice::from_raw_parts(raw_data_ptr as *const T, n_blocks) };
-    super::QTensor::new(data.to_vec(), dims)
+    let data: QStorage = match device {
+        Device::Cpu => QStorage::Cpu(Box::new(data.to_vec())),
+        Device::Metal(metal) => super::metal::load_quantized(metal, data)?,
+        Device::Cuda(cuda) => super::cuda::load_quantized(cuda, data)?,
+    };
+    super::QTensor::new(data, dims)
 }
 
 /// Creates a [Tensor] from a raw GGML tensor.
@@ -133,29 +139,50 @@ pub fn qtensor_from_ggml(
     ggml_dtype: GgmlDType,
     raw_data: &[u8],
     dims: Vec<usize>,
+    device: &Device,
 ) -> Result<super::QTensor> {
     let tensor_elems = dims.iter().product::<usize>();
-    let blck_size = ggml_dtype.blck_size();
-    if tensor_elems % blck_size != 0 {
+    let block_size = ggml_dtype.block_size();
+    if tensor_elems % block_size != 0 {
         crate::bail!(
-            "the number of elements {tensor_elems} is not divisible by the block size {blck_size}"
+            "the number of elements {tensor_elems} is not divisible by the block size {block_size}"
         )
     }
-    let size_in_bytes = tensor_elems / blck_size * ggml_dtype.type_size();
+    let size_in_bytes = tensor_elems / block_size * ggml_dtype.type_size();
 
     match ggml_dtype {
-        GgmlDType::F32 => from_raw_data::<f32>(raw_data, size_in_bytes, dims),
-        GgmlDType::F16 => from_raw_data::<half::f16>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q4_0 => from_raw_data::<k_quants::BlockQ4_0>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q4_1 => from_raw_data::<k_quants::BlockQ4_1>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q5_0 => from_raw_data::<k_quants::BlockQ5_0>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q5_1 => from_raw_data::<k_quants::BlockQ5_1>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q8_0 => from_raw_data::<k_quants::BlockQ8_0>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q2K => from_raw_data::<k_quants::BlockQ2K>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q3K => from_raw_data::<k_quants::BlockQ3K>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q4K => from_raw_data::<k_quants::BlockQ4K>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q5K => from_raw_data::<k_quants::BlockQ5K>(raw_data, size_in_bytes, dims),
-        GgmlDType::Q6K => from_raw_data::<k_quants::BlockQ6K>(raw_data, size_in_bytes, dims),
+        GgmlDType::F32 => from_raw_data::<f32>(raw_data, size_in_bytes, dims, device),
+        GgmlDType::F16 => from_raw_data::<half::f16>(raw_data, size_in_bytes, dims, device),
+        GgmlDType::Q4_0 => {
+            from_raw_data::<k_quants::BlockQ4_0>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q4_1 => {
+            from_raw_data::<k_quants::BlockQ4_1>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q5_0 => {
+            from_raw_data::<k_quants::BlockQ5_0>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q5_1 => {
+            from_raw_data::<k_quants::BlockQ5_1>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q8_0 => {
+            from_raw_data::<k_quants::BlockQ8_0>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q2K => {
+            from_raw_data::<k_quants::BlockQ2K>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q3K => {
+            from_raw_data::<k_quants::BlockQ3K>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q4K => {
+            from_raw_data::<k_quants::BlockQ4K>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q5K => {
+            from_raw_data::<k_quants::BlockQ5K>(raw_data, size_in_bytes, dims, device)
+        }
+        GgmlDType::Q6K => {
+            from_raw_data::<k_quants::BlockQ6K>(raw_data, size_in_bytes, dims, device)
+        }
         _ => crate::bail!("quantized type {ggml_dtype:?} is not supported yet"),
     }
 }
@@ -163,6 +190,7 @@ pub fn qtensor_from_ggml(
 fn read_one_tensor<R: std::io::Seek + std::io::Read>(
     reader: &mut R,
     magic: VersionedMagic,
+    device: &Device,
 ) -> Result<(String, super::QTensor)> {
     let n_dims = reader.read_u32::<LittleEndian>()?;
     let name_len = reader.read_u32::<LittleEndian>()?;
@@ -183,11 +211,11 @@ fn read_one_tensor<R: std::io::Seek + std::io::Read>(
     }
     let dims = dims.iter().map(|&u| u as usize).collect::<Vec<_>>();
     let tensor_elems = dims.iter().product::<usize>();
-    let size_in_bytes = tensor_elems * ggml_dtype.type_size() / ggml_dtype.blck_size();
+    let size_in_bytes = tensor_elems * ggml_dtype.type_size() / ggml_dtype.block_size();
     // TODO: Mmap version to avoid copying the data around?
     let mut raw_data = vec![0u8; size_in_bytes];
     reader.read_exact(&mut raw_data)?;
-    match qtensor_from_ggml(ggml_dtype, &raw_data, dims) {
+    match qtensor_from_ggml(ggml_dtype, &raw_data, dims, device) {
         Ok(tensor) => Ok((name, tensor)),
         Err(e) => crate::bail!("Error creating tensor {name}: {e}"),
     }
@@ -198,10 +226,14 @@ pub struct Content {
     pub hparams: HParams,
     pub vocab: Vocab,
     pub tensors: HashMap<String, super::QTensor>,
+    pub device: Device,
 }
 
 impl Content {
-    pub fn read<R: std::io::Seek + std::io::Read>(reader: &mut R) -> Result<Content> {
+    pub fn read<R: std::io::Seek + std::io::Read>(
+        reader: &mut R,
+        device: &Device,
+    ) -> Result<Content> {
         // https://github.com/ggerganov/llama.cpp/blob/468ea24fb4633a0d681f7ac84089566c1c6190cb/llama.cpp#L505
         let last_position = reader.seek(std::io::SeekFrom::End(0))?;
         reader.seek(std::io::SeekFrom::Start(0))?;
@@ -211,14 +243,16 @@ impl Content {
         let mut tensors = HashMap::new();
 
         while reader.stream_position()? != last_position {
-            let (name, tensor) = read_one_tensor(reader, magic)?;
+            let (name, tensor) = read_one_tensor(reader, magic, device)?;
             tensors.insert(name, tensor);
         }
+        let device = device.clone();
         Ok(Self {
             magic,
             hparams,
             vocab,
             tensors,
+            device,
         })
     }
 

candle-core/src/quantized/gguf_file.rs

@@ -3,7 +3,7 @@
 //! Spec: https://github.com/philpax/ggml/blob/gguf-spec/docs/gguf.md
 
 use super::{GgmlDType, QTensor};
-use crate::Result;
+use crate::{Device, Result};
 use byteorder::{LittleEndian, ReadBytesExt, WriteBytesExt};
 use std::collections::HashMap;
 
@@ -41,7 +41,7 @@ impl VersionedMagic {
             (Magic::Gguf, 1) => Self::GgufV1,
             (Magic::Gguf, 2) => Self::GgufV2,
             (Magic::Gguf, 3) => Self::GgufV3,
-            _ => crate::bail!("ggml: unsupported magic/version {magic:?}/{version}"),
+            _ => crate::bail!("gguf: unsupported magic/version {magic:?}/{version}"),
         };
         Ok(versioned_magic)
     }
@@ -59,19 +59,25 @@ impl TensorInfo {
         &self,
         reader: &mut R,
         tensor_data_offset: u64,
+        device: &Device,
     ) -> Result<QTensor> {
         let tensor_elems = self.shape.elem_count();
-        let blck_size = self.ggml_dtype.blck_size();
-        if tensor_elems % blck_size != 0 {
+        let block_size = self.ggml_dtype.block_size();
+        if tensor_elems % block_size != 0 {
             crate::bail!(
-            "the number of elements {tensor_elems} is not divisible by the block size {blck_size}"
+            "the number of elements {tensor_elems} is not divisible by the block size {block_size}"
         )
         }
-        let size_in_bytes = tensor_elems / blck_size * self.ggml_dtype.type_size();
+        let size_in_bytes = tensor_elems / block_size * self.ggml_dtype.type_size();
         let mut raw_data = vec![0u8; size_in_bytes];
         reader.seek(std::io::SeekFrom::Start(tensor_data_offset + self.offset))?;
         reader.read_exact(&mut raw_data)?;
-        super::ggml_file::qtensor_from_ggml(self.ggml_dtype, &raw_data, self.shape.dims().to_vec())
+        super::ggml_file::qtensor_from_ggml(
+            self.ggml_dtype,
+            &raw_data,
+            self.shape.dims().to_vec(),
+            device,
+        )
     }
 }
 
@@ -129,7 +135,6 @@ 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,
 }
@@ -460,12 +465,13 @@ impl Content {
         &self,
         reader: &mut R,
         name: &str,
+        device: &Device,
     ) -> Result<QTensor> {
         let tensor_info = match self.tensor_infos.get(name) {
             Some(tensor_info) => tensor_info,
-            None => crate::bail!("cannot find tensor-infor for {name}"),
+            None => crate::bail!("cannot find tensor info for {name}"),
         };
-        tensor_info.read(reader, self.tensor_data_offset)
+        tensor_info.read(reader, self.tensor_data_offset, device)
     }
 }
 
@@ -517,10 +523,9 @@ pub fn write<W: std::io::Seek + std::io::Write>(
                 "internal error, unexpected current position {tensor_start_pos} {offset} {pos}"
             )
         }
-        let data_ptr = tensor.as_ptr();
-        let size_in_bytes = tensor.storage_size_in_bytes();
-        let data = unsafe { std::slice::from_raw_parts(data_ptr, size_in_bytes) };
-        w.write_all(data)?;
+        let data = tensor.data()?;
+        let size_in_bytes = data.len();
+        w.write_all(&data)?;
         let padding = 31 - (31 + size_in_bytes) % 32;
         w.write_all(&vec![0u8; padding])?;
     }

candle-core/src/quantized/k_quants.rs

@@ -1545,13 +1545,13 @@ impl GgmlType for BlockQ5K {
                 let d2 = d * sc as f32;
                 let m2 = min * m as f32;
                 for (ql, qh) in ql.iter().zip(qh) {
-                    let to_add = if qh & u1 != 0 { 16 } else { 1 };
-                    y[ys_index] = d1 * ((ql & 0xF) + to_add) as f32 - m1;
+                    let to_add = if qh & u1 != 0 { 16f32 } else { 0f32 };
+                    y[ys_index] = d1 * ((ql & 0xF) as f32 + to_add) - m1;
                     ys_index += 1;
                 }
                 for (ql, qh) in ql.iter().zip(qh) {
-                    let to_add = if qh & u2 != 0 { 16 } else { 1 };
-                    y[ys_index] = d2 * ((ql >> 4) + to_add) as f32 - m2;
+                    let to_add = if qh & u2 != 0 { 16f32 } else { 0f32 };
+                    y[ys_index] = d2 * ((ql >> 4) as f32 + to_add) - m2;
                     ys_index += 1;
                 }
                 is += 2;

candle-core/src/quantized/metal.rs

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

candle-core/src/quantized/mod.rs

@@ -1,23 +1,134 @@
-use crate::{Device, Result, Shape, Tensor};
+use crate::{CpuStorage, DType, Device, Result, Shape, Storage, Tensor};
+use k_quants::*;
+use std::borrow::Cow;
 
 #[cfg(target_feature = "avx")]
 pub mod avx;
+mod dummy_cuda;
+mod dummy_metal;
 pub mod ggml_file;
 pub mod gguf_file;
 pub mod k_quants;
+#[cfg(feature = "metal")]
+pub mod metal;
+#[cfg(not(feature = "metal"))]
+mod metal {
+    pub use super::dummy_metal::*;
+}
+#[cfg(feature = "cuda")]
+pub mod cuda;
+#[cfg(not(feature = "cuda"))]
+mod cuda {
+    pub use super::dummy_cuda::*;
+}
+
 #[cfg(target_feature = "neon")]
 pub mod neon;
 #[cfg(target_feature = "simd128")]
 pub mod simd128;
 pub mod utils;
+use half::f16;
 
 pub use k_quants::GgmlType;
 
 pub struct QTensor {
-    data: Box<dyn QuantizedType>,
+    storage: QStorage,
     shape: Shape,
 }
 
+impl Device {
+    fn qzeros(&self, elem_count: usize, dtype: GgmlDType) -> Result<QStorage> {
+        match self {
+            Device::Cpu => {
+                let storage = dtype.cpu_zeros(elem_count);
+                Ok(QStorage::Cpu(storage))
+            }
+            Device::Metal(metal) => {
+                let storage = metal::QMetalStorage::zeros(metal, elem_count, dtype)?;
+                Ok(QStorage::Metal(storage))
+            }
+            Device::Cuda(cuda) => {
+                let storage = cuda::QCudaStorage::zeros(cuda, elem_count, dtype)?;
+                Ok(QStorage::Cuda(storage))
+            }
+        }
+    }
+}
+
+pub enum QStorage {
+    Cpu(Box<dyn QuantizedType>),
+    Metal(metal::QMetalStorage),
+    Cuda(cuda::QCudaStorage),
+}
+
+impl QStorage {
+    fn block_size(&self) -> usize {
+        match self {
+            QStorage::Cpu(storage) => storage.block_size(),
+            QStorage::Metal(storage) => storage.dtype().block_size(),
+            QStorage::Cuda(storage) => storage.dtype().block_size(),
+        }
+    }
+
+    fn dtype(&self) -> GgmlDType {
+        match self {
+            QStorage::Cpu(storage) => storage.dtype(),
+            QStorage::Metal(storage) => storage.dtype(),
+            QStorage::Cuda(storage) => storage.dtype(),
+        }
+    }
+
+    fn device(&self) -> Device {
+        match self {
+            QStorage::Cpu(_storage) => Device::Cpu,
+            QStorage::Metal(storage) => Device::Metal(storage.device().clone()),
+            QStorage::Cuda(storage) => Device::Cuda(storage.device().clone()),
+        }
+    }
+
+    fn size_in_bytes(&self) -> usize {
+        match self {
+            QStorage::Cpu(storage) => storage.storage_size_in_bytes(),
+            QStorage::Metal(storage) => storage.storage_size_in_bytes(),
+            QStorage::Cuda(storage) => storage.storage_size_in_bytes(),
+        }
+    }
+
+    fn quantize(&mut self, src: &Storage) -> Result<()> {
+        match (self, src) {
+            (QStorage::Cpu(storage), Storage::Cpu(src)) => {
+                storage.from_float(src.as_slice::<f32>()?)?;
+            }
+            (QStorage::Metal(storage), Storage::Metal(src)) => storage.quantize(src)?,
+            (QStorage::Cuda(storage), Storage::Cuda(src)) => storage.quantize(src)?,
+            _ => crate::bail!("Invalid dequantize storage locations do not match"),
+        }
+        Ok(())
+    }
+
+    fn dequantize(&self, elem_count: usize) -> Result<Storage> {
+        match self {
+            QStorage::Cpu(storage) => Ok(Storage::Cpu(storage.dequantize(elem_count)?)),
+            QStorage::Metal(storage) => Ok(Storage::Metal(storage.dequantize(elem_count)?)),
+            QStorage::Cuda(storage) => Ok(Storage::Cuda(storage.dequantize(elem_count)?)),
+        }
+    }
+
+    fn data(&self) -> Result<Cow<[u8]>> {
+        match self {
+            QStorage::Cpu(storage) => {
+                let data_ptr = storage.as_ptr();
+                let size_in_bytes = storage.storage_size_in_bytes();
+                let data = unsafe { std::slice::from_raw_parts(data_ptr, size_in_bytes) };
+                Ok(Cow::from(data))
+            }
+            QStorage::Metal(_) | QStorage::Cuda(_) => {
+                crate::bail!("not implemented");
+            }
+        }
+    }
+}
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
 pub enum GgmlDType {
     F32,
@@ -77,6 +188,25 @@ impl GgmlDType {
         }
     }
 
+    /// The block dtype
+    pub fn cpu_zeros(&self, elem_count: usize) -> Box<dyn QuantizedType> {
+        match self {
+            Self::F32 => Box::new(vec![f32::zeros(); elem_count]),
+            Self::F16 => Box::new(vec![f16::zeros(); elem_count]),
+            Self::Q4_0 => Box::new(vec![BlockQ4_0::zeros(); elem_count / BlockQ4_0::BLCK_SIZE]),
+            Self::Q4_1 => Box::new(vec![BlockQ4_1::zeros(); elem_count / BlockQ4_1::BLCK_SIZE]),
+            Self::Q5_0 => Box::new(vec![BlockQ5_0::zeros(); elem_count / BlockQ5_0::BLCK_SIZE]),
+            Self::Q5_1 => Box::new(vec![BlockQ5_1::zeros(); elem_count / BlockQ5_1::BLCK_SIZE]),
+            Self::Q8_0 => Box::new(vec![BlockQ8_0::zeros(); elem_count / BlockQ8_0::BLCK_SIZE]),
+            Self::Q8_1 => Box::new(vec![BlockQ8_1::zeros(); elem_count / BlockQ8_1::BLCK_SIZE]),
+            Self::Q2K => Box::new(vec![BlockQ2K::zeros(); elem_count / BlockQ2K::BLCK_SIZE]),
+            Self::Q3K => Box::new(vec![BlockQ3K::zeros(); elem_count / BlockQ3K::BLCK_SIZE]),
+            Self::Q4K => Box::new(vec![BlockQ4K::zeros(); elem_count / BlockQ4K::BLCK_SIZE]),
+            Self::Q5K => Box::new(vec![BlockQ5K::zeros(); elem_count / BlockQ5K::BLCK_SIZE]),
+            Self::Q6K => Box::new(vec![BlockQ6K::zeros(); elem_count / BlockQ6K::BLCK_SIZE]),
+            Self::Q8K => Box::new(vec![BlockQ8K::zeros(); elem_count / BlockQ8K::BLCK_SIZE]),
+        }
+    }
     /// The type size for blocks in bytes.
     pub fn type_size(&self) -> usize {
         use k_quants::*;
@@ -100,7 +230,7 @@ impl GgmlDType {
     }
 
     /// The block size, i.e. the number of elements stored in each block.
-    pub fn blck_size(&self) -> usize {
+    pub fn block_size(&self) -> usize {
         match self {
             Self::F32 => 1,
             Self::F16 => 1,
@@ -119,9 +249,13 @@ impl GgmlDType {
 pub trait QuantizedType: Send + Sync {
     fn dtype(&self) -> GgmlDType;
     fn matmul_t(&self, mkn: (usize, usize, usize), lhs: &[f32], dst: &mut [f32]) -> Result<()>;
-    fn to_float(&self, ys: &mut [f32]) -> Result<()>;
+    fn dequantize(&self, elem_count: usize) -> Result<CpuStorage>;
     fn storage_size_in_bytes(&self) -> usize;
     fn as_ptr(&self) -> *const u8;
+    fn block_size(&self) -> usize;
+    #[allow(clippy::wrong_self_convention)]
+    fn from_float(&mut self, xs: &[f32]) -> Result<()>;
+    fn size(&self) -> usize;
 }
 
 impl<T: k_quants::GgmlType + Send + Sync> QuantizedType for Vec<T> {
@@ -129,12 +263,26 @@ impl<T: k_quants::GgmlType + Send + Sync> QuantizedType for Vec<T> {
         k_quants::matmul(mkn, lhs, self.as_slice(), dst)
     }
 
+    fn size(&self) -> usize {
+        self.len() * core::mem::size_of::<T>()
+    }
+
+    fn from_float(&mut self, xs: &[f32]) -> Result<()> {
+        T::from_float(xs, self)
+    }
+
     fn dtype(&self) -> GgmlDType {
         T::DTYPE
     }
 
-    fn to_float(&self, ys: &mut [f32]) -> Result<()> {
-        T::to_float(self.as_slice(), ys)
+    fn block_size(&self) -> usize {
+        T::BLCK_SIZE
+    }
+
+    fn dequantize(&self, elem_count: usize) -> Result<CpuStorage> {
+        let mut ys = vec![0.0f32; elem_count];
+        T::to_float(self.as_slice(), &mut ys)?;
+        Ok(CpuStorage::F32(ys))
     }
 
     fn storage_size_in_bytes(&self) -> usize {
@@ -152,56 +300,53 @@ impl std::fmt::Debug for QTensor {
     }
 }
 
-fn check_shape<T: k_quants::GgmlType>(shape: &Shape) -> Result<()> {
+fn check_shape(shape: &Shape, block_size: usize) -> Result<()> {
     let dims = shape.dims();
     if dims.is_empty() {
         crate::bail!("scalar tensor cannot be quantized {shape:?}")
     }
-    if dims[dims.len() - 1] % T::BLCK_SIZE != 0 {
+    if dims[dims.len() - 1] % block_size != 0 {
         crate::bail!(
             "quantized tensor must have their last dim divisible by block size {shape:?} {}",
-            T::BLCK_SIZE
+            block_size
         )
     }
     Ok(())
 }
 
 impl QTensor {
-    pub fn new<S: Into<Shape>, T: k_quants::GgmlType + Send + Sync + 'static>(
-        data: Vec<T>,
-        shape: S,
-    ) -> Result<Self> {
+    pub fn new<S: Into<Shape>>(storage: QStorage, shape: S) -> Result<Self> {
         let shape = shape.into();
-        check_shape::<T>(&shape)?;
-        Ok(Self {
-            data: Box::new(data),
-            shape,
-        })
+        check_shape(&shape, storage.block_size())?;
+        Ok(Self { storage, shape })
     }
 
-    pub fn quantize<T: k_quants::GgmlType + Send + Sync + 'static>(src: &Tensor) -> Result<Self> {
+    pub fn quantize(src: &Tensor, dtype: GgmlDType) -> Result<Self> {
         let shape = src.shape();
-        check_shape::<T>(shape)?;
-        let src = src
-            .to_dtype(crate::DType::F32)?
-            .flatten_all()?
-            .to_vec1::<f32>()?;
-        if src.len() % T::BLCK_SIZE != 0 {
+        let block_size = dtype.block_size();
+        check_shape(shape, block_size)?;
+        let src = src.to_dtype(crate::DType::F32)?.flatten_all()?;
+        let elem_count = shape.elem_count();
+        if elem_count % block_size != 0 {
             crate::bail!(
                 "tensor size ({shape:?}) is not divisible by block size {}",
-                T::BLCK_SIZE
+                block_size
             )
         }
-        let mut data = vec![T::zeros(); src.len() / T::BLCK_SIZE];
-        T::from_float(&src, &mut data)?;
+        let mut storage = src.device().qzeros(elem_count, dtype)?;
+        storage.quantize(&src.storage())?;
         Ok(Self {
-            data: Box::new(data),
+            storage,
             shape: shape.clone(),
         })
     }
 
     pub fn dtype(&self) -> GgmlDType {
-        self.data.dtype()
+        self.storage.dtype()
+    }
+
+    pub fn device(&self) -> Device {
+        self.storage.device()
     }
 
     pub fn rank(&self) -> usize {
@@ -213,21 +358,34 @@ impl QTensor {
     }
 
     pub fn dequantize(&self, device: &Device) -> Result<Tensor> {
-        let mut f32_data = vec![0f32; self.shape.elem_count()];
-        self.data.to_float(&mut f32_data)?;
-        Tensor::from_vec(f32_data, &self.shape, device)
+        let storage = self.storage.dequantize(self.shape.elem_count())?;
+        let none = crate::op::BackpropOp::none();
+        crate::tensor::from_storage(storage, self.shape.clone(), none, false).to_device(device)
     }
 
-    pub fn matmul_t(&self, mkn: (usize, usize, usize), lhs: &[f32], dst: &mut [f32]) -> Result<()> {
-        self.data.matmul_t(mkn, lhs, dst)
+    pub fn dequantize_f16(&self, device: &Device) -> Result<Tensor> {
+        // In the CUDA case, we have a specialized kernel as this can be useful for volta
+        // architectures. https://github.com/huggingface/candle/issues/2136
+        match &self.storage {
+            QStorage::Cuda(s) => {
+                let s = s.dequantize_f16(self.shape.elem_count())?;
+                let none = crate::op::BackpropOp::none();
+                crate::tensor::from_storage(Storage::Cuda(s), self.shape.clone(), none, false)
+                    .to_device(device)
+            }
+            _ => {
+                let s = self.dequantize(device)?.to_dtype(crate::DType::F16)?;
+                Ok(s)
+            }
+        }
     }
 
     pub fn storage_size_in_bytes(&self) -> usize {
-        self.data.storage_size_in_bytes()
+        self.storage.size_in_bytes()
     }
 
-    pub fn as_ptr(&self) -> *const u8 {
-        self.data.as_ptr()
+    pub fn data(&self) -> Result<Cow<'_, [u8]>> {
+        self.storage.data()
     }
 }
 
@@ -235,6 +393,7 @@ impl QTensor {
 pub enum QMatMul {
     QTensor(std::sync::Arc<QTensor>),
     Tensor(Tensor),
+    TensorF16(Tensor),
 }
 
 thread_local! {
@@ -248,6 +407,17 @@ 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() {
@@ -255,8 +425,11 @@ impl QMatMul {
             _ => DEQUANTIZE_ALL.with(|b| *b),
         };
         let t = if dequantize {
-            let tensor = qtensor.dequantize(&Device::Cpu)?;
+            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)
         };
@@ -266,6 +439,25 @@ 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 {
@@ -294,21 +486,45 @@ impl crate::CustomOp1 for QTensor {
         }
         dst_shape.push(n);
         let dst_shape = Shape::from(dst_shape);
-        let storage = storage.as_slice::<f32>()?;
-        let storage =
-            &storage[layout.start_offset()..layout.start_offset() + src_shape.elem_count()];
+        #[allow(clippy::infallible_destructuring_match)]
+        let self_storage = match &self.storage {
+            QStorage::Cpu(storage) => storage,
+            QStorage::Metal(_) | QStorage::Cuda(_) => crate::bail!("Invalid storage"),
+        };
+        let slice = storage.as_slice::<f32>()?;
+        let slice = &slice[layout.start_offset()..layout.start_offset() + src_shape.elem_count()];
         let mut dst_storage = vec![0f32; dst_shape.elem_count()];
-        self.matmul_t(
-            (dst_shape.elem_count() / n, k, n),
-            storage,
-            &mut dst_storage,
-        )?;
+        self_storage.matmul_t((dst_shape.elem_count() / n, k, n), slice, &mut dst_storage)?;
         Ok((crate::CpuStorage::F32(dst_storage), dst_shape))
     }
+
+    fn metal_fwd(
+        &self,
+        storage: &crate::MetalStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::MetalStorage, Shape)> {
+        let self_storage = match &self.storage {
+            QStorage::Metal(metal) => metal,
+            _ => unreachable!("Cannot call metal matmul on non metal QTensor"),
+        };
+        self_storage.fwd(&self.shape, storage, layout)
+    }
+
+    fn cuda_fwd(
+        &self,
+        storage: &crate::CudaStorage,
+        layout: &crate::Layout,
+    ) -> Result<(crate::CudaStorage, Shape)> {
+        let self_storage = match &self.storage {
+            QStorage::Cuda(cuda) => cuda,
+            _ => unreachable!("Cannot call cuda matmul on non cuda QTensor"),
+        };
+        self_storage.fwd(&self.shape, storage, layout)
+    }
 }
 
-impl QMatMul {
-    pub fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+impl crate::Module for QMatMul {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
         match self {
             Self::QTensor(t) => xs.apply_op1_no_bwd(t.as_ref()),
             Self::Tensor(w) => {
@@ -319,6 +535,15 @@ impl 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/quantized/neon.rs

@@ -12,6 +12,14 @@ use core::arch::arm::*;
 #[cfg(target_arch = "aarch64")]
 use core::arch::aarch64::*;
 
+#[inline(always)]
+unsafe fn vdotq_s32(a: int8x16_t, b: int8x16_t) -> int32x4_t {
+    // TODO: dotprod
+    let p0 = vmull_s8(vget_low_s8(a), vget_low_s8(b));
+    let p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+    vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1))
+}
+
 #[inline(always)]
 pub(crate) fn vec_dot_q4_0_q8_0(n: usize, xs: &[BlockQ4_0], ys: &[BlockQ8_0]) -> Result<f32> {
     let qk = QK8_0;
@@ -43,15 +51,8 @@ pub(crate) fn vec_dot_q4_0_q8_0(n: usize, xs: &[BlockQ4_0], ys: &[BlockQ8_0]) ->
             let v1_0l = vld1q_s8(y0.qs.as_ptr());
             let v1_0h = vld1q_s8(y0.qs.as_ptr().add(16));
 
-            // TODO: Support dotprod when it's available outside of nightly.
-            let pl0l = vmull_s8(vget_low_s8(v0_0ls), vget_low_s8(v1_0l));
-            let pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l));
-            let ph0l = vmull_s8(vget_low_s8(v0_0hs), vget_low_s8(v1_0h));
-            let ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h));
-
-            let pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
-            let ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
-
+            let pl0 = vdotq_s32(v0_0ls, v1_0l);
+            let ph0 = vdotq_s32(v0_0hs, v1_0h);
             sumv0 = vmlaq_n_f32(
                 sumv0,
                 vcvtq_f32_s32(vaddq_s32(pl0, ph0)),
@@ -82,14 +83,8 @@ pub(crate) fn vec_dot_q8_0_q8_0(n: usize, xs: &[BlockQ8_0], ys: &[BlockQ8_0]) ->
             let y0_0 = vld1q_s8(y0.qs.as_ptr());
             let y0_1 = vld1q_s8(y0.qs.as_ptr().add(16));
 
-            // TODO dotprod once this is the intrinsics are.
-            let p0_0 = vmull_s8(vget_low_s8(x0_0), vget_low_s8(y0_0));
-            let p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0));
-            let p0_2 = vmull_s8(vget_low_s8(x0_1), vget_low_s8(y0_1));
-            let p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1));
-
-            let p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1));
-            let p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3));
+            let p0 = vdotq_s32(x0_0, y0_0);
+            let p1 = vdotq_s32(x0_1, y0_1);
 
             sumv0 = vmlaq_n_f32(
                 sumv0,
@@ -118,10 +113,7 @@ pub(crate) fn vec_dot_q8k_q8k(n: usize, xs: &[BlockQ8K], ys: &[BlockQ8K]) -> Res
             for i in (0..QK_K).step_by(16) {
                 let xs = vld1q_s8(xs.add(i));
                 let ys = vld1q_s8(ys.add(i));
-                let xy_lo = vmull_s8(vget_low_s8(xs), vget_low_s8(ys));
-                let xy_up = vmull_s8(vget_high_s8(xs), vget_high_s8(ys));
-
-                let xy = vaddq_s32(vpaddlq_s16(xy_lo), vpaddlq_s16(xy_up));
+                let xy = vdotq_s32(xs, ys);
                 sum_i = vaddq_s32(sum_i, xy)
             }
             sumf += vaddvq_s32(sum_i) as f32 * scale
@@ -191,30 +183,16 @@ pub(crate) fn vec_dot_q6k_q8k(n: usize, xs: &[BlockQ6K], ys: &[BlockQ8K]) -> Res
                 let q6bytes_2 = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.2, m4b), q6h_2));
                 let q6bytes_3 = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.3, m4b), q6h_3));
 
-                // TODO: dotprod
-
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q6bytes_0), vget_high_s8(q8bytes.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q6bytes_1), vget_high_s8(q8bytes.1)),
-                );
+                let p0 = vdotq_s32(q6bytes_0, q8bytes.0);
+                let p1 = vdotq_s32(q6bytes_1, q8bytes.1);
                 let (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
-                isum += vaddvq_s16(p0) as i32 * scale0 + vaddvq_s16(p1) as i32 * scale1;
+                isum += vaddvq_s32(p0) * scale0 + vaddvq_s32(p1) * scale1;
                 scale = scale.add(2);
 
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_2), vget_low_s8(q8bytes.2)),
-                    vmull_s8(vget_high_s8(q6bytes_2), vget_high_s8(q8bytes.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_3), vget_low_s8(q8bytes.3)),
-                    vmull_s8(vget_high_s8(q6bytes_3), vget_high_s8(q8bytes.3)),
-                );
+                let p2 = vdotq_s32(q6bytes_2, q8bytes.2);
+                let p3 = vdotq_s32(q6bytes_3, q8bytes.3);
                 let (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
-                isum += vaddvq_s16(p2) as i32 * scale0 + vaddvq_s16(p3) as i32 * scale1;
+                isum += vaddvq_s32(p2) * scale0 + vaddvq_s32(p3) * scale1;
                 scale = scale.add(2);
 
                 let q8bytes = vld1q_s8_x4(q8);
@@ -234,29 +212,16 @@ pub(crate) fn vec_dot_q6k_q8k(n: usize, xs: &[BlockQ6K], ys: &[BlockQ8K]) -> Res
                 let q6bytes_2 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.2, 4), q6h_2));
                 let q6bytes_3 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.3, 4), q6h_3));
 
-                // TODO: dotprod case.
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q6bytes_0), vget_high_s8(q8bytes.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q6bytes_1), vget_high_s8(q8bytes.1)),
-                );
+                let p0 = vdotq_s32(q6bytes_0, q8bytes.0);
+                let p1 = vdotq_s32(q6bytes_1, q8bytes.1);
                 let (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
-                isum += vaddvq_s16(p0) as i32 * scale0 + vaddvq_s16(p1) as i32 * scale1;
+                isum += vaddvq_s32(p0) * scale0 + vaddvq_s32(p1) * scale1;
                 scale = scale.add(2);
 
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_2), vget_low_s8(q8bytes.2)),
-                    vmull_s8(vget_high_s8(q6bytes_2), vget_high_s8(q8bytes.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q6bytes_3), vget_low_s8(q8bytes.3)),
-                    vmull_s8(vget_high_s8(q6bytes_3), vget_high_s8(q8bytes.3)),
-                );
+                let p2 = vdotq_s32(q6bytes_2, q8bytes.2);
+                let p3 = vdotq_s32(q6bytes_3, q8bytes.3);
                 let (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
-                isum += vaddvq_s16(p2) as i32 * scale0 + vaddvq_s16(p3) as i32 * scale1;
+                isum += vaddvq_s32(p2) * scale0 + vaddvq_s32(p3) * scale1;
                 scale = scale.add(2);
             }
             sum += d_all * y.d * ((isum - 32 * isum_mins) as f32);
@@ -333,28 +298,14 @@ pub(crate) fn vec_dot_q5k_q8k(n: usize, xs: &[BlockQ5K], ys: &[BlockQ8K]) -> Res
                 let q5bytes_2 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.0, 4), q5h_2));
                 let q5bytes_3 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.1, 4), q5h_3));
 
-                // TODO: dotprod
-
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q5bytes_0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q5bytes_0), vget_high_s8(q8bytes.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q5bytes_1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q5bytes_1), vget_high_s8(q8bytes.1)),
-                );
-                sumi += vaddvq_s16(vaddq_s16(p0, p1)) as i32 * *scales as i32;
+                let p0 = vdotq_s32(q5bytes_0, q8bytes.0);
+                let p1 = vdotq_s32(q5bytes_1, q8bytes.1);
+                sumi += vaddvq_s32(vaddq_s32(p0, p1)) * *scales as i32;
                 scales = scales.add(1);
 
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q5bytes_2), vget_low_s8(q8bytes.2)),
-                    vmull_s8(vget_high_s8(q5bytes_2), vget_high_s8(q8bytes.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q5bytes_3), vget_low_s8(q8bytes.3)),
-                    vmull_s8(vget_high_s8(q5bytes_3), vget_high_s8(q8bytes.3)),
-                );
-                sumi += vaddvq_s16(vaddq_s16(p2, p3)) as i32 * *scales as i32;
+                let p2 = vdotq_s32(q5bytes_2, q8bytes.2);
+                let p3 = vdotq_s32(q5bytes_3, q8bytes.3);
+                sumi += vaddvq_s32(vaddq_s32(p2, p3)) * *scales as i32;
                 scales = scales.add(1);
             }
             sumf += d * sumi as f32 - dmin * sumi_mins as f32;
@@ -417,22 +368,15 @@ pub(crate) fn vec_dot_q4k_q8k(n: usize, xs: &[BlockQ4K], ys: &[BlockQ8K]) -> Res
             for j in 0..QK_K / 64 {
                 let q4bits = vld1q_u8_x2(q4);
                 q4 = q4.add(32);
-                // TODO: dotprod
                 let q8bytes = vld1q_s8_x2(q8);
                 q8 = q8.add(32);
                 let q4bytes = int8x16x2_t(
                     vreinterpretq_s8_u8(vandq_u8(q4bits.0, m4b)),
                     vreinterpretq_s8_u8(vandq_u8(q4bits.1, m4b)),
                 );
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q4bytes.0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q4bytes.0), vget_high_s8(q8bytes.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q4bytes.1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q4bytes.1), vget_high_s8(q8bytes.1)),
-                );
-                sumi1 += vaddvq_s16(vaddq_s16(p0, p1)) as i32 * scales[2 * j] as i32;
+                let p0 = vdotq_s32(q4bytes.0, q8bytes.0);
+                let p1 = vdotq_s32(q4bytes.1, q8bytes.1);
+                sumi1 += vaddvq_s32(vaddq_s32(p0, p1)) * scales[2 * j] as i32;
 
                 let q8bytes = vld1q_s8_x2(q8);
                 q8 = q8.add(32);
@@ -440,15 +384,9 @@ pub(crate) fn vec_dot_q4k_q8k(n: usize, xs: &[BlockQ4K], ys: &[BlockQ8K]) -> Res
                     vreinterpretq_s8_u8(vshrq_n_u8(q4bits.0, 4)),
                     vreinterpretq_s8_u8(vshrq_n_u8(q4bits.1, 4)),
                 );
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q4bytes.0), vget_low_s8(q8bytes.0)),
-                    vmull_s8(vget_high_s8(q4bytes.0), vget_high_s8(q8bytes.0)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q4bytes.1), vget_low_s8(q8bytes.1)),
-                    vmull_s8(vget_high_s8(q4bytes.1), vget_high_s8(q8bytes.1)),
-                );
-                sumi2 += vaddvq_s16(vaddq_s16(p2, p3)) as i32 * scales[2 * j + 1] as i32;
+                let p2 = vdotq_s32(q4bytes.0, q8bytes.0);
+                let p3 = vdotq_s32(q4bytes.1, q8bytes.1);
+                sumi2 += vaddvq_s32(vaddq_s32(p2, p3)) * scales[2 * j + 1] as i32;
             }
             sumf += d * (sumi1 + sumi2) as f32;
         }
@@ -526,27 +464,14 @@ pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Res
                     vreinterpretq_s8_u8(q3h_3),
                 );
 
-                // TODO: dotprod
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_0), vget_low_s8(q8bytes_1.0)),
-                    vmull_s8(vget_high_s8(q3bytes_0), vget_high_s8(q8bytes_1.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_1), vget_low_s8(q8bytes_1.1)),
-                    vmull_s8(vget_high_s8(q3bytes_1), vget_high_s8(q8bytes_1.1)),
-                );
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_2), vget_low_s8(q8bytes_1.2)),
-                    vmull_s8(vget_high_s8(q3bytes_2), vget_high_s8(q8bytes_1.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_3), vget_low_s8(q8bytes_1.3)),
-                    vmull_s8(vget_high_s8(q3bytes_3), vget_high_s8(q8bytes_1.3)),
-                );
-                isum += vaddvq_s16(p0) as i32 * *scale as i32
-                    + vaddvq_s16(p1) as i32 * *scale.add(1) as i32
-                    + vaddvq_s16(p2) as i32 * *scale.add(2) as i32
-                    + vaddvq_s16(p3) as i32 * *scale.add(3) as i32;
+                let p0 = vdotq_s32(q3bytes_0, q8bytes_1.0);
+                let p1 = vdotq_s32(q3bytes_1, q8bytes_1.1);
+                let p2 = vdotq_s32(q3bytes_2, q8bytes_1.2);
+                let p3 = vdotq_s32(q3bytes_3, q8bytes_1.3);
+                isum += vaddvq_s32(p0) * *scale as i32
+                    + vaddvq_s32(p1) * *scale.add(1) as i32
+                    + vaddvq_s32(p2) * *scale.add(2) as i32
+                    + vaddvq_s32(p3) * *scale.add(3) as i32;
                 scale = scale.add(4);
 
                 let q3h_0 = vbicq_u8(m2, qhbits.0);
@@ -571,27 +496,14 @@ pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Res
                     vreinterpretq_s8_u8(q3h_3),
                 );
 
-                // TODO: dotprod
-                let p0 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_0), vget_low_s8(q8bytes_2.0)),
-                    vmull_s8(vget_high_s8(q3bytes_0), vget_high_s8(q8bytes_2.0)),
-                );
-                let p1 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_1), vget_low_s8(q8bytes_2.1)),
-                    vmull_s8(vget_high_s8(q3bytes_1), vget_high_s8(q8bytes_2.1)),
-                );
-                let p2 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_2), vget_low_s8(q8bytes_2.2)),
-                    vmull_s8(vget_high_s8(q3bytes_2), vget_high_s8(q8bytes_2.2)),
-                );
-                let p3 = vaddq_s16(
-                    vmull_s8(vget_low_s8(q3bytes_3), vget_low_s8(q8bytes_2.3)),
-                    vmull_s8(vget_high_s8(q3bytes_3), vget_high_s8(q8bytes_2.3)),
-                );
-                isum += vaddvq_s16(p0) as i32 * *scale as i32
-                    + vaddvq_s16(p1) as i32 * *scale.add(1) as i32
-                    + vaddvq_s16(p2) as i32 * *scale.add(2) as i32
-                    + vaddvq_s16(p3) as i32 * *scale.add(3) as i32;
+                let p0 = vdotq_s32(q3bytes_0, q8bytes_2.0);
+                let p1 = vdotq_s32(q3bytes_1, q8bytes_2.1);
+                let p2 = vdotq_s32(q3bytes_2, q8bytes_2.2);
+                let p3 = vdotq_s32(q3bytes_3, q8bytes_2.3);
+                isum += vaddvq_s32(p0) * *scale as i32
+                    + vaddvq_s32(p1) * *scale.add(1) as i32
+                    + vaddvq_s32(p2) * *scale.add(2) as i32
+                    + vaddvq_s32(p3) * *scale.add(3) as i32;
                 scale = scale.add(4);
 
                 if j == 0 {
@@ -649,7 +561,6 @@ pub(crate) fn vec_dot_q2k_q8k(n: usize, xs: &[BlockQ2K], ys: &[BlockQ8K]) -> Res
             let mut is = 0usize;
 
             // TODO: dotprod
-
             for _j in 0..QK_K / 128 {
                 let q2bits = vld1q_u8_x2(q2);
                 q2 = q2.add(32);
@@ -696,14 +607,7 @@ unsafe fn multiply_accum_with_scale(
     q2bytes: int8x16x2_t,
     q8bytes: int8x16x2_t,
 ) -> i32 {
-    let p1 = vaddq_s16(
-        vmull_s8(vget_low_s8(q2bytes.0), vget_low_s8(q8bytes.0)),
-        vmull_s8(vget_high_s8(q2bytes.0), vget_high_s8(q8bytes.0)),
-    );
-    let p2 = vaddq_s16(
-        vmull_s8(vget_low_s8(q2bytes.1), vget_low_s8(q8bytes.1)),
-        vmull_s8(vget_high_s8(q2bytes.1), vget_high_s8(q8bytes.1)),
-    );
-    vaddvq_s16(p1) as i32 * aux[is + index] as i32
-        + vaddvq_s16(p2) as i32 * aux[is + 1 + index] as i32
+    let p1 = vdotq_s32(q2bytes.0, q8bytes.0);
+    let p2 = vdotq_s32(q2bytes.1, q8bytes.1);
+    vaddvq_s32(p1) * aux[is + index] as i32 + vaddvq_s32(p2) * aux[is + 1 + index] as i32
 }

candle-core/src/safetensors.rs

@@ -349,6 +349,30 @@ 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/shape.rs

@@ -171,7 +171,7 @@ impl Shape {
         }
         let mut acc = 1;
         for (&stride, &dim) in stride.iter().zip(self.0.iter()).rev() {
-            if stride != acc {
+            if dim > 1 && stride != acc {
                 return false;
             }
             acc *= dim;
@@ -186,7 +186,7 @@ impl Shape {
         }
         let mut acc = 1;
         for (&stride, &dim) in stride.iter().zip(self.0.iter()) {
-            if stride != acc {
+            if dim > 1 && stride != acc {
                 return false;
             }
             acc *= dim;
@@ -478,23 +478,6 @@ extract_dims!(
     (usize, usize, usize, usize, usize)
 );
 
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn stride() {
-        let shape = Shape::from(());
-        assert_eq!(shape.stride_contiguous(), Vec::<usize>::new());
-        let shape = Shape::from(42);
-        assert_eq!(shape.stride_contiguous(), [1]);
-        let shape = Shape::from((42, 1337));
-        assert_eq!(shape.stride_contiguous(), [1337, 1]);
-        let shape = Shape::from((299, 792, 458));
-        assert_eq!(shape.stride_contiguous(), [458 * 792, 458, 1]);
-    }
-}
-
 pub trait ShapeWithOneHole {
     fn into_shape(self, el_count: usize) -> Result<Shape>;
 }
@@ -627,3 +610,20 @@ impl ShapeWithOneHole for (usize, usize, usize, usize, ()) {
         Ok((d1, d2, d3, d4, d).into())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn stride() {
+        let shape = Shape::from(());
+        assert_eq!(shape.stride_contiguous(), Vec::<usize>::new());
+        let shape = Shape::from(42);
+        assert_eq!(shape.stride_contiguous(), [1]);
+        let shape = Shape::from((42, 1337));
+        assert_eq!(shape.stride_contiguous(), [1337, 1]);
+        let shape = Shape::from((299, 792, 458));
+        assert_eq!(shape.stride_contiguous(), [458 * 792, 458, 1]);
+    }
+}

candle-core/src/sort.rs

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

candle-core/src/storage.rs

@@ -1,6 +1,7 @@
 use crate::backend::BackendStorage;
-use crate::op::{self, CmpOp, CustomOp1, CustomOp2, CustomOp3, ReduceOp};
-use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, Result, Shape};
+use crate::op::{self, CmpOp, ReduceOp};
+use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, MetalStorage, Result, Shape};
+use crate::{CustomOp1, CustomOp2, CustomOp3, InplaceOp1, InplaceOp2, InplaceOp3};
 
 // We do not want to implement Clone on Storage as cloning may fail because of
 // out of memory. Instead try_clone should be used.
@@ -8,6 +9,7 @@ use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, Result, Shape
 pub enum Storage {
     Cpu(CpuStorage),
     Cuda(CudaStorage),
+    Metal(MetalStorage),
 }
 
 impl Storage {
@@ -18,6 +20,10 @@ impl Storage {
                 let storage = storage.try_clone(layout)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.try_clone(layout)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -25,6 +31,7 @@ impl Storage {
         match self {
             Self::Cpu(_) => Device::Cpu,
             Self::Cuda(storage) => Device::Cuda(storage.device().clone()),
+            Self::Metal(storage) => Device::Metal(storage.device().clone()),
         }
     }
 
@@ -32,13 +39,24 @@ impl Storage {
         match self {
             Self::Cpu(storage) => storage.dtype(),
             Self::Cuda(storage) => storage.dtype(),
+            Self::Metal(storage) => storage.dtype(),
         }
     }
 
     pub(crate) fn same_device(&self, rhs: &Self, op: &'static str) -> Result<()> {
-        let lhs = self.device().location();
-        let rhs = rhs.device().location();
-        if lhs != rhs {
+        let lhs_device = self.device();
+        let rhs_device = rhs.device();
+        let lhs = lhs_device.location();
+        let rhs = rhs_device.location();
+        let same_device = if self.device().is_metal() {
+            // On metal, we require the device to be exactly the same rather than
+            // having the same location. In cuda this is not necessary as all CudaDevice on the
+            // same GPU will use the same cuda stream.
+            lhs_device.same_device(&rhs_device)
+        } else {
+            lhs == rhs
+        };
+        if !same_device {
             Err(Error::DeviceMismatchBinaryOp { lhs, rhs, op }.bt())
         } else {
             Ok(())
@@ -65,6 +83,10 @@ impl Storage {
                 let storage = storage.affine(layout, mul, add)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.affine(layout, mul, add)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -78,6 +100,10 @@ impl Storage {
                 let storage = storage.powf(layout, alpha)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.powf(layout, alpha)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -91,6 +117,10 @@ impl Storage {
                 let storage = storage.elu(layout, alpha)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.elu(layout, alpha)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -112,6 +142,10 @@ impl Storage {
                 let storage = lhs.cmp(op, rhs, lhs_layout, rhs_layout)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(lhs), Self::Metal(rhs)) => {
+                let storage = lhs.cmp(op, rhs, lhs_layout, rhs_layout)?;
+                Ok(Self::Metal(storage))
+            }
             (lhs, rhs) => {
                 // Should not happen because of the same device check above but we're defensive
                 // anyway.
@@ -135,6 +169,10 @@ impl Storage {
                 let storage = storage.reduce_op(op, layout, s)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.reduce_op(op, layout, s)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -148,6 +186,10 @@ impl Storage {
                 let storage = storage.to_dtype(layout, dtype)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.to_dtype(layout, dtype)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -161,6 +203,10 @@ impl Storage {
                 let (storage, shape) = c.cuda_fwd(storage, l)?;
                 Ok((Self::Cuda(storage), shape))
             }
+            Self::Metal(storage) => {
+                let (storage, shape) = c.metal_fwd(storage, l)?;
+                Ok((Self::Metal(storage), shape))
+            }
         }
     }
 
@@ -181,6 +227,10 @@ impl Storage {
                 let (s, shape) = c.cuda_fwd(s1, l1, s2, l2)?;
                 Ok((Self::Cuda(s), shape))
             }
+            (Self::Metal(s1), Self::Metal(s2)) => {
+                let (s, shape) = c.metal_fwd(s1, l1, s2, l2)?;
+                Ok((Self::Metal(s), shape))
+            }
             _ => unreachable!(),
         }
     }
@@ -205,6 +255,55 @@ impl Storage {
                 let (s, shape) = c.cuda_fwd(s1, l1, s2, l2, s3, l3)?;
                 Ok((Self::Cuda(s), shape))
             }
+            (Self::Metal(s1), Self::Metal(s2), Self::Metal(s3)) => {
+                let (s, shape) = c.metal_fwd(s1, l1, s2, l2, s3, l3)?;
+                Ok((Self::Metal(s), shape))
+            }
+            _ => unreachable!(),
+        }
+    }
+
+    pub(crate) fn inplace_op1(&mut self, l: &Layout, c: &dyn InplaceOp1) -> Result<()> {
+        match self {
+            Self::Cpu(storage) => c.cpu_fwd(storage, l),
+            Self::Cuda(storage) => c.cuda_fwd(storage, l),
+            Self::Metal(storage) => c.metal_fwd(storage, l),
+        }
+    }
+
+    pub(crate) fn inplace_op2(
+        &mut self,
+        l1: &Layout,
+        t2: &Self,
+        l2: &Layout,
+        c: &dyn InplaceOp2,
+    ) -> Result<()> {
+        self.same_device(t2, c.name())?;
+        match (self, t2) {
+            (Self::Cpu(s1), Self::Cpu(s2)) => c.cpu_fwd(s1, l1, s2, l2),
+            (Self::Cuda(s1), Self::Cuda(s2)) => c.cuda_fwd(s1, l1, s2, l2),
+            (Self::Metal(s1), Self::Metal(s2)) => c.metal_fwd(s1, l1, s2, l2),
+            _ => unreachable!(),
+        }
+    }
+
+    pub(crate) fn inplace_op3(
+        &mut self,
+        l1: &Layout,
+        t2: &Self,
+        l2: &Layout,
+        t3: &Self,
+        l3: &Layout,
+        c: &dyn InplaceOp3,
+    ) -> Result<()> {
+        self.same_device(t2, c.name())?;
+        self.same_device(t3, c.name())?;
+        match (self, t2, t3) {
+            (Self::Cpu(s1), Self::Cpu(s2), Self::Cpu(s3)) => c.cpu_fwd(s1, l1, s2, l2, s3, l3),
+            (Self::Cuda(s1), Self::Cuda(s2), Self::Cuda(s3)) => c.cuda_fwd(s1, l1, s2, l2, s3, l3),
+            (Self::Metal(s1), Self::Metal(s2), Self::Metal(s3)) => {
+                c.metal_fwd(s1, l1, s2, l2, s3, l3)
+            }
             _ => unreachable!(),
         }
     }
@@ -219,6 +318,10 @@ impl Storage {
                 let storage = storage.unary_impl::<B>(layout)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.unary_impl::<B>(layout)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -239,6 +342,10 @@ impl Storage {
                 let storage = lhs.binary_impl::<B>(rhs, lhs_layout, rhs_layout)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(lhs), Self::Metal(rhs)) => {
+                let storage = lhs.binary_impl::<B>(rhs, lhs_layout, rhs_layout)?;
+                Ok(Self::Metal(storage))
+            }
             (lhs, rhs) => {
                 // Should not happen because of the same device check above but we're defensive
                 // anyway.
@@ -270,6 +377,10 @@ impl Storage {
                 let s = inp.conv1d(l, kernel, kernel_l, params)?;
                 Ok(Self::Cuda(s))
             }
+            (Storage::Metal(inp), Storage::Metal(kernel)) => {
+                let s = inp.conv1d(l, kernel, kernel_l, params)?;
+                Ok(Self::Metal(s))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -297,6 +408,10 @@ impl Storage {
                 let s = inp.conv_transpose1d(l, kernel, kernel_l, params)?;
                 Ok(Self::Cuda(s))
             }
+            (Storage::Metal(inp), Storage::Metal(kernel)) => {
+                let s = inp.conv_transpose1d(l, kernel, kernel_l, params)?;
+                Ok(Self::Metal(s))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -324,6 +439,10 @@ impl Storage {
                 let s = inp.conv2d(l, kernel, kernel_l, params)?;
                 Ok(Self::Cuda(s))
             }
+            (Storage::Metal(inp), Storage::Metal(kernel)) => {
+                let s = inp.conv2d(l, kernel, kernel_l, params)?;
+                Ok(Self::Metal(s))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -351,6 +470,10 @@ impl Storage {
                 let s = inp.conv_transpose2d(l, kernel, kernel_l, params)?;
                 Ok(Self::Cuda(s))
             }
+            (Storage::Metal(inp), Storage::Metal(kernel)) => {
+                let s = inp.conv_transpose2d(l, kernel, kernel_l, params)?;
+                Ok(Self::Metal(s))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -375,6 +498,10 @@ impl Storage {
                 let storage = storage.avg_pool2d(layout, kernel_size, stride)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.avg_pool2d(layout, kernel_size, stride)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -393,6 +520,10 @@ impl Storage {
                 let storage = storage.max_pool2d(layout, kernel_size, stride)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.max_pool2d(layout, kernel_size, stride)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -406,6 +537,10 @@ impl Storage {
                 let storage = storage.upsample_nearest1d(layout, sz)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.upsample_nearest1d(layout, sz)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -419,6 +554,10 @@ impl Storage {
                 let storage = storage.upsample_nearest2d(layout, h, w)?;
                 Ok(Self::Cuda(storage))
             }
+            Self::Metal(storage) => {
+                let storage = storage.upsample_nearest2d(layout, h, w)?;
+                Ok(Self::Metal(storage))
+            }
         }
     }
 
@@ -442,6 +581,10 @@ impl Storage {
                 let storage = cond.where_cond(layout, t, layout_t, f, layout_f)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(cond), Self::Metal(t), Self::Metal(f)) => {
+                let storage = cond.where_cond(layout, t, layout_t, f, layout_f)?;
+                Ok(Self::Metal(storage))
+            }
             (_, lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -468,6 +611,10 @@ impl Storage {
                 let storage = s.gather(l, indexes, indexes_l, d)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(s), Self::Metal(indexes)) => {
+                let storage = s.gather(l, indexes, indexes_l, d)?;
+                Ok(Self::Metal(storage))
+            }
             _ => unreachable!(),
         }
     }
@@ -492,6 +639,10 @@ impl Storage {
                 let storage = s.scatter_add(l, indexes, indexes_l, source, source_l, d)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(s), Self::Metal(indexes), Self::Metal(source)) => {
+                let storage = s.scatter_add(l, indexes, indexes_l, source, source_l, d)?;
+                Ok(Self::Metal(storage))
+            }
             _ => unreachable!(),
         }
     }
@@ -516,6 +667,10 @@ impl Storage {
                 let storage = s.index_add(l, indexes, indexes_l, source, source_l, d)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(s), Self::Metal(indexes), Self::Metal(source)) => {
+                let storage = s.index_add(l, indexes, indexes_l, source, source_l, d)?;
+                Ok(Self::Metal(storage))
+            }
             _ => unreachable!(),
         }
     }
@@ -537,6 +692,10 @@ impl Storage {
                 let storage = lhs.index_select(rhs, lhs_l, rhs_l, d)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(lhs), Self::Metal(rhs)) => {
+                let storage = lhs.index_select(rhs, lhs_l, rhs_l, d)?;
+                Ok(Self::Metal(storage))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -564,6 +723,10 @@ impl Storage {
                 let storage = lhs.matmul(rhs, bmnk, lhs_layout, rhs_layout)?;
                 Ok(Self::Cuda(storage))
             }
+            (Self::Metal(lhs), Self::Metal(rhs)) => {
+                let storage = lhs.matmul(rhs, bmnk, lhs_layout, rhs_layout)?;
+                Ok(Self::Metal(storage))
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -583,6 +746,9 @@ impl Storage {
         match (self, dst) {
             (Self::Cpu(src), Self::Cpu(dst)) => src.copy_strided_src(dst, dst_offset, src_l),
             (Self::Cuda(src), Self::Cuda(dst)) => Ok(src.copy_strided_src(dst, dst_offset, src_l)?),
+            (Self::Metal(src), Self::Metal(dst)) => {
+                Ok(src.copy_strided_src(dst, dst_offset, src_l)?)
+            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -591,4 +757,32 @@ impl Storage {
             .bt()),
         }
     }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(crate) fn copy2d(
+        &self,
+        dst: &mut Self,
+        d1: usize,
+        d2: usize,
+        src_s: usize,
+        dst_s: usize,
+        src_o: usize,
+        dst_o: usize,
+    ) -> Result<()> {
+        match (self, dst) {
+            (Self::Cpu(src), Self::Cpu(dst)) => src.copy2d(dst, d1, d2, src_s, dst_s, src_o, dst_o),
+            (Self::Cuda(src), Self::Cuda(dst)) => {
+                Ok(src.copy2d(dst, d1, d2, src_s, dst_s, src_o, dst_o)?)
+            }
+            (Self::Metal(src), Self::Metal(dst)) => {
+                Ok(src.copy2d(dst, d1, d2, src_s, dst_s, src_o, dst_o)?)
+            }
+            (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
+                lhs: lhs.device().location(),
+                rhs: rhs.device().location(),
+                op: "copy2d",
+            }
+            .bt()),
+        }
+    }
 }

candle-core/src/tensor.rs

@@ -1,12 +1,10 @@
-//! Tensors are N-dimenional matrixes of elements using a single data type.
+//! Tensors are N-dimensional matrixes of elements using a single data type.
 #![allow(clippy::redundant_closure_call)]
 use crate::backend::{BackendDevice, BackendStorage};
-use crate::op::{
-    BackpropOp, BinaryOp, CmpOp, CustomOp1, CustomOp2, CustomOp3, Op, ReduceOp, UnaryOp,
-};
+use crate::op::{BackpropOp, BinaryOp, CmpOp, Op, ReduceOp, UnaryOp};
 use crate::scalar::TensorOrScalar;
 use crate::shape::{Dim, Dims};
-use crate::{storage::Storage, DType, Device, Error, Layout, Result, Shape};
+use crate::{bail, storage::Storage, DType, Device, Error, Layout, Result, Shape};
 use std::sync::{Arc, RwLock};
 
 /// Unique identifier for tensors.
@@ -81,6 +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())?;
@@ -94,6 +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::$op_name>(
                 &*rhs.storage(),
                 self.layout(),
@@ -116,6 +120,9 @@ macro_rules! binary_op_scalar {
                     .broadcast_as(self.shape())?,
             };
             let shape = self.same_shape_binary_op(&rhs, stringify!($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(),
@@ -361,6 +368,16 @@ impl Tensor {
         Self::new_impl(array, shape, device, false)
     }
 
+    /// Returns a new tensor with all the elements having the same specified value. Note that
+    /// the tensor is not contiguous so you would have to call `.contiguous()` on it if needed.
+    pub fn full<D: crate::WithDType, S: Into<Shape>>(
+        value: D,
+        shape: S,
+        device: &Device,
+    ) -> Result<Self> {
+        Self::from_vec_impl(vec![value], (), device, false)?.broadcast_as(shape)
+    }
+
     /// Creates a new 1D tensor from an iterator.
     pub fn from_iter<D: crate::WithDType>(
         iter: impl IntoIterator<Item = D>,
@@ -386,7 +403,7 @@ impl Tensor {
         device: &Device,
     ) -> Result<Self> {
         if D::is_zero(&step) {
-            crate::bail!("step cannot be zero")
+            bail!("step cannot be zero")
         }
         let mut data = vec![];
         let mut current = start;
@@ -439,7 +456,15 @@ impl Tensor {
         shape: S,
         device: &Device,
     ) -> Result<Self> {
-        Self::new_impl(array, shape.into(), device, false)
+        let shape = shape.into();
+        let n: usize = shape.elem_count();
+        let buffer_size: usize = array.len();
+        if buffer_size != n {
+            return Err(Error::ShapeMismatch { buffer_size, shape }.bt());
+        }
+        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> {
@@ -498,9 +523,11 @@ impl Tensor {
     unary_op!(gelu_erf, GeluErf);
     unary_op!(erf, Erf);
     unary_op!(relu, Relu);
+    unary_op!(silu, Silu);
     unary_op!(ceil, Ceil);
     unary_op!(floor, Floor);
     unary_op!(round, Round);
+    unary_op!(sign, Sign);
 
     /// Round element of the input tensor to the nearest integer.
     ///
@@ -529,6 +556,7 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(cpu_storage) => from_cpu_storage(cpu_storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
+            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -635,6 +663,9 @@ 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))
@@ -642,6 +673,9 @@ 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))
@@ -649,12 +683,15 @@ 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))
     }
 
-    fn check_dim(&self, dim: usize, op: &'static str) -> Result<()> {
+    pub(crate) fn check_dim(&self, dim: usize, op: &'static str) -> Result<()> {
         if dim >= self.dims().len() {
             Err(Error::DimOutOfRange {
                 shape: self.shape().clone(),
@@ -668,7 +705,7 @@ impl Tensor {
     }
 
     /// Split a tensor into the specified number of chunks, this may return less chunks than
-    /// specificed.
+    /// specified.
     pub fn chunk<D: Dim>(&self, chunks: usize, dim: D) -> Result<Vec<Self>> {
         let dim = dim.to_index(self.shape(), "chunk")?;
         let size = self.dim(dim)?;
@@ -793,6 +830,35 @@ impl Tensor {
         }
     }
 
+    /// Roll the tensor input along the given dimension.
+    /// Elements that are shifted beyond the last position are re-introduced at the first position.
+    ///
+    /// ```rust
+    /// # use candle_core::{Tensor, Device};
+    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
+    /// let tensor = tensor.roll(1, 0)?;
+    /// assert_eq!(tensor.to_vec2::<f32>()?, &[[4., 5.], [0., 1.], [2., 3.]]);
+    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
+    /// let tensor = tensor.roll(-1, 0)?;
+    /// assert_eq!(tensor.to_vec2::<f32>()?, &[[2., 3.], [4., 5.], [0., 1.]]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn roll<D>(&self, shift: i32, dim: D) -> Result<Self>
+    where
+        D: Dim + Clone,
+    {
+        let dim = dim.to_index(self.shape(), "roll")?;
+        let dim_size = self.dim(dim)?;
+        let shift = shift.rem_euclid(dim_size as i32) as usize;
+        if shift == 0 {
+            Ok(self.clone())
+        } else {
+            let a = self.narrow(dim, 0, dim_size - shift)?;
+            let b = self.narrow(dim, dim_size - shift, shift)?;
+            Tensor::cat(&[&b, &a], dim)
+        }
+    }
+
     /// Returns the sum of all elements in the input tensor. The sum is performed over all the
     /// input dimensions.
     ///
@@ -855,6 +921,20 @@ impl Tensor {
         self.sum_impl(mean_dims, false)? * scale
     }
 
+    /// Returns the unbiased variance over the selected dimension.
+    pub fn var_keepdim<D: Dim>(&self, dim: D) -> Result<Self> {
+        let dim = dim.to_index(self.shape(), "var")?;
+        let mean = self.mean_keepdim(dim)?;
+        let squares = self.broadcast_sub(&mean)?.sqr()?;
+        squares.sum_impl(dim, true)? / (self.dim(dim)? - 1) as f64
+    }
+
+    /// Returns the unbiased variance over the selected dimension.
+    pub fn var<D: Dim>(&self, dim: D) -> Result<Self> {
+        let dim = dim.to_index(self.shape(), "var")?;
+        self.var_keepdim(dim)?.squeeze(dim)
+    }
+
     /// Gathers the maximum value across the selected dimension. The resulting shape has the same
     /// number of dimensions as the original tensor and the select dimension has a single element.
     pub fn max_keepdim<D: Dim>(&self, dim: D) -> Result<Self> {
@@ -960,7 +1040,7 @@ impl Tensor {
     /// tensor also has three dimensions, `(batch, channels, target_size)`.
     pub fn interpolate1d(&self, target_size: usize) -> Result<Self> {
         let (n, c, _l) = self.dims3()?;
-        let op = BackpropOp::new1(self, Op::UpsampleNearest1D);
+        let op = BackpropOp::new1(self, |arg| Op::UpsampleNearest1D { arg, target_size });
         let storage = self
             .storage()
             .upsample_nearest1d(self.layout(), target_size)?;
@@ -979,7 +1059,11 @@ impl Tensor {
     /// tensor also has four dimensions, `(batch, channels, target_h, target_w)`.
     pub fn interpolate2d(&self, target_h: usize, target_w: usize) -> Result<Self> {
         let (n, c, _h, _w) = self.dims4()?;
-        let op = BackpropOp::new1(self, Op::UpsampleNearest2D);
+        let op = BackpropOp::new1(self, |arg| Op::UpsampleNearest2D {
+            arg,
+            target_h,
+            target_w,
+        });
         let storage = self
             .storage()
             .upsample_nearest2d(self.layout(), target_h, target_w)?;
@@ -1012,6 +1096,9 @@ impl Tensor {
         let kernel_size = kernel_size.to_usize2();
         let stride = stride.to_usize2();
         let (n, c, h, w) = self.dims4()?;
+        if h < kernel_size.0 || w < kernel_size.1 {
+            bail!("kernel-size {kernel_size:?} is larger than the input size {h},{w}")
+        }
         // https://pytorch.org/docs/stable/generated/torch.nn.AvgPool2d.html#torch.nn.AvgPool2d
         let h_out = (h - kernel_size.0) / stride.0 + 1;
         let w_out = (w - kernel_size.1) / stride.1 + 1;
@@ -1047,6 +1134,9 @@ impl Tensor {
         let kernel_size = kernel_size.to_usize2();
         let stride = stride.to_usize2();
         let (n, c, h, w) = self.dims4()?;
+        if h < kernel_size.0 || w < kernel_size.1 {
+            bail!("kernel-size {kernel_size:?} is larger than the input size {h},{w}")
+        }
         // https://pytorch.org/docs/stable/generated/torch.nn.MaxPool2d.html#torch.nn.MaxPool2d
         let h_out = (h - kernel_size.0) / stride.0 + 1;
         let w_out = (w - kernel_size.1) / stride.1 + 1;
@@ -1090,6 +1180,9 @@ 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 {
@@ -1286,7 +1379,7 @@ impl Tensor {
             }
             .bt())?
         }
-        let mut storage = self.device().zeros(self.shape(), self.dtype())?;
+        let mut storage = unsafe { self.device().alloc_uninit(self.shape(), self.dtype())? };
         self.storage()
             .copy_strided_src(&mut storage, 0, self.layout())?;
         let offset = start * src.dims()[1..].iter().product::<usize>();
@@ -1454,6 +1547,7 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(storage) => from_cpu_storage(storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
+            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -1484,6 +1578,7 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(storage) => from_cpu_storage(storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
+            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -1524,6 +1619,7 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(storage) => from_cpu_storage(storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
+            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -1766,7 +1862,7 @@ impl Tensor {
         let is_permutation =
             dims.len() == self.rank() && (0..dims.len()).all(|i| dims.contains(&i));
         if !is_permutation {
-            crate::bail!(
+            bail!(
                 "dimension mismatch in permute, tensor {:?}, dims: {:?}",
                 self.dims(),
                 dims
@@ -1815,9 +1911,9 @@ impl Tensor {
     /// this new node. The storage of this tensor is shared with the initial tensor.
     ///
     /// If the tensor is already detached from the computation graph, the same tensor is returned.
-    pub fn detach(&self) -> Result<Tensor> {
+    pub fn detach(&self) -> Tensor {
         if self.op.is_none() && !self.is_variable {
-            Ok(self.clone())
+            self.clone()
         } else {
             let tensor_ = Tensor_ {
                 id: TensorId::new(),
@@ -1828,7 +1924,7 @@ impl Tensor {
                 dtype: self.dtype,
                 device: self.device.clone(),
             };
-            Ok(Tensor(Arc::new(tensor_)))
+            Tensor(Arc::new(tensor_))
         }
     }
 
@@ -1841,7 +1937,11 @@ impl Tensor {
                 (Storage::Cpu(storage), Device::Cuda(cuda)) => {
                     Storage::Cuda(cuda.storage_from_cpu_storage(storage)?)
                 }
+                (Storage::Cpu(storage), Device::Metal(metal)) => {
+                    Storage::Metal(metal.storage_from_cpu_storage(storage)?)
+                }
                 (Storage::Cuda(storage), Device::Cpu) => Storage::Cpu(storage.to_cpu_storage()?),
+                (Storage::Metal(storage), Device::Cpu) => Storage::Cpu(storage.to_cpu_storage()?),
                 (Storage::Cuda(storage), Device::Cuda(cuda)) => {
                     // TODO: Avoid passing through the cpu storage here, especially if the gpu ids
                     // are the same.
@@ -1849,6 +1949,9 @@ impl Tensor {
                     Storage::Cuda(cuda.storage_from_cpu_storage(&cpu_storage)?)
                 }
                 (Storage::Cpu(storage), Device::Cpu) => Storage::Cpu(storage.clone()),
+                _ => {
+                    bail!("not implemented yet")
+                }
             };
             let op = BackpropOp::new1(self, Op::ToDevice);
             let tensor_ = Tensor_ {
@@ -1926,7 +2029,7 @@ impl Tensor {
             Ok(self.clone())
         } else {
             let shape = self.shape();
-            let mut storage = self.device().zeros(shape, self.dtype())?;
+            let mut storage = unsafe { self.device().alloc_uninit(shape, self.dtype())? };
             self.storage()
                 .copy_strided_src(&mut storage, 0, self.layout())?;
             let op = BackpropOp::new1(self, Op::Copy);
@@ -1934,11 +2037,21 @@ impl Tensor {
         }
     }
 
+    /// Returns a tensor that is in row major order. This always makes a copy.
+    pub fn force_contiguous(&self) -> Result<Tensor> {
+        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 op = BackpropOp::new1(self, Op::Copy);
+        Ok(from_storage(storage, shape.clone(), op, false))
+    }
+
     /// Create a variable based on the values currently stored in a tensor. The storage is always
     /// copied.
     pub(crate) fn make_var(&self) -> Result<Tensor> {
         let shape = self.shape().clone();
-        let mut storage = self.device().zeros(&shape, self.dtype())?;
+        let mut storage = unsafe { self.device().alloc_uninit(&shape, self.dtype())? };
         self.storage()
             .copy_strided_src(&mut storage, 0, self.layout())?;
         Ok(from_storage(storage, shape, BackpropOp::none(), true))
@@ -1991,7 +2104,7 @@ impl Tensor {
             };
             Ok(Tensor(Arc::new(tensor_)))
         } else {
-            let mut storage = self.device().zeros(&shape, self.dtype())?;
+            let mut storage = unsafe { self.device().alloc_uninit(&shape, self.dtype())? };
             self.storage()
                 .copy_strided_src(&mut storage, 0, self.layout())?;
             Ok(from_storage(storage, shape, op, false))
@@ -2018,8 +2131,19 @@ impl Tensor {
         let dim = dim.to_index(self.shape(), "squeeze")?;
         if dims[dim] == 1 {
             let mut dims = dims.to_vec();
+            let mut strides = self.stride().to_vec();
             dims.remove(dim);
-            self.reshape(dims)
+            strides.remove(dim);
+            let tensor_ = Tensor_ {
+                id: TensorId::new(),
+                storage: self.storage.clone(),
+                layout: Layout::new(dims.into(), strides, self.layout.start_offset()),
+                op: BackpropOp::new1(self, Op::Reshape),
+                is_variable: false,
+                dtype: self.dtype,
+                device: self.device.clone(),
+            };
+            Ok(Tensor(Arc::new(tensor_)))
         } else {
             Ok(self.clone())
         }
@@ -2040,10 +2164,24 @@ impl Tensor {
     /// ```
     pub fn unsqueeze<D: Dim>(&self, dim: D) -> Result<Self> {
         let mut dims = self.dims().to_vec();
+        let mut strides = self.stride().to_vec();
         let dim = dim.to_index_plus_one(self.shape(), "unsqueeze")?;
         // Cannot panic because to_index_plus_one already checks dimensions
         dims.insert(dim, 1);
-        self.reshape(dims)
+        // Any stride would work here, but we pick one so as to maximize the probability to remain
+        // C contiguous.
+        let stride = if dim < strides.len() { strides[dim] } else { 1 };
+        strides.insert(dim, stride);
+        let tensor_ = Tensor_ {
+            id: TensorId::new(),
+            storage: self.storage.clone(),
+            layout: Layout::new(dims.into(), strides, self.layout.start_offset()),
+            op: BackpropOp::new1(self, Op::Reshape),
+            is_variable: false,
+            dtype: self.dtype,
+            device: self.device.clone(),
+        };
+        Ok(Tensor(Arc::new(tensor_)))
     }
 
     /// Stacks two or more tensors along a particular dimension.
@@ -2074,152 +2212,6 @@ impl Tensor {
         Self::cat(&args, dim)
     }
 
-    /// Concatenates two or more tensors along a particular dimension.
-    ///
-    /// All tensors must of the same rank, and the output will have
-    /// the same rank
-    ///
-    /// ```rust
-    /// # use candle_core::{Tensor, DType, Device};
-    /// let a = Tensor::zeros((2, 3), DType::F32, &Device::Cpu)?;
-    /// let b = Tensor::zeros((2, 3), DType::F32, &Device::Cpu)?;
-    ///
-    /// let c = Tensor::cat(&[&a, &b], 0)?;
-    /// assert_eq!(c.shape().dims(), &[4, 3]);
-    ///
-    /// let c = Tensor::cat(&[&a, &b], 1)?;
-    /// assert_eq!(c.shape().dims(), &[2, 6]);
-    /// # Ok::<(), candle_core::Error>(())
-    /// ```
-    pub fn cat<A: AsRef<Tensor>, D: Dim>(args: &[A], dim: D) -> Result<Self> {
-        if args.is_empty() {
-            Err(Error::OpRequiresAtLeastOneTensor { op: "cat" }.bt())?
-        }
-        let arg0 = args[0].as_ref();
-        if args.len() == 1 {
-            return Ok(arg0.clone());
-        }
-        let dim = dim.to_index(arg0.shape(), "cat")?;
-        for arg in args {
-            arg.as_ref().check_dim(dim, "cat")?;
-        }
-        for (arg_idx, arg) in args.iter().enumerate() {
-            let arg = arg.as_ref();
-            if arg0.rank() != arg.rank() {
-                Err(Error::UnexpectedNumberOfDims {
-                    expected: arg0.rank(),
-                    got: arg.rank(),
-                    shape: arg.shape().clone(),
-                }
-                .bt())?
-            }
-            for (dim_idx, (v1, v2)) in arg0
-                .shape()
-                .dims()
-                .iter()
-                .zip(arg.shape().dims().iter())
-                .enumerate()
-            {
-                if dim_idx != dim && v1 != v2 {
-                    Err(Error::ShapeMismatchCat {
-                        dim: dim_idx,
-                        first_shape: arg0.shape().clone(),
-                        n: arg_idx + 1,
-                        nth_shape: arg.shape().clone(),
-                    }
-                    .bt())?
-                }
-            }
-        }
-        if dim == 0 {
-            Self::cat0(args)
-        } else {
-            // TODO: Avoid these transpositions and have an implementation that works
-            // for dim != 0...
-            let args: Vec<Tensor> = args
-                .iter()
-                .map(|a| a.as_ref().transpose(0, dim))
-                .collect::<Result<Vec<_>>>()?;
-            let cat = Self::cat0(&args)?;
-            cat.transpose(0, dim)
-        }
-    }
-
-    fn cat0<A: AsRef<Tensor>>(args: &[A]) -> Result<Self> {
-        if args.is_empty() {
-            Err(Error::OpRequiresAtLeastOneTensor { op: "cat" }.bt())?
-        }
-        let arg0 = args[0].as_ref();
-        if args.len() == 1 {
-            return Ok(arg0.clone());
-        }
-        let rank = arg0.rank();
-        let device = arg0.device();
-        let dtype = arg0.dtype();
-        let first_dims = arg0.shape().dims();
-        let mut cat_dims = first_dims.to_vec();
-        cat_dims[0] = 0;
-        let mut offsets = vec![0usize];
-        for (arg_idx, arg) in args.iter().enumerate() {
-            let arg = arg.as_ref();
-            if arg.dtype() != dtype {
-                Err(Error::DTypeMismatchBinaryOp {
-                    lhs: dtype,
-                    rhs: arg.dtype(),
-                    op: "cat",
-                }
-                .bt())?
-            }
-            if arg.device().location() != device.location() {
-                Err(Error::DeviceMismatchBinaryOp {
-                    lhs: device.location(),
-                    rhs: arg.device().location(),
-                    op: "cat",
-                }
-                .bt())?
-            }
-            if rank != arg.rank() {
-                Err(Error::UnexpectedNumberOfDims {
-                    expected: rank,
-                    got: arg.rank(),
-                    shape: arg.shape().clone(),
-                }
-                .bt())?
-            }
-            for (dim_idx, (v1, v2)) in arg0
-                .shape()
-                .dims()
-                .iter()
-                .zip(arg.shape().dims().iter())
-                .enumerate()
-            {
-                if dim_idx == 0 {
-                    cat_dims[0] += v2;
-                }
-                if dim_idx != 0 && v1 != v2 {
-                    Err(Error::ShapeMismatchCat {
-                        dim: dim_idx,
-                        first_shape: arg0.shape().clone(),
-                        n: arg_idx + 1,
-                        nth_shape: arg.shape().clone(),
-                    }
-                    .bt())?
-                }
-            }
-            let next_offset = offsets.last().unwrap() + arg.elem_count();
-            offsets.push(next_offset);
-        }
-        let shape = Shape::from(cat_dims);
-        let op = BackpropOp::new(args, |args| Op::Cat(args, 0));
-        let mut storage = device.zeros(&shape, dtype)?;
-        for (arg, &offset) in args.iter().zip(offsets.iter()) {
-            let arg = arg.as_ref();
-            arg.storage()
-                .copy_strided_src(&mut storage, offset, arg.layout())?;
-        }
-        Ok(from_storage(storage, shape, op, false))
-    }
-
     /// Pad the input tensor using 0s along dimension `dim`. This adds `left` elements before the
     /// input tensor values and `right` elements after.
     pub fn pad_with_zeros<D: Dim>(&self, dim: D, left: usize, right: usize) -> Result<Self> {
@@ -2254,7 +2246,7 @@ impl Tensor {
         if left == 0 && right == 0 {
             Ok(self.clone())
         } else if self.elem_count() == 0 {
-            crate::bail!("cannot use pad_with_same on an empty tensor")
+            bail!("cannot use pad_with_same on an empty tensor")
         } else if left == 0 {
             let dim = dim.to_index(self.shape(), "pad_with_same")?;
             let r = self.narrow(dim, self.dim(dim)? - 1, 1)?;
@@ -2302,6 +2294,10 @@ impl Tensor {
         self.storage.read().unwrap()
     }
 
+    pub(crate) fn storage_mut(&self) -> std::sync::RwLockWriteGuard<'_, Storage> {
+        self.storage.write().unwrap()
+    }
+
     // If we extend the visibility of this function to be usable outside of this crate, we should
     // make it unsafe.
     pub(crate) fn storage_mut_and_layout(
@@ -2323,112 +2319,141 @@ impl Tensor {
         std::ptr::eq(lhs, rhs)
     }
 
-    /// Applies a unary custom op without backward support
-    pub fn apply_op1_no_bwd<C: CustomOp1>(&self, c: &C) -> Result<Self> {
-        let (storage, shape) = self.storage().apply_op1(self.layout(), c)?;
-        Ok(from_storage(storage, shape, BackpropOp::none(), false))
-    }
-
-    /// Applies a binary custom op without backward support
-    pub fn apply_op2_no_bwd<C: CustomOp2>(&self, rhs: &Self, c: &C) -> Result<Self> {
-        let (storage, shape) =
-            self.storage()
-                .apply_op2(self.layout(), &rhs.storage(), rhs.layout(), c)?;
-        Ok(from_storage(storage, shape, BackpropOp::none(), false))
-    }
-
-    /// Applies a ternary custom op without backward support
-    pub fn apply_op3_no_bwd<C: CustomOp3>(&self, t2: &Self, t3: &Self, c: &C) -> Result<Self> {
-        let (storage, shape) = self.storage().apply_op3(
-            self.layout(),
-            &t2.storage(),
-            t2.layout(),
-            &t3.storage(),
-            t3.layout(),
-            c,
-        )?;
-        Ok(from_storage(storage, shape, BackpropOp::none(), false))
-    }
-
-    /// Applies a unary custom op.
-    pub fn apply_op1_arc(&self, c: Arc<Box<dyn CustomOp1 + Send + Sync>>) -> Result<Self> {
-        let (storage, shape) = self
-            .storage()
-            .apply_op1(self.layout(), c.as_ref().as_ref())?;
-        let op = BackpropOp::new1(self, |s| Op::CustomOp1(s, c.clone()));
-        Ok(from_storage(storage, shape, op, false))
-    }
-
-    pub fn apply_op1<C: 'static + CustomOp1 + Send + Sync>(&self, c: C) -> Result<Self> {
-        self.apply_op1_arc(Arc::new(Box::new(c)))
-    }
-
-    /// Applies a binary custom op.
-    pub fn apply_op2_arc(
-        &self,
-        rhs: &Self,
-        c: Arc<Box<dyn CustomOp2 + Send + Sync>>,
-    ) -> Result<Self> {
-        let (storage, shape) = self.storage().apply_op2(
-            self.layout(),
-            &rhs.storage(),
-            rhs.layout(),
-            c.as_ref().as_ref(),
-        )?;
-        let op = BackpropOp::new2(self, rhs, |t1, t2| Op::CustomOp2(t1, t2, c.clone()));
-        Ok(from_storage(storage, shape, op, false))
-    }
-
-    pub fn apply_op2<C: 'static + CustomOp2 + Send + Sync>(&self, r: &Self, c: C) -> Result<Self> {
-        self.apply_op2_arc(r, Arc::new(Box::new(c)))
-    }
-
-    /// Applies a ternary custom op.
-    pub fn apply_op3_arc(
-        &self,
-        t2: &Self,
-        t3: &Self,
-        c: Arc<Box<dyn CustomOp3 + Send + Sync>>,
-    ) -> Result<Self> {
-        let (storage, shape) = self.storage().apply_op3(
-            self.layout(),
-            &t2.storage(),
-            t2.layout(),
-            &t3.storage(),
-            t3.layout(),
-            c.as_ref().as_ref(),
-        )?;
-        let op = BackpropOp::new3(self, t2, t3, |t1, t2, t3| {
-            Op::CustomOp3(t1, t2, t3, c.clone())
-        });
-        Ok(from_storage(storage, shape, op, false))
-    }
-
-    pub fn apply_op3<C: 'static + CustomOp3 + Send + Sync>(
-        &self,
-        t2: &Self,
-        t3: &Self,
-        c: C,
-    ) -> Result<Self> {
-        self.apply_op3_arc(t2, t3, Arc::new(Box::new(c)))
-    }
-
     /// Normalize a 'relative' axis value: positive values are kept, negative
     /// values means counting the dimensions from the back.
     pub fn normalize_axis(&self, axis: i64) -> Result<usize> {
         let rank = self.rank() as i64;
         if rank <= axis {
-            crate::bail!("axis {axis} is too large, tensor rank {rank}")
+            bail!("axis {axis} is too large, tensor rank {rank}")
         } else if 0 <= axis {
             Ok(axis as usize)
         } else {
             let naxis = rank + axis;
             if naxis < 0 {
-                crate::bail!("axis {axis} is too small, tensor rank {rank}")
+                bail!("axis {axis} is too small, tensor rank {rank}")
             }
             Ok(naxis as usize)
         }
     }
+
+    /// Returns a lower triangular matrix of ones of size n by n.
+    pub fn tril2(n: usize, dtype: DType, device: &Device) -> Result<Self> {
+        let t = Tensor::arange(0u32, n as u32, device)?;
+        let t1 = t.reshape((1, n))?.broadcast_as((n, n))?;
+        let t2 = t.reshape((n, 1))?.broadcast_as((n, n))?;
+        t1.le(&t2)?.to_dtype(dtype)
+    }
+
+    /// Returns an upper triangular matrix of ones of size n by n.
+    pub fn triu2(n: usize, dtype: DType, device: &Device) -> Result<Self> {
+        let t = Tensor::arange(0u32, n as u32, device)?;
+        let t1 = t.reshape((1, n))?.broadcast_as((n, n))?;
+        let t2 = t.reshape((n, 1))?.broadcast_as((n, n))?;
+        t1.ge(&t2)?.to_dtype(dtype)
+    }
+
+    /// Returns a matrix with a diagonal of ones of size n by n.
+    pub fn eye(n: usize, dtype: DType, device: &Device) -> Result<Self> {
+        let t = Tensor::arange(0u32, n as u32, device)?;
+        let t1 = t.reshape((1, n))?.broadcast_as((n, n))?;
+        let t2 = t.reshape((n, 1))?.broadcast_as((n, n))?;
+        t1.eq(&t2)?.to_dtype(dtype)
+    }
+
+    /// Returns the cumulative sum of elements of the input tensor summed over the specified
+    /// dimension.
+    ///
+    /// This operation is most efficient when dim is the last dimension of the tensor.
+    pub fn cumsum<D: Dim>(&self, dim: D) -> Result<Self> {
+        let dim = dim.to_index(self.shape(), "cumsum")?;
+        let rank = self.rank();
+        if rank == 0 {
+            return Ok(self.clone());
+        }
+        let n_axis = self.dim(dim)?;
+        let triu = Tensor::triu2(n_axis, self.dtype(), self.device())?;
+        if rank == 1 {
+            self.unsqueeze(0)?.matmul(&triu)?.squeeze(0)
+        } else {
+            let last = rank - 1;
+            let t = self.transpose(dim, last)?;
+            let t = t.broadcast_matmul(&triu)?;
+            t.transpose(dim, last)
+        }
+    }
+
+    /// Returns a copy of `self` where the values within `ranges` have been replaced with the
+    /// content of `src`.
+    pub fn slice_assign<D: std::ops::RangeBounds<usize>>(
+        &self,
+        ranges: &[D],
+        src: &Tensor,
+    ) -> Result<Self> {
+        let src_dims = src.dims();
+        let self_dims = self.dims();
+        if self_dims.len() != src_dims.len() {
+            bail!(
+                "slice-assign requires input with the same rank {} <> {}",
+                self_dims.len(),
+                src_dims.len()
+            )
+        }
+        if self_dims.len() != ranges.len() {
+            bail!(
+                "slice-assign requires input with the same rank as there are ranges {} <> {}",
+                self_dims.len(),
+                ranges.len()
+            )
+        }
+        let mut src = src.clone();
+        let mut mask = Self::ones(src.shape(), DType::U8, src.device())?;
+        for (i, range) in ranges.iter().enumerate() {
+            let start_included = match range.start_bound() {
+                std::ops::Bound::Unbounded => 0,
+                std::ops::Bound::Included(v) => *v,
+                std::ops::Bound::Excluded(v) => *v + 1,
+            };
+            let end_excluded = match range.end_bound() {
+                std::ops::Bound::Unbounded => self_dims[i],
+                std::ops::Bound::Included(v) => *v + 1,
+                std::ops::Bound::Excluded(v) => *v,
+            };
+            if end_excluded <= start_included {
+                bail!("slice-assign: empty range for dim {i}, {start_included} {end_excluded}")
+            }
+            if self_dims[i] < end_excluded {
+                bail!(
+                    "slice-assign: upper bound is out of range for dim {i}, {end_excluded} {}",
+                    self_dims[i]
+                )
+            }
+            if end_excluded - start_included != src_dims[i] {
+                bail!(
+                    "slice-assign: the range for dim {i} ({start_included}..{end_excluded}) does not match the size of src {}", src_dims[i]
+                )
+            }
+            src = src.pad_with_zeros(i, start_included, self_dims[i] - end_excluded)?;
+            mask = mask.pad_with_zeros(i, start_included, self_dims[i] - end_excluded)?
+        }
+        mask.where_cond(/* on_true= */ &src, /* on_false= */ self)
+    }
+
+    /// Returns log(sum(exp(tensor), dim)).
+    pub fn log_sum_exp<D: Dims>(&self, sum_dims: D) -> Result<Self> {
+        let exp = self.exp()?;
+        let sum = exp.sum(sum_dims)?;
+        sum.log()
+    }
+
+    /// Pointwise pow operation.
+    pub fn pow(&self, rhs: &Tensor) -> Result<Self> {
+        rhs.mul(&self.log()?)?.exp()
+    }
+
+    /// Broadcasting version of `pow`.
+    pub fn broadcast_pow(&self, rhs: &Tensor) -> Result<Self> {
+        rhs.broadcast_mul(&self.log()?)?.exp()
+    }
 }
 
 macro_rules! bin_trait {

candle-core/src/tensor_cat.rs

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

candle-core/src/test_utils.rs

@@ -4,7 +4,7 @@ use crate::{Result, Tensor};
 macro_rules! test_device {
     // TODO: Switch to generating the two last arguments automatically once concat_idents is
     // stable. https://github.com/rust-lang/rust/issues/29599
-    ($fn_name: ident, $test_cpu: ident, $test_cuda: ident) => {
+    ($fn_name: ident, $test_cpu: ident, $test_cuda: ident, $test_metal: ident) => {
         #[test]
         fn $test_cpu() -> Result<()> {
             $fn_name(&Device::Cpu)
@@ -15,6 +15,12 @@ macro_rules! test_device {
         fn $test_cuda() -> Result<()> {
             $fn_name(&Device::new_cuda(0)?)
         }
+
+        #[cfg(feature = "metal")]
+        #[test]
+        fn $test_metal() -> Result<()> {
+            $fn_name(&Device::new_metal(0)?)
+        }
     };
 }
 

candle-core/src/utils.rs

@@ -23,6 +23,10 @@ pub fn cuda_is_available() -> bool {
     cfg!(feature = "cuda")
 }
 
+pub fn metal_is_available() -> bool {
+    cfg!(feature = "metal")
+}
+
 pub fn with_avx() -> bool {
     cfg!(target_feature = "avx")
 }

candle-core/src/variable.rs

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

candle-core/tests/conv_tests.rs

@@ -18,6 +18,9 @@ w_t = w.transpose(0, 1)
 res = torch.nn.functional.conv_transpose1d(t, w_t)
 print(res.shape)
 print(res)
+res = torch.nn.functional.conv_transpose1d(t, w_t, groups=2)
+print(res.shape)
+print(res)
 */
 fn conv1d(dev: &Device) -> Result<()> {
     let t = Tensor::new(
@@ -50,8 +53,11 @@ 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]
     );
-    if dev.is_cpu() {
-        let res = t.conv_transpose1d(&w.transpose(0, 1)?, 0, 0, 1, 1)?;
+
+    let w = w.transpose(0, 1)?;
+    // The CPU kernels applied in the contiguous and non contiguous cases are different.
+    for w in [w.clone(), w.contiguous()?] {
+        let res = t.conv_transpose1d(&w, 0, 0, 1, 1, 1)?;
         assert_eq!(res.dims(), [1, 2, 7]);
         assert_eq!(
             test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
@@ -60,6 +66,17 @@ fn conv1d(dev: &Device) -> Result<()> {
                 4.7076, -5.9745, -0.8276, 1.621
             ],
         );
+        let res = t.conv_transpose1d(&w, 0, 0, 1, 1, 2)?;
+        assert_eq!(res.dims(), [1, 4, 7]);
+        assert_eq!(
+            test_utils::to_vec2_round(&res.squeeze(0)?, 4)?,
+            [
+                [-1.5596, -1.8099, 2.0407, 4.8764, -0.1743, -0.735, -0.7819],
+                [0.7816, 3.8152, -0.5926, 2.2515, -5.1844, -0.3157, 1.4721],
+                [1.6295, 0.52, 6.2611, 0.7109, 2.6315, -1.8793, 0.7113],
+                [1.0949, 1.0166, 1.7464, 2.4561, -0.79, -0.5119, 0.1488]
+            ]
+        );
     }
     Ok(())
 }
@@ -118,7 +135,7 @@ fn conv2d(dev: &Device) -> Result<()> {
             0.6466, -0.5042, -0.0603, -1.6538, -1.2429, 1.8357, 1.6052, -1.3844, 0.3323, -1.3712,
             0.9634, -0.4799, -0.6451, -0.0840, -1.4247, 0.5512, -0.1747, -0.5509, -0.3742, 0.3790,
             -0.4431, -0.4720, -0.7890, 0.2620, 0.7875, 0.5377, -0.6779, -0.8088, 1.9098, 1.2006,
-            -0.8000, -0.4983, 1.5480, 0.8265, -0.1025, 0.5138, 0.5748, 0.3821, -0.4607, 0.0085,
+            -0.8, -0.4983, 1.5480, 0.8265, -0.1025, 0.5138, 0.5748, 0.3821, -0.4607, 0.0085,
         ],
         dev,
     )?;
@@ -146,7 +163,9 @@ fn conv2d(dev: &Device) -> Result<()> {
             10.389, 3.6023, -4.2808, 0.2672, 5.3646, -5.2023, -2.1955, -9.4075
         ]
     );
+
     let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 1)?;
+
     assert_eq!(res.dims(), [1, 2, 7, 7]);
     assert_eq!(
         test_utils::to_vec3_round(&res.i(0)?, 4)?,
@@ -171,6 +190,7 @@ fn conv2d(dev: &Device) -> Result<()> {
             ]
         ]
     );
+
     // Dilations.
     let res = t.conv2d(&w, 0, 1, 2, 1)?;
     assert_eq!(res.dims(), [1, 2, 1, 1]);
@@ -209,6 +229,7 @@ fn conv2d(dev: &Device) -> Result<()> {
             ]
         ]
     );
+
     Ok(())
 }
 
@@ -255,13 +276,13 @@ fn conv2d_small(dev: &Device) -> Result<()> {
     assert_eq!(
         test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
         [
-            0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
-            0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.1640, -0.0111, -0.1742, 0.0000, 0.0000,
-            0.0000, 0.0000, 2.6437, -2.0268, 1.1823, 0.0000, 0.0000, 0.0000, 0.0000, 3.2855,
-            -1.0324, 0.2539, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
-            0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000
+            0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1640,
+            -0.0111, -0.1742, 0.0, 0.0, 0.0, 0.0, 2.6437, -2.0268, 1.1823, 0.0, 0.0, 0.0, 0.0,
+            3.2855, -1.0324, 0.2539, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
+            0.0, 0.0, 0.0, 0.0
         ]
     );
+
     let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 1)?;
     assert_eq!(res.dims(), [1, 1, 3, 3]);
     assert_eq!(
@@ -363,6 +384,7 @@ print(w.grad.shape)
 print(w.grad[0])
 */
 fn conv2d_grad(dev: &Device) -> Result<()> {
+    // conv-transposes are not implemented for metal
     use candle_core::Var;
     let t = Var::from_slice(
         &[
@@ -375,7 +397,7 @@ fn conv2d_grad(dev: &Device) -> Result<()> {
             0.6466, -0.5042, -0.0603, -1.6538, -1.2429, 1.8357, 1.6052, -1.3844, 0.3323, -1.3712,
             0.9634, -0.4799, -0.6451, -0.0840, -1.4247, 0.5512, -0.1747, -0.5509, -0.3742, 0.3790,
             -0.4431, -0.4720, -0.7890, 0.2620, 0.7875, 0.5377, -0.6779, -0.8088, 1.9098, 1.2006,
-            -0.8000, -0.4983, 1.5480, 0.8265, -0.1025, 0.5138, 0.5748, 0.3821, -0.4607, 0.0085,
+            -0.8, -0.4983, 1.5480, 0.8265, -0.1025, 0.5138, 0.5748, 0.3821, -0.4607, 0.0085,
         ],
         (1, 4, 5, 5),
         dev,
@@ -560,17 +582,186 @@ fn conv2d_grad(dev: &Device) -> Result<()> {
         ]
     );
 
+    // Conv Transpose 2d Test
+    //tested against following python
+
+    // import torch
+    // torch.manual_seed(4242)
+    // padding = 4
+    // outpadding = 2
+    // dilation = 3
+    // stride = 3
+    // input = torch.randn((1, 4, 7, 5), requires_grad=True)
+    // kernel = torch.randn((4, 2, 3, 5), requires_grad=True)
+    // print("input", input.flatten())
+    // print("kernel", kernel.flatten())
+    // res = torch.nn.functional.conv_transpose2d(
+    //     input,
+    //     kernel,
+    //     stride=stride,
+    //     padding=padding,
+    //     dilation=dilation,
+    //     output_padding=outpadding,
+    // )
+    // res.retain_grad()
+    // print(res.shape)
+    // loss = (res**2).sum()
+    // print(loss)
+    // loss.backward()
+    // print(input.grad.shape)
+    // print("input grad", torch.round(input.grad, decimals=1))
+    // print(kernel.grad.shape)
+    // print("kernel grad", torch.round(kernel.grad.flatten(), decimals=1))
+
+    let padding = 4;
+    let outpadding = 2;
+    let dilation = 3;
+    let stride = 3;
+
+    let t = Var::from_slice(
+        &[
+            0.4056_f32, -0.8689, -0.0773, -1.5630, -2.8012, -1.5059, 0.3972, 1.0852, 0.4997,
+            3.0616, 1.6541, 0.0964, -0.8338, -1.6523, -0.8323, -0.1699, 0.0823, 0.3526, 0.6843,
+            0.2395, 1.2279, -0.9287, -1.7030, 0.1370, 0.6047, 0.3770, -0.6266, 0.3529, 2.2013,
+            -0.6836, 0.2477, 1.3127, -0.2260, 0.2622, -1.2974, -0.8140, -0.8404, -0.3490, 0.0130,
+            1.3123, 1.7569, -0.3956, -1.8255, 0.1727, -0.3538, 2.6941, 1.0529, 0.4219, -0.2071,
+            1.1586, 0.4717, 0.3865, -0.5690, -0.5010, -0.1310, 0.7796, 0.6630, -0.2021, 2.6090,
+            0.2049, 0.6466, -0.5042, -0.0603, -1.6538, -1.2429, 1.8357, 1.6052, -1.3844, 0.3323,
+            -1.3712, 0.9634, -0.4799, -0.6451, -0.0840, -1.4247, 0.5512, -0.1747, -0.5509, -0.3742,
+            0.3790, -0.4431, -0.4720, -0.7890, 0.2620, 0.5411, -1.1715, -2.4997, 2.3249, -0.8912,
+            -0.4733, -0.5701, -2.8888, -1.4112, -0.5471, -0.9234, -1.1660, 0.4189, -0.7465,
+            -0.6473, 0.1402, 0.7875, 0.5377, -0.6779, -0.8088, -0.4864, -0.2312, 0.9279, 0.1264,
+            1.5480, 0.8265, -0.1025, 0.5138, -0.2512, 0.1576, 1.2705, 0.3641, -0.9325, 0.6451,
+            -0.8537, 0.2378, 0.1794, 0.2752, -0.3687, -1.1149, -0.1410, -0.5829, -0.0892, 1.4258,
+            -2.2789, 0.5270, 0.1825, 1.7007, -0.5263, -0.2954, 0.4440, 0.5537, 0.3492, 0.6186,
+            1.6475, 0.2219,
+        ],
+        (1, 4, 7, 5),
+        dev,
+    )?;
+
+    #[rustfmt::skip]
+    let w = Var::from_slice(
+        &[
+            -1.1744_f32, 0.3266, 2.5893, 1.0142, 0.1763, 0.7752, 0.6604, 0.2029, -0.2145, 0.7234,
+            -0.3441, -1.5400, -0.6333, 0.6613, 0.2083, 0.6230, -1.7002, 0.3393, 0.4049, 1.0762,
+            0.2723, 1.4181, 0.0029, -0.2122, 1.7668, 1.4168, 0.3320, -0.2719, 0.7932, -0.7204,
+            0.4447, 0.1211, 0.5908, 1.0089, -0.1646, 1.8033, -0.6286, 0.2016, -0.3370, 1.2555,
+            0.8009, -0.6488, -0.4652, -1.5685, 1.5860, 0.5583, 0.4623, 0.6026, 0.8828, 2.4990,
+            0.6811, -0.3369, 1.3320, 1.7669, -1.1067, 1.2958, -0.9415, -0.9655, -0.4462, 0.7181,
+            0.5181, -1.1658, -1.8467, -0.7763, 1.2769, 0.8651, 0.9890, 1.5092, 0.7207, -0.8481,
+            0.7417, 0.3375, -1.2685, 1.4572, 1.0915, 0.1093, -0.8550, -0.5831, -0.6309, -0.2509,
+            0.5220, -0.0914, 0.7900, 0.1096, 0.3258, 0.2723, -1.0942, -0.3393, -0.1653, 0.5732,
+            -0.8014, 1.8194, -1.9023, 0.2127, 1.8636, -0.8979, 0.1927, -0.2778, 0.3105, 0.0071,
+            -1.1823, 0.2476, -0.7178, -1.3821, 1.0769, -0.4376, -0.9967, -0.1227, 1.6197, -1.0604,
+            0.1372, 0.8141, -0.6163, 0.7304, -0.8285, 2.0636, -0.7176, 0.2495, -0.2581, -0.4478,
+        ],
+        (4, 2, 3, 5),
+        dev,
+    )?;
+    let res = t.conv_transpose2d(&w, padding, outpadding, stride, dilation)?;
+    let loss = res.sqr()?.sum_all()?;
+    assert_eq!(test_utils::to_vec0_round(&loss, 0)?, 2904.0);
+    let grads = loss.backward()?;
+
+    let grad_t = grads.get(&t).unwrap();
+    let grad_w = grads.get(&w).unwrap();
+    assert_eq!(grad_t.dims(), [1, 4, 7, 5]);
+    assert_eq!(grad_w.dims(), [4, 2, 3, 5]);
+
+    assert_eq!(
+        test_utils::to_vec1_round(&grad_w.flatten_all()?, 1)?,
+        [
+            // torch gets 89.1
+            -89.0, -135.3, 136.7, 102.0, -53.4, 117.9, 118.6, -43.9, -218.0, -58.5, -114.3, -150.0,
+            -15.6, 172.1, 66.3, -64.3, -27.9, -19.8, 31.7, 62.1, 5.5, 92.6, 28.2, -29.6, 55.9,
+            52.7, -72.7, -119.8, 53.8, -25.5, 128.8, 19.3, 68.0, 190.9, -64.1, -86.2, -111.2,
+            106.6, -67.7, 37.8, 115.9, 50.4, -77.7, -54.9, 22.3, -4.6, 89.8, 61.7, 122.4, 192.6,
+            -27.8, -104.6, 57.0, 166.4, 27.1, 6.1, 18.7, -93.2, 31.5, 168.2, -3.7, -99.5, -55.5,
+            -10.8, 17.5, 20.8, 16.9, 43.8, 42.0, -89.2, 18.8, -9.6, -84.1, 212.6, 19.7, -50.0,
+            -52.0, -40.0, -166.6, -73.2, -10.8, -73.3, 31.5, -23.4, -79.3, -27.0, -84.4, -42.9,
+            -20.3, 51.8, -16.7, 76.3, -120.5, -65.8, 96.5, -10.7, -45.9, -88.1, 65.4, -7.0, -1.5,
+            92.8, -25.1, -114.2, -5.8, -14.8, -51.2, -20.7, 54.2, -79.8, 47.7, -29.2, -8.8, 53.5,
+            -28.4, 85.0, -18.3, 107.0, 28.3, -71.8
+        ]
+    );
+
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_t.i(0)?, 1)?,
+        [
+            [
+                [32.3, -41.6, -24.0, 14.1, 17.6],
+                [-11.8, 72.5, 87.6, 46.4, 61.5],
+                [115.0, 108.5, -48.6, -63.4, -50.0],
+                [51.3, 5.4, 31.3, 91.1, -30.9],
+                [52.7, 92.8, -68.0, -47.0, 83.0],
+                // pytorch gets -107.1
+                [-10.2, -107.0, -5.4, 213.1, -31.4],
+                [-2.4, 65.1, 9.2, -146.2, -24.2]
+            ],
+            [
+                [-72.6, -63.9, -61.9, 45.3, 33.0],
+                [79.3, -0.5, -26.2, 78.2, 42.7],
+                [90.9, 141.6, 40.1, -62.7, 37.0],
+                [32.8, 198.2, -0.8, -31.1, 27.3],
+                // torch gets 48.0
+                [34.5, 34.9, -47.9, 127.6, -12.3],
+                [-61.4, -3.2, -2.9, -10.9, -16.6],
+                [74.6, 60.1, -68.9, 34.5, -50.4]
+            ],
+            [
+                [37.5, -56.9, -43.6, -13.5, -9.9],
+                [40.0, 97.3, 28.6, 14.2, -30.1],
+                [-22.3, -126.3, -68.8, -8.2, 26.1],
+                [-32.9, 37.3, 108.5, -54.8, 29.6],
+                [34.9, -176.9, -125.0, -28.3, -13.9],
+                [-54.9, 142.6, 62.1, -80.4, -65.6],
+                [7.4, -91.1, -67.6, 35.0, 39.7]
+            ],
+            [
+                [-57.2, -40.9, -10.1, 32.6, 29.4],
+                [18.7, -18.0, 29.5, -1.2, 59.2],
+                [-14.0, -74.4, 19.8, -117.0, 58.2],
+                [-21.8, 163.5, -71.1, -99.0, 80.9],
+                [-58.9, -10.9, 93.8, -139.6, 98.0],
+                // torch gets 54.5
+                [-54.4, 135.3, 6.0, -79.1, 134.6],
+                [27.5, -76.0, 43.4, -2.8, -7.8]
+            ]
+        ]
+    );
     Ok(())
 }
 
-test_device!(conv1d, conv1d_cpu, conv1d_gpu);
-test_device!(conv1d_small, conv1d_small_cpu, conv1d_small_gpu);
-test_device!(conv2d, conv2d_cpu, conv2d_gpu);
+test_device!(conv1d, conv1d_cpu, conv1d_gpu, conv1d_metal);
+test_device!(
+    conv1d_small,
+    conv1d_small_cpu,
+    conv1d_small_gpu,
+    conv1d_small_metal
+);
+test_device!(conv2d, conv2d_cpu, conv2d_gpu, conv2d_metal);
 test_device!(
     conv2d_non_square,
     conv2d_non_square_cpu,
-    conv2d_non_square_gpu
+    conv2d_non_square_gpu,
+    conv2d_non_square_metal
+);
+test_device!(
+    conv2d_small,
+    conv2d_small_cpu,
+    conv2d_small_gpu,
+    conv2d_small_metal
+);
+test_device!(
+    conv2d_smaller,
+    conv2d_smaller_cpu,
+    conv2d_smaller_gpu,
+    conv2d_smaller_metal
+);
+test_device!(
+    conv2d_grad,
+    conv2d_grad_cpu,
+    conv2d_grad_gpu,
+    conv2_grad_metal
 );
-test_device!(conv2d_small, conv2d_small_cpu, conv2d_small_gpu);
-test_device!(conv2d_smaller, conv2d_smaller_cpu, conv2d_smaller_gpu);
-test_device!(conv2d_grad, conv2d_grad_cpu, conv2d_grad_gpu);

candle-core/tests/custom_op_tests.rs

@@ -112,3 +112,34 @@ fn custom_op1_with_backward() -> Result<()> {
 
     Ok(())
 }
+
+impl candle_core::InplaceOp1 for Elu {
+    fn name(&self) -> &'static str {
+        "elu"
+    }
+
+    fn cpu_fwd(&self, s: &mut CpuStorage, _l: &Layout) -> Result<()> {
+        let alpha = self.alpha;
+        match s {
+            CpuStorage::BF16(s) => s.iter_mut().for_each(|v| *v = fwd(*v, alpha)),
+            CpuStorage::F16(s) => s.iter_mut().for_each(|v| *v = fwd(*v, alpha)),
+            CpuStorage::F32(s) => s.iter_mut().for_each(|v| *v = fwd(*v, alpha)),
+            CpuStorage::F64(s) => s.iter_mut().for_each(|v| *v = fwd(*v, alpha)),
+            _ => candle_core::bail!("unsupported dtype for inplace elu"),
+        }
+        Ok(())
+    }
+}
+
+#[test]
+fn inplace_op1() -> Result<()> {
+    let cpu = &Device::Cpu;
+    let t = Tensor::arange(0u32, 12u32, cpu)?.to_dtype(DType::F32)?;
+    let t = (t - 5.)?;
+    t.inplace_op1(&Elu { alpha: 1. })?;
+    assert_eq!(
+        to_vec1_round(&t, 4)?,
+        &[-0.9933, -0.9817, -0.9502, -0.8647, -0.6321, 0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
+    );
+    Ok(())
+}

candle-core/tests/grad_tests.rs

@@ -1,3 +1,4 @@
+#![allow(clippy::approx_constant)]
 use anyhow::{Context, Result};
 use candle_core::{test_device, test_utils, Device, Shape, Tensor, Var};
 
@@ -96,24 +97,24 @@ fn unary_grad(device: &Device) -> Result<()> {
     let grads = y.backward()?;
     let grad_x = grads.get(x).context("no grad for x")?;
     assert_eq!(
-        y.to_vec1::<f32>()?,
-        [20.085537, 2.7182817, 54.59815, 1.1618342]
+        test_utils::to_vec1_round(&y, 4)?,
+        [20.0855, 2.7183, 54.5982, 1.1618]
     );
     assert_eq!(
-        grad_x.to_vec1::<f32>()?,
-        [20.085537, 2.7182817, 54.59815, 1.1618342]
+        test_utils::to_vec1_round(grad_x, 4)?,
+        [20.0855, 2.7183, 54.5982, 1.1618]
     );
     let y = x.exp()?.sqr()?;
     let grads = y.backward()?;
     let grad_x = grads.get(x).context("no grad for x")?;
     assert_eq!(
-        y.to_vec1::<f32>()?,
-        [403.4288, 7.3890557, 2980.9578, 1.3498588]
+        test_utils::to_vec1_round(&y, 3)?,
+        [403.429, 7.389, 2980.958, 1.35]
     );
     // exp(x)^2 = exp(2*x)
     assert_eq!(
-        grad_x.to_vec1::<f32>()?,
-        [806.8576, 14.778111, 5961.9155, 2.6997175]
+        test_utils::to_vec1_round(grad_x, 2)?,
+        [806.86, 14.78, 5961.92, 2.7]
     );
     let y = x.sin()?;
     let grads = y.backward()?;
@@ -261,6 +262,7 @@ fn unary_grad(device: &Device) -> Result<()> {
     let y = elu_x.elu(2.)?;
     let grads = y.backward()?;
     let grad_x = grads.get(&elu_x).context("no grad for x")?;
+
     assert_eq!(
         test_utils::to_vec1_round(&y, 4)?,
         [-1.2642, 0.0000, -1.7293, 3.0000]
@@ -270,6 +272,194 @@ fn unary_grad(device: &Device) -> Result<()> {
         [0.7358, 2.0000, 0.2707, 1.0000]
     );
 
+    // testing compared to pytorch nn.Silu()
+    let y = x.silu()?;
+    let grads = y.backward()?;
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [2.8577, 0.7311, 3.9281, 0.0806]
+    );
+    assert_eq!(
+        test_utils::to_vec1_round(grad_x, 4)?,
+        [1.0881, 0.9277, 1.0527, 0.5747],
+    );
+
+    if device.is_cpu() {
+        let x = Var::new(&[[[1f32, 2., 3.], [4., 5., 6.], [7., 8., 9.]]], device)?;
+        let y = x.interpolate1d(12)?.reshape(36)?;
+
+        let z = Tensor::new(
+            &[
+                1_f32, 02., 03., 04., 05., 06., 07., 08., 09., 10., 11., 12., 13., 14., 15., 16.,
+                17., 18., 19., 20., 21., 22., 23., 24., 25., 26., 27., 28., 29., 30., 31., 32.,
+                33., 34., 35., 36.,
+            ],
+            device,
+        )?;
+
+        let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+        let grads = loss.backward()?;
+        let grad_x = grads.get(&x).context("no grad for x")?;
+
+        assert_eq!(
+            test_utils::to_vec3_round(grad_x, 4)?,
+            [[[10_f32, 26., 42.], [58., 74., 90.], [106., 122., 138.]]]
+        );
+    }
+
+    // manually checked: see comments
+    let x = Var::new(&[[[[1f32, 2., 3.], [4., 5., 6.], [7., 8., 9.]]]], device)?;
+    let y = x.interpolate2d(6, 6)?.reshape(36)?;
+
+    let z = Tensor::new(
+        &[
+            1_f32, 02., 03., 04., 05., 06., 07., 08., 09., 10., 11., 12., 13., 14., 15., 16., 17.,
+            18., 19., 20., 21., 22., 23., 24., 25., 26., 27., 28., 29., 30., 31., 32., 33., 34.,
+            35., 36.,
+        ],
+        device,
+    )?;
+    // gradient should be
+    // row 1
+    // 1+2+7+8 = 18
+    // 3+4+9+10 = 26
+    // 5+6+11+12 = 34
+    // row 2
+    // 13+14+19+20 = 66
+    // 15+16+21+22 = 74
+    // 17+18+23+24 = 82
+    // row 3
+    // 25+26+31+32 = 114
+    // 27+28+33+34 = 122
+    // 29+30+35+36 = 130
+    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+
+    let grads = loss.backward()?;
+
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(
+        test_utils::to_vec2_round(&grad_x.flatten(0, 2)?, 4)?,
+        [[18_f32, 26., 34.], [66., 74., 82.], [114., 122., 130.]]
+    );
+
+    // manually checked: see comments
+    let x = Var::new(&[[[[1f32, 2.], [4., 5.]]]], device)?;
+    let y = x.interpolate2d(6, 6)?.reshape(36)?;
+
+    let z = Tensor::new(
+        &[
+            1_f32, 02., 03., 04., 05., 06., 07., 08., 09., 10., 11., 12., 13., 14., 15., 16., 17.,
+            18., 19., 20., 21., 22., 23., 24., 25., 26., 27., 28., 29., 30., 31., 32., 33., 34.,
+            35., 36.,
+        ],
+        device,
+    )?;
+    // gradient should be
+    // row 1
+    // 1+2+3+7+8+9+13+14+15 = 72
+    // 4+5+6+10+11+12+16+17+18 = 99
+    // row 2
+    // 19+20+21+25+26+27+31+32+33 = 234
+    // 22+23+24+28+29+30+34+35+36 = 243
+    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+
+    let grads = loss.backward()?;
+
+    let grad_x = grads.get(&x).context("no grad for x")?;
+    assert_eq!(
+        test_utils::to_vec2_round(&grad_x.flatten(0, 2)?, 4)?,
+        [[72_f32, 99.], [234., 261.]]
+    );
+
+    // manually checked: see comments
+    let x = Var::new(&[[[[1f32, 2.], [4., 5.]], [[6f32, 7.], [8., 9.]]]], device)?;
+
+    let y = x.interpolate2d(4, 4)?.reshape(32)?;
+
+    #[rustfmt::skip]
+    let z = Tensor::new(
+        &[
+            1_f32, 02., 03., 04.,
+            05.,   06., 07., 08.,
+            09.,   10., 11., 12.,
+            13.,   14., 15., 16.,
+            17.,   18., 19., 20.,
+            21.,   22., 23., 24.,
+            25.,   26., 27., 28.,
+            29.,   30., 31., 32.
+        ],
+        device,
+    )?;
+    // gradient should be
+    // m1r1
+    // 1+2+5+6=14
+    // 3+4+7+8=22
+    // m1r2
+    // 9+10+13+14=46
+    // 11+12+15+16=54
+    // m2r1
+    // 17+18+21+22=78
+    // 19+20+23+24=86
+    // m2r2
+    // 25+26+29+30=110
+    // 27+28+31+32=118
+    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+
+    let grads = loss.backward()?;
+
+    let grad_x = grads.get(&x).context("no grad for x")?;
+
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_x.flatten(0, 1)?, 4)?,
+        [[[14_f32, 22.], [46., 54.]], [[78., 86.], [110., 118.]]]
+    );
+
+    // manually checked: see comments
+    let x = Var::new(
+        &[[[[1f32, 2.], [4., 5.]]], [[[6f32, 7.], [8., 9.]]]],
+        device,
+    )?;
+
+    let y = x.interpolate2d(4, 4)?.reshape(32)?;
+
+    #[rustfmt::skip]
+       let z = Tensor::new(
+           &[
+               1_f32, 02., 03., 04.,
+               05.,   06., 07., 08.,
+               09.,   10., 11., 12.,
+               13.,   14., 15., 16.,
+               17.,   18., 19., 20.,
+               21.,   22., 23., 24.,
+               25.,   26., 27., 28.,
+               29.,   30., 31., 32.
+           ],
+           device,
+       )?;
+    // gradient should be
+    // m1r1
+    // 1+2+5+6=14
+    // 3+4+7+8=22
+    // m1r2
+    // 9+10+13+14=46
+    // 11+12+15+16=54
+    // m2r1
+    // 17+18+21+22=78
+    // 19+20+23+24=86
+    // m2r2
+    // 25+26+29+30=110
+    // 27+28+31+32=118
+    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
+
+    let grads = loss.backward()?;
+
+    let grad_x = grads.get(&x).context("no grad for x")?;
+
+    assert_eq!(
+        test_utils::to_vec3_round(&grad_x.flatten(0, 1)?, 4)?,
+        [[[14_f32, 22.], [46., 54.]], [[78., 86.], [110., 118.]]]
+    );
     Ok(())
 }
 
@@ -315,9 +505,29 @@ fn binary_grad(device: &Device) -> Result<()> {
     Ok(())
 }
 
-test_device!(simple_grad, simple_grad_cpu, simple_grad_gpu);
-test_device!(sum_grad, sum_grad_cpu, sum_grad_gpu);
-test_device!(matmul_grad, matmul_grad_cpu, matmul_grad_gpu);
-test_device!(grad_descent, grad_descent_cpu, grad_descent_gpu);
-test_device!(unary_grad, unary_grad_cpu, unary_grad_gpu);
-test_device!(binary_grad, binary_grad_cpu, binary_grad_gpu);
+test_device!(
+    simple_grad,
+    simple_grad_cpu,
+    simple_grad_gpu,
+    simple_grad_metal
+);
+test_device!(sum_grad, sum_grad_cpu, sum_grad_gpu, sum_grad_metal);
+test_device!(
+    matmul_grad,
+    matmul_grad_cpu,
+    matmul_grad_gpu,
+    matmul_grad_metal
+);
+test_device!(
+    grad_descent,
+    grad_descent_cpu,
+    grad_descent_gpu,
+    grad_descent_metal
+);
+test_device!(unary_grad, unary_grad_cpu, unary_grad_gpu, unary_grad_metal);
+test_device!(
+    binary_grad,
+    binary_grad_cpu,
+    binary_grad_gpu,
+    binary_grad_metal
+);

candle-core/tests/indexing_tests.rs

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

candle-core/tests/layout_tests.rs

@@ -49,7 +49,7 @@ fn contiguous(device: &Device) -> Result<()> {
     Ok(())
 }
 
-test_device!(contiguous, contiguous_cpu, contiguous_gpu);
+test_device!(contiguous, contiguous_cpu, contiguous_gpu, contiguous_metal);
 
 #[test]
 fn strided_blocks() -> Result<()> {
@@ -88,7 +88,7 @@ fn strided_blocks() -> Result<()> {
         }
     };
     let tensor = Tensor::arange(0u32, 24u32, &Cpu)?.reshape((2, 3, 4))?;
-    let tensor = tensor.i((.., 1))?;
+    let tensor = tensor.i((.., 1))?.contiguous()?;
     match tensor.strided_blocks() {
         candle::StridedBlocks::SingleBlock { start_offset, len } => {
             assert_eq!(start_offset, 0);
@@ -100,6 +100,20 @@ fn strided_blocks() -> Result<()> {
         }
     };
     let tensor = Tensor::arange(0u32, 24u32, &Cpu)?.reshape((2, 3, 4))?;
+    let tensor = tensor.i((.., 1))?;
+    match tensor.strided_blocks() {
+        candle::StridedBlocks::SingleBlock { .. } => {
+            panic!("unexpected block structure")
+        }
+        candle::StridedBlocks::MultipleBlocks {
+            block_len,
+            block_start_index,
+        } => {
+            assert_eq!(block_len, 4);
+            assert_eq!(block_start_index.collect::<Vec<_>>(), &[4, 16])
+        }
+    };
+    let tensor = Tensor::arange(0u32, 24u32, &Cpu)?.reshape((2, 3, 4))?;
     match tensor.t()?.strided_blocks() {
         candle::StridedBlocks::SingleBlock { .. } => {
             panic!("unexpected block structure")

candle-core/tests/matmul_tests.rs

@@ -0,0 +1,106 @@
+use candle_core::{test_device, DType, Device, IndexOp, Result, Tensor};
+
+fn matmul(device: &Device) -> Result<()> {
+    let data = vec![1.0f32, 2.0, 3.0, 4.0];
+    let a = Tensor::from_slice(&data, (2, 2), device)?;
+    let data = vec![1.0f32, 2.0, 3.0, 4.0];
+    let b = Tensor::from_slice(&data, (2, 2), device)?;
+
+    let c = a.matmul(&b)?;
+    assert_eq!(c.to_vec2::<f32>()?, &[[7.0f32, 10.0], [15.0, 22.0]]);
+
+    let data = vec![1.0f32, 2.0];
+    let a = Tensor::from_slice(&data, (2, 1), device)?;
+    let data = vec![3.0f32, 4.0];
+    let b = Tensor::from_slice(&data, (1, 2), device)?;
+    let c = a.matmul(&b)?;
+    assert_eq!(c.to_vec2::<f32>()?, &[&[3.0, 4.0], &[6.0, 8.0]]);
+
+    let data: Vec<_> = (0..6).map(|i| i as f32).collect();
+    let a = Tensor::from_slice(&data, (2, 3), device)?;
+    let data: Vec<_> = (0..6).map(|i| (i + 2) as f32).collect();
+    let b = Tensor::from_slice(&data, (3, 2), device)?;
+    let c = a.matmul(&b)?;
+    assert_eq!(c.to_vec2::<f32>()?, &[&[16., 19.], &[52., 64.]]);
+
+    let data: Vec<_> = (0..12).map(|i| i as f32).collect();
+    let a = Tensor::from_slice(&data, (2, 2, 3), device)?;
+    let data: Vec<_> = (0..12).map(|i| (i + 2) as f32).collect();
+    let b = Tensor::from_slice(&data, (2, 3, 2), device)?;
+    let expected = [[[16., 19.], [52., 64.]], [[214., 235.], [304., 334.]]];
+
+    let c = a.matmul(&b)?;
+    assert_eq!(c.to_vec3::<f32>()?, &expected);
+
+    // Also perform the matmul on contiguous transposed versions.
+    let a_tt = a.t()?.contiguous()?.t()?;
+    assert!(!a_tt.is_contiguous());
+    assert_eq!(a.dims(), a_tt.dims());
+    assert_eq!(a_tt.stride(), &[6, 1, 2]);
+
+    let b_tt = b.t()?.contiguous()?.t()?;
+    assert!(!b_tt.is_contiguous());
+    assert_eq!(b.dims(), b_tt.dims());
+    assert_eq!(b_tt.stride(), &[6, 1, 3]);
+
+    assert_eq!(a_tt.matmul(&b)?.to_vec3::<f32>()?, &expected);
+    assert_eq!(a.matmul(&b_tt)?.to_vec3::<f32>()?, &expected);
+    assert_eq!(a_tt.matmul(&b_tt)?.to_vec3::<f32>()?, &expected);
+    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)?;
+    let out = lhs.broadcast_matmul(&rhs)?;
+    assert_eq!(out.dims(), &[3, 6, 4, 2]);
+    for idx1 in 0..3 {
+        for idx2 in 0..6 {
+            let out = out.i((idx1, idx2))?;
+            let lhs = lhs.i((idx1, 0))?;
+            let rhs = rhs.i(idx2)?;
+            let out2 = lhs.matmul(&rhs);
+            let sum_diff2 = (out - out2)?.sqr()?.sum_all()?;
+            // With cuda, we see errors of up to ~1e-12.
+            assert!(sum_diff2.to_vec0::<f32>()? < 1e-6)
+        }
+    }
+    Ok(())
+}
+
+// https://github.com/huggingface/candle/issues/1948
+fn squeeze_mm(device: &Device) -> Result<()> {
+    let seq_len = 8_usize;
+    let a = Tensor::zeros((1, seq_len, 16), DType::F32, device)?;
+    let x = a.i((.., seq_len - 1, ..))?;
+    let w = Tensor::zeros((32, 16), DType::F32, device)?.t()?;
+    let x = x.matmul(&w)?;
+    assert_eq!(x.dims(), &[1, 32]);
+    Ok(())
+}
+
+// https://github.com/huggingface/candle/issues/1992
+fn mm_layout(device: &Device) -> Result<()> {
+    let a = Tensor::arange(0f32, 16f32, device)?.reshape((1, 1, 4, 4))?;
+    let b = Tensor::arange(0f32, 8f32, device)?.reshape((1, 1, 4, 2))?;
+    let mm1 = a.matmul(&b)?;
+    // Forces the layout to be:
+    // shape: [1, 1, 4, 2], stride: [8, 2, 2, 1], start_offset: 0
+    // This is still a contiguous matrix but matmul checks are only the two last dimensions have
+    // non 1 sizes but matmul check may be reluctant to handle it.
+    let b = b.transpose(1, 2)?.force_contiguous()?.transpose(1, 2)?;
+    let mm2 = a.matmul(&b)?;
+    let diff = (mm1 - mm2)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
+    Ok(())
+}
+
+test_device!(matmul, matmul_cpu, matmul_gpu, matmul_metal);
+test_device!(
+    broadcast_matmul,
+    broadcast_matmul_cpu,
+    broadcast_matmul_gpu,
+    broadcast_matmul_metal
+);
+test_device!(squeeze_mm, squeeze_mm_cpu, squeeze_mm_gpu, squeeze_mm_metal);
+test_device!(mm_layout, mm_layout_cpu, mm_layout_gpu, mm_layout_metal);

candle-core/tests/pool_tests.rs

@@ -43,6 +43,9 @@ res = torch.nn.functional.avg_pool2d(t, 2)
 print(res)
 */
 fn avg_pool2d_pytorch(dev: &Device) -> Result<()> {
+    if dev.is_metal() {
+        return Ok(());
+    }
     let t = Tensor::new(
         &[
             0.4056f32, -0.8689, -0.0773, -1.5630, -2.8012, -1.5059, 0.3972, 1.0852, 0.4997, 3.0616,
@@ -98,15 +101,17 @@ fn upsample_nearest2d(dev: &Device) -> Result<()> {
     Ok(())
 }
 
-test_device!(avg_pool2d, avg_pool2d_cpu, avg_pool2d_gpu);
+test_device!(avg_pool2d, avg_pool2d_cpu, avg_pool2d_gpu, avg_pool2d_metal);
 test_device!(
     avg_pool2d_pytorch,
     avg_pool2d_pytorch_cpu,
-    avg_pool2d_pytorch_gpu
+    avg_pool2d_pytorch_gpu,
+    avg_pool2d_pytorch_metal
 );
-test_device!(max_pool2d, max_pool2d_cpu, max_pool2d_gpu);
+test_device!(max_pool2d, max_pool2d_cpu, max_pool2d_gpu, max_pool2d_metal);
 test_device!(
     upsample_nearest2d,
     upsample_nearest2d_cpu,
-    upsample_nearest2d_gpu
+    upsample_nearest2d_gpu,
+    upsample_nearest2d_metal
 );

candle-core/tests/pth.py

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

candle-core/tests/pth_tests.rs

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

candle-core/tests/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

@@ -1,4 +1,4 @@
-use candle_core::{test_device, test_utils, DType, Device, IndexOp, Result, Tensor};
+use candle_core::{test_device, test_utils, DType, Device, IndexOp, Result, Tensor, D};
 
 fn zeros(device: &Device) -> Result<()> {
     let tensor = Tensor::zeros((5, 2), DType::F32, device)?;
@@ -32,6 +32,14 @@ fn ones(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn full(device: &Device) -> Result<()> {
+    assert_eq!(
+        Tensor::full(42u32, (2, 3), device)?.to_vec2::<u32>()?,
+        [[42, 42, 42], [42, 42, 42]],
+    );
+    Ok(())
+}
+
 fn arange(device: &Device) -> Result<()> {
     assert_eq!(
         Tensor::arange(0u8, 5u8, device)?.to_vec1::<u8>()?,
@@ -88,6 +96,40 @@ 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)?;
@@ -98,6 +140,9 @@ fn unary_op(device: &Device) -> Result<()> {
             [2.6911, -0.0647, -0.1091, 1.7353, 2.7933]
         ]
     );
+    let t_f16 = tensor.to_dtype(DType::F16)?.gelu()?.to_dtype(DType::F32)?;
+    let max_diff = (tensor.gelu()? - t_f16)?.flatten_all()?.max(0)?;
+    assert!(max_diff.to_vec0::<f32>()? < 5e-3);
     assert_eq!(
         test_utils::to_vec2_round(&tensor.gelu_erf()?, 4)?,
         [
@@ -112,6 +157,13 @@ fn unary_op(device: &Device) -> Result<()> {
             [0.9999, -0.9891, -0.3079, 0.9891, 0.9999]
         ]
     );
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.silu()?, 4)?,
+        [
+            [-0.1423, 0.7311, 3.9281, -0.0475, 0.3112],
+            [2.53, -0.2553, -0.1205, 1.5447, 2.6395]
+        ]
+    );
     assert_eq!(
         test_utils::to_vec2_round(&tensor.ceil()?, 4)?,
         [[-3.0, 1.0, 4.0, -0.0, 1.0], [3.0, -1.0, -0.0, 2.0, 3.0]]
@@ -133,6 +185,14 @@ fn unary_op(device: &Device) -> Result<()> {
         test_utils::to_vec1_round(&tensor.round_to(-2)?, 4)?,
         [3000.0, 300.]
     );
+    let tensor = Tensor::new(
+        &[-1.01f32, -0.9, -0.1, 0.0, -0.0, 0.1, 0.9, 1.0, 1.1],
+        device,
+    )?;
+    assert_eq!(
+        tensor.sign()?.to_vec1::<f32>()?,
+        [-1., -1., -1., 0., 0., 1., 1., 1., 1.]
+    );
     Ok(())
 }
 
@@ -180,6 +240,22 @@ fn transpose(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn var(device: &Device) -> Result<()> {
+    // Values taken from https://pytorch.org/docs/stable/generated/torch.var.html
+    let data = &[
+        [0.2035f32, 1.2959, 1.8101, -0.4644],
+        [1.5027, -0.3270, 0.5905, 0.6538],
+        [-1.5745, 1.3330, -0.5596, -0.6548],
+        [0.1264, -0.5080, 1.6420, 0.1992],
+    ];
+    let tensor = Tensor::new(data, device)?;
+    assert_eq!(
+        test_utils::to_vec2_round(&tensor.var_keepdim(1)?, 4)?,
+        &[[1.0631], [0.559], [1.4893], [0.8258]]
+    );
+    Ok(())
+}
+
 fn sum(device: &Device) -> Result<()> {
     let data = &[[[3u32, 1, 4], [1, 5, 9]], [[2, 1, 7], [8, 2, 8]]];
     let tensor = Tensor::new(data, device)?;
@@ -589,6 +665,30 @@ 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.);
+    Ok(())
+}
+
 fn cat(device: &Device) -> Result<()> {
     // 1D
     let t1 = Tensor::new(&[3f32, 1., 4.], device)?;
@@ -641,6 +741,31 @@ fn cat(device: &Device) -> Result<()> {
             [2.0, 7.0, 1.0, 8.0, 2.0, 2.0, 7.0, 1.0, 8.0, 2.0]
         ]
     );
+
+    // 3D
+    let t1 = Tensor::arange(0, 48i64, device)?.reshape((2, 6, 4))?;
+    let t2 = Tensor::arange(100, 124i64, device)?.reshape((2, 3, 4))?;
+    let t3 = Tensor::arange(10000, 10032i64, device)?.reshape((2, 4, 4))?;
+
+    let t_cat = Tensor::cat(&[&t1, &t2, &t3], 1)?;
+
+    let t1 = t1.t()?.contiguous()?.t()?;
+    let t2 = t2.t()?.contiguous()?.t()?;
+    let t3 = t3.t()?.contiguous()?.t()?;
+    let t_cat2 = Tensor::cat(&[&t1, &t2, &t3], 1)?;
+
+    let diff = t_cat.eq(&t_cat2)?.to_dtype(DType::F32)?.sum_all()?;
+    assert_eq!(diff.to_vec0::<f32>()?, 104.0);
+    assert_eq!(t_cat.i((0, 0, 0))?.to_vec0::<i64>()?, 0);
+    assert_eq!(t_cat.i((0, 4, 0))?.to_vec0::<i64>()?, 16);
+    assert_eq!(t_cat.i((0, 5, 0))?.to_vec0::<i64>()?, 20);
+    assert_eq!(t_cat.i((1, 5, 0))?.to_vec0::<i64>()?, 44);
+    assert_eq!(t_cat.i((0, 6, 0))?.to_vec0::<i64>()?, 100);
+    assert_eq!(t_cat.i((1, 6, 0))?.to_vec0::<i64>()?, 112);
+    assert_eq!(t_cat.i((0, 6, 1))?.to_vec0::<i64>()?, 101);
+    assert_eq!(t_cat.i((0, 7, 1))?.to_vec0::<i64>()?, 105);
+    assert_eq!(t_cat.i((0, 12, 1))?.to_vec0::<i64>()?, 10013);
+    assert_eq!(t_cat.i((1, 12, 3))?.to_vec0::<i64>()?, 10031);
     Ok(())
 }
 
@@ -651,6 +776,8 @@ fn embeddings(device: &Device) -> Result<()> {
     assert_eq!(hs.to_vec2::<f32>()?, &[[0.0, 1.0], [4.0, 5.0], [2.0, 3.0]]);
     let hs = t.index_select(&ids, 0)?;
     assert_eq!(hs.to_vec2::<f32>()?, &[[0.0, 1.0], [4.0, 5.0], [2.0, 3.0]]);
+    let hs = t.index_select(&ids.to_dtype(DType::I64)?, 0)?;
+    assert_eq!(hs.to_vec2::<f32>()?, &[[0.0, 1.0], [4.0, 5.0], [2.0, 3.0]]);
     Ok(())
 }
 
@@ -678,44 +805,47 @@ fn index_select(device: &Device) -> Result<()> {
             [9.0, 10.0, 11.0]
         ]
     );
-    let hs = t.index_select(&ids, 1)?;
-    assert_eq!(
-        hs.to_vec2::<f32>()?,
-        &[
-            [0.0, 2.0, 1.0],
-            [3.0, 5.0, 4.0],
-            [6.0, 8.0, 7.0],
-            [9.0, 11.0, 10.0]
-        ]
-    );
-    let hs = t.index_select(&ids, 0)?;
-    assert_eq!(
-        hs.to_vec2::<f32>()?,
-        &[[0.0, 1.0, 2.0], [6.0, 7.0, 8.0], [3.0, 4.0, 5.0]]
-    );
-    // Prior to https://github.com/huggingface/candle/pull/1022
-    // There would be a bug where the last values in the result tensor would be set to 0.
-    let ids = Tensor::new(&[0u32, 2u32, 1u32, 0u32, 2u32, 1u32], device)?;
-    let hs = t.index_select(&ids, 0)?;
-    assert_eq!(
-        hs.to_vec2::<f32>()?,
-        &[
-            [0.0, 1.0, 2.0],
-            [6.0, 7.0, 8.0],
-            [3.0, 4.0, 5.0],
-            [0.0, 1.0, 2.0],
-            [6.0, 7.0, 8.0],
-            [3.0, 4.0, 5.0],
-        ]
-    );
+    for dtype in [DType::U8, DType::U32, DType::I64] {
+        let ids = ids.to_dtype(dtype)?;
+        let hs = t.index_select(&ids, 1)?;
+        assert_eq!(
+            hs.to_vec2::<f32>()?,
+            &[
+                [0.0, 2.0, 1.0],
+                [3.0, 5.0, 4.0],
+                [6.0, 8.0, 7.0],
+                [9.0, 11.0, 10.0]
+            ]
+        );
+        let hs = t.index_select(&ids, 0)?;
+        assert_eq!(
+            hs.to_vec2::<f32>()?,
+            &[[0.0, 1.0, 2.0], [6.0, 7.0, 8.0], [3.0, 4.0, 5.0]]
+        );
+        // Prior to https://github.com/huggingface/candle/pull/1022
+        // There would be a bug where the last values in the result tensor would be set to 0.
+        let ids = Tensor::new(&[0u32, 2u32, 1u32, 0u32, 2u32, 1u32], device)?;
+        let hs = t.index_select(&ids, 0)?;
+        assert_eq!(
+            hs.to_vec2::<f32>()?,
+            &[
+                [0.0, 1.0, 2.0],
+                [6.0, 7.0, 8.0],
+                [3.0, 4.0, 5.0],
+                [0.0, 1.0, 2.0],
+                [6.0, 7.0, 8.0],
+                [3.0, 4.0, 5.0],
+            ]
+        );
 
-    // Test when selecting dim > 0 with ids size different from elem count of
-    // target dim in source/input.
-    let ids = Tensor::new(&[1u32, 0u32, 1u32], device)?;
-    let t = Tensor::arange(1f32, 5f32, device)?.reshape((2, 2))?;
-    assert_eq!(t.to_vec2::<f32>()?, &[[1.0, 2.0], [3.0, 4.0]]);
-    let hs = t.index_select(&ids, 1)?;
-    assert_eq!(hs.to_vec2::<f32>()?, &[[2.0, 1.0, 2.0], [4.0, 3.0, 4.0]]);
+        // Test when selecting dim > 0 with ids size different from elem count of
+        // target dim in source/input.
+        let ids = Tensor::new(&[1u32, 0u32, 1u32], device)?;
+        let t = Tensor::arange(1f32, 5f32, device)?.reshape((2, 2))?;
+        assert_eq!(t.to_vec2::<f32>()?, &[[1.0, 2.0], [3.0, 4.0]]);
+        let hs = t.index_select(&ids, 1)?;
+        assert_eq!(hs.to_vec2::<f32>()?, &[[2.0, 1.0, 2.0], [4.0, 3.0, 4.0]]);
+    }
 
     Ok(())
 }
@@ -877,74 +1007,6 @@ fn gather(device: &Device) -> Result<()> {
     Ok(())
 }
 
-fn matmul(device: &Device) -> Result<()> {
-    let data = vec![1.0f32, 2.0, 3.0, 4.0];
-    let a = Tensor::from_slice(&data, (2, 2), device)?;
-    let data = vec![1.0f32, 2.0, 3.0, 4.0];
-    let b = Tensor::from_slice(&data, (2, 2), device)?;
-
-    let c = a.matmul(&b)?;
-    assert_eq!(c.to_vec2::<f32>()?, &[[7.0f32, 10.0], [15.0, 22.0]]);
-
-    let data = vec![1.0f32, 2.0];
-    let a = Tensor::from_slice(&data, (2, 1), device)?;
-    let data = vec![3.0f32, 4.0];
-    let b = Tensor::from_slice(&data, (1, 2), device)?;
-    let c = a.matmul(&b)?;
-    assert_eq!(c.to_vec2::<f32>()?, &[&[3.0, 4.0], &[6.0, 8.0]]);
-
-    let data: Vec<_> = (0..6).map(|i| i as f32).collect();
-    let a = Tensor::from_slice(&data, (2, 3), device)?;
-    let data: Vec<_> = (0..6).map(|i| (i + 2) as f32).collect();
-    let b = Tensor::from_slice(&data, (3, 2), device)?;
-    let c = a.matmul(&b)?;
-    assert_eq!(c.to_vec2::<f32>()?, &[&[16., 19.], &[52., 64.]]);
-
-    let data: Vec<_> = (0..12).map(|i| i as f32).collect();
-    let a = Tensor::from_slice(&data, (2, 2, 3), device)?;
-    let data: Vec<_> = (0..12).map(|i| (i + 2) as f32).collect();
-    let b = Tensor::from_slice(&data, (2, 3, 2), device)?;
-    let expected = [[[16., 19.], [52., 64.]], [[214., 235.], [304., 334.]]];
-
-    let c = a.matmul(&b)?;
-    assert_eq!(c.to_vec3::<f32>()?, &expected);
-
-    // Also perform the matmul on contiguous transposed versions.
-    let a_tt = a.t()?.contiguous()?.t()?;
-    assert!(!a_tt.is_contiguous());
-    assert_eq!(a.dims(), a_tt.dims());
-    assert_eq!(a_tt.stride(), &[6, 1, 2]);
-
-    let b_tt = b.t()?.contiguous()?.t()?;
-    assert!(!b_tt.is_contiguous());
-    assert_eq!(b.dims(), b_tt.dims());
-    assert_eq!(b_tt.stride(), &[6, 1, 3]);
-
-    assert_eq!(a_tt.matmul(&b)?.to_vec3::<f32>()?, &expected);
-    assert_eq!(a.matmul(&b_tt)?.to_vec3::<f32>()?, &expected);
-    assert_eq!(a_tt.matmul(&b_tt)?.to_vec3::<f32>()?, &expected);
-    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)?;
-    let out = lhs.broadcast_matmul(&rhs)?;
-    assert_eq!(out.dims(), &[3, 6, 4, 2]);
-    for idx1 in 0..3 {
-        for idx2 in 0..6 {
-            let out = out.i((idx1, idx2))?;
-            let lhs = lhs.i((idx1, 0))?;
-            let rhs = rhs.i(idx2)?;
-            let out2 = lhs.matmul(&rhs);
-            let sum_diff2 = (out - out2)?.sqr()?.sum_all()?;
-            // With cuda, we see errors of up to ~1e-12.
-            assert!(sum_diff2.to_vec0::<f32>()? < 1e-6)
-        }
-    }
-    Ok(())
-}
-
 fn broadcasting(device: &Device) -> Result<()> {
     let t1 = Tensor::arange(0f32, 24f32, device)?.reshape((4, 2, 3))?;
     let t2 = Tensor::new(&[100f32, 200f32], device)?;
@@ -1049,39 +1111,108 @@ fn broadcasting(device: &Device) -> Result<()> {
 fn randn(device: &Device) -> Result<()> {
     let tensor = Tensor::randn(0f32, 1f32, (5, 3), device)?;
     assert_eq!(tensor.dims(), [5, 3]);
+    // Check that the seed gets updated by checking that
+    // a new series of numbers is generated each time
+    let tensor2 = Tensor::randn(0f32, 1f32, (5, 3), device)?;
+    assert_ne!(tensor.to_vec2::<f32>()?, tensor2.to_vec2::<f32>()?);
     let tensor = Tensor::rand(0f32, 1f32, (5, 3), device)?;
     assert_eq!(tensor.dims(), [5, 3]);
+    // Check that the seed gets updated by checking that
+    // a new series of numbers is generated each time
+    let tensor2 = Tensor::rand(0f32, 1f32, (5, 3), device)?;
+    assert_ne!(tensor.to_vec2::<f32>()?, tensor2.to_vec2::<f32>()?);
+    // We do not expect deterministic elements at any index.
+    // There once was a bug that had a deterministic zero element in evenly sized tensors.
+    const N: usize = 2;
+    let v = (0..100)
+        .map(|_| Tensor::randn(0f32, 1f32, N, device).and_then(|t| t.to_vec1::<f32>()))
+        .collect::<Result<Vec<_>>>()?;
+    assert!(
+        (0..N).all(|i| v.windows(2).any(|pair| pair[0][i] != pair[1][i])),
+        "There are deterministic values in the randn tensors"
+    );
+    let v = (0..100)
+        .map(|_| Tensor::rand(0f32, 1f32, N, device).and_then(|t| t.to_vec1::<f32>()))
+        .collect::<Result<Vec<_>>>()?;
+    assert!(
+        (0..N).all(|i| v.windows(2).any(|pair| pair[0][i] != pair[1][i])),
+        "There are deterministic values in the rand tensors"
+    );
     Ok(())
 }
 
-test_device!(zeros, zeros_cpu, zeros_gpu);
-test_device!(ones, ones_cpu, ones_gpu);
-test_device!(arange, arange_cpu, arange_gpu);
-test_device!(add_mul, add_mul_cpu, add_mul_gpu);
-test_device!(tensor_2d, tensor_2d_cpu, tensor_2d_gpu);
-test_device!(narrow, narrow_cpu, narrow_gpu);
-test_device!(broadcast, broadcast_cpu, broadcast_gpu);
-test_device!(cat, cat_cpu, cat_gpu);
-test_device!(sum, sum_cpu, sum_gpu);
-test_device!(min, min_cpu, min_gpu);
-test_device!(max, max_cpu, max_gpu);
-test_device!(argmax, argmax_cpu, argmax_gpu);
-test_device!(argmin, argmin_cpu, argmin_gpu);
-test_device!(transpose, transpose_cpu, transpose_gpu);
-test_device!(unary_op, unary_op_cpu, unary_op_gpu);
-test_device!(binary_op, binary_op_cpu, binary_op_gpu);
-test_device!(embeddings, embeddings_cpu, embeddings_gpu);
-test_device!(cmp, cmp_cpu, cmp_gpu);
-test_device!(matmul, matmul_cpu, matmul_gpu);
-test_device!(broadcast_matmul, broadcast_matmul_cpu, broadcast_matmul_gpu);
-test_device!(broadcasting, broadcasting_cpu, broadcasting_gpu);
-test_device!(index_select, index_select_cpu, index_select_gpu);
-test_device!(index_add, index_add_cpu, index_add_gpu);
-test_device!(gather, gather_cpu, gather_gpu);
-test_device!(scatter_add, scatter_add_cpu, scatter_add_gpu);
-test_device!(slice_scatter, slice_scatter_cpu, slice_scatter_gpu);
-test_device!(randn, randn_cpu, randn_gpu);
-test_device!(clamp, clamp_cpu, clamp_gpu);
+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!(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);
+test_device!(max, max_cpu, max_gpu, max_metal);
+test_device!(argmax, argmax_cpu, argmax_gpu, argmax_metal);
+test_device!(argmin, argmin_cpu, argmin_gpu, argmin_metal);
+test_device!(transpose, transpose_cpu, transpose_gpu, transpose_metal);
+test_device!(unary_op, unary_op_cpu, unary_op_gpu, unary_op_metal);
+test_device!(binary_op, binary_op_cpu, binary_op_gpu, binary_op_metal);
+test_device!(embeddings, embeddings_cpu, embeddings_gpu, embeddings_metal);
+test_device!(cmp, cmp_cpu, cmp_gpu, cmp_metal);
+test_device!(
+    broadcasting,
+    broadcasting_cpu,
+    broadcasting_gpu,
+    broadcasting_metal
+);
+test_device!(
+    index_select,
+    index_select_cpu,
+    index_select_gpu,
+    index_select_metal
+);
+test_device!(index_add, index_add_cpu, index_add_gpu, index_add_metal);
+test_device!(gather, gather_cpu, gather_gpu, gather_metal);
+test_device!(
+    scatter_add,
+    scatter_add_cpu,
+    scatter_add_gpu,
+    scatter_add_metal
+);
+test_device!(
+    slice_scatter,
+    slice_scatter_cpu,
+    slice_scatter_gpu,
+    slice_scatter_metal
+);
+test_device!(randn, randn_cpu, randn_gpu, randn_metal);
+test_device!(clamp, clamp_cpu, clamp_gpu, clamp_metal);
+test_device!(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
@@ -1117,3 +1248,100 @@ fn i64_abs() -> Result<()> {
     assert_eq!(t.to_vec1::<i64>()?, [42, 1337]);
     Ok(())
 }
+
+#[test]
+fn tril_triu_eye() -> Result<()> {
+    let t = Tensor::tril2(4, DType::F32, &Device::Cpu)?;
+    assert_eq!(
+        t.to_vec2::<f32>()?,
+        [
+            [1.0, 0.0, 0.0, 0.0],
+            [1.0, 1.0, 0.0, 0.0],
+            [1.0, 1.0, 1.0, 0.0],
+            [1.0, 1.0, 1.0, 1.0]
+        ],
+    );
+    let t = Tensor::triu2(4, DType::F32, &Device::Cpu)?;
+    assert_eq!(
+        t.to_vec2::<f32>()?,
+        [
+            [1.0, 1.0, 1.0, 1.0],
+            [0.0, 1.0, 1.0, 1.0],
+            [0.0, 0.0, 1.0, 1.0],
+            [0.0, 0.0, 0.0, 1.0]
+        ]
+    );
+    let t = Tensor::eye(4, DType::F32, &Device::Cpu)?;
+    assert_eq!(
+        t.to_vec2::<f32>()?,
+        [
+            [1.0, 0.0, 0.0, 0.0],
+            [0.0, 1.0, 0.0, 0.0],
+            [0.0, 0.0, 1.0, 0.0],
+            [0.0, 0.0, 0.0, 1.0]
+        ]
+    );
+    Ok(())
+}
+
+#[test]
+fn cumsum() -> Result<()> {
+    let t = &[3f32, 1., 4., 1., 5.];
+    let t = Tensor::new(t, &Device::Cpu)?;
+    assert_eq!(t.cumsum(0)?.to_vec1::<f32>()?, [3., 4., 8., 9., 14.]);
+    let t = t.unsqueeze(1)?;
+    assert_eq!(
+        t.cumsum(0)?.to_vec2::<f32>()?,
+        [[3.0], [4.0], [8.0], [9.0], [14.0]]
+    );
+    assert_eq!(
+        t.cumsum(1)?.to_vec2::<f32>()?,
+        [[3.0], [1.0], [4.0], [1.0], [5.0]]
+    );
+    let t = &[[3f32, 1., 4., 1., 5.], [2., 1., 7., 8., 2.]];
+    let t = Tensor::new(t, &Device::Cpu)?;
+    assert_eq!(
+        t.cumsum(1)?.to_vec2::<f32>()?,
+        [[3.0, 4.0, 8.0, 9.0, 14.0], [2.0, 3.0, 10.0, 18.0, 20.0]],
+    );
+    assert_eq!(
+        t.cumsum(0)?.to_vec2::<f32>()?,
+        [[3.0, 1.0, 4.0, 1.0, 5.0], [5.0, 2.0, 11.0, 9.0, 7.0]]
+    );
+    Ok(())
+}
+
+/// A helper function for floating point comparison. Both a and b must be 1D Tensor and contains the same amount of data.
+/// Assertion passes if the difference of all pairs of a and b is smaller than epsilon.
+fn assert_close(a: &Tensor, b: &Tensor, epsilon: f64) -> Result<()> {
+    let a_vec: Vec<f64> = a.to_vec1()?;
+    let b_vec: Vec<f64> = b.to_vec1()?;
+
+    assert_eq!(a_vec.len(), b_vec.len());
+    for (a, b) in a_vec.iter().zip(b_vec.iter()) {
+        assert!((a - b).abs() < epsilon);
+    }
+    Ok(())
+}
+
+#[test]
+fn log_sum_exp() -> Result<()> {
+    let input = Tensor::new(&[[1f64, 2., 3.], [4., 5., 6.]], &Device::Cpu)?;
+    let output = input.log_sum_exp(D::Minus1)?;
+    // The expectations obtained from pytorch.
+    let expected = Tensor::new(&[3.4076, 6.4076], &Device::Cpu)?;
+    assert_close(&output, &expected, 0.00001)?;
+    Ok(())
+}
+
+#[test]
+fn pow() -> Result<()> {
+    let lhs = Tensor::new(&[[1f32, 2., 3.], [4., 5., 6.]], &Device::Cpu)?;
+    let rhs = (&lhs - 2.)?;
+    let res = lhs.pow(&rhs)?;
+    assert_eq!(
+        test_utils::to_vec2_round(&res, 3)?,
+        [[1.0, 1.0, 3.0], [16.0, 125.0, 1296.0]]
+    );
+    Ok(())
+}

candle-datasets/Cargo.toml

@@ -11,8 +11,8 @@ readme = "README.md"
 
 [dependencies]
 byteorder = { workspace = true }
-candle = { path = "../candle-core", version = "0.3.0", package = "candle-core" }
-candle-nn = { path = "../candle-nn", version = "0.3.0" }
+candle = { workspace = true }
+candle-nn = { workspace = true }
 hf-hub = { workspace = true}
 intel-mkl-src = { workspace = true, optional = true }
 memmap2 = { workspace = true }

candle-examples/Cargo.toml

@@ -11,52 +11,60 @@ readme = "README.md"
 
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
-candle = { path = "../candle-core", version = "0.3.0", package = "candle-core" }
-candle-datasets = { path = "../candle-datasets", version = "0.3.0" }
-candle-nn = { path = "../candle-nn", version = "0.3.0" }
-candle-transformers = { path = "../candle-transformers", version = "0.3.0" }
-candle-flash-attn = { path = "../candle-flash-attn", version = "0.3.0", optional = true }
-candle-onnx = { path = "../candle-onnx", version = "0.3.0", optional = true }
+candle = { workspace = true }
+candle-datasets = { workspace = true, optional = true }
+candle-nn = { workspace = true }
+candle-transformers = { workspace = true }
+candle-flash-attn = { workspace = true, optional = true }
+candle-onnx = { workspace = true, optional = true }
+
+csv = "1.3.0"
 cudarc = { workspace = true, optional = true }
 half = { workspace = true, optional = true }
+hf-hub = { workspace = true, features = ["tokio"] }
 image = { workspace = true }
 intel-mkl-src = { workspace = true, optional = true }
 num-traits = { workspace = true }
-pyo3 = { version = "0.20.0", features = ["auto-initialize"], optional = true }
+pyo3 = { version = "0.21.0", features = ["auto-initialize"], optional = true }
 rayon = { workspace = true }
+rubato = { version = "0.15.0", optional = true }
 safetensors = { workspace = true }
 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 }
 
 [dev-dependencies]
 anyhow = { workspace = true }
 byteorder = { workspace = true }
 clap = { workspace = true }
-hf-hub = { workspace = true, features=["tokio"]}
 imageproc = { workspace = true }
 memmap2 = { workspace = true }
 rand = { workspace = true }
-rusttype = { workspace = true }
+ab_glyph = { workspace = true }
 tracing = { workspace = true }
 tracing-chrome = { workspace = true }
 tracing-subscriber = { workspace = true }
-wav = { workspace = true }
 # Necessary to disambiguate with tokio in wasm examples which are 1.28.1
 tokio = "1.29.1"
 
 [build-dependencies]
 anyhow = { workspace = true }
+bindgen_cuda = { version = "0.1.1", optional = true }
 
 [features]
 default = []
 accelerate = ["dep:accelerate-src", "candle/accelerate", "candle-nn/accelerate", "candle-transformers/accelerate"]
-cuda = ["candle/cuda", "candle-nn/cuda", "candle-transformers/cuda"]
+cuda = ["candle/cuda", "candle-nn/cuda", "candle-transformers/cuda", "dep:bindgen_cuda"]
 cudnn = ["candle/cudnn"]
 flash-attn = ["cuda", "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"]
+encodec = ["cpal", "symphonia", "rubato"]
 
 [[example]]
 name = "llama_multiprocess"
@@ -73,3 +81,23 @@ required-features = ["onnx"]
 [[example]]
 name = "onnx_basics"
 required-features = ["onnx"]
+
+[[example]]
+name = "whisper"
+required-features = ["symphonia"]
+
+[[example]]
+name = "whisper-microphone"
+required-features = ["microphone"]
+
+[[example]]
+name = "mnist-training"
+required-features = ["candle-datasets"]
+
+[[example]]
+name = "llama2-c"
+required-features = ["candle-datasets"]
+
+[[example]]
+name = "encodec"
+required-features = ["encodec"]

candle-examples/build.rs

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

candle-examples/examples/bert/README.md

@@ -2,10 +2,10 @@
 
 Bert is a general large language model. In this example it can be used for two
 different tasks:
+
 - Compute sentence embeddings for a prompt.
 - Compute similarities between a set of sentences.
 
-
 ## Sentence embeddings
 
 Bert is used to compute the sentence embeddings for a prompt. The model weights
@@ -24,6 +24,48 @@ cargo run --example bert --release -- --prompt "Here is a test sentence"
 > Tensor[[1, 7, 384], f32]
 ```
 
+### Custom models
+
+You can specify different models, such as BGE, with the `--model-id` flag:
+
+```bash
+cargo run  --example bert --release -- \
+--model-id BAAI/bge-large-zh-v1.5 \
+--prompt "Here is a test sentence"
+Loaded and encoded 435.70775ms
+[[[ 3.0944e-1, -7.8455e-5,  -1.2768e0, ...,  1.3755e-2, -3.2371e-1,  2.3819e-1],
+  [-2.8506e-1,  1.9953e-1,  -1.3076e0, ...,  6.9819e-2,  1.0833e-2,  -1.1512e0],
+  [ 3.9892e-1,  2.0000e-1, -9.3178e-1, ..., -4.1393e-1, -4.9644e-2, -3.3786e-1],
+  ...
+  [ 6.0345e-1,  3.5744e-1,  -1.2672e0, ..., -6.9165e-1, -3.4973e-3, -8.4214e-1],
+  [ 3.9218e-1, -3.2735e-1,  -1.3123e0, ..., -4.9318e-1, -5.1334e-1, -3.6391e-1],
+  [ 3.0978e-1,  2.5662e-4,  -1.2773e0, ...,  1.3357e-2, -3.2390e-1,  2.3858e-1]]]
+Tensor[[1, 9, 1024], f32]
+Took 176.744667ms
+```
+
+### Gelu approximation
+
+You can get a speedup by using an approximation of the gelu activation, with a
+small loss of precision, by passing the `--approximate-gelu` flag:
+
+```bash
+$ cargo run  --example bert --release -- \
+--model-id BAAI/bge-large-zh-v1.5 \
+--prompt "Here is a test sentence" \
+--approximate-gelu
+Loaded and encoded 244.388042ms
+[[[ 3.1048e-1, -6.0339e-4,  -1.2758e0, ...,  1.3718e-2, -3.2362e-1,  2.3775e-1],
+  [-2.8354e-1,  1.9984e-1,  -1.3077e0, ...,  6.9390e-2,  9.9681e-3,  -1.1531e0],
+  [ 3.9947e-1,  1.9917e-1, -9.3178e-1, ..., -4.1301e-1, -5.0719e-2, -3.3955e-1],
+  ...
+  [ 6.0499e-1,  3.5664e-1,  -1.2642e0, ..., -6.9134e-1, -3.4581e-3, -8.4471e-1],
+  [ 3.9311e-1, -3.2812e-1,  -1.3105e0, ..., -4.9291e-1, -5.1270e-1, -3.6543e-1],
+  [ 3.1082e-1, -2.6737e-4,  -1.2762e0, ...,  1.3319e-2, -3.2381e-1,  2.3815e-1]]]
+Tensor[[1, 9, 1024], f32]
+Took 116.840791ms
+```
+
 ## Similarities
 
 In this example, Bert is used to compute the sentence embeddings for a set of

candle-examples/examples/bert/main.rs

@@ -3,7 +3,7 @@ extern crate intel_mkl_src;
 
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
-use candle_transformers::models::bert::{BertModel, Config, DTYPE};
+use candle_transformers::models::bert::{BertModel, Config, HiddenAct, DTYPE};
 
 use anyhow::{Error as E, Result};
 use candle::Tensor;
@@ -45,6 +45,10 @@ struct Args {
     /// L2 normalization for embeddings.
     #[arg(long, default_value = "true")]
     normalize_embeddings: bool,
+
+    /// Use tanh based approximation for Gelu instead of erf implementation.
+    #[arg(long, default_value = "false")]
+    approximate_gelu: bool,
 }
 
 impl Args {
@@ -73,7 +77,7 @@ impl Args {
             (config, tokenizer, weights)
         };
         let config = std::fs::read_to_string(config_filename)?;
-        let config: Config = serde_json::from_str(&config)?;
+        let mut config: Config = serde_json::from_str(&config)?;
         let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
 
         let vb = if self.use_pth {
@@ -81,6 +85,9 @@ impl Args {
         } else {
             unsafe { VarBuilder::from_mmaped_safetensors(&[weights_filename], DTYPE, &device)? }
         };
+        if self.approximate_gelu {
+            config.hidden_act = HiddenAct::GeluApproximate;
+        }
         let model = BertModel::load(vb, &config)?;
         Ok((model, tokenizer))
     }

candle-examples/examples/blip/main.rs

@@ -106,17 +106,17 @@ pub fn main() -> anyhow::Result<()> {
 
     let config = blip::Config::image_captioning_large();
 
+    let device = candle_examples::device(args.cpu)?;
     let (image_embeds, device, mut model) = if args.quantized {
         let device = Device::Cpu;
         let image = load_image(args.image)?.to_device(&device)?;
         println!("loaded image {image:?}");
 
-        let vb = quantized_blip::VarBuilder::from_gguf(model_file)?;
+        let vb = quantized_blip::VarBuilder::from_gguf(model_file, &device)?;
         let model = quantized_blip::BlipForConditionalGeneration::new(&config, vb)?;
         let image_embeds = image.unsqueeze(0)?.apply(model.vision_model())?;
         (image_embeds, device, Model::Q(model))
     } else {
-        let device = candle_examples::device(args.cpu)?;
         let image = load_image(args.image)?.to_device(&device)?;
         println!("loaded image {image:?}");
 

candle-examples/examples/chatglm/main.rs

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

candle-examples/examples/clip/README.md

@@ -0,0 +1,46 @@
+Contrastive Language-Image Pre-Training
+
+Contrastive Language-Image Pre-Training (CLIP) is an architecture trained on
+pairs of images with related texts.
+
+https://github.com/openai/CLIP
+
+https://github.com/huggingface/transformers/tree/f6fa0f0bf0796ac66f201f23bdb8585de1609add/src/transformers/models/clip
+
+## Running on an example on cpu
+
+```
+$ cargo run --example clip --release -- --images "candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg","candle-examples/examples/yolo-v8/assets/bike.jpg" --cpu --sequences  "a cycling race","a photo of two cats","a robot holding a candle"
+
+
+Results for image: candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg
+
+INFO clip: Probability: 0.0000% Text: a cycling race
+INFO clip: Probability: 0.0000% Text: a photo of two cats
+INFO clip: Probability: 100.0000% Text: a robot holding a candle
+
+Results for image: candle-examples/examples/yolo-v8/assets/bike.jpg
+
+INFO clip: Probability: 99.9999% Text: a cycling race
+INFO clip: Probability: 0.0001% Text: a photo of two cats
+INFO clip: Probability: 0.0000% Text: a robot holding a candle
+```
+
+## Running on an example with metal feature (mac)
+
+```
+$ cargo run --features metal --example clip --release -- --images "candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg","candle-examples/examples/yolo-v8/assets/bike.jpg" --cpu --sequences "a cycling race","a photo of two cats","a robot holding a candle"
+
+
+Results for image: candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg
+
+INFO clip: Probability: 0.0000% Text: a cycling race
+INFO clip: Probability: 0.0000% Text: a photo of two cats
+INFO clip: Probability: 100.0000% Text: a robot holding a candle
+
+Results for image: candle-examples/examples/yolo-v8/assets/bike.jpg
+
+INFO clip: Probability: 99.9999% Text: a cycling race
+INFO clip: Probability: 0.0001% Text: a photo of two cats
+INFO clip: Probability: 0.0000% Text: a robot holding a candle
+```

candle-examples/examples/clip/main.rs

@@ -0,0 +1,202 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::Error as E;
+use clap::Parser;
+
+use candle::{DType, Device, Tensor};
+use candle_nn::{ops::softmax, VarBuilder};
+use candle_transformers::models::clip;
+
+use tokenizers::Tokenizer;
+use tracing::info;
+
+#[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 load_image<T: AsRef<std::path::Path>>(path: T, image_size: usize) -> anyhow::Result<Tensor> {
+    let img = image::io::Reader::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)
+}
+
+fn load_images<T: AsRef<std::path::Path>>(
+    paths: &Vec<T>,
+    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()?;
+
+            let api = api.repo(hf_hub::Repo::with_revision(
+                "openai/clip-vit-base-patch32".to_string(),
+                hf_hub::RepoType::Model,
+                "refs/pr/15".to_string(),
+            ));
+
+            api.get("model.safetensors")?
+        }
+        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![
+            "candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg".to_string(),
+            "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);
+
+    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);
+
+        for (i, p) in prob.iter().enumerate() {
+            info!("Probability: {:.4}% Text: {} ", p, vec_seq[i]);
+        }
+    }
+
+    Ok(())
+}
+
+pub fn get_tokenizer(tokenizer: Option<String>) -> anyhow::Result<Tokenizer> {
+    let tokenizer = match tokenizer {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.repo(hf_hub::Repo::with_revision(
+                "openai/clip-vit-base-patch32".to_string(),
+                hf_hub::RepoType::Model,
+                "refs/pr/15".to_string(),
+            ));
+            api.get("tokenizer.json")?
+        }
+        Some(file) => file.into(),
+    };
+
+    Tokenizer::from_file(tokenizer).map_err(E::msg)
+}
+
+pub fn tokenize_sequences(
+    sequences: Option<Vec<String>>,
+    tokenizer: &Tokenizer,
+    device: &Device,
+) -> anyhow::Result<(Tensor, Vec<String>)> {
+    let pad_id = *tokenizer
+        .get_vocab(true)
+        .get("<|endoftext|>")
+        .ok_or(E::msg("No pad token"))?;
+
+    let vec_seq = match sequences {
+        Some(seq) => seq,
+        None => vec![
+            "a cycling race".to_string(),
+            "a photo of two cats".to_string(),
+            "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();
+        if len_diff > 0 {
+            token_vec.extend(vec![pad_id; len_diff]);
+        }
+    }
+
+    let input_ids = Tensor::new(tokens, device)?;
+
+    Ok((input_ids, vec_seq))
+}

candle-examples/examples/convmixer/main.rs

@@ -28,7 +28,7 @@ pub fn main() -> anyhow::Result<()> {
 
     let device = candle_examples::device(args.cpu)?;
 
-    let image = candle_examples::imagenet::load_image224(args.image)?;
+    let image = candle_examples::imagenet::load_image224(args.image)?.to_device(&device)?;
     println!("loaded image {image:?}");
 
     let model_file = match args.model {

candle-examples/examples/convnext/README.md

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

candle-examples/examples/convnext/main.rs

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

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

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

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

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

candle-examples/examples/dinov2/main.rs

@@ -31,7 +31,7 @@ pub fn main() -> anyhow::Result<()> {
 
     let device = candle_examples::device(args.cpu)?;
 
-    let image = candle_examples::imagenet::load_image224(args.image)?;
+    let image = candle_examples::imagenet::load_image224(args.image)?.to_device(&device)?;
     println!("loaded image {image:?}");
 
     let model_file = match args.model {

candle-examples/examples/distilbert/README.md

@@ -0,0 +1,22 @@
+# candle-distilbert
+
+DistilBert is a distiled version of the Bert model.
+
+## Sentence embeddings
+
+DistilBert is used to compute the sentence embeddings for a prompt. The model weights
+are downloaded from the hub on the first run.
+
+```bash
+cargo run --example distilbert --release -- --prompt "Here is a test sentence"
+
+> [[[ 0.5109,  0.1280, -0.2635, ...,  0.3462, -1.0434,  0.1441],
+>   [ 0.1735,  0.0818, -0.5549, ...,  0.3472, -0.8264, -0.0244],
+>   [ 0.0702, -0.1311, -0.4914, ...,  0.3483, -0.6194,  0.1829],
+>   ...
+>   [ 0.2993, -0.0106, -0.4640, ...,  0.2844, -0.6732,  0.0042],
+>   [ 0.1066, -0.0081, -0.4299, ...,  0.3435, -0.7729,  0.0190],
+>   [ 0.8903,  0.2055, -0.2541, ...,  0.3208, -0.6585,  0.0586]]]
+> Tensor[[1, 7, 768], f32]
+
+```

candle-examples/examples/distilbert/main.rs

@@ -0,0 +1,135 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+use candle_transformers::models::distilbert::{Config, DistilBertModel, DTYPE};
+
+use anyhow::{Error as E, Result};
+use candle::{Device, Tensor};
+use candle_nn::VarBuilder;
+use clap::Parser;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::Tokenizer;
+
+#[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,
+
+    /// The model to use, check out available models: https://huggingface.co/models?library=sentence-transformers&sort=trending
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long)]
+    revision: Option<String>,
+
+    /// When set, compute embeddings for this prompt.
+    #[arg(long)]
+    prompt: String,
+
+    /// Use the pytorch weights rather than the safetensors ones
+    #[arg(long)]
+    use_pth: bool,
+
+    /// The number of times to run the prompt.
+    #[arg(long, default_value = "1")]
+    n: usize,
+
+    /// L2 normalization for embeddings.
+    #[arg(long, default_value = "true")]
+    normalize_embeddings: bool,
+}
+
+impl Args {
+    fn build_model_and_tokenizer(&self) -> Result<(DistilBertModel, Tokenizer)> {
+        let device = candle_examples::device(self.cpu)?;
+        let default_model = "distilbert-base-uncased".to_string();
+        let default_revision = "main".to_string();
+        let (model_id, revision) = match (self.model_id.to_owned(), self.revision.to_owned()) {
+            (Some(model_id), Some(revision)) => (model_id, revision),
+            (Some(model_id), None) => (model_id, "main".to_string()),
+            (None, Some(revision)) => (default_model, revision),
+            (None, None) => (default_model, default_revision),
+        };
+
+        let repo = Repo::with_revision(model_id, RepoType::Model, revision);
+        let (config_filename, tokenizer_filename, weights_filename) = {
+            let api = Api::new()?;
+            let api = api.repo(repo);
+            let config = api.get("config.json")?;
+            let tokenizer = api.get("tokenizer.json")?;
+            let weights = if self.use_pth {
+                api.get("pytorch_model.bin")?
+            } else {
+                api.get("model.safetensors")?
+            };
+            (config, tokenizer, weights)
+        };
+        let config = std::fs::read_to_string(config_filename)?;
+        let config: Config = serde_json::from_str(&config)?;
+        let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
+
+        let vb = if self.use_pth {
+            VarBuilder::from_pth(&weights_filename, DTYPE, &device)?
+        } else {
+            unsafe { VarBuilder::from_mmaped_safetensors(&[weights_filename], DTYPE, &device)? }
+        };
+        let model = DistilBertModel::load(vb, &config)?;
+        Ok((model, tokenizer))
+    }
+}
+
+fn get_mask(size: usize, device: &Device) -> Tensor {
+    let mask: Vec<_> = (0..size)
+        .flat_map(|i| (0..size).map(move |j| u8::from(j > i)))
+        .collect();
+    Tensor::from_slice(&mask, (size, size), device).unwrap()
+}
+
+fn main() -> Result<()> {
+    use tracing_chrome::ChromeLayerBuilder;
+    use tracing_subscriber::prelude::*;
+
+    let args = Args::parse();
+    let _guard = if args.tracing {
+        println!("tracing...");
+        let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
+        tracing_subscriber::registry().with(chrome_layer).init();
+        Some(guard)
+    } else {
+        None
+    };
+    let (model, mut tokenizer) = args.build_model_and_tokenizer()?;
+    let device = &model.device;
+
+    let tokenizer = tokenizer
+        .with_padding(None)
+        .with_truncation(None)
+        .map_err(E::msg)?;
+    let tokens = tokenizer
+        .encode(args.prompt, true)
+        .map_err(E::msg)?
+        .get_ids()
+        .to_vec();
+    let token_ids = Tensor::new(&tokens[..], device)?.unsqueeze(0)?;
+    let mask = get_mask(tokens.len(), device);
+
+    println!("token_ids: {:?}", token_ids.to_vec2::<u32>());
+    println!("mask: {:?}", mask.to_vec2::<u8>());
+
+    let ys = model.forward(&token_ids, &mask)?;
+    println!("{ys}");
+
+    Ok(())
+}
+
+pub fn normalize_l2(v: &Tensor) -> Result<Tensor> {
+    Ok(v.broadcast_div(&v.sqr()?.sum_keepdim(1)?.sqrt()?)?)
+}

candle-examples/examples/efficientnet/main.rs

@@ -47,7 +47,7 @@ pub fn main() -> anyhow::Result<()> {
 
     let device = candle_examples::device(args.cpu)?;
 
-    let image = candle_examples::imagenet::load_image224(args.image)?;
+    let image = candle_examples::imagenet::load_image224(args.image)?.to_device(&device)?;
     println!("loaded image {image:?}");
 
     let model_file = match args.model {