Merge remote-tracking branch 'origin/main' into cuda-conv-tr1d

* Add a specialized kernel for copy2d.

* Move the cat operations.

* Avoid transpositions in cat.

* Bugfix.

* Bugfix for the cuda kernel.

* Add a benchmark.

* Add more testing.

* Test fix.

* Faster kernel.

* Add the missing kernel.

* Tweak the test.

* Add a metal kernel.

* Fix for the metal kernel.

* Get the tests to pass on metal.

* Also use this opportunity to fix the metal kernel for ELU.

* Add some bf16 kernels.

* Clippy fixes.

.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

@@ -19,7 +19,7 @@ exclude = [
 resolver = "2"
 
 [workspace.package]
-version = "0.3.1"
+version = "0.4.2"
 edition = "2021"
 description = "Minimalist ML framework."
 repository = "https://github.com/huggingface/candle"
@@ -31,9 +31,19 @@ license = "MIT OR Apache-2.0"
 accelerate-src = { version = "0.3.2" }
 anyhow = { version = "1", features = ["backtrace"] }
 byteorder = "1.4.3"
+candle = { path = "./candle-core", package = "candle-core", version = "0.4.2" }
+candle-datasets = { path = "./candle-datasets", version = "0.4.2" }
+candle-flash-attn = { path = "./candle-flash-attn", version = "0.4.2" }
+candle-kernels = { path = "./candle-kernels", version = "0.4.2" }
+candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.4.2" }
+candle-nn = { path = "./candle-nn", version = "0.4.2" }
+candle-onnx = { path = "./candle-onnx", version = "0.4.2" }
+candle-transformers = { path = "./candle-transformers", version = "0.4.2" }
 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 = "0.10.0", features = ["f16"] }
+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"] }
@@ -41,27 +51,27 @@ imageproc = { version = "0.23.0", default-features = false }
 intel-mkl-src = { version = "0.8.1", features = ["mkl-static-lp64-iomp"] }
 libc = { version = "0.2.147" }
 log = "0.4"
-memmap2 = { 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 = "50.0.0" }
 rand = "0.8.5"
 rand_distr = "0.4.3"
 rayon = "1.7.0"
 rusttype = { version = "0.9", default-features = false }
-safetensors = "0.3.1"
+safetensors = "0.4.1"
 serde = { version = "1.0.171", features = ["derive"] }
 serde_plain = "1.0.2"
 serde_json = "1.0.99"
 thiserror = "1"
-tokenizers = { version = "0.13.4", default-features = false }
+tokenizers = { version = "0.15.0", default-features = false }
 tracing = "0.1.37"
 tracing-chrome = "0.7.1"
 tracing-subscriber = "0.3.7"
 wav = "1.0.0"
 yoke = { version = "0.7.2", features = ["derive"] }
 zip = { version = "0.6.6", default-features = false }
-metal = { version = "0.27.1", features = ["mps"], package="candle-metal" }
+metal = { version = "0.27.0", features = ["mps"]}
 
 [profile.release-with-debug]
 inherits = "release"

README.md

@@ -54,20 +54,33 @@ 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 and LLaMA-v2](./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.
+- [Phi-1, Phi-1.5, and Phi-2](./candle-examples/examples/phi/): 1.3b and 2.7b general LLMs with performance on par with LLaMA-v2 7b.
 - [StableLM-3B-4E1T](./candle-examples/examples/stable-lm/): a 3b general LLM
-  pre-trained on 1T tokens of English and code datasets.
+  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.
@@ -78,7 +91,7 @@ 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">
 
@@ -97,14 +110,23 @@ We also provide a some command line based examples using state of the art models
 
 <img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/segment-anything/assets/sam_merged.jpg" width="200">
 
+- [SegFormer](./candle-examples/examples/segformer/): transformer based semantic segmantation model.
 - [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.
+- [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.
 
@@ -122,7 +144,7 @@ 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
@@ -141,8 +163,10 @@ And then head over to
 ## Useful External Resources
 - [`candle-tutorial`](https://github.com/ToluClassics/candle-tutorial): A
   very detailed tutorial showing how to convert a PyTorch model to Candle.
-- [`candle-lora`](https://github.com/EricLBuehler/candle-lora): Efficient and ergonomic LoRA implemenation 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).
+- [`candle-lora`](https://github.com/EricLBuehler/candle-lora): Efficient and
+  ergonomic LoRA implementation for Candle. `candle-lora` has      
+  out-of-the-box LoRA support for many models from Candle, which can be found
+  [here](https://github.com/EricLBuehler/candle-lora/tree/master/candle-lora-transformers/examples).
 - [`optimisers`](https://github.com/KGrewal1/optimisers): A collection of optimisers
   including SGD with momentum, AdaGrad, AdaDelta, AdaMax, NAdam, RAdam, and RMSprop.
 - [`candle-vllm`](https://github.com/EricLBuehler/candle-vllm): Efficient platform for inference and
@@ -150,6 +174,8 @@ And then head over to
 - [`candle-ext`](https://github.com/mokeyish/candle-ext): An extension library to Candle that provides PyTorch functions not currently available in Candle.
 - [`kalosm`](https://github.com/floneum/floneum/tree/master/interfaces/kalosm): A multi-modal meta-framework in Rust for interfacing with local pre-trained models with support for controlled generation, custom samplers, in-memory vector databases, audio transcription, and more.
 - [`candle-sampling`](https://github.com/EricLBuehler/candle-sampling): Sampling techniques for Candle.
+- [`gpt-from-scratch-rs`](https://github.com/jeroenvlek/gpt-from-scratch-rs): A port of Andrej Karpathy's _Let's build GPT_ tutorial on YouTube showcasing the Candle API on a toy problem.
+- [`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.
 
@@ -168,33 +194,45 @@ 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 and v2 with variants such as SOLAR-10.7B.
         - Falcon.
-        - StarCoder.
-        - Phi v1.5.
+        - StarCoder, StarCoder2.
+        - Phi 1, 1.5, and 2.
+        - 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.
+        - RWKV v5 and v6.
     - Quantized LLMs.
         - Llama 7b, 13b, 70b, as well as the chat and code variants.
         - Mistral 7b, and 7b instruct.
-        - Zephyr 7b a and b (Mistral based).
-        - OpenChat 3.5 (Mistral based).
+        - Mixtral 8x7b.
+        - Zephyr 7b a and b (Mistral-7b based).
+        - OpenChat 3.5 (Mistral-7b based).
     - Text to text.
         - T5 and its variants: FlanT5, UL2, MADLAD400 (translation), CoEdit (Grammar correction).
         - Marian MT (Machine Translation).
-    - Whisper (multi-lingual support).
     - 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.

candle-book/Cargo.toml

@@ -11,11 +11,11 @@ readme = "README.md"
 
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
-candle = { path = "../candle-core", version = "0.3.1", package = "candle-core" }
-candle-datasets = { path = "../candle-datasets", version = "0.3.1" }
-candle-nn = { path = "../candle-nn", version = "0.3.1" }
-candle-transformers = { path = "../candle-transformers", version = "0.3.1" }
-candle-flash-attn = { path = "../candle-flash-attn", version = "0.3.1", optional = true }
+candle = { workspace = true }
+candle-datasets = { workspace = true }
+candle-nn = { workspace = true }
+candle-transformers = { workspace = true }
+candle-flash-attn = { workspace = true, optional = true }
 safetensors = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }

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;
@@ -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

@@ -12,8 +12,8 @@ readme = "README.md"
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
 byteorder = { workspace = true }
-candle-kernels = { path = "../candle-kernels", version = "0.3.1", optional = true }
-candle-metal-kernels = { path = "../candle-metal-kernels", version = "0.3.1", optional = true }
+candle-kernels = { workspace = true, optional = true }
+candle-metal-kernels = { workspace = true, optional = true }
 metal = { workspace = true, optional = true}
 cudarc = { workspace = true, optional = true }
 gemm = { workspace = true }
@@ -34,6 +34,8 @@ zip = { workspace = true }
 [dev-dependencies]
 anyhow = { workspace = true }
 clap = { workspace = true }
+criterion = { workspace = true }
+
 
 [features]
 default = []
@@ -42,3 +44,7 @@ cudnn = ["cuda", "cudarc/cudnn"]
 mkl = ["dep:libc", "dep:intel-mkl-src"]
 accelerate = ["dep:libc", "dep:accelerate-src"]
 metal = ["dep:metal", "dep:candle-metal-kernels"]
+
+[[bench]]
+name = "bench_main"
+harness = false

candle-core/benches/bench_main.rs

@@ -0,0 +1,9 @@
+mod benchmarks;
+
+use criterion::criterion_main;
+criterion_main!(
+    benchmarks::affine::benches,
+    benchmarks::matmul::benches,
+    benchmarks::random::benches,
+    benchmarks::where_cond::benches
+);

candle-core/benches/benchmarks/affine.rs

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

candle-core/benches/benchmarks/matmul.rs

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

candle-core/benches/benchmarks/mod.rs

@@ -0,0 +1,66 @@
+pub(crate) mod affine;
+pub(crate) mod matmul;
+pub(crate) mod random;
+pub(crate) mod where_cond;
+
+use candle_core::{Device, Result};
+
+pub(crate) trait BenchDevice {
+    fn sync(&self) -> Result<()>;
+
+    fn bench_name<S: Into<String>>(&self, name: S) -> String;
+}
+
+impl BenchDevice for Device {
+    fn sync(&self) -> Result<()> {
+        match self {
+            Device::Cpu => Ok(()),
+            Device::Cuda(device) => {
+                #[cfg(feature = "cuda")]
+                return Ok(device.synchronize()?);
+                #[cfg(not(feature = "cuda"))]
+                panic!("Cuda device without cuda feature enabled: {:?}", device)
+            }
+            Device::Metal(device) => {
+                #[cfg(feature = "metal")]
+                return Ok(device.wait_until_completed()?);
+                #[cfg(not(feature = "metal"))]
+                panic!("Metal device without metal feature enabled: {:?}", device)
+            }
+        }
+    }
+
+    fn bench_name<S: Into<String>>(&self, name: S) -> String {
+        match self {
+            Device::Cpu => {
+                let cpu_type = if cfg!(feature = "accelerate") {
+                    "accelerate"
+                } else if cfg!(feature = "mkl") {
+                    "mkl"
+                } else {
+                    "cpu"
+                };
+                format!("{}_{}", cpu_type, name.into())
+            }
+            Device::Cuda(_) => format!("cuda_{}", name.into()),
+            Device::Metal(_) => format!("metal_{}", name.into()),
+        }
+    }
+}
+
+struct BenchDeviceHandler {
+    devices: Vec<Device>,
+}
+
+impl BenchDeviceHandler {
+    pub fn new() -> Result<Self> {
+        let mut devices = Vec::new();
+        if cfg!(feature = "metal") {
+            devices.push(Device::new_metal(0)?);
+        } else if cfg!(feature = "cuda") {
+            devices.push(Device::new_cuda(0)?);
+        }
+        devices.push(Device::Cpu);
+        Ok(Self { devices })
+    }
+}

candle-core/benches/benchmarks/random.rs

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

candle-core/benches/benchmarks/where_cond.rs

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

candle-core/examples/cuda_basics.rs

@@ -5,25 +5,32 @@ extern crate accelerate_src;
 extern crate intel_mkl_src;
 
 use anyhow::Result;
-use candle_core::{Device, Tensor};
+use candle_core::{Device, Module, Tensor};
+
+use candle_core::quantized::{QMatMul, QTensor};
 
 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 q = Tensor::randn(0f32, 1.0, (72, 256), &device)?;
+    let q_cpu = q.to_device(&Device::Cpu)?;
+    let q = QTensor::quantize(&q, candle_core::quantized::GgmlDType::Q8K)?;
+    let q = QMatMul::from_qtensor(q)?;
+    let x = Tensor::randn(0f32, 1.0, (5, 256), &device)?;
+    let res_q_cuda = q.forward(&x)?;
+    println!("{res_q_cuda}");
 
-    let 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 q_cpu = QTensor::quantize(&q_cpu, candle_core::quantized::GgmlDType::Q8K)?;
+    let q_cpu_tensor = q_cpu.dequantize(&Device::Cpu)?;
+    let q_cpu = QMatMul::from_qtensor(q_cpu)?;
+    let x_cpu = x.to_device(&Device::Cpu)?;
+    let res_q_cpu = q_cpu.forward(&x_cpu)?;
+    println!("{res_q_cpu}");
+
+    let res_mm = x_cpu.matmul(&q_cpu_tensor.t()?)?;
+    let diff = (res_mm - res_q_cuda.to_device(&Device::Cpu))?
+        .abs()?
+        .flatten_all()?
+        .max(0)?;
+    println!("{diff}");
     Ok(())
 }

candle-core/examples/tensor-tools.rs

@@ -1,5 +1,5 @@
-use candle_core::quantized::{gguf_file, k_quants, QTensor};
-use candle_core::{Device, Result, Tensor};
+use candle_core::quantized::{gguf_file, GgmlDType, QTensor};
+use candle_core::{Device, Result};
 use clap::{Parser, Subcommand, ValueEnum};
 use rayon::prelude::*;
 
@@ -11,12 +11,7 @@ enum QuantizationMode {
 }
 
 impl QuantizationMode {
-    fn quantize(
-        &self,
-        name: &str,
-        tensor: QTensor,
-        default: fn(&Tensor) -> Result<QTensor>,
-    ) -> Result<QTensor> {
+    fn quantize(&self, name: &str, tensor: QTensor, dtype: GgmlDType) -> Result<QTensor> {
         match self {
             Self::Llama => {
                 // Same behavior as the llama.cpp quantization.
@@ -24,9 +19,9 @@ impl QuantizationMode {
                 if should_quantize {
                     let tensor = tensor.dequantize(&Device::Cpu)?;
                     if name == "output.weight" {
-                        QTensor::quantize::<k_quants::BlockQ6K>(&tensor)
+                        QTensor::quantize(&tensor, GgmlDType::Q6K)
                     } else {
-                        default(&tensor)
+                        QTensor::quantize(&tensor, dtype)
                     }
                 } else {
                     Ok(tensor)
@@ -60,6 +55,27 @@ enum Quantization {
     F32,
 }
 
+impl Quantization {
+    fn dtype(&self) -> GgmlDType {
+        match self {
+            Quantization::Q4_0 => GgmlDType::Q4_0,
+            Quantization::Q4_1 => GgmlDType::Q4_1,
+            Quantization::Q5_0 => GgmlDType::Q5_0,
+            Quantization::Q5_1 => GgmlDType::Q5_1,
+            Quantization::Q8_0 => GgmlDType::Q8_0,
+            Quantization::Q8_1 => GgmlDType::Q8_1,
+            Quantization::Q2k => GgmlDType::Q2K,
+            Quantization::Q3k => GgmlDType::Q3K,
+            Quantization::Q4k => GgmlDType::Q4K,
+            Quantization::Q5k => GgmlDType::Q5K,
+            Quantization::Q6k => GgmlDType::Q6K,
+            Quantization::Q8k => GgmlDType::Q8K,
+            Quantization::F16 => GgmlDType::F16,
+            Quantization::F32 => GgmlDType::F32,
+        }
+    }
+}
+
 #[derive(ValueEnum, Debug, Clone)]
 enum Format {
     Safetensors,
@@ -102,7 +118,7 @@ enum Command {
     },
 
     Quantize {
-        /// The input file, in gguf format.
+        /// The input file(s), in safetensors format.
         in_file: Vec<std::path::PathBuf>,
 
         /// The output file, in gguf format.
@@ -117,6 +133,15 @@ enum Command {
         #[arg(long, value_enum, default_value_t = QuantizationMode::Llama)]
         mode: QuantizationMode,
     },
+
+    Dequantize {
+        /// The input file, in gguf format.
+        in_file: std::path::PathBuf,
+
+        /// The output file, in safetensors format.
+        #[arg(long)]
+        out_file: std::path::PathBuf,
+    },
 }
 
 #[derive(Parser, Debug, Clone)]
@@ -125,7 +150,12 @@ struct Args {
     command: Command,
 }
 
-fn run_ls(file: &std::path::PathBuf, format: Option<Format>, verbose: bool) -> Result<()> {
+fn run_ls(
+    file: &std::path::PathBuf,
+    format: Option<Format>,
+    verbose: bool,
+    device: &Device,
+) -> Result<()> {
     let format = match format {
         Some(format) => format,
         None => match Format::infer(file) {
@@ -166,7 +196,7 @@ fn run_ls(file: &std::path::PathBuf, format: Option<Format>, verbose: bool) -> R
             }
         }
         Format::Pth => {
-            let mut tensors = candle_core::pickle::read_pth_tensor_info(file, verbose)?;
+            let mut tensors = candle_core::pickle::read_pth_tensor_info(file, verbose, None)?;
             tensors.sort_by(|a, b| a.name.cmp(&b.name));
             for tensor_info in tensors.iter() {
                 println!(
@@ -191,7 +221,7 @@ fn run_ls(file: &std::path::PathBuf, format: Option<Format>, verbose: bool) -> R
         }
         Format::Ggml => {
             let mut file = std::fs::File::open(file)?;
-            let content = candle_core::quantized::ggml_file::Content::read(&mut file)?;
+            let content = candle_core::quantized::ggml_file::Content::read(&mut file, device)?;
             let mut tensors = content.tensors.into_iter().collect::<Vec<_>>();
             tensors.sort_by(|a, b| a.0.cmp(&b.0));
             for (name, qtensor) in tensors.iter() {
@@ -232,37 +262,8 @@ fn run_quantize_safetensors(
     }
     println!("tensors: {}", tensors.len());
 
-    let quantize_fn = match q {
-        Quantization::Q4_0 => QTensor::quantize::<k_quants::BlockQ4_0>,
-        Quantization::Q4_1 => QTensor::quantize::<k_quants::BlockQ4_1>,
-        Quantization::Q5_0 => QTensor::quantize::<k_quants::BlockQ5_0>,
-        Quantization::Q5_1 => QTensor::quantize::<k_quants::BlockQ5_1>,
-        Quantization::Q8_0 => QTensor::quantize::<k_quants::BlockQ8_0>,
-        Quantization::Q8_1 => QTensor::quantize::<k_quants::BlockQ8_1>,
-        Quantization::Q2k => QTensor::quantize::<k_quants::BlockQ2K>,
-        Quantization::Q3k => QTensor::quantize::<k_quants::BlockQ3K>,
-        Quantization::Q4k => QTensor::quantize::<k_quants::BlockQ4K>,
-        Quantization::Q5k => QTensor::quantize::<k_quants::BlockQ5K>,
-        Quantization::Q6k => QTensor::quantize::<k_quants::BlockQ6K>,
-        Quantization::Q8k => QTensor::quantize::<k_quants::BlockQ8K>,
-        Quantization::F16 => QTensor::quantize::<half::f16>,
-        Quantization::F32 => QTensor::quantize::<f32>,
-    };
-    let block_size = match q {
-        Quantization::Q4_0 => k_quants::QK4_0,
-        Quantization::Q4_1 => k_quants::QK4_1,
-        Quantization::Q5_0 => k_quants::QK5_0,
-        Quantization::Q5_1 => k_quants::QK5_1,
-        Quantization::Q8_0 => k_quants::QK8_0,
-        Quantization::Q8_1 => k_quants::QK8_1,
-        Quantization::Q2k
-        | Quantization::Q3k
-        | Quantization::Q4k
-        | Quantization::Q5k
-        | Quantization::Q6k
-        | Quantization::Q8k => k_quants::QK_K,
-        Quantization::F16 | Quantization::F32 => 1,
-    };
+    let dtype = q.dtype();
+    let block_size = dtype.block_size();
 
     let qtensors = tensors
         .into_par_iter()
@@ -270,9 +271,9 @@ fn run_quantize_safetensors(
             let should_quantize = tensor.rank() == 2 && tensor.dim(1)? % block_size == 0;
             println!("  quantizing {name} {tensor:?} {should_quantize}");
             let tensor = if should_quantize {
-                quantize_fn(&tensor)?
+                QTensor::quantize(&tensor, dtype)?
             } else {
-                QTensor::quantize::<f32>(&tensor)?
+                QTensor::quantize(&tensor, GgmlDType::F32)?
             };
             Ok((name, tensor))
         })
@@ -285,11 +286,29 @@ fn run_quantize_safetensors(
     Ok(())
 }
 
+fn run_dequantize(
+    in_file: std::path::PathBuf,
+    out_file: std::path::PathBuf,
+    device: &Device,
+) -> Result<()> {
+    let mut in_file = std::fs::File::open(in_file)?;
+    let content = gguf_file::Content::read(&mut in_file)?;
+    let mut tensors = std::collections::HashMap::new();
+    for (tensor_name, _) in content.tensor_infos.iter() {
+        let tensor = content.tensor(&mut in_file, tensor_name, device)?;
+        let tensor = tensor.dequantize(device)?;
+        tensors.insert(tensor_name.to_string(), tensor);
+    }
+    candle_core::safetensors::save(&tensors, out_file)?;
+    Ok(())
+}
+
 fn run_quantize(
     in_files: &[std::path::PathBuf],
     out_file: std::path::PathBuf,
     q: Quantization,
     qmode: QuantizationMode,
+    device: &Device,
 ) -> Result<()> {
     if in_files.is_empty() {
         candle_core::bail!("no specified input files")
@@ -315,31 +334,15 @@ fn run_quantize(
     let content = gguf_file::Content::read(&mut in_)?;
     println!("tensors: {}", content.tensor_infos.len());
 
-    let quantize_fn = match q {
-        Quantization::Q4_0 => QTensor::quantize::<k_quants::BlockQ4_0>,
-        Quantization::Q4_1 => QTensor::quantize::<k_quants::BlockQ4_1>,
-        Quantization::Q5_0 => QTensor::quantize::<k_quants::BlockQ5_0>,
-        Quantization::Q5_1 => QTensor::quantize::<k_quants::BlockQ5_1>,
-        Quantization::Q8_0 => QTensor::quantize::<k_quants::BlockQ8_0>,
-        Quantization::Q8_1 => QTensor::quantize::<k_quants::BlockQ8_1>,
-        Quantization::Q2k => QTensor::quantize::<k_quants::BlockQ2K>,
-        Quantization::Q3k => QTensor::quantize::<k_quants::BlockQ3K>,
-        Quantization::Q4k => QTensor::quantize::<k_quants::BlockQ4K>,
-        Quantization::Q5k => QTensor::quantize::<k_quants::BlockQ5K>,
-        Quantization::Q6k => QTensor::quantize::<k_quants::BlockQ6K>,
-        Quantization::Q8k => QTensor::quantize::<k_quants::BlockQ8K>,
-        Quantization::F16 => QTensor::quantize::<half::f16>,
-        Quantization::F32 => QTensor::quantize::<f32>,
-    };
-
+    let dtype = q.dtype();
     let qtensors = content
         .tensor_infos
         .par_iter()
         .map(|(name, _)| {
             println!("  quantizing {name}");
             let mut in_file = std::fs::File::open(&in_files[0])?;
-            let tensor = content.tensor(&mut in_file, name)?;
-            let tensor = qmode.quantize(name, tensor, quantize_fn)?;
+            let tensor = content.tensor(&mut in_file, name, device)?;
+            let tensor = qmode.quantize(name, tensor, dtype)?;
             Ok((name, tensor))
         })
         .collect::<Result<Vec<_>>>()?;
@@ -359,6 +362,7 @@ fn run_quantize(
 
 fn main() -> anyhow::Result<()> {
     let args = Args::parse();
+    let device = Device::Cpu;
     match args.command {
         Command::Ls {
             files,
@@ -370,7 +374,7 @@ fn main() -> anyhow::Result<()> {
                 if multiple_files {
                     println!("--- {file:?} ---");
                 }
-                run_ls(file, format.clone(), verbose)?
+                run_ls(file, format.clone(), verbose, &device)?
             }
         }
         Command::Quantize {
@@ -378,7 +382,8 @@ fn main() -> anyhow::Result<()> {
             out_file,
             quantization,
             mode,
-        } => run_quantize(&in_file, out_file, quantization, mode)?,
+        } => run_quantize(&in_file, out_file, quantization, mode, &device)?,
+        Command::Dequantize { in_file, out_file } => run_dequantize(in_file, out_file, &device)?,
     }
     Ok(())
 }

candle-core/src/accelerate.rs

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

candle-core/src/backend.rs

@@ -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 {

candle-core/src/backprop.rs

@@ -113,8 +113,8 @@ impl Tensor {
                     | Op::Unary(_node, UnaryOp::Floor)
                     | Op::Unary(_node, UnaryOp::Round) => 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 +175,7 @@ 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 +250,7 @@ impl Tensor {
                             out_padding,
                             *stride,
                             *dilation,
+                            /* groups */ 1,
                         )?;
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&grad_arg)?;
@@ -347,12 +348,39 @@ 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)?)?;
@@ -571,6 +599,13 @@ 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 = (*node / arg)?;
+                        let silu_grad = (&sigmoid_arg * (1. + (arg * (1. - &sigmoid_arg)?)?)?)?;
+                        *sum_grad = sum_grad.add(&(&grad * silu_grad)?)?
+                    }
                     Op::Elu(arg, alpha) => {
                         // d/dx elu(x) = 1 for x > 0, alpha * e^x for x <= 0
                         let sum_grad = grads.or_insert(arg)?;

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_backend.rs

@@ -5,6 +5,7 @@ use half::{bf16, f16};
 use rayon::prelude::*;
 
 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 +
@@ -1022,6 +1023,26 @@ 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 +1277,34 @@ 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 +1312,7 @@ 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();
@@ -2422,6 +2472,48 @@ 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),
@@ -2510,7 +2602,52 @@ 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(
@@ -2574,7 +2711,7 @@ 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 +2720,7 @@ 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 +2737,7 @@ 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()),
         }
     }
 

candle-core/src/cuda_backend.rs

@@ -608,6 +608,34 @@ impl Map1 for Elu {
     }
 }
 
+struct Col2Im1D {
+    stride: usize,
+}
+
+impl Map1 for Col2Im1D {
+    fn f<T: DeviceRepr + WithDType>(
+        &self,
+        src: &CudaSlice<T>,
+        dev: &CudaDevice,
+        layout: &Layout,
+    ) -> Result<CudaSlice<T>> {
+        let (b_size, l_in, c_out, k_size) = layout.shape().dims4()?;
+        let stride = self.stride;
+        let l_out = (l_in - 1) * stride + k_size;
+
+        let dst_el = b_size * c_out * l_out;
+        let cfg = LaunchConfig::for_num_elems(dst_el as u32);
+        let src = &src.slice(layout.start_offset()..);
+        let func = dev.get_or_load_func(&kernel_name::<T>("col2im1d"), kernels::CONV)?;
+        // SAFETY: Set later by running the kernel.
+        let dst = unsafe { dev.alloc::<T>(dst_el) }.w()?;
+        let params = (l_in, l_out, c_out, k_size, b_size, stride, src, &dst);
+        // SAFETY: ffi.
+        unsafe { func.launch(cfg, params) }.w()?;
+        Ok(dst)
+    }
+}
+
 struct Im2Col1D {
     l_k: usize,
     stride: usize,
@@ -1149,6 +1177,55 @@ impl<'a> Map2 for Conv2D<'a> {
     }
 }
 
+struct ConvTranspose1D<'a>(&'a crate::conv::ParamsConvTranspose1D);
+impl<'a> Map2 for ConvTranspose1D<'a> {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        inp: &CudaSlice<T>,
+        inp_l: &Layout,
+        k: &CudaSlice<T>,
+        k_l: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<CudaSlice<T>> {
+        // Kernel shape: (c_in_k, c_out, l_k)
+        // Input shape: (b_size, c_in, l_in)
+        let p = &self.0;
+        let l_out = p.l_out();
+        let dst_el = p.c_out * l_out * p.b_size;
+        let inp = &inp.slice(inp_l.start_offset()..);
+        let k = &k.slice(k_l.start_offset()..);
+        let shape = inp_l.shape();
+        let dims = shape.dims();
+        let el = shape.elem_count();
+
+        // SAFETY: Set later by running the kernel.
+        let out = unsafe { dev.alloc::<T>(dst_el) }.w()?;
+        let cfg = LaunchConfig::for_num_elems(dst_el as u32);
+        let func = dev.get_or_load_func(&kernel_name::<T>("conv_transpose1d"), kernels::CONV)?;
+        let ds = if dims.len() == 3 {
+            [dims, inp_l.stride(), k_l.dims(), k_l.stride()].concat()
+        } else {
+            crate::bail!("unexpected input shape for conv_transpose1d {dims:?}")
+        };
+        let ds = dev.htod_copy(ds).w()?;
+        let params = (
+            el,
+            l_out,
+            p.stride,
+            p.padding,
+            p.output_padding,
+            p.dilation,
+            &ds,
+            inp,
+            k,
+            &out,
+        );
+        // SAFETY: ffi.
+        unsafe { func.launch(cfg, params) }.w()?;
+        Ok(out)
+    }
+}
+
 struct ConvTranspose2D<'a>(&'a crate::conv::ParamsConvTranspose2D);
 impl<'a> Map2 for ConvTranspose2D<'a> {
     fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
@@ -1810,12 +1887,61 @@ impl BackendStorage for CudaStorage {
 
     fn conv_transpose1d(
         &self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: &crate::conv::ParamsConvTranspose1D,
+        l: &Layout,
+        kernel: &Self,
+        kernel_l: &Layout,
+        params: &crate::conv::ParamsConvTranspose1D,
     ) -> Result<Self> {
-        todo!()
+        let device = self.device().clone();
+        const USE_COL2IM_CONV1D_TR: bool = true;
+
+        let can_use_col2im = kernel_l.is_contiguous()
+            && params.dilation == 1
+            && params.padding == 0
+            && params.output_padding == 0;
+        if !can_use_col2im || !USE_COL2IM_CONV1D_TR {
+            let slice =
+                ConvTranspose1D(params).map(&self.slice, l, &kernel.slice, kernel_l, &device)?;
+            return Ok(Self { slice, device });
+        }
+
+        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));
+        let slice = Col2Im1D {
+            stride: params.stride,
+        }
+        .map(&col.slice, &device, &col_l)?;
+        Ok(Self { slice, device })
     }
 
     #[cfg(not(feature = "cudnn"))]
@@ -2093,6 +2219,67 @@ impl BackendStorage for CudaStorage {
         Ok(Self { slice, device })
     }
 
+    fn copy2d(
+        &self,
+        dst: &mut Self,
+        d1: usize,
+        d2: usize,
+        src_s: usize,
+        dst_s: usize,
+        src_o: usize,
+        dst_o: usize,
+    ) -> Result<()> {
+        let dev = &self.device;
+        let d1 = d1 as u32;
+        let d2 = d2 as u32;
+        let dst_s = dst_s as u32;
+        let src_s = src_s as u32;
+        let (src, dst, kname) = match (&self.slice, &mut dst.slice) {
+            (S::U8(s), S::U8(d)) => (
+                *s.slice(src_o..).device_ptr(),
+                *d.slice(dst_o..).device_ptr(),
+                "copy2d_u8",
+            ),
+            (S::U32(s), S::U32(d)) => (
+                *s.slice(src_o..).device_ptr(),
+                *d.slice(dst_o..).device_ptr(),
+                "copy2d_u32",
+            ),
+            (S::I64(s), S::I64(d)) => (
+                *s.slice(src_o..).device_ptr(),
+                *d.slice(dst_o..).device_ptr(),
+                "copy2d_i64",
+            ),
+            (S::BF16(s), S::BF16(d)) => (
+                *s.slice(src_o..).device_ptr(),
+                *d.slice(dst_o..).device_ptr(),
+                "copy2d_bf16",
+            ),
+            (S::F16(s), S::F16(d)) => (
+                *s.slice(src_o..).device_ptr(),
+                *d.slice(dst_o..).device_ptr(),
+                "copy2d_f16",
+            ),
+            (S::F32(s), S::F32(d)) => (
+                *s.slice(src_o..).device_ptr(),
+                *d.slice(dst_o..).device_ptr(),
+                "copy2d_f32",
+            ),
+            (S::F64(s), S::F64(d)) => (
+                *s.slice(src_o..).device_ptr(),
+                *d.slice(dst_o..).device_ptr(),
+                "copy2d_f64",
+            ),
+            _ => Err(CudaError::InternalError("dtype mismatch in copy2d"))?,
+        };
+        let func = dev.get_or_load_func(kname, kernels::FILL)?;
+        let cfg = LaunchConfig::for_num_elems(d1 * d2);
+        let params = (src, dst, d1, d2, src_s, dst_s);
+        // SAFETY: ffi.
+        unsafe { func.launch(cfg, params) }.w()?;
+        Ok(())
+    }
+
     fn copy_strided_src(&self, dst: &mut Self, dst_offset: usize, src_l: &Layout) -> Result<()> {
         let src_shape = src_l.shape();
         let dims = src_shape.dims();

candle-core/src/device.rs

@@ -201,10 +201,9 @@ impl Device {
                     Ok(Storage::Cuda(storage))
                 }
             }
-            Device::Metal(_device) => {
-                // let storage = device.rand_uniform(shape, dtype, lo, up)?;
-                // Ok(Storage::Metal(storage))
-                crate::bail!("Metal rand_uniform not implemented")
+            Device::Metal(device) => {
+                let storage = device.rand_uniform(shape, dtype, lo, up)?;
+                Ok(Storage::Metal(storage))
             }
         }
     }

candle-core/src/display.rs

@@ -65,12 +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> =
@@ -89,6 +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
 }
@@ -117,6 +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,
 }

candle-core/src/dtype.rs

@@ -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())),
         }
     }
 }

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)
     }

candle-core/src/dummy_metal_backend.rs

@@ -166,6 +166,19 @@ impl crate::backend::BackendStorage for MetalStorage {
         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)
     }

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>),

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

@@ -67,12 +67,13 @@ pub mod shape;
 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 device::{Device, DeviceLocation, NdArray};
 pub use dtype::{DType, FloatDType, IntDType, WithDType};
 pub use error::{Error, Result};
 pub use indexer::IndexOp;
@@ -129,6 +130,15 @@ impl<T: Fn(&Tensor) -> Result<Tensor>> Module for T {
     }
 }
 
+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/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/op.rs

@@ -61,6 +61,7 @@ pub enum UnaryOp {
     GeluErf,
     Erf,
     Relu,
+    Silu,
     Tanh,
     Floor,
     Ceil,
@@ -131,8 +132,15 @@ 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),
 
@@ -386,6 +394,7 @@ 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;
@@ -720,6 +729,77 @@ 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";

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 +1,343 @@
+use super::{GgmlDType, QStorage};
+use crate::{backend::BackendDevice, cuda_backend::WrapErr};
+use crate::{CudaDevice, CudaStorage, Result};
+
+use cudarc::driver::{CudaSlice, DeviceSlice};
+
+pub struct QCudaStorage {
+    data: CudaSlice<u8>,
+    dtype: GgmlDType,
+    device: CudaDevice,
+}
+
+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_DEQUANTIZE_BLOCK_SIZE: usize = 256;
+
+fn dequantize(
+    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", false, 32, nb),
+        GgmlDType::Q4_1 => ("dequantize_block_q4_1", false, 32, nb),
+        GgmlDType::Q5_0 => {
+            let nb = (elem_count + 2 * CUDA_DEQUANTIZE_BLOCK_SIZE - 1)
+                / (2 * CUDA_DEQUANTIZE_BLOCK_SIZE);
+            (
+                "dequantize_block_q5_0",
+                false,
+                CUDA_DEQUANTIZE_BLOCK_SIZE,
+                nb,
+            )
+        }
+        GgmlDType::Q5_1 => {
+            let nb = (elem_count + 2 * CUDA_DEQUANTIZE_BLOCK_SIZE - 1)
+                / (2 * CUDA_DEQUANTIZE_BLOCK_SIZE);
+            (
+                "dequantize_block_q5_1",
+                false,
+                CUDA_DEQUANTIZE_BLOCK_SIZE,
+                nb,
+            )
+        }
+        GgmlDType::Q8_0 => ("dequantize_block_q8_0", false, 32, nb),
+        GgmlDType::Q2K => ("dequantize_block_q2_K", true, 64, nb),
+        GgmlDType::Q3K => ("dequantize_block_q3_K", true, 64, nb),
+        GgmlDType::Q4K => ("dequantize_block_q4_K", true, 32, nb),
+        GgmlDType::Q5K => ("dequantize_block_q5_K", true, 64, nb),
+        GgmlDType::Q6K => ("dequantize_block_q6_K", true, 64, nb),
+        GgmlDType::Q8K => ("dequantize_block_q8_K", true, 32, nb),
+        _ => crate::bail!("unsupported dtype for dequantize {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = dev.alloc_zeros::<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_mut_mal_vec(
+    data: &CudaSlice<u8>,
+    y: &cudarc::driver::CudaView<f32>,
+    dtype: GgmlDType,
+    ncols: usize,
+    nrows: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    use cudarc::driver::LaunchAsync;
+
+    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 = dev.alloc_zeros::<f32>(nrows).w()?;
+    let block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / 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()))
+}
+
+impl QCudaStorage {
+    pub fn zeros(device: &CudaDevice, el_count: usize, dtype: GgmlDType) -> Result<Self> {
+        let size_in_bytes = el_count * dtype.type_size() / dtype.block_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> {
+        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(&self.data, self.dtype, elem_count, self.device());
+        }
+        // Run the dequantization on cpu.
+        use crate::quantized::k_quants::GgmlType;
+
+        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 => {
+                let slice =
+                    unsafe { std::slice::from_raw_parts(buffer.as_ptr() as *const f32, block_len) };
+                out.copy_from_slice(slice)
+            }
+            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)?;
+            }
+        }
+
+        self.device
+            .storage_from_cpu_storage(&crate::CpuStorage::F32(out))
+    }
+
+    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)> {
+        if matches!(layout.shape().dims(), [1, 1, _] | [1, _]) {
+            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 (with_batch, k) = match rhs_l.shape().dims() {
+            [1, 1, k] => (true, k),
+            [1, k] => (false, 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 =
+            dequantize_mut_mal_vec(&self.data, &rhs, self.dtype, ncols, nrows, self.device())?;
+        let out_shape = if with_batch {
+            vec![1, 1, nrows]
+        } else {
+            vec![1, 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 data_f32 = self.dequantize(n * k)?;
+        let rhs_l = crate::Layout::new((k, n).into(), vec![1, k], 0).broadcast_as((b, k, n))?;
+        let out = storage.matmul(&data_f32, (b, m, n, k), layout, &rhs_l)?;
+        let mut out_shape = layout.shape().dims().to_vec();
+        out_shape.pop();
+        out_shape.push(n);
+        Ok((out, out_shape.into()))
+    }
+}
+
+fn read_to_vec<T: Clone>(buffer: &[u8], n: usize) -> Vec<T> {
+    let slice = unsafe { std::slice::from_raw_parts(buffer.as_ptr() as *const T, n) };
+    slice.to_vec()
+}
+
+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,
+    }))
+}

candle-core/src/quantized/dummy_cuda.rs

@@ -0,0 +1,50 @@
+#![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 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,
+        )
     }
 }
 
@@ -460,12 +466,13 @@ impl Content {
         &self,
         reader: &mut R,
         name: &str,
+        device: &Device,
     ) -> Result<QTensor> {
         let tensor_info = match self.tensor_infos.get(name) {
             Some(tensor_info) => tensor_info,
-            None => crate::bail!("cannot find tensor-infor for {name}"),
+            None => crate::bail!("cannot find tensor info for {name}"),
         };
-        tensor_info.read(reader, self.tensor_data_offset)
+        tensor_info.read(reader, self.tensor_data_offset, device)
     }
 }
 
@@ -517,10 +524,9 @@ pub fn write<W: std::io::Seek + std::io::Write>(
                 "internal error, unexpected current position {tensor_start_pos} {offset} {pos}"
             )
         }
-        let data_ptr = tensor.as_ptr();
-        let size_in_bytes = tensor.storage_size_in_bytes();
-        let data = unsafe { std::slice::from_raw_parts(data_ptr, size_in_bytes) };
-        w.write_all(data)?;
+        let data = tensor.data()?;
+        let size_in_bytes = data.len();
+        w.write_all(&data)?;
         let padding = 31 - (31 + size_in_bytes) % 32;
         w.write_all(&vec![0u8; padding])?;
     }

candle-core/src/quantized/k_quants.rs

@@ -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 +1,222 @@
+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();
+
+        let (b, m) = match dst_shape.len() {
+            3 => (dst_shape[0], dst_shape[1]),
+            2 => (1, dst_shape[0]),
+            n => crate::bail!("Invalid rank {n} for quantized matmul metal"),
+        };
+        let last_k = dst_shape.pop().unwrap();
+        if last_k != k {
+            crate::bail!("input tensor {layout:?} incompatible with {:?}", self_shape)
+        }
+        dst_shape.push(n);
+        let dst_shape = Shape::from(dst_shape);
+        let device = storage.device().clone();
+        let dst = device.new_buffer(dst_shape.elem_count(), DType::F32, "qmatmul")?;
+        let command_buffer = device.command_buffer()?;
+        candle_metal_kernels::call_quantized_matmul_t(
+            device.device(),
+            &command_buffer,
+            device.kernels(),
+            self.dtype.into(),
+            (b, m, n, k),
+            storage.buffer(),
+            layout.start_offset() * storage.dtype().size_in_bytes(),
+            &self.buffer,
+            &dst,
+        )
+        .map_err(MetalError::from)?;
+        let dst_storage = crate::MetalStorage::new(dst, device, 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, 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 +358,19 @@ impl QTensor {
     }
 
     pub fn dequantize(&self, device: &Device) -> Result<Tensor> {
-        let mut f32_data = vec![0f32; self.shape.elem_count()];
-        self.data.to_float(&mut f32_data)?;
-        Tensor::from_vec(f32_data, &self.shape, device)
-    }
-
-    pub fn matmul_t(&self, mkn: (usize, usize, usize), lhs: &[f32], dst: &mut [f32]) -> Result<()> {
-        self.data.matmul_t(mkn, lhs, dst)
+        let storage = self.storage.dequantize(self.shape.elem_count())?;
+        let none = crate::op::BackpropOp::none();
+        let is_variable = false;
+        crate::tensor::from_storage(storage, self.shape.clone(), none, is_variable)
+            .to_device(device)
     }
 
     pub fn storage_size_in_bytes(&self) -> usize {
-        self.data.storage_size_in_bytes()
+        self.storage.size_in_bytes()
     }
 
-    pub fn as_ptr(&self) -> *const u8 {
-        self.data.as_ptr()
+    pub fn data(&self) -> Result<Cow<'_, [u8]>> {
+        self.storage.data()
     }
 }
 
@@ -255,7 +398,7 @@ 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 {
             Self::QTensor(qtensor)
@@ -294,17 +437,41 @@ 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 crate::Module for QMatMul {

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/shape.rs

@@ -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/storage.rs

@@ -352,6 +352,10 @@ 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(),
@@ -697,4 +701,32 @@ 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

@@ -1,4 +1,4 @@
-//! Tensors are N-dimenional matrixes of elements using a single data type.
+//! Tensors are N-dimensional matrixes of elements using a single data type.
 #![allow(clippy::redundant_closure_call)]
 use crate::backend::{BackendDevice, BackendStorage};
 use crate::op::{
@@ -361,6 +361,16 @@ impl Tensor {
         Self::new_impl(array, shape, device, false)
     }
 
+    /// Returns a new tensor with all the elements having the same specified value. Note that
+    /// the tensor is not contiguous so you would have to call `.contiguous()` on it if needed.
+    pub fn full<D: crate::WithDType, S: Into<Shape>>(
+        value: D,
+        shape: S,
+        device: &Device,
+    ) -> Result<Self> {
+        Self::from_vec_impl(vec![value], (), device, false)?.broadcast_as(shape)
+    }
+
     /// Creates a new 1D tensor from an iterator.
     pub fn from_iter<D: crate::WithDType>(
         iter: impl IntoIterator<Item = D>,
@@ -386,7 +396,7 @@ 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;
@@ -498,6 +508,7 @@ 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);
@@ -655,7 +666,7 @@ impl Tensor {
         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(),
@@ -669,7 +680,7 @@ impl Tensor {
     }
 
     /// Split a tensor into the specified number of chunks, this may return less chunks than
-    /// specificed.
+    /// specified.
     pub fn chunk<D: Dim>(&self, chunks: usize, dim: D) -> Result<Vec<Self>> {
         let dim = dim.to_index(self.shape(), "chunk")?;
         let size = self.dim(dim)?;
@@ -794,6 +805,35 @@ impl Tensor {
         }
     }
 
+    /// Roll the tensor input along the given dimension.
+    /// Elements that are shifted beyond the last position are re-introduced at the first position.
+    ///
+    /// ```rust
+    /// # use candle_core::{Tensor, Device};
+    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
+    /// let tensor = tensor.roll(1, 0)?;
+    /// assert_eq!(tensor.to_vec2::<f32>()?, &[[4., 5.], [0., 1.], [2., 3.]]);
+    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
+    /// let tensor = tensor.roll(-1, 0)?;
+    /// assert_eq!(tensor.to_vec2::<f32>()?, &[[2., 3.], [4., 5.], [0., 1.]]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
+    pub fn roll<D>(&self, shift: i32, dim: D) -> Result<Self>
+    where
+        D: Dim + Clone,
+    {
+        let dim = dim.to_index(self.shape(), "roll")?;
+        let dim_size = self.dim(dim)?;
+        let shift = shift.rem_euclid(dim_size as i32) as usize;
+        if shift == 0 {
+            Ok(self.clone())
+        } else {
+            let a = self.narrow(dim, 0, dim_size - shift)?;
+            let b = self.narrow(dim, dim_size - shift, shift)?;
+            Tensor::cat(&[&b, &a], dim)
+        }
+    }
+
     /// Returns the sum of all elements in the input tensor. The sum is performed over all the
     /// input dimensions.
     ///
@@ -975,7 +1015,7 @@ 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)?;
@@ -994,7 +1034,11 @@ 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)?;
@@ -1027,6 +1071,9 @@ impl Tensor {
         let kernel_size = kernel_size.to_usize2();
         let stride = stride.to_usize2();
         let (n, c, h, w) = self.dims4()?;
+        if h < kernel_size.0 || w < kernel_size.1 {
+            bail!("kernel-size {kernel_size:?} is larger than the input size {h},{w}")
+        }
         // https://pytorch.org/docs/stable/generated/torch.nn.AvgPool2d.html#torch.nn.AvgPool2d
         let h_out = (h - kernel_size.0) / stride.0 + 1;
         let w_out = (w - kernel_size.1) / stride.1 + 1;
@@ -1062,6 +1109,9 @@ impl Tensor {
         let kernel_size = kernel_size.to_usize2();
         let stride = stride.to_usize2();
         let (n, c, h, w) = self.dims4()?;
+        if h < kernel_size.0 || w < kernel_size.1 {
+            bail!("kernel-size {kernel_size:?} is larger than the input size {h},{w}")
+        }
         // https://pytorch.org/docs/stable/generated/torch.nn.MaxPool2d.html#torch.nn.MaxPool2d
         let h_out = (h - kernel_size.0) / stride.0 + 1;
         let w_out = (w - kernel_size.1) / stride.1 + 1;
@@ -1784,7 +1834,7 @@ impl Tensor {
         let is_permutation =
             dims.len() == self.rank() && (0..dims.len()).all(|i| dims.contains(&i));
         if !is_permutation {
-            crate::bail!(
+            bail!(
                 "dimension mismatch in permute, tensor {:?}, dims: {:?}",
                 self.dims(),
                 dims
@@ -1833,9 +1883,9 @@ 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(),
@@ -1846,7 +1896,7 @@ impl Tensor {
                 dtype: self.dtype,
                 device: self.device.clone(),
             };
-            Ok(Tensor(Arc::new(tensor_)))
+            Tensor(Arc::new(tensor_))
         }
     }
 
@@ -1863,10 +1913,7 @@ impl Tensor {
                     Storage::Metal(metal.storage_from_cpu_storage(storage)?)
                 }
                 (Storage::Cuda(storage), Device::Cpu) => Storage::Cpu(storage.to_cpu_storage()?),
-                (Storage::Metal(storage), Device::Cpu) => {
-                    // println!("{storage:?} - {:?}", storage.to_cpu_storage()?);
-                    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.
@@ -2102,152 +2149,6 @@ 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> {
@@ -2282,7 +2183,7 @@ 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)?;
@@ -2446,13 +2347,13 @@ impl Tensor {
     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)
         }
@@ -2514,14 +2415,14 @@ impl Tensor {
         let src_dims = src.dims();
         let self_dims = self.dims();
         if self_dims.len() != src_dims.len() {
-            crate::bail!(
+            bail!(
                 "slice-assign requires input with the same rank {} <> {}",
                 self_dims.len(),
                 src_dims.len()
             )
         }
         if self_dims.len() != ranges.len() {
-            crate::bail!(
+            bail!(
                 "slice-assign requires input with the same rank as there are ranges {} <> {}",
                 self_dims.len(),
                 ranges.len()
@@ -2541,18 +2442,16 @@ impl Tensor {
                 std::ops::Bound::Excluded(v) => *v,
             };
             if end_excluded <= start_included {
-                crate::bail!(
-                    "slice-assign: empty range for dim {i}, {start_included} {end_excluded}"
-                )
+                bail!("slice-assign: empty range for dim {i}, {start_included} {end_excluded}")
             }
             if self_dims[i] < end_excluded {
-                crate::bail!(
+                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] {
-                crate::bail!(
+                bail!(
                     "slice-assign: the range for dim {i} ({start_included}..{end_excluded}) does not match the size of src {}", src_dims[i]
                 )
             }
@@ -2561,6 +2460,23 @@ impl Tensor {
         }
         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 +1,240 @@
+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())?
+                }
+            }
+        }
+        if dim == 0 {
+            Self::cat0(args)
+        } else {
+            let all_contiguous = args.iter().all(|v| v.as_ref().is_contiguous());
+            if all_contiguous {
+                Self::cat_contiguous(args, dim)
+            } 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 = 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(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 = device.zeros(&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))
+    }
+}

candle-core/src/variable.rs

@@ -107,6 +107,10 @@ impl Var {
         Ok(Self(inner))
     }
 
+    pub fn as_detached_tensor(&self) -> Tensor {
+        self.0.detach()
+    }
+
     pub fn as_tensor(&self) -> &Tensor {
         &self.0
     }

candle-core/tests/conv_tests.rs

@@ -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 +53,16 @@ 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)?;
+
+    // conv-transposes are not implemented for metal.
+    if dev.is_metal() {
+        return Ok(());
+    }
+
+    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 +71,17 @@ 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(())
 }
@@ -146,31 +168,33 @@ 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)?,
-        [
+    if !dev.is_metal() {
+        let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 1)?;
+        assert_eq!(res.dims(), [1, 2, 7, 7]);
+        assert_eq!(
+            test_utils::to_vec3_round(&res.i(0)?, 4)?,
             [
-                [-1.9918, 2.6797, -0.4599, -1.6037, 1.4131, -2.4012, 2.9277],
-                [1.8016, -3.5361, 1.0757, 3.5395, -8.2168, -3.2023, 0.5375],
-                [0.8243, 1.8675, 7.8929, -4.0746, -6.4415, 5.1139, 1.6889],
-                [0.2722, 8.9679, 3.3477, 1.8514, -4.2896, -3.8228, -7.5632],
-                [-8.5412, -5.8142, -7.1587, -1.6095, 0.4651, 0.2748, -2.0985],
-                [2.0833, -0.6482, -12.1692, -4.1284, -2.9765, -0.0656, -4.5114],
-                [5.307, 2.6957, 2.3087, 1.0478, 0.7808, -1.1519, -0.9579]
-            ],
-            [
-                [1.089, 0.1872, -0.6408, -0.9897, 0.8503, 1.1019, -0.9211],
-                [-0.1741, -0.2915, 4.2472, 1.9417, 1.65, 0.6303, -4.7131],
-                [1.6555, 2.4026, -2.9293, 2.9953, 0.5328, 3.5873, -0.9621],
-                [-1.4289, -3.2787, 4.1747, -6.0341, -4.6341, -5.7945, 4.142],
-                [7.5973, 6.4431, 5.9872, 2.1639, -8.6566, 3.3143, -3.4059],
-                [-0.8775, -3.048, 11.6543, 0.6442, 2.3218, -0.4765, 1.1516],
-                [-5.5423, -2.5188, 1.0754, -0.0563, -2.9386, -1.1504, 1.0171]
+                [
+                    [-1.9918, 2.6797, -0.4599, -1.6037, 1.4131, -2.4012, 2.9277],
+                    [1.8016, -3.5361, 1.0757, 3.5395, -8.2168, -3.2023, 0.5375],
+                    [0.8243, 1.8675, 7.8929, -4.0746, -6.4415, 5.1139, 1.6889],
+                    [0.2722, 8.9679, 3.3477, 1.8514, -4.2896, -3.8228, -7.5632],
+                    [-8.5412, -5.8142, -7.1587, -1.6095, 0.4651, 0.2748, -2.0985],
+                    [2.0833, -0.6482, -12.1692, -4.1284, -2.9765, -0.0656, -4.5114],
+                    [5.307, 2.6957, 2.3087, 1.0478, 0.7808, -1.1519, -0.9579]
+                ],
+                [
+                    [1.089, 0.1872, -0.6408, -0.9897, 0.8503, 1.1019, -0.9211],
+                    [-0.1741, -0.2915, 4.2472, 1.9417, 1.65, 0.6303, -4.7131],
+                    [1.6555, 2.4026, -2.9293, 2.9953, 0.5328, 3.5873, -0.9621],
+                    [-1.4289, -3.2787, 4.1747, -6.0341, -4.6341, -5.7945, 4.142],
+                    [7.5973, 6.4431, 5.9872, 2.1639, -8.6566, 3.3143, -3.4059],
+                    [-0.8775, -3.048, 11.6543, 0.6442, 2.3218, -0.4765, 1.1516],
+                    [-5.5423, -2.5188, 1.0754, -0.0563, -2.9386, -1.1504, 1.0171]
+                ]
             ]
-        ]
-    );
+        );
+    }
     // Dilations.
     let res = t.conv2d(&w, 0, 1, 2, 1)?;
     assert_eq!(res.dims(), [1, 2, 1, 1]);
@@ -179,36 +203,44 @@ fn conv2d(dev: &Device) -> Result<()> {
         [2.45, -2.3504],
     );
 
-    // Transpose and dilations.
-    let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 2)?;
-    assert_eq!(res.dims(), [1, 2, 9, 9]);
-    assert_eq!(
-        test_utils::to_vec3_round(&res.i(0)?, 4)?,
-        [
+    if !dev.is_metal() {
+        // Transpose and dilations.
+        let res = t.conv_transpose2d(&w.transpose(0, 1)?, 0, 0, 1, 2)?;
+        assert_eq!(res.dims(), [1, 2, 9, 9]);
+        assert_eq!(
+            test_utils::to_vec3_round(&res.i(0)?, 4)?,
             [
-                [-1.9918, 3.1652, -0.6778, -4.3442, 4.4351, 0.6652, -3.0124, -0.6031, 2.9277],
-                [2.7036, -1.7156, -0.3969, 1.0516, 1.6381, -2.8886, -0.205, 2.4682, -1.0499],
-                [-0.9459, 3.1631, 3.707, -4.8369, -8.5166, -1.4496, -2.7559, -3.2698, 1.4376],
-                [-0.2157, 3.7786, -2.0252, -4.2633, 3.6731, -1.5142, 5.9391, -0.2622, -0.141],
-                [-6.8121, -3.1744, 1.5945, 3.0637, -9.6088, 1.4446, 2.9489, -3.0082, -7.3822],
-                [0.2371, 3.3303, 0.3861, 2.2646, -4.6784, 4.1235, -0.0109, 0.3176, -0.03],
-                [-2.5339, -2.9564, -3.4518, -4.4594, -9.1873, -1.9709, -0.4676, 0.51, -3.5024],
-                [4.007, 0.3067, -2.2954, 1.1105, -0.1992, 1.6372, -2.9268, 0.2807, -1.2787],
-                [5.307, 1.1317, 1.3518, 0.9049, 3.8116, -0.4075, -0.8874, -0.2241, -0.9579]
-            ],
-            [
-                [1.089, -0.6483, 0.0726, -0.4752, -1.3283, 1.7103, 1.0703, 0.1076, -0.9211],
-                [-0.8629, 0.1376, 0.3202, 2.0955, 0.9696, 2.8988, -1.0012, 1.5049, -0.1278],
-                [1.9286, -1.5255, -2.9563, 2.4589, 3.3611, -0.6951, 0.3525, -1.7724, -5.9861],
-                [1.1226, 2.1561, 3.6417, 4.7546, -0.692, 4.4126, -5.1902, 6.0805, 2.3185],
-                [1.0111, 0.3604, 0.6432, -3.6605, 7.9517, -9.2955, -5.2988, -3.7803, -2.0642],
-                [3.3172, -1.7967, -3.6576, -2.0942, 1.3158, 0.112, -1.7405, 2.9167, 0.7957],
-                [5.1001, 1.8995, -1.8639, 1.1262, 9.9629, 2.683, -3.6319, -1.1607, 0.5856],
-                [-4.8445, -0.5642, 4.2317, 0.0856, 1.2267, -0.5712, 1.736, 1.0997, 0.6908],
-                [-5.5423, -1.1831, -1.2176, 0.0843, 0.0446, -0.7545, -2.4798, -0.0827, 1.0171]
+                [
+                    [-1.9918, 3.1652, -0.6778, -4.3442, 4.4351, 0.6652, -3.0124, -0.6031, 2.9277],
+                    [2.7036, -1.7156, -0.3969, 1.0516, 1.6381, -2.8886, -0.205, 2.4682, -1.0499],
+                    [-0.9459, 3.1631, 3.707, -4.8369, -8.5166, -1.4496, -2.7559, -3.2698, 1.4376],
+                    [-0.2157, 3.7786, -2.0252, -4.2633, 3.6731, -1.5142, 5.9391, -0.2622, -0.141],
+                    [-6.8121, -3.1744, 1.5945, 3.0637, -9.6088, 1.4446, 2.9489, -3.0082, -7.3822],
+                    [0.2371, 3.3303, 0.3861, 2.2646, -4.6784, 4.1235, -0.0109, 0.3176, -0.03],
+                    [
+                        -2.5339, -2.9564, -3.4518, -4.4594, -9.1873, -1.9709, -0.4676, 0.51,
+                        -3.5024
+                    ],
+                    [4.007, 0.3067, -2.2954, 1.1105, -0.1992, 1.6372, -2.9268, 0.2807, -1.2787],
+                    [5.307, 1.1317, 1.3518, 0.9049, 3.8116, -0.4075, -0.8874, -0.2241, -0.9579]
+                ],
+                [
+                    [1.089, -0.6483, 0.0726, -0.4752, -1.3283, 1.7103, 1.0703, 0.1076, -0.9211],
+                    [-0.8629, 0.1376, 0.3202, 2.0955, 0.9696, 2.8988, -1.0012, 1.5049, -0.1278],
+                    [1.9286, -1.5255, -2.9563, 2.4589, 3.3611, -0.6951, 0.3525, -1.7724, -5.9861],
+                    [1.1226, 2.1561, 3.6417, 4.7546, -0.692, 4.4126, -5.1902, 6.0805, 2.3185],
+                    [1.0111, 0.3604, 0.6432, -3.6605, 7.9517, -9.2955, -5.2988, -3.7803, -2.0642],
+                    [3.3172, -1.7967, -3.6576, -2.0942, 1.3158, 0.112, -1.7405, 2.9167, 0.7957],
+                    [5.1001, 1.8995, -1.8639, 1.1262, 9.9629, 2.683, -3.6319, -1.1607, 0.5856],
+                    [-4.8445, -0.5642, 4.2317, 0.0856, 1.2267, -0.5712, 1.736, 1.0997, 0.6908],
+                    [
+                        -5.5423, -1.1831, -1.2176, 0.0843, 0.0446, -0.7545, -2.4798, -0.0827,
+                        1.0171
+                    ]
+                ]
             ]
-        ]
-    );
+        );
+    }
     Ok(())
 }
 
@@ -262,6 +294,12 @@ fn conv2d_small(dev: &Device) -> Result<()> {
             0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000
         ]
     );
+
+    // conv-transposes are not implemented for metal
+    if dev.is_metal() {
+        return Ok(());
+    }
+
     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 +401,10 @@ print(w.grad.shape)
 print(w.grad[0])
 */
 fn conv2d_grad(dev: &Device) -> Result<()> {
+    // conv-transposes are not implemented for metal
+    if dev.is_metal() {
+        return Ok(());
+    }
     use candle_core::Var;
     let t = Var::from_slice(
         &[

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(())
 }
 

candle-core/tests/pool_tests.rs

@@ -2,6 +2,9 @@ use candle_core::{test_device, test_utils, Device, IndexOp, Result, Tensor};
 
 // https://github.com/huggingface/candle/issues/364
 fn avg_pool2d(dev: &Device) -> Result<()> {
+    if dev.is_metal() {
+        return Ok(());
+    }
     let data: Vec<f32> = vec![
         1., 1., 1., 1., 0., 0., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
     ];
@@ -19,6 +22,9 @@ fn avg_pool2d(dev: &Device) -> Result<()> {
 }
 
 fn max_pool2d(dev: &Device) -> Result<()> {
+    if dev.is_metal() {
+        return Ok(());
+    }
     let data: Vec<f32> = vec![
         1., 2., 1., 3., 0., 0., 1., 1., 1., 1., 1., 1., 5., 1., 1., 1.,
     ];
@@ -43,6 +49,9 @@ 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,

candle-core/tests/pth.py

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

candle-core/tests/pth_tests.rs

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

candle-core/tests/quantized_tests.rs

@@ -1,5 +1,7 @@
 use candle_core::{
+    bail,
     quantized::{self, GgmlDType},
+    test_device,
     test_utils::to_vec2_round,
     Device, Module, Result, Tensor,
 };
@@ -13,16 +15,48 @@ const GGML_MAX_QUANTIZATION_TOTAL_ERROR_2BITS: f32 = 0.0075;
 const GGML_MAX_QUANTIZATION_TOTAL_ERROR_3BITS: f32 = 0.0040;
 const GGML_MAX_DOT_PRODUCT_ERROR: f32 = 0.02;
 
-#[test]
-fn quantized_matmul() -> Result<()> {
-    let cpu = &Device::Cpu;
+fn test_matmul(
+    device: &Device,
+    (b, m, n, k): (usize, usize, usize, usize),
+    dtype: GgmlDType,
+) -> Result<()> {
+    let lhs = (0..(m * k))
+        .map(|v| v as f32 / (m * k) as f32)
+        .collect::<Vec<_>>();
+    let rhs = (0..(k * n))
+        .map(|v| v as f32 / (n * k) as f32)
+        .collect::<Vec<_>>();
+
+    let lhs = Tensor::from_slice(&lhs, (m, k), device)?;
+    let rhs = Tensor::from_slice(&rhs, (k, n), device)?;
+    let mm = lhs.matmul(&rhs)?;
+    let qtensor = quantized::QTensor::quantize(&rhs.t()?, dtype)?;
+    let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
+    let res = matmul.forward(&lhs)?;
+
+    let error: f32 = ((&mm - &res)?.abs()? / &mm.abs()?)?
+        .sum_all()?
+        .to_scalar()?;
+    let error = error / (b * m * n) as f32;
+    assert!(
+        error <= 0.02,
+        "Error {error} is too big. \nExpected:\n {mm} \nFound:\n {res}\n for {dtype:?}"
+    );
+
+    Ok(())
+}
+
+fn quantized_matmul(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
     let (m, k, n) = (3, 64, 4);
     let lhs = (0..(m * k)).map(|v| v as f32).collect::<Vec<_>>();
-    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), cpu)?;
+    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), device)?;
     let mut dst = vec![42.; 3 * 4];
     let mut rhs_t = vec![k_quants::BlockQ4_0::zeros(); 8];
     let rhs = (0..(k * n)).map(|v| v as f32).collect::<Vec<_>>();
-    let tensor_rhs = Tensor::from_slice(&rhs, (n, k), cpu)?.t()?;
     k_quants::BlockQ4_0::from_float(&rhs, &mut rhs_t)?;
     k_quants::matmul((m, k, n), &lhs, &rhs_t, &mut dst)?;
     assert_eq!(
@@ -32,6 +66,7 @@ fn quantized_matmul() -> Result<()> {
             341876.0, 994283.0, 1655709.0, 2301518.0
         ]
     );
+    let tensor_rhs = Tensor::from_slice(&rhs, (n, k), device)?.t()?;
     let mm = tensor_lhs.matmul(&tensor_rhs)?;
     assert_eq!(
         mm.to_vec2::<f32>()?,
@@ -42,35 +77,49 @@ fn quantized_matmul() -> Result<()> {
         ]
     );
 
-    let qtensor = quantized::QTensor::new(rhs_t, (4, 64))?;
+    let qtensor = quantized::QTensor::quantize(&tensor_rhs.t()?, GgmlDType::Q4_0)?;
     let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
     let res = matmul.forward(&tensor_lhs)?;
-    assert_eq!(
-        to_vec2_round(&res, 0)?,
-        &[
-            [85120.0, 214562.0, 345455.0, 474748.0],
-            [213475.0, 604465.0, 1000686.0, 1388317.0],
-            [341876.0, 994283.0, 1655709.0, 2301518.0]
-        ]
-    );
+    match device {
+        Device::Metal(_) => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [84946.0, 214126.0, 344757.0, 473798.0],
+                [213458.0, 604350.0, 1000469.0, 1387990.0],
+                [341970.0, 994574.0, 1656181.0, 2302182.0]
+            ]
+        ),
+        _ => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [85120.0, 214562.0, 345455.0, 474748.0],
+                [213475.0, 604465.0, 1000686.0, 1388317.0],
+                [341876.0, 994283.0, 1655709.0, 2301518.0]
+            ]
+        ),
+    }
+
+    test_matmul(device, (1, 3, 4, 256), GgmlDType::Q4_0)?;
 
     Ok(())
 }
 
-#[test]
-fn quantized_matmul_neg() -> Result<()> {
-    let cpu = &Device::Cpu;
+fn quantized_matmul_neg(device: &Device) -> Result<()> {
+    // TODO Enable this later when we enable cuda.
+    if device.is_cuda() {
+        return Ok(());
+    }
     let (m, k, n) = (3, 64, 4);
     let lhs = (0..(m * k))
         .map(|v| v as f32 - (m * k) as f32 / 2.0)
         .collect::<Vec<_>>();
-    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), cpu)?;
+    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), device)?;
     let mut dst = vec![42.; 3 * 4];
     let mut rhs_t = vec![k_quants::BlockQ4_0::zeros(); 8];
     let rhs = (0..k * n)
         .map(|v| v as f32 - (k * n) as f32 / 3.0)
         .collect::<Vec<_>>();
-    let tensor_rhs = Tensor::from_slice(&rhs, (n, k), cpu)?.t()?;
+    let tensor_rhs = Tensor::from_slice(&rhs, (n, k), device)?.t()?;
     k_quants::BlockQ4_0::from_float(&rhs, &mut rhs_t)?;
     k_quants::matmul((m, k, n), &lhs, &rhs_t, &mut dst)?;
     assert_eq!(
@@ -90,32 +139,52 @@ fn quantized_matmul_neg() -> Result<()> {
         ]
     );
 
-    let qtensor = quantized::QTensor::new(rhs_t, (4, 64))?;
+    let qtensor = quantized::QTensor::quantize(&tensor_rhs.t()?, GgmlDType::Q4_0)?;
     let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
     let res = matmul.forward(&tensor_lhs)?;
-    assert_eq!(
-        to_vec2_round(&res, 0)?,
-        &[
-            [243524.0, -19596.0, -285051.0, -549815.0],
-            [23777.0, 21651.0, 19398.0, 18367.0],
-            [-196472.0, 63012.0, 324585.0, 587902.0]
-        ]
-    );
+    match device {
+        Device::Metal(_) => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [243666.0, -19714.0, -285433.0, -550453.0],
+                [23782.0, 21654.0, 19400.0, 18369.0],
+                [-196102.0, 63022.0, 324233.0, 587191.0]
+            ]
+        ),
+        _ => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [243524.0, -19596.0, -285051.0, -549815.0],
+                [23777.0, 21651.0, 19398.0, 18367.0],
+                [-196472.0, 63012.0, 324585.0, 587902.0]
+            ]
+        ),
+    }
 
     Ok(())
 }
 
-#[test]
-fn quantize_q4_0() -> Result<()> {
-    use k_quants::BlockQ4_0;
+test_device!(
+    quantized_matmul,
+    quantized_matmul_cpu,
+    quantized_matmul_cuda,
+    quantized_matmul_metal
+);
+test_device!(
+    quantized_matmul_neg,
+    quantized_matmul_neg_cpu,
+    quantized_matmul_neg_cuda,
+    quantized_matmul_neg_metal
+);
 
+fn quantize_q4_0(device: &Device) -> Result<()> {
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
-    let mut dst = vec![0f32; 32 * 4];
-    let mut quant = vec![BlockQ4_0::zeros(); 4];
-    BlockQ4_0::from_float(&src, &mut quant)?;
-    BlockQ4_0::to_float(&quant, dst.as_mut_slice())?;
+
+    let src = Tensor::from_slice(&src, (32 * 4,), device)?;
+    let quant = quantized::QTensor::quantize(&src, GgmlDType::Q4_0)?;
+    let dst = quant.dequantize(device)?;
     assert_eq!(
-        dst,
+        dst.to_vec1::<f32>()?,
         &[
             -0.0, -0.0, 3.875, 3.875, 3.875, 3.875, 7.75, 7.75, 7.75, 7.75, 11.625, 11.625, 11.625,
             11.625, 15.5, 15.5, 15.5, 15.5, 19.375, 19.375, 19.375, 19.375, 23.25, 23.25, 23.25,
@@ -131,21 +200,17 @@ fn quantize_q4_0() -> Result<()> {
             127.0, 127.0
         ]
     );
-    ggml_quantization_error_test::<BlockQ4_0>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(GgmlDType::Q4_0, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q4_1() -> Result<()> {
-    use k_quants::BlockQ4_1;
-
+fn quantize_q4_1(device: &Device) -> Result<()> {
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
-    let mut dst = vec![0f32; 32 * 4];
-    let mut quant = vec![BlockQ4_1::zeros(); 4];
-    BlockQ4_1::from_float(&src, &mut quant)?;
-    BlockQ4_1::to_float(&quant, dst.as_mut_slice())?;
+    let src = Tensor::from_slice(&src, (32 * 4,), device)?;
+    let quant = quantized::QTensor::quantize(&src, GgmlDType::Q4_1)?;
+    let dst = quant.dequantize(device)?;
     assert_eq!(
-        round_vector(&dst),
+        round_vector(&dst.to_vec1::<f32>()?),
         &[
             0.0, 0.0, 2.066, 2.066, 4.133, 4.133, 6.199, 6.199, 8.266, 8.266, 10.332, 10.332,
             12.398, 12.398, 14.465, 14.465, 16.531, 16.531, 18.598, 18.598, 20.664, 20.664, 22.73,
@@ -161,21 +226,17 @@ fn quantize_q4_1() -> Result<()> {
             118.73, 118.73, 120.797, 120.797, 122.863, 122.863, 124.93, 124.93, 126.996, 126.996
         ]
     );
-    ggml_quantization_error_test::<BlockQ4_1>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(GgmlDType::Q4_1, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q5_0() -> Result<()> {
-    use k_quants::BlockQ5_0;
-
+fn quantize_q5_0(device: &Device) -> Result<()> {
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
-    let mut dst = vec![0f32; 32 * 4];
-    let mut quant = vec![BlockQ5_0::zeros(); 4];
-    BlockQ5_0::from_float(&src, &mut quant)?;
-    BlockQ5_0::to_float(&quant, dst.as_mut_slice())?;
+    let src = Tensor::from_slice(&src, (32 * 4,), device)?;
+    let quant = quantized::QTensor::quantize(&src, GgmlDType::Q5_0)?;
+    let dst = quant.dequantize(device)?;
     assert_eq!(
-        round_vector(&dst),
+        round_vector(&dst.to_vec1::<f32>()?),
         &[
             -0.0, 1.938, 1.938, 3.875, 3.875, 5.813, 5.813, 7.75, 7.75, 9.688, 9.688, 11.625,
             11.625, 13.563, 13.563, 15.5, 15.5, 17.438, 17.438, 19.375, 19.375, 21.313, 21.313,
@@ -191,21 +252,17 @@ fn quantize_q5_0() -> Result<()> {
             119.063, 119.063, 119.063, 119.063, 127.0, 127.0, 127.0, 127.0
         ]
     );
-    ggml_quantization_error_test::<BlockQ5_0>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(GgmlDType::Q5_0, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q5_1() -> Result<()> {
-    use k_quants::BlockQ5_1;
-
+fn quantize_q5_1(device: &Device) -> Result<()> {
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
-    let mut dst = vec![0f32; 32 * 4];
-    let mut quant = vec![BlockQ5_1::zeros(); 4];
-    BlockQ5_1::from_float(&src, &mut quant)?;
-    BlockQ5_1::to_float(&quant, dst.as_mut_slice())?;
+    let src = Tensor::from_slice(&src, (32 * 4,), device)?;
+    let quant = quantized::QTensor::quantize(&src, GgmlDType::Q5_1)?;
+    let dst = quant.dequantize(device)?;
     assert_eq!(
-        dst,
+        round_vector(&dst.to_vec1::<f32>()?),
         &[
             0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0,
             16.0, 17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0, 25.0, 26.0, 27.0, 28.0, 29.0,
@@ -219,13 +276,11 @@ fn quantize_q5_1() -> Result<()> {
             124.0, 125.0, 126.0, 127.0
         ]
     );
-
-    ggml_quantization_error_test::<BlockQ5_1>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(GgmlDType::Q5_1, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-/// Generates a small test vector ranging from -`bound` to `bound` with `size` steps
-fn get_test_vector(bound: f32, size: usize) -> (Vec<f32>, Vec<f32>) {
+fn get_test_vector2(bound: f32, size: usize, device: &Device) -> Result<Tensor> {
     assert!(
         size % crate::quantized::k_quants::QK_K == 0,
         "size must be a multiple of {}",
@@ -235,10 +290,8 @@ fn get_test_vector(bound: f32, size: usize) -> (Vec<f32>, Vec<f32>) {
     let src = (0..size)
         .map(|v| (v as f32 - size as f32 / 2.) * bound / (size as f32 / 2.))
         .collect::<Vec<_>>();
-
-    let dst = vec![0f32; size];
     assert_eq!([src[0], src[size / 2]], [-bound, 0.0]);
-    (src, dst)
+    Tensor::from_vec(src, (size,), device)
 }
 
 /// Round a vector
@@ -265,7 +318,8 @@ fn compare_with_error(values: &[f32], expected: &[f32], tolerance: f32) {
     }
 }
 
-/// Creates a vector simillarly to the one used in GGML unit tests: https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L26-L30
+/// Creates a vector similar to the ones used in GGML unit tests:
+/// https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L26-L30
 fn create_ggml_like_vector(offset: f32) -> Vec<f32> {
     (0..GGML_TEST_SIZE)
         .map(|i| 0.1 + 2.0 * (i as f32 + offset).cos())
@@ -284,14 +338,16 @@ fn calculate_rmse(a: &[f32], b: &[f32]) -> f32 {
     sum / a.len() as f32
 }
 
-/// Mirrores the GGML quanitzation unit test: https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L43-L50
-fn ggml_quantization_error_test<T: GgmlType>(max_error: f32) -> Result<()> {
+/// Similar to the GGML quantization unit test:
+/// https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L43-L50
+fn ggml_quantization_error_test(dtype: GgmlDType, device: &Device, max_error: f32) -> Result<()> {
     let src = create_ggml_like_vector(0.0);
-    let mut dst = vec![0.0; GGML_TEST_SIZE];
-    let _quant = quantize_roundtrip::<T>(src.as_slice(), dst.as_mut_slice())?;
-    let error = calculate_rmse(src.as_slice(), dst.as_slice());
+    let src = Tensor::from_slice(&src, (GGML_TEST_SIZE,), device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
+    let error = calculate_rmse(&src.to_vec1::<f32>()?, &dst.to_vec1::<f32>()?);
     if error > max_error {
-        candle_core::bail!(
+        bail!(
             "Quantization error {} exceeds max error {}",
             error,
             max_error
@@ -300,19 +356,15 @@ fn ggml_quantization_error_test<T: GgmlType>(max_error: f32) -> Result<()> {
     Ok(())
 }
 
-fn quantize_roundtrip<T: GgmlType>(src: &[f32], dst: &mut [f32]) -> Result<Vec<T>> {
-    let mut quant = vec![T::zeros(); src.len() / T::BLCK_SIZE];
-    T::from_float(src, &mut quant)?;
-    T::to_float(&quant, dst)?;
-    Ok(quant)
-}
+fn quantize_q2k(device: &Device) -> Result<()> {
+    let dtype = GgmlDType::Q2K;
 
-#[test]
-fn quantize_q2k() -> Result<()> {
-    use k_quants::BlockQ2K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ2K>(src.as_slice(), dst.as_mut_slice())?;
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
     compare_with_error(dst.as_slice(), src.as_slice(), 0.1);
 
     // Test some specific values
@@ -326,20 +378,26 @@ fn quantize_q2k() -> Result<()> {
         [-0.499, -0.366, -0.249, 0.0, 0.295, 0.492]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ2K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 6.0);
 
-    ggml_quantization_error_test::<BlockQ2K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR_2BITS)?;
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR_2BITS)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q3k() -> Result<()> {
-    use k_quants::BlockQ3K;
+fn quantize_q3k(device: &Device) -> Result<()> {
+    let dtype = GgmlDType::Q3K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ3K>(src.as_slice(), dst.as_mut_slice())?;
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
     compare_with_error(dst.as_slice(), src.as_slice(), 0.03);
 
     // Test some specific values
@@ -353,20 +411,26 @@ fn quantize_q3k() -> Result<()> {
         [-0.493, -0.37, -0.243, -0.0, 0.292, 0.492]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ3K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 3.5);
 
-    ggml_quantization_error_test::<BlockQ3K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR_3BITS)?;
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR_3BITS)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q4k() -> Result<()> {
-    use k_quants::BlockQ4K;
+fn quantize_q4k(device: &Device) -> Result<()> {
+    let dtype = GgmlDType::Q4K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ4K>(src.as_slice(), dst.as_mut_slice())?;
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
     compare_with_error(dst.as_slice(), src.as_slice(), 0.017);
 
     // Test some specific values
@@ -380,21 +444,27 @@ fn quantize_q4k() -> Result<()> {
         [-0.5, -0.373, -0.25, 0.0, 0.288, 0.498]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ4K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 4.5);
 
-    ggml_quantization_error_test::<BlockQ4K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q5k() -> Result<()> {
-    use k_quants::BlockQ5K;
+fn quantize_q5k(device: &Device) -> Result<()> {
+    let dtype = GgmlDType::Q5K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ5K>(src.as_slice(), dst.as_mut_slice())?;
-    compare_with_error(dst.as_slice(), src.as_slice(), 0.008);
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
+    compare_with_error(dst.as_slice(), src.as_slice(), 0.009);
 
     // Test some specific values
     assert_eq!(
@@ -404,24 +474,29 @@ fn quantize_q5k() -> Result<()> {
     let dst = round_vector(&dst);
     assert_eq!(
         [dst[0], dst[128], dst[256], dst[512], dst[800], dst[1023]],
-        [-0.499, -0.372, -0.249, 0.001, 0.279, 0.499]
+        [-0.5, -0.373, -0.25, 0.0, 0.279, 0.499]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ5K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 2.5);
 
-    ggml_quantization_error_test::<BlockQ5K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
-
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q6k() -> Result<()> {
-    use k_quants::BlockQ6K;
+fn quantize_q6k(device: &Device) -> Result<()> {
+    let dtype = GgmlDType::Q6K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ6K>(src.as_slice(), dst.as_mut_slice())?;
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
     compare_with_error(dst.as_slice(), src.as_slice(), 0.008);
 
     // Test some specific values
@@ -435,22 +510,27 @@ fn quantize_q6k() -> Result<()> {
         [-0.497, -0.372, -0.25, -0.0, 0.284, 0.5]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ6K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 2.0);
 
-    ggml_quantization_error_test::<BlockQ6K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
-
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
-#[test]
-fn quantize_q8k() -> Result<()> {
-    use k_quants::BlockQ8K;
+fn quantize_q8k(device: &Device) -> Result<()> {
+    let dtype = GgmlDType::Q8K;
+    let src = get_test_vector2(0.5, 1024, device)?;
+    let quant = quantized::QTensor::quantize(&src, dtype)?;
+    let dst = quant.dequantize(device)?;
 
-    let (src, mut dst) = get_test_vector(0.5, 1024);
-    let _quant = quantize_roundtrip::<BlockQ8K>(src.as_slice(), dst.as_mut_slice())?;
-    compare_with_error(dst.as_slice(), src.as_slice(), 0.003);
+    let src = src.to_vec1::<f32>()?;
+    let dst = dst.to_vec1::<f32>()?;
+    compare_with_error(dst.as_slice(), src.as_slice(), 0.008);
 
     // Test some specific values
     assert_eq!(
@@ -463,15 +543,79 @@ fn quantize_q8k() -> Result<()> {
         [-0.5, -0.375, -0.25, -0.0, 0.281, 0.499]
     );
 
-    let (src_big, mut dst_big) = get_test_vector(128.0, 1024);
-    let _quant_big = quantize_roundtrip::<BlockQ8K>(src_big.as_slice(), dst_big.as_mut_slice())?;
+    let src_big = get_test_vector2(128.0, 1024, device)?;
+    let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
+    let dst_big = quant_big.dequantize(device)?;
+
+    let src_big = src_big.to_vec1::<f32>()?;
+    let dst_big = dst_big.to_vec1::<f32>()?;
     compare_with_error(dst_big.as_slice(), src_big.as_slice(), 0.6);
 
-    ggml_quantization_error_test::<BlockQ8K>(GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
-
+    ggml_quantization_error_test(dtype, device, GGML_MAX_QUANTIZATION_TOTAL_ERROR)?;
     Ok(())
 }
 
+test_device!(
+    quantize_q4_0,
+    quantize_q4_0_cpu,
+    quantize_q4_0_cuda,
+    quantize_q4_0_metal
+);
+test_device!(
+    quantize_q4_1,
+    quantize_q4_1_cpu,
+    quantize_q4_1_cuda,
+    quantize_q4_1_metal
+);
+test_device!(
+    quantize_q5_0,
+    quantize_q5_0_cpu,
+    quantize_q5_0_cuda,
+    quantize_q5_0_metal
+);
+test_device!(
+    quantize_q5_1,
+    quantize_q5_1_cpu,
+    quantize_q5_1_cuda,
+    quantize_q5_1_metal
+);
+test_device!(
+    quantize_q2k,
+    quantize_q2k_cpu,
+    quantize_q2k_cuda,
+    quantize_q2k_metal
+);
+test_device!(
+    quantize_q3k,
+    quantize_q3k_cpu,
+    quantize_q3k_cuda,
+    quantize_q3k_metal
+);
+test_device!(
+    quantize_q4k,
+    quantize_q4k_cpu,
+    quantize_q4k_cuda,
+    quantize_q4k_metal
+);
+test_device!(
+    quantize_q5k,
+    quantize_q5k_cpu,
+    quantize_q5k_cuda,
+    quantize_q5k_metal
+);
+test_device!(
+    quantize_q6k,
+    quantize_q6k_cpu,
+    quantize_q6k_cuda,
+    quantize_q6k_metal
+);
+test_device!(
+    quantize_q8k,
+    quantize_q8k_cpu,
+    quantize_q8k_cuda,
+    quantize_q8k_metal
+);
+
 /// Very simple dot product implementation
 fn vec_dot_reference(a: &[f32], b: &[f32]) -> f32 {
     a.iter().zip(b).map(|(a, b)| a * b).sum()
@@ -487,54 +631,66 @@ fn ggml_reference_matmul_error(dtype: GgmlDType) -> Result<f32> {
         GgmlDType::Q5K => 0.000740,
         GgmlDType::Q6K => 0.000952,
         GgmlDType::Q4_0 => 0.001143,
-        GgmlDType::Q4_1 => 0.007784,
+        GgmlDType::Q4_1 => 0.008,
         GgmlDType::Q5_0 => 0.001353,
-        GgmlDType::Q5_1 => 0.001363,
+        GgmlDType::Q5_1 => 0.00149,
         GgmlDType::Q8_0 => 0.000092,
 
         // Not from the ggml repo.
         GgmlDType::Q8K => 0.00065,
-        _ => candle_core::bail!("No GGML results for quantization type {dtype:?}",),
+        _ => bail!("No GGML results for quantization type {dtype:?}",),
     };
     Ok(err)
 }
 
-/// Mirrores the GGML matmul unit test: https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L76-L91
+/// Similar to the GGML matmul unit test:
+/// https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L76-L91
 fn ggml_matmul_error_test<T: GgmlType>() -> Result<()> {
     let a = create_ggml_like_vector(0.0);
     let b = create_ggml_like_vector(1.0);
+    ggml_matmul_error_test_::<T>(a.as_slice(), b.as_slice(), 1.0)?;
+    // Another example that is more likely to trigger the overflow reported in #1526
+    let a = (0..GGML_TEST_SIZE)
+        .map(|i| i as f32 / GGML_TEST_SIZE as f32)
+        .collect::<Vec<_>>();
+    let b = (0..GGML_TEST_SIZE)
+        .map(|i| i as f32 / GGML_TEST_SIZE as f32)
+        .collect::<Vec<_>>();
+    ggml_matmul_error_test_::<T>(a.as_slice(), b.as_slice(), 2.0)?;
+    Ok(())
+}
+
+fn ggml_matmul_error_test_<T: GgmlType>(a: &[f32], b: &[f32], err_m: f32) -> Result<()> {
     let length = a.len();
 
     let mut a_quant = vec![T::zeros(); length / T::BLCK_SIZE];
     let mut b_quant = vec![T::VecDotType::zeros(); length / T::VecDotType::BLCK_SIZE];
-    T::from_float(&a, &mut a_quant)?;
-    T::VecDotType::from_float(&b, &mut b_quant)?;
+    T::from_float(a, &mut a_quant)?;
+    T::VecDotType::from_float(b, &mut b_quant)?;
 
     let result = T::vec_dot(length, &a_quant, &b_quant)?;
     let result_unopt = T::vec_dot_unopt(length, &a_quant, &b_quant)?;
-    let reference_result = vec_dot_reference(&a, &b);
+    let reference_result = vec_dot_reference(a, b);
 
     if (result - result_unopt).abs() / length as f32 > 1e-6 {
-        candle_core::bail!(
+        bail!(
             "the opt and unopt vec-dot returned different values, opt {result}, unopt {result_unopt}"
         )
     }
 
     let error = (result - reference_result).abs() / length as f32;
 
-    let ggml_error = ggml_reference_matmul_error(T::DTYPE)?;
+    let ggml_error = ggml_reference_matmul_error(T::DTYPE)? * err_m;
 
     if !error.is_finite() || error > GGML_MAX_DOT_PRODUCT_ERROR {
-        candle_core::bail!(
-            "Dot product error {error} exceeds max error {GGML_MAX_DOT_PRODUCT_ERROR}",
-        );
+        bail!("Dot product error {error} exceeds max error {GGML_MAX_DOT_PRODUCT_ERROR}",);
     }
 
     // We diverge slightly due to different rounding behavior / f16 to f32 conversions in GGML
     // => we use a slightly higher error threshold
     const ERROR_LENIENCY: f32 = 0.00001;
     if error - ERROR_LENIENCY > ggml_error {
-        candle_core::bail!(
+        bail!(
             "Dot product error {} exceeds ggml reference error {}",
             error,
             ggml_error
@@ -543,6 +699,16 @@ fn ggml_matmul_error_test<T: GgmlType>() -> Result<()> {
     Ok(())
 }
 
+#[test]
+fn quantized_mm() -> Result<()> {
+    ggml_matmul_error_test::<k_quants::BlockQ4_0>()?;
+    ggml_matmul_error_test::<k_quants::BlockQ4_1>()?;
+    ggml_matmul_error_test::<k_quants::BlockQ5_0>()?;
+    ggml_matmul_error_test::<k_quants::BlockQ5_1>()?;
+    ggml_matmul_error_test::<k_quants::BlockQ8_0>()?;
+    Ok(())
+}
+
 /// generates random tensors of size `m x k` and `n x k` and calculates their expected matrix multiplication result.
 fn get_random_tensors(
     m: usize,
@@ -566,6 +732,108 @@ fn get_random_tensors(
     Ok((lhs, rhs, mm))
 }
 
+#[macro_export]
+macro_rules! quantized_matmul {
+    // TODO: Switch to generating the two last arguments automatically once concat_idents is
+    // stable. https://github.com/rust-lang/rust/issues/29599
+    ($fn_name: ident, $fn_name_cpu: ident, $fn_name_cuda: ident, $fn_name_metal: ident, $dtype: expr) => {
+        fn $fn_name(device: &Device) -> Result<()> {
+            test_matmul(device, (1, 3, 4, 256), $dtype)?;
+            Ok(())
+        }
+
+        test_device!($fn_name, $fn_name_cpu, $fn_name_cuda, $fn_name_metal);
+    };
+}
+
+quantized_matmul!(
+    quantized_matmul_q4_0_bis,
+    quantized_matmul_q4_0_cpu,
+    quantized_matmul_q4_0_cuda,
+    quantized_matmul_q4_0_metal,
+    GgmlDType::Q4_0
+);
+quantized_matmul!(
+    quantized_matmul_q4_1_bis,
+    quantized_matmul_q4_1_cpu,
+    quantized_matmul_q4_1_cuda,
+    quantized_matmul_q4_1_metal,
+    GgmlDType::Q4_1
+);
+quantized_matmul!(
+    quantized_matmul_q5_0_bis,
+    quantized_matmul_q5_0_cpu,
+    quantized_matmul_q5_0_cuda,
+    quantized_matmul_q5_0_metal,
+    GgmlDType::Q5_0
+);
+quantized_matmul!(
+    quantized_matmul_q5_1_bis,
+    quantized_matmul_q5_1_cpu,
+    quantized_matmul_q5_1_cuda,
+    quantized_matmul_q5_1_metal,
+    GgmlDType::Q5_1
+);
+quantized_matmul!(
+    quantized_matmul_q8_0_bis,
+    quantized_matmul_q8_0_cpu,
+    quantized_matmul_q8_0_cuda,
+    quantized_matmul_q8_0_metal,
+    GgmlDType::Q8_0
+);
+// Not implemented in Ggml
+// quantized_matmul!(
+//     quantized_matmul_q8_1_bis,
+//     quantized_matmul_q8_1_cpu,
+//     quantized_matmul_q8_1_cuda,
+//     quantized_matmul_q8_1_metal,
+//     GgmlDType::Q8_1
+// );
+// TODO This is bugged (also bugged in GGML
+quantized_matmul!(
+    quantized_matmul_q2k_bis,
+    quantized_matmul_q2k_cpu,
+    quantized_matmul_q2k_cuda,
+    quantized_matmul_q2k_metal,
+    GgmlDType::Q2K
+);
+quantized_matmul!(
+    quantized_matmul_q3k_bis,
+    quantized_matmul_q3k_cpu,
+    quantized_matmul_q3k_cuda,
+    quantized_matmul_q3k_metal,
+    GgmlDType::Q3K
+);
+quantized_matmul!(
+    quantized_matmul_q4k_bis,
+    quantized_matmul_q4k_cpu,
+    quantized_matmul_q4k_cuda,
+    quantized_matmul_q4k_metal,
+    GgmlDType::Q4K
+);
+quantized_matmul!(
+    quantized_matmul_q5k_bis,
+    quantized_matmul_q5k_cpu,
+    quantized_matmul_q5k_cuda,
+    quantized_matmul_q5k_metal,
+    GgmlDType::Q5K
+);
+quantized_matmul!(
+    quantized_matmul_q6k_bis,
+    quantized_matmul_q6k_cpu,
+    quantized_matmul_q6k_cuda,
+    quantized_matmul_q6k_metal,
+    GgmlDType::Q6K
+);
+// Not implemented on metal
+// quantized_matmul!(
+//     quantized_matmul_q8k_bis,
+//     quantized_matmul_q8k_cpu,
+//     quantized_matmul_q8k_cuda,
+//     quantized_matmul_q8k_metal,
+//     GgmlDType::Q8K
+// );
+
 #[test]
 fn quantized_matmul_q2k() -> Result<()> {
     use k_quants::BlockQ2K;
@@ -578,7 +846,7 @@ fn quantized_matmul_q2k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ2K>(&rhs)?;
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q2K)?;
     let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
@@ -604,7 +872,7 @@ fn quantized_matmul_q3k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ3K>(&rhs)?;
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q3K)?;
     let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
@@ -630,7 +898,7 @@ fn quantized_matmul_q4k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ4K>(&rhs)?;
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q4K)?;
     let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
@@ -656,7 +924,7 @@ fn quantized_matmul_q5k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ5K>(&rhs)?;
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q5K)?;
     let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
@@ -683,7 +951,7 @@ fn quantized_matmul_q6k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ6K>(&rhs)?;
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q6K)?;
     let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 
@@ -708,7 +976,7 @@ fn quantized_matmul_q8k() -> Result<()> {
     let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
     assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
-    let rhs = quantized::QTensor::quantize::<BlockQ8K>(&rhs)?;
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q8K)?;
     let rhs = quantized::QMatMul::from_qtensor(rhs)?;
     let mm = rhs.forward(&lhs)?;
 

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>()?,
@@ -112,6 +120,13 @@ 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]]
@@ -657,6 +672,31 @@ 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(())
 }
 
@@ -1065,13 +1105,39 @@ 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, 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);
@@ -1221,3 +1287,38 @@ fn cumsum() -> Result<()> {
     );
     Ok(())
 }
+
+/// A helper function for floating point comparison. Both a and b must be 1D Tensor and contains the same amount of data.
+/// Assertion passes if the difference of all pairs of a and b is smaller than epsilon.
+fn assert_close(a: &Tensor, b: &Tensor, epsilon: f64) -> Result<()> {
+    let a_vec: Vec<f64> = a.to_vec1()?;
+    let b_vec: Vec<f64> = b.to_vec1()?;
+
+    assert_eq!(a_vec.len(), b_vec.len());
+    for (a, b) in a_vec.iter().zip(b_vec.iter()) {
+        assert!((a - b).abs() < epsilon);
+    }
+    Ok(())
+}
+
+#[test]
+fn log_sum_exp() -> Result<()> {
+    let input = Tensor::new(&[[1f64, 2., 3.], [4., 5., 6.]], &Device::Cpu)?;
+    let output = input.log_sum_exp(D::Minus1)?;
+    // The expectations obtained from pytorch.
+    let expected = Tensor::new(&[3.4076, 6.4076], &Device::Cpu)?;
+    assert_close(&output, &expected, 0.00001)?;
+    Ok(())
+}
+
+#[test]
+fn pow() -> Result<()> {
+    let lhs = Tensor::new(&[[1f32, 2., 3.], [4., 5., 6.]], &Device::Cpu)?;
+    let rhs = (&lhs - 2.)?;
+    let res = lhs.pow(&rhs)?;
+    assert_eq!(
+        test_utils::to_vec2_round(&res, 4)?,
+        [[1.0, 1.0, 3.0], [16.0, 125.0, 1296.0001]]
+    );
+    Ok(())
+}

candle-datasets/Cargo.toml

@@ -11,8 +11,8 @@ readme = "README.md"
 
 [dependencies]
 byteorder = { workspace = true }
-candle = { path = "../candle-core", version = "0.3.1", package = "candle-core" }
-candle-nn = { path = "../candle-nn", version = "0.3.1" }
+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

@@ -11,14 +11,17 @@ readme = "README.md"
 
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
-candle = { path = "../candle-core", version = "0.3.1", package = "candle-core" }
-candle-datasets = { path = "../candle-datasets", version = "0.3.1" }
-candle-nn = { path = "../candle-nn", version = "0.3.1" }
-candle-transformers = { path = "../candle-transformers", version = "0.3.1" }
-candle-flash-attn = { path = "../candle-flash-attn", version = "0.3.1", optional = true }
-candle-onnx = { path = "../candle-onnx", version = "0.3.1", 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 }
@@ -27,13 +30,14 @@ rayon = { workspace = 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 }
@@ -41,23 +45,24 @@ rusttype = { 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"]
 
 [[example]]
 name = "llama_multiprocess"
@@ -74,3 +79,25 @@ 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 = ["symphonia"]
+
+

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/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/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 {

candle-examples/examples/efficientvit/README.md

@@ -0,0 +1,20 @@
+# candle-efficientvit
+
+[EfficientViT: Memory Efficient Vision Transformer with Cascaded Group Attention](https://arxiv.org/abs/2305.07027).
+
+This candle implementation uses a pre-trained EfficientViT (from Microsoft Research Asia) 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 efficientvit --release -- --image candle-examples/examples/yolo-v8/assets/bike.jpg --which m1
+
+loaded image Tensor[dims 3, 224, 224; f32]
+model built
+mountain bike, all-terrain bike, off-roader: 69.80%
+unicycle, monocycle     : 13.03%
+bicycle-built-for-two, tandem bicycle, tandem: 9.28%
+crash helmet            : 2.25%
+alp                     : 0.46%
+```

candle-examples/examples/efficientvit/main.rs

@@ -0,0 +1,99 @@
+#[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::efficientvit;
+
+#[derive(Clone, Copy, Debug, ValueEnum)]
+enum Which {
+    M0,
+    M1,
+    M2,
+    M3,
+    M4,
+    M5,
+}
+
+impl Which {
+    fn model_filename(&self) -> String {
+        let name = match self {
+            Self::M0 => "m0",
+            Self::M1 => "m1",
+            Self::M2 => "m2",
+            Self::M3 => "m3",
+            Self::M4 => "m4",
+            Self::M5 => "m5",
+        };
+        format!("timm/efficientvit_{}.r224_in1k", name)
+    }
+
+    fn config(&self) -> efficientvit::Config {
+        match self {
+            Self::M0 => efficientvit::Config::m0(),
+            Self::M1 => efficientvit::Config::m1(),
+            Self::M2 => efficientvit::Config::m2(),
+            Self::M3 => efficientvit::Config::m3(),
+            Self::M4 => efficientvit::Config::m4(),
+            Self::M5 => efficientvit::Config::m5(),
+        }
+    }
+}
+
+#[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::M0)]
+    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 = efficientvit::efficientvit(&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/encodec/README.md

@@ -0,0 +1,20 @@
+# candle-endocec
+
+[EnCodec](https://huggingface.co/facebook/encodec_24khz) is a high-quality audio
+compression model using an encoder/decoder architecture with residual vector
+quantization.
+
+## Running one example
+
+```bash
+cargo run --example encodec --features symphonia --release -- code-to-audio \
+    candle-examples/examples/encodec/jfk-codes.safetensors \
+    jfk.wav
+```
+
+This decodes the EnCodec tokens stored in `jfk-codes.safetensors` and generates
+an output wav file containing the audio data. Instead of `code-to-audio` one
+can use:
+- `audio-to-audio in.mp3 out.wav`: encodes the input audio file then decodes it to a wav file.
+- `audio-to-code in.mp3 out.safetensors`: generates a safetensors file
+  containing EnCodec tokens for the input audio file.

candle-examples/examples/encodec/main.rs

@@ -0,0 +1,142 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::Result;
+use candle::{DType, IndexOp, Tensor};
+use candle_nn::VarBuilder;
+use candle_transformers::models::encodec::{Config, Model};
+use clap::{Parser, ValueEnum};
+use hf_hub::api::sync::Api;
+
+fn conv<T>(samples: &mut Vec<f32>, data: std::borrow::Cow<symphonia::core::audio::AudioBuffer<T>>)
+where
+    T: symphonia::core::sample::Sample,
+    f32: symphonia::core::conv::FromSample<T>,
+{
+    use symphonia::core::audio::Signal;
+    use symphonia::core::conv::FromSample;
+    samples.extend(data.chan(0).iter().map(|v| f32::from_sample(*v)))
+}
+
+fn pcm_decode<P: AsRef<std::path::Path>>(path: P) -> anyhow::Result<(Vec<f32>, u32)> {
+    use symphonia::core::audio::{AudioBufferRef, Signal};
+
+    let src = std::fs::File::open(path)?;
+    let mss = symphonia::core::io::MediaSourceStream::new(Box::new(src), Default::default());
+    let hint = symphonia::core::probe::Hint::new();
+    let meta_opts: symphonia::core::meta::MetadataOptions = Default::default();
+    let fmt_opts: symphonia::core::formats::FormatOptions = Default::default();
+    let probed = symphonia::default::get_probe().format(&hint, mss, &fmt_opts, &meta_opts)?;
+    let mut format = probed.format;
+    let track = format
+        .tracks()
+        .iter()
+        .find(|t| t.codec_params.codec != symphonia::core::codecs::CODEC_TYPE_NULL)
+        .expect("no supported audio tracks");
+    let mut decoder = symphonia::default::get_codecs()
+        .make(&track.codec_params, &Default::default())
+        .expect("unsupported codec");
+    let track_id = track.id;
+    let sample_rate = track.codec_params.sample_rate.unwrap_or(0);
+    let mut pcm_data = Vec::new();
+    while let Ok(packet) = format.next_packet() {
+        while !format.metadata().is_latest() {
+            format.metadata().pop();
+        }
+        if packet.track_id() != track_id {
+            continue;
+        }
+        match decoder.decode(&packet)? {
+            AudioBufferRef::F32(buf) => pcm_data.extend(buf.chan(0)),
+            AudioBufferRef::U8(data) => conv(&mut pcm_data, data),
+            AudioBufferRef::U16(data) => conv(&mut pcm_data, data),
+            AudioBufferRef::U24(data) => conv(&mut pcm_data, data),
+            AudioBufferRef::U32(data) => conv(&mut pcm_data, data),
+            AudioBufferRef::S8(data) => conv(&mut pcm_data, data),
+            AudioBufferRef::S16(data) => conv(&mut pcm_data, data),
+            AudioBufferRef::S24(data) => conv(&mut pcm_data, data),
+            AudioBufferRef::S32(data) => conv(&mut pcm_data, data),
+            AudioBufferRef::F64(data) => conv(&mut pcm_data, data),
+        }
+    }
+    Ok((pcm_data, sample_rate))
+}
+
+#[derive(Clone, Debug, Copy, PartialEq, Eq, ValueEnum)]
+enum Action {
+    AudioToAudio,
+    AudioToCode,
+    CodeToAudio,
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// The action to be performed, specifies the format for the input and output data.
+    action: Action,
+
+    /// The input file, either an audio file or some encodec tokens stored as safetensors.
+    in_file: String,
+
+    /// The output file, either a wave audio file or some encodec tokens stored as safetensors.
+    out_file: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// The model weight file, in safetensor format.
+    #[arg(long)]
+    model: Option<String>,
+}
+
+fn main() -> Result<()> {
+    let args = Args::parse();
+    let device = candle_examples::device(args.cpu)?;
+    let model = match args.model {
+        Some(model) => std::path::PathBuf::from(model),
+        None => Api::new()?
+            .model("facebook/encodec_24khz".to_string())
+            .get("model.safetensors")?,
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model], DType::F32, &device)? };
+    let config = Config::default();
+    let model = Model::new(&config, vb)?;
+
+    let codes = match args.action {
+        Action::CodeToAudio => {
+            let codes = candle::safetensors::load(args.in_file, &device)?;
+            codes.get("codes").expect("no codes in input file").clone()
+        }
+        Action::AudioToCode | Action::AudioToAudio => {
+            let (pcm, sample_rate) = pcm_decode(args.in_file)?;
+            if sample_rate != 24_000 {
+                println!("WARNING: encodec uses a 24khz sample rate, input uses {sample_rate}")
+            }
+            let pcm_len = pcm.len();
+            let pcm = Tensor::from_vec(pcm, (1, 1, pcm_len), &device)?;
+            println!("input pcm shape: {:?}", pcm.shape());
+            model.encode(&pcm)?
+        }
+    };
+    println!("codes shape: {:?}", codes.shape());
+
+    match args.action {
+        Action::AudioToCode => {
+            codes.save_safetensors("codes", &args.out_file)?;
+        }
+        Action::AudioToAudio | Action::CodeToAudio => {
+            let pcm = model.decode(&codes)?;
+            println!("output pcm shape: {:?}", pcm.shape());
+            let pcm = pcm.i(0)?.i(0)?;
+            let pcm = candle_examples::audio::normalize_loudness(&pcm, 24_000, true)?;
+            let pcm = pcm.to_vec1::<f32>()?;
+            let mut output = std::fs::File::create(&args.out_file)?;
+            candle_examples::wav::write_pcm_as_wav(&mut output, &pcm, 24_000)?;
+        }
+    }
+    Ok(())
+}

candle-examples/examples/falcon/main.rs

@@ -165,14 +165,7 @@ fn main() -> Result<()> {
         args.revision,
     ));
     let tokenizer_filename = repo.get("tokenizer.json")?;
-    let mut filenames = vec![];
-    for rfilename in [
-        "model-00001-of-00002.safetensors",
-        "model-00002-of-00002.safetensors",
-    ] {
-        let filename = repo.get(rfilename)?;
-        filenames.push(filename);
-    }
+    let filenames = 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)?;
 

candle-examples/examples/gemma/README.md

@@ -0,0 +1,27 @@
+# candle-gemma: 2b and 7b LLMs from Google DeepMind
+
+[Gemma](https://ai.google.dev/gemma/docs) is a collection of lightweight open
+models published by Google Deepmind with a 2b and a 7b variant.
+
+In order to use the example below, you have to accept the license on the
+[HuggingFace Hub Gemma repo](https://huggingface.co/google/gemma-7b) and set up
+your access token via the [HuggingFace cli login
+command](https://huggingface.co/docs/huggingface_hub/guides/cli#huggingface-cli-login).
+
+## Running the example
+
+```bash
+$ cargo run --example gemma --release -- --prompt "fn count_primes(max_n: usize)"
+fn count_primes(max_n: usize) -> usize {
+    let mut primes = vec![true; max_n];
+    for i in 2..=max_n {
+        if primes[i] {
+            for j in i * i..max_n {
+                primes[j] = false;
+             }
+         }
+    }
+    primes.len()
+}
+```
+

candle-examples/examples/gemma/main.rs

@@ -0,0 +1,256 @@
+#[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::gemma::{Config, Model};
+
+use candle::{DType, Device, Tensor};
+use candle_examples::token_output_stream::TokenOutputStream;
+use candle_nn::VarBuilder;
+use candle_transformers::generation::LogitsProcessor;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::Tokenizer;
+
+struct TextGeneration {
+    model: Model,
+    device: Device,
+    tokenizer: TokenOutputStream,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+}
+
+impl TextGeneration {
+    #[allow(clippy::too_many_arguments)]
+    fn new(
+        model: Model,
+        tokenizer: Tokenizer,
+        seed: u64,
+        temp: Option<f64>,
+        top_p: Option<f64>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        device: &Device,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            tokenizer: TokenOutputStream::new(tokenizer),
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            device: device.clone(),
+        }
+    }
+
+    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
+        use std::io::Write;
+        self.tokenizer.clear();
+        let mut tokens = self
+            .tokenizer
+            .tokenizer()
+            .encode(prompt, true)
+            .map_err(E::msg)?
+            .get_ids()
+            .to_vec();
+        for &t in tokens.iter() {
+            if let Some(t) = self.tokenizer.next_token(t)? {
+                print!("{t}")
+            }
+        }
+        std::io::stdout().flush()?;
+
+        let mut generated_tokens = 0usize;
+        let eos_token = match self.tokenizer.get_token("<eos>") {
+            Some(token) => token,
+            None => anyhow::bail!("cannot find the <eos> token"),
+        };
+        let start_gen = std::time::Instant::now();
+        for index in 0..sample_len {
+            let context_size = if index > 0 { 1 } else { tokens.len() };
+            let start_pos = tokens.len().saturating_sub(context_size);
+            let ctxt = &tokens[start_pos..];
+            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+            let logits = self.model.forward(&input, start_pos)?;
+            let logits = logits.squeeze(0)?.squeeze(0)?.to_dtype(DType::F32)?;
+            let logits = if self.repeat_penalty == 1. {
+                logits
+            } else {
+                let start_at = tokens.len().saturating_sub(self.repeat_last_n);
+                candle_transformers::utils::apply_repeat_penalty(
+                    &logits,
+                    self.repeat_penalty,
+                    &tokens[start_at..],
+                )?
+            };
+
+            let next_token = self.logits_processor.sample(&logits)?;
+            tokens.push(next_token);
+            generated_tokens += 1;
+            if next_token == eos_token {
+                break;
+            }
+            if let Some(t) = self.tokenizer.next_token(next_token)? {
+                print!("{t}");
+                std::io::stdout().flush()?;
+            }
+        }
+        let dt = start_gen.elapsed();
+        if let Some(rest) = self.tokenizer.decode_rest().map_err(E::msg)? {
+            print!("{rest}");
+        }
+        std::io::stdout().flush()?;
+        println!(
+            "\n{generated_tokens} tokens generated ({:.2} token/s)",
+            generated_tokens as f64 / dt.as_secs_f64(),
+        );
+        Ok(())
+    }
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    #[arg(long)]
+    prompt: String,
+
+    /// The temperature used to generate samples.
+    #[arg(long)]
+    temperature: Option<f64>,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// The seed to use when generating random samples.
+    #[arg(long, default_value_t = 299792458)]
+    seed: u64,
+
+    /// The length of the sample to generate (in tokens).
+    #[arg(long, short = 'n', default_value_t = 10000)]
+    sample_len: usize,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long, default_value = "main")]
+    revision: String,
+
+    #[arg(long)]
+    tokenizer_file: Option<String>,
+
+    #[arg(long)]
+    config_file: Option<String>,
+
+    #[arg(long)]
+    weight_files: Option<String>,
+
+    /// Penalty to be applied for repeating tokens, 1. means no penalty.
+    #[arg(long, default_value_t = 1.1)]
+    repeat_penalty: f32,
+
+    /// The context size to consider for the repeat penalty.
+    #[arg(long, default_value_t = 64)]
+    repeat_last_n: usize,
+}
+
+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) => match model_id.as_str() {
+            "7b-it" => "google/gemma-7b-it".to_string(),
+            "7b" => "google/gemma-7b".to_string(),
+            "2b-it" => "google/gemma-2b-it".to_string(),
+            "2b" => "google/gemma-2b".to_string(),
+            _ => model_id.to_string(),
+        },
+        None => "google/gemma-2b".to_string(),
+    };
+    let repo = api.repo(Repo::with_revision(
+        model_id,
+        RepoType::Model,
+        args.revision,
+    ));
+    let tokenizer_filename = match args.tokenizer_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("tokenizer.json")?,
+    };
+    let config_filename = match args.config_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("config.json")?,
+    };
+    let filenames = match args.weight_files {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        None => 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 config: Config = serde_json::from_reader(std::fs::File::open(config_filename)?)?;
+
+    let start = std::time::Instant::now();
+    let device = candle_examples::device(args.cpu)?;
+    let dtype = if device.is_cuda() {
+        DType::BF16
+    } else {
+        DType::F32
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &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,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}

candle-examples/examples/llama/main.rs

@@ -13,7 +13,7 @@ extern crate accelerate_src;
 extern crate intel_mkl_src;
 
 use anyhow::{bail, Error as E, Result};
-use clap::Parser;
+use clap::{Parser, ValueEnum};
 
 use candle::{DType, Tensor};
 use candle_nn::VarBuilder;
@@ -22,11 +22,21 @@ use hf_hub::{api::sync::Api, Repo, RepoType};
 use std::io::Write;
 
 use candle_transformers::models::llama as model;
-use model::{Config, Llama, LlamaConfig};
+use model::{Llama, LlamaConfig};
 
 const EOS_TOKEN: &str = "</s>";
 const DEFAULT_PROMPT: &str = "My favorite theorem is ";
 
+#[derive(Clone, Debug, Copy, PartialEq, Eq, ValueEnum)]
+enum Which {
+    V1,
+    V2,
+    #[value(name = "solar-10.7b")]
+    Solar10_7B,
+    #[value(name = "tiny-llama-1.1b-chat")]
+    TinyLlama1_1BChat,
+}
+
 #[derive(Parser, Debug)]
 #[command(author, version, about, long_about = None)]
 struct Args {
@@ -34,10 +44,6 @@ struct Args {
     #[arg(long)]
     cpu: bool,
 
-    /// Use npy instead of safetensors
-    #[arg(long)]
-    npy: Option<String>,
-
     /// The temperature used to generate samples.
     #[arg(long)]
     temperature: Option<f64>,
@@ -51,7 +57,7 @@ struct Args {
     seed: u64,
 
     /// The length of the sample to generate (in tokens).
-    #[arg(long, default_value_t = 100)]
+    #[arg(long, default_value_t = 10000)]
     sample_len: usize,
 
     /// Disable the key-value cache.
@@ -76,17 +82,13 @@ struct Args {
     #[arg(long)]
     revision: Option<String>,
 
-    #[arg(long)]
-    v1: bool,
+    /// The model size to use.
+    #[arg(long, default_value = "v2")]
+    which: Which,
 
     #[arg(long)]
     use_flash_attn: bool,
 
-    /// The folder name that contains safetensor weights and json files
-    /// (same structure as huggingface online)
-    #[arg(long)]
-    local_weights: Option<String>,
-
     /// Penalty to be applied for repeating tokens, 1. means no penalty.
     #[arg(long, default_value_t = 1.0)]
     repeat_penalty: f32,
@@ -118,65 +120,33 @@ fn main() -> Result<()> {
         Some(dtype) => bail!("Unsupported dtype {dtype}"),
         None => DType::F16,
     };
-    let (llama, tokenizer_filename, cache) = match args.npy {
-        Some(filename) => {
-            let config = if args.v1 {
-                Config::config_7b_v1(args.use_flash_attn)
-            } else {
-                Config::config_7b_v2(args.use_flash_attn)
-            };
-            let cache = model::Cache::new(!args.no_kv_cache, dtype, &config, &device)?;
-            let vb = VarBuilder::from_npz(filename, dtype, &device)?;
-            let tokenizer = std::path::PathBuf::from("llama-tokenizer.json");
-            (Llama::load(vb, &cache, &config)?, tokenizer, cache)
-        }
-        None => {
-            let api = Api::new()?;
-            let model_id = args.model_id.unwrap_or_else(|| {
-                if args.v1 {
-                    "Narsil/amall-7b".to_string()
-                } else {
-                    "meta-llama/Llama-2-7b-hf".to_string()
-                }
-            });
-            println!("loading the model weights from {model_id}");
-            let revision = args.revision.unwrap_or("main".to_string());
-            let api = api.repo(Repo::with_revision(model_id, RepoType::Model, revision));
+    let (llama, tokenizer_filename, mut cache) = {
+        let api = Api::new()?;
+        let model_id = args.model_id.unwrap_or_else(|| match args.which {
+            Which::V1 => "Narsil/amall-7b".to_string(),
+            Which::V2 => "meta-llama/Llama-2-7b-hf".to_string(),
+            Which::Solar10_7B => "upstage/SOLAR-10.7B-v1.0".to_string(),
+            Which::TinyLlama1_1BChat => "TinyLlama/TinyLlama-1.1B-Chat-v1.0".to_string(),
+        });
+        println!("loading the model weights from {model_id}");
+        let revision = args.revision.unwrap_or("main".to_string());
+        let api = api.repo(Repo::with_revision(model_id, RepoType::Model, revision));
 
-            let tokenizer_filename = match &args.local_weights {
-                Some(path) => (path.to_owned() + "tokenizer.json").into(),
-                _ => api.get("tokenizer.json")?,
-            };
+        let tokenizer_filename = api.get("tokenizer.json")?;
+        let config_filename = api.get("config.json")?;
+        let config: LlamaConfig = serde_json::from_slice(&std::fs::read(config_filename)?)?;
+        let config = config.into_config(args.use_flash_attn);
 
-            let config_filename = match &args.local_weights {
-                Some(path) => (path.to_owned() + "config.json").into(),
-                _ => api.get("config.json")?,
-            };
-            let config: LlamaConfig = serde_json::from_slice(&std::fs::read(config_filename)?)?;
-            let config = config.into_config(args.use_flash_attn);
-
-            let mut filenames = vec![];
-            for rfilename in [
-                "model-00001-of-00002.safetensors",
-                "model-00002-of-00002.safetensors",
-            ] {
-                match &args.local_weights {
-                    Some(path) => {
-                        filenames.push((path.to_owned() + rfilename).into());
-                    }
-                    _ => {
-                        let filename = api.get(rfilename)?;
-                        filenames.push(filename);
-                    }
-                };
+        let filenames = match args.which {
+            Which::V1 | Which::V2 | Which::Solar10_7B => {
+                candle_examples::hub_load_safetensors(&api, "model.safetensors.index.json")?
             }
+            Which::TinyLlama1_1BChat => vec![api.get("model.safetensors")?],
+        };
+        let cache = model::Cache::new(!args.no_kv_cache, dtype, &config, &device)?;
 
-            println!("building the model");
-            let cache = model::Cache::new(!args.no_kv_cache, dtype, &config, &device)?;
-
-            let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
-            (Llama::load(vb, &cache, &config)?, tokenizer_filename, cache)
-        }
+        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
+        (Llama::load(vb, &config)?, tokenizer_filename, cache)
     };
     let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
     let eos_token_id = tokenizer.token_to_id(EOS_TOKEN);
@@ -186,6 +156,7 @@ fn main() -> Result<()> {
         .map_err(E::msg)?
         .get_ids()
         .to_vec();
+    let mut tokenizer = candle_examples::token_output_stream::TokenOutputStream::new(tokenizer);
 
     println!("starting the inference loop");
     print!("{prompt}");
@@ -194,14 +165,14 @@ fn main() -> Result<()> {
     let mut index_pos = 0;
     let mut token_generated = 0;
     for index in 0..args.sample_len {
-        let context_size = if cache.use_kv_cache && index > 0 {
-            1
+        let (context_size, context_index) = if cache.use_kv_cache && index > 0 {
+            (1, index_pos)
         } else {
-            tokens.len()
+            (tokens.len(), 0)
         };
         let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
         let input = Tensor::new(ctxt, &device)?.unsqueeze(0)?;
-        let logits = llama.forward(&input, index_pos)?;
+        let logits = llama.forward(&input, context_index, &mut cache)?;
         let logits = logits.squeeze(0)?;
         let logits = if args.repeat_penalty == 1. {
             logits
@@ -219,18 +190,16 @@ fn main() -> Result<()> {
         token_generated += 1;
         tokens.push(next_token);
 
-        // Extracting the last token as a string is complicated, here we just apply some simple
-        // heuristics as it seems to work well enough for this example. See the following for more
-        // details:
-        // https://github.com/huggingface/tokenizers/issues/1141#issuecomment-1562644141
-        if let Some(text) = tokenizer.id_to_token(next_token) {
-            let text = text.replace('▁', " ").replace("<0x0A>", "\n");
-            print!("{text}");
-            std::io::stdout().flush()?;
-        }
         if Some(next_token) == eos_token_id {
             break;
         }
+        if let Some(t) = tokenizer.next_token(next_token)? {
+            print!("{t}");
+            std::io::stdout().flush()?;
+        }
+    }
+    if let Some(rest) = tokenizer.decode_rest().map_err(E::msg)? {
+        print!("{rest}");
     }
     let dt = start_gen.elapsed();
     println!(

candle-examples/examples/llama2-c/main.rs

@@ -19,7 +19,7 @@ use candle_transformers::generation::LogitsProcessor;
 use std::io::Write;
 use tokenizers::Tokenizer;
 
-use model::{Config, Llama};
+use model::{Cache, Config, Llama};
 use qmodel::QLlama;
 use weights::TransformerWeights;
 
@@ -160,10 +160,10 @@ enum Model {
 }
 
 impl Model {
-    fn forward(&self, xs: &Tensor, pos: usize) -> anyhow::Result<Tensor> {
+    fn forward(&self, xs: &Tensor, pos: usize, cache: &mut Cache) -> anyhow::Result<Tensor> {
         match self {
-            Self::Llama(l) => Ok(l.forward(xs, pos)?),
-            Self::QLlama(l) => Ok(l.forward(xs, pos)?),
+            Self::Llama(l) => Ok(l.forward(xs, pos, cache)?),
+            Self::QLlama(l) => Ok(l.forward(xs, pos, cache)?),
         }
     }
 }
@@ -188,8 +188,8 @@ fn run_eval(args: &EvaluationCmd, common_args: &Args) -> Result<()> {
     let config = Config::from_reader(&mut file)?;
     let weights = TransformerWeights::from_reader(&mut file, &config, &device)?;
     let vb = weights.var_builder(&config, &device)?;
-    let cache = model::Cache::new(false, &config, vb.pp("rot"))?;
-    let model = Llama::load(vb, &cache, config)?;
+    let mut cache = Cache::new(false, &config, vb.pp("rot"))?;
+    let model = Llama::load(vb, config)?;
 
     let tokens = match &args.pretokenized_dir {
         None => {
@@ -235,7 +235,7 @@ fn run_eval(args: &EvaluationCmd, common_args: &Args) -> Result<()> {
     let batch_iter = candle_datasets::Batcher::new_r2(iter).batch_size(args.batch_size);
     for inp_tgt in batch_iter {
         let (inp, tgt) = inp_tgt?;
-        let logits = model.forward(&inp, 0)?;
+        let logits = model.forward(&inp, 0, &mut cache)?;
         let loss = candle_nn::loss::cross_entropy(&logits.flatten_to(1)?, &tgt.flatten_to(1)?)?;
         println!("{}", loss.to_vec0::<f32>()?);
     }
@@ -261,8 +261,8 @@ fn run_inference(args: &InferenceCmd, common_args: &Args) -> Result<()> {
     let is_safetensors = config_path
         .extension()
         .map_or(false, |v| v == "safetensors");
-    let (model, config) = if is_gguf {
-        let vb = qmodel::VarBuilder::from_gguf(config_path)?;
+    let (model, config, mut cache) = if is_gguf {
+        let vb = qmodel::VarBuilder::from_gguf(config_path, &device)?;
         let (_vocab_size, dim) = vb
             .get_no_shape("model.embed_tokens.weight")?
             .shape()
@@ -279,13 +279,13 @@ fn run_inference(args: &InferenceCmd, common_args: &Args) -> Result<()> {
                 (config.seq_len, config.head_size() / 2),
                 "rot.freq_cis_real",
             )?
-            .dequantize(&candle::Device::Cpu)?;
+            .dequantize(&device)?;
         let freq_cis_imag = vb
             .get(
                 (config.seq_len, config.head_size() / 2),
                 "rot.freq_cis_imag",
             )?
-            .dequantize(&candle::Device::Cpu)?;
+            .dequantize(&device)?;
 
         let fake_vb = candle_nn::VarBuilder::from_tensors(
             [
@@ -295,18 +295,18 @@ fn run_inference(args: &InferenceCmd, common_args: &Args) -> Result<()> {
             .into_iter()
             .collect(),
             candle::DType::F32,
-            &candle::Device::Cpu,
+            &device,
         );
         let cache = model::Cache::new(true, &config, fake_vb)?;
-        let model = Model::QLlama(QLlama::load(vb, &cache, config.clone())?);
-        (model, config)
+        let model = Model::QLlama(QLlama::load(vb, config.clone())?);
+        (model, config, cache)
     } else if is_safetensors {
         let config = Config::tiny_15m();
         let tensors = candle::safetensors::load(config_path, &device)?;
         let vb = candle_nn::VarBuilder::from_tensors(tensors, candle::DType::F32, &device);
         let cache = model::Cache::new(true, &config, vb.pp("rot"))?;
-        let model = Model::Llama(Llama::load(vb, &cache, config.clone())?);
-        (model, config)
+        let model = Model::Llama(Llama::load(vb, config.clone())?);
+        (model, config, cache)
     } else {
         let mut file = std::fs::File::open(config_path)?;
         let config = Config::from_reader(&mut file)?;
@@ -314,8 +314,8 @@ fn run_inference(args: &InferenceCmd, common_args: &Args) -> Result<()> {
         let weights = TransformerWeights::from_reader(&mut file, &config, &device)?;
         let vb = weights.var_builder(&config, &device)?;
         let cache = model::Cache::new(true, &config, vb.pp("rot"))?;
-        let model = Model::Llama(Llama::load(vb, &cache, config.clone())?);
-        (model, config)
+        let model = Model::Llama(Llama::load(vb, config.clone())?);
+        (model, config, cache)
     };
 
     println!("starting the inference loop");
@@ -328,6 +328,7 @@ fn run_inference(args: &InferenceCmd, common_args: &Args) -> Result<()> {
         .map_err(E::msg)?
         .get_ids()
         .to_vec();
+    let mut tokenizer = candle_examples::token_output_stream::TokenOutputStream::new(tokenizer);
 
     let start_gen = std::time::Instant::now();
     for index in 0.. {
@@ -337,7 +338,7 @@ fn run_inference(args: &InferenceCmd, common_args: &Args) -> Result<()> {
         let context_size = if index > 0 { 1 } else { tokens.len() };
         let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
         let input = Tensor::new(ctxt, &device)?.unsqueeze(0)?;
-        let logits = model.forward(&input, index_pos)?;
+        let logits = model.forward(&input, index_pos, &mut cache)?;
         let logits = logits.i((0, logits.dim(1)? - 1))?;
         let logits = if common_args.repeat_penalty == 1. || tokens.is_empty() {
             logits
@@ -353,16 +354,14 @@ fn run_inference(args: &InferenceCmd, common_args: &Args) -> Result<()> {
 
         let next_token = logits_processor.sample(&logits)?;
         tokens.push(next_token);
-        // Extracting the last token as a string is complicated, here we just apply some simple
-        // heuristics as it seems to work well enough for this example. See the following for more
-        // details:
-        // https://github.com/huggingface/tokenizers/issues/1141#issuecomment-1562644141
-        if let Some(text) = tokenizer.id_to_token(next_token) {
-            let text = text.replace('▁', " ").replace("<0x0A>", "\n");
-            print!("{text}");
+        if let Some(t) = tokenizer.next_token(next_token)? {
+            print!("{t}");
             std::io::stdout().flush()?;
         }
     }
+    if let Some(rest) = tokenizer.decode_rest().map_err(E::msg)? {
+        print!("{rest}");
+    }
     let dt = start_gen.elapsed();
     println!(
         "\n{} tokens generated ({:.2} token/s)\n",

candle-examples/examples/llama2-c/training.rs

@@ -8,6 +8,7 @@ fn valid_loss(
     model: &Llama,
     args: &crate::TrainingCmd,
     device: &Device,
+    cache: &mut Cache,
 ) -> Result<f64> {
     let iter = DatasetRandomIter::new(dataset, true, model.config.seq_len, device.clone());
     let batch_iter = candle_datasets::Batcher::new_r2(iter).batch_size(args.batch_size);
@@ -15,7 +16,7 @@ fn valid_loss(
     let mut cnt = 0usize;
     for inp_tgt in batch_iter.take(50) {
         let (inp, tgt) = inp_tgt?;
-        let logits = model.forward(&inp, 0)?;
+        let logits = model.forward(&inp, 0, cache)?;
         let loss = candle_nn::loss::cross_entropy(&logits.flatten_to(1)?, &tgt.flatten_to(1)?)?;
         sum_ce += loss.to_vec0::<f32>()? as f64;
         cnt += 1;
@@ -37,8 +38,8 @@ pub fn run(args: &crate::TrainingCmd, common_args: &crate::Args) -> Result<()> {
     let iter = DatasetRandomIter::new(&dataset, false, config.seq_len, device.clone());
     let batch_iter = candle_datasets::Batcher::new_r2(iter).batch_size(args.batch_size);
 
-    let cache = Cache::new(false, &config, vb.pp("rot"))?;
-    let model = Llama::load(vb, &cache, config)?;
+    let mut cache = Cache::new(false, &config, vb.pp("rot"))?;
+    let model = Llama::load(vb, config)?;
     let params = candle_nn::ParamsAdamW {
         lr: args.learning_rate,
         ..Default::default()
@@ -46,14 +47,14 @@ pub fn run(args: &crate::TrainingCmd, common_args: &crate::Args) -> Result<()> {
     let mut opt = candle_nn::AdamW::new(varmap.all_vars(), params)?;
     for (batch_index, batch) in batch_iter.enumerate() {
         let (inp, tgt) = batch?;
-        let logits = model.forward(&inp, 0)?;
+        let logits = model.forward(&inp, 0, &mut cache)?;
         let loss = candle_nn::loss::cross_entropy(&logits.flatten_to(1)?, &tgt.flatten_to(1)?)?;
         opt.backward_step(&loss)?;
 
         if batch_index > 0 && batch_index % 100 == 0 {
             // TODO: Add a way to deactivate the backprop graph tracking when computing the
             // validation loss.
-            let loss = valid_loss(&dataset, &model, args, &device)?;
+            let loss = valid_loss(&dataset, &model, args, &device, &mut cache)?;
             println!("{batch_index} {loss}");
         }
         if batch_index > 0 && batch_index % 1000 == 0 {

candle-examples/examples/llama_multiprocess/main.rs

@@ -143,14 +143,7 @@ fn main() -> Result<()> {
     let config_filename = api.get("config.json")?;
     let config: Config = serde_json::from_slice(&std::fs::read(config_filename)?)?;
     let tokenizer_filename = api.get("tokenizer.json")?;
-    let mut filenames = vec![];
-    for rfilename in [
-        "model-00001-of-00002.safetensors",
-        "model-00002-of-00002.safetensors",
-    ] {
-        let filename = api.get(rfilename)?;
-        filenames.push(filename);
-    }
+    let filenames = candle_examples::hub_load_safetensors(&api, "model.safetensors.index.json")?;
 
     if args.rank.is_none() {
         let children: Vec<_> = (0..args.num_shards)

candle-examples/examples/mamba-minimal/README.md

@@ -0,0 +1,15 @@
+# candle-mamba-minimal: minimal implementation of Mamba
+
+This is based on [mamba-minimal](https://github.com/johnma2006/mamba-minimal).
+
+Compared to the mamba example, this version can handle training but is much
+slower.
+
+## Running the example
+
+```bash
+$ cargo run --example mamba-minimal --release -- --prompt "Mamba is the"
+Mamba is the most popular and best-selling game in the world. It has been downloaded more than 1,000 times by over 1 million people worldwide since its release on March 18th 2016.
+
+The Mamba series of games are a collection that combines elements from all genres including action, adventure, strategy & puzzle games with some unique gameplay features such as stealth and survival. The game is also known for its innovative graphics and the ability to play in a variety of different modes like single player or multiplayer.
+```

candle-examples/examples/mamba-minimal/main.rs

@@ -0,0 +1,287 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::{Error as E, Result};
+use clap::{Parser, ValueEnum};
+
+mod model;
+use model::{Config, Model};
+
+use candle::{DType, Device, Module, Tensor};
+use candle_examples::token_output_stream::TokenOutputStream;
+use candle_nn::VarBuilder;
+use candle_transformers::generation::LogitsProcessor;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::Tokenizer;
+
+struct TextGeneration {
+    model: Model,
+    device: Device,
+    tokenizer: TokenOutputStream,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+}
+
+impl TextGeneration {
+    #[allow(clippy::too_many_arguments)]
+    fn new(
+        model: Model,
+        tokenizer: Tokenizer,
+        seed: u64,
+        temp: Option<f64>,
+        top_p: Option<f64>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        device: &Device,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            tokenizer: TokenOutputStream::new(tokenizer),
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            device: device.clone(),
+        }
+    }
+
+    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
+        use std::io::Write;
+        self.tokenizer.clear();
+        let mut tokens = self
+            .tokenizer
+            .tokenizer()
+            .encode(prompt, true)
+            .map_err(E::msg)?
+            .get_ids()
+            .to_vec();
+        for &t in tokens.iter() {
+            if let Some(t) = self.tokenizer.next_token(t)? {
+                print!("{t}")
+            }
+        }
+        std::io::stdout().flush()?;
+
+        let mut generated_tokens = 0usize;
+        let eos_token = match self.tokenizer.get_token("<|endoftext|>") {
+            Some(token) => token,
+            None => anyhow::bail!("cannot find the </s> token"),
+        };
+        let start_gen = std::time::Instant::now();
+        for _ in 0..sample_len {
+            let input = Tensor::new(tokens.as_slice(), &self.device)?.unsqueeze(0)?;
+            let logits = self.model.forward(&input)?;
+            let logits = logits.squeeze(0)?.squeeze(0)?.to_dtype(DType::F32)?;
+            let logits = if self.repeat_penalty == 1. {
+                logits
+            } else {
+                let start_at = tokens.len().saturating_sub(self.repeat_last_n);
+                candle_transformers::utils::apply_repeat_penalty(
+                    &logits,
+                    self.repeat_penalty,
+                    &tokens[start_at..],
+                )?
+            };
+
+            let next_token = self.logits_processor.sample(&logits)?;
+            tokens.push(next_token);
+            generated_tokens += 1;
+            if next_token == eos_token {
+                break;
+            }
+            if let Some(t) = self.tokenizer.next_token(next_token)? {
+                print!("{t}");
+                std::io::stdout().flush()?;
+            }
+        }
+        let dt = start_gen.elapsed();
+        if let Some(rest) = self.tokenizer.decode_rest().map_err(E::msg)? {
+            print!("{rest}");
+        }
+        std::io::stdout().flush()?;
+        println!(
+            "\n{generated_tokens} tokens generated ({:.2} token/s)",
+            generated_tokens as f64 / dt.as_secs_f64(),
+        );
+        Ok(())
+    }
+}
+
+#[derive(Parser, ValueEnum, Clone, Copy, PartialEq, Eq, Debug)]
+enum Which {
+    Mamba130m,
+    Mamba370m,
+    Mamba790m,
+    Mamba1_4b,
+    Mamba2_8b,
+    Mamba2_8bSlimPj,
+}
+
+impl std::fmt::Display for Which {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{:?}", self)
+    }
+}
+
+impl Which {
+    fn model_id(&self) -> &'static str {
+        match self {
+            Self::Mamba130m => "state-spaces/mamba-130m",
+            Self::Mamba370m => "state-spaces/mamba-370m",
+            Self::Mamba790m => "state-spaces/mamba-790m",
+            Self::Mamba1_4b => "state-spaces/mamba-1.4b",
+            Self::Mamba2_8b => "state-spaces/mamba-2.8b",
+            Self::Mamba2_8bSlimPj => "state-spaces/mamba-2.8b-slimpj'",
+        }
+    }
+
+    fn revision(&self) -> &'static str {
+        match self {
+            Self::Mamba130m
+            | Self::Mamba370m
+            | Self::Mamba790m
+            | Self::Mamba1_4b
+            | Self::Mamba2_8bSlimPj => "refs/pr/1",
+            Self::Mamba2_8b => "refs/pr/4",
+        }
+    }
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    #[arg(long)]
+    prompt: String,
+
+    /// The temperature used to generate samples.
+    #[arg(long)]
+    temperature: Option<f64>,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// The seed to use when generating random samples.
+    #[arg(long, default_value_t = 299792458)]
+    seed: u64,
+
+    /// The length of the sample to generate (in tokens).
+    #[arg(long, short = 'n', default_value_t = 5000)]
+    sample_len: usize,
+
+    #[arg(long, default_value = "mamba130m")]
+    which: Which,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long)]
+    revision: Option<String>,
+
+    #[arg(long)]
+    tokenizer_file: Option<String>,
+
+    #[arg(long)]
+    weight_files: Option<String>,
+
+    #[arg(long)]
+    config_file: 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 repo = api.repo(Repo::with_revision(
+        args.model_id
+            .unwrap_or_else(|| args.which.model_id().to_string()),
+        RepoType::Model,
+        args.revision
+            .unwrap_or_else(|| args.which.revision().to_string()),
+    ));
+    let tokenizer_filename = match args.tokenizer_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => api
+            .model("EleutherAI/gpt-neox-20b".to_string())
+            .get("tokenizer.json")?,
+    };
+    let config_filename = match args.config_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("config.json")?,
+    };
+    let filenames = match args.weight_files {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        None => {
+            vec![repo.get("model.safetensors")?]
+        }
+    };
+    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 = serde_json::from_slice(&std::fs::read(config_filename)?)?;
+    let device = candle_examples::device(args.cpu)?;
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, DType::F32, &device)? };
+    let model = Model::new(&config, vb.pp("backbone"))?;
+    println!("loaded the model in {:?}", start.elapsed());
+
+    let mut pipeline = TextGeneration::new(
+        model,
+        tokenizer,
+        args.seed,
+        args.temperature,
+        args.top_p,
+        args.repeat_penalty,
+        args.repeat_last_n,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}

candle-examples/examples/mamba-minimal/model.rs

@@ -0,0 +1,204 @@
+/// This follows the lines of:
+/// https://github.com/johnma2006/mamba-minimal/blob/master/model.py
+/// Simple, minimal implementation of Mamba in one file of PyTorch.
+use candle::{IndexOp, Module, Result, Tensor, D};
+use candle_nn::{RmsNorm, VarBuilder};
+
+use candle_transformers::models::with_tracing::{linear, linear_no_bias, Linear};
+
+#[derive(Debug, Clone, serde::Deserialize)]
+pub struct Config {
+    d_model: usize,
+    n_layer: usize,
+    vocab_size: usize,
+    pad_vocab_size_multiple: usize,
+}
+
+impl Config {
+    fn vocab_size(&self) -> usize {
+        let pad = self.pad_vocab_size_multiple;
+        (self.vocab_size + pad - 1) / pad * pad
+    }
+
+    fn dt_rank(&self) -> usize {
+        (self.d_model + 15) / 16
+    }
+
+    fn d_conv(&self) -> usize {
+        4
+    }
+
+    fn d_state(&self) -> usize {
+        16
+    }
+
+    fn d_inner(&self) -> usize {
+        self.d_model * 2
+    }
+}
+
+// https://github.com/johnma2006/mamba-minimal/blob/61f01953ca153f8c4a850d7111beecbf4be9cee1/model.py#L177
+#[derive(Clone, Debug)]
+pub struct MambaBlock {
+    in_proj: Linear,
+    conv1d: candle_nn::Conv1d,
+    x_proj: Linear,
+    dt_proj: Linear,
+    a_log: Tensor,
+    d: Tensor,
+    out_proj: Linear,
+    dt_rank: usize,
+}
+
+impl MambaBlock {
+    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let d_inner = cfg.d_inner();
+        let d_conv = cfg.d_conv();
+        let d_state = cfg.d_state();
+        let dt_rank = cfg.dt_rank();
+        let in_proj = linear_no_bias(cfg.d_model, d_inner * 2, vb.pp("in_proj"))?;
+        let conv_cfg = candle_nn::Conv1dConfig {
+            groups: d_inner,
+            padding: d_conv - 1,
+            ..Default::default()
+        };
+        let conv1d = candle_nn::conv1d(d_inner, d_inner, d_conv, conv_cfg, vb.pp("conv1d"))?;
+        let x_proj = linear_no_bias(d_inner, dt_rank + d_state * 2, vb.pp("x_proj"))?;
+        let dt_proj = linear(dt_rank, d_inner, vb.pp("dt_proj"))?;
+        let a_log = vb.get((d_inner, d_state), "A_log")?;
+        let d = vb.get(d_inner, "D")?;
+        let out_proj = linear_no_bias(d_inner, cfg.d_model, vb.pp("out_proj"))?;
+        Ok(Self {
+            in_proj,
+            conv1d,
+            x_proj,
+            dt_proj,
+            a_log,
+            d,
+            out_proj,
+            dt_rank,
+        })
+    }
+
+    fn ssm(&self, xs: &Tensor) -> Result<Tensor> {
+        let (_d_in, n) = self.a_log.dims2()?;
+        let a = self.a_log.to_dtype(candle::DType::F32)?.exp()?.neg()?;
+        let d = self.d.to_dtype(candle::DType::F32)?;
+        let x_dbl = xs.apply(&self.x_proj)?;
+        let delta = x_dbl.narrow(D::Minus1, 0, self.dt_rank)?;
+        let b = x_dbl.narrow(D::Minus1, self.dt_rank, n)?;
+        let c = x_dbl.narrow(D::Minus1, self.dt_rank + n, n)?;
+        let delta = delta.contiguous()?.apply(&self.dt_proj)?;
+        // softplus without threshold
+        let delta = (delta.exp()? + 1.)?.log()?;
+        let ss = selective_scan(xs, &delta, &a, &b, &c, &d)?;
+        Ok(ss)
+    }
+}
+
+// https://github.com/johnma2006/mamba-minimal/blob/61f01953ca153f8c4a850d7111beecbf4be9cee1/model.py#L275
+fn selective_scan(
+    u: &Tensor,
+    delta: &Tensor,
+    a: &Tensor,
+    b: &Tensor,
+    c: &Tensor,
+    d: &Tensor,
+) -> Result<Tensor> {
+    let (b_sz, l, d_in) = u.dims3()?;
+    let n = a.dim(1)?;
+    let delta = delta.t()?.reshape((b_sz, d_in, l, 1))?; // b d_in l 1
+    let delta_a = delta.broadcast_mul(&a.reshape((1, d_in, 1, n))?)?.exp()?;
+    let delta_b_u = delta
+        .broadcast_mul(&b.reshape((b_sz, 1, l, n))?)?
+        .broadcast_mul(&u.t()?.reshape((b_sz, d_in, l, 1))?)?;
+    let mut xs = Tensor::zeros((b_sz, d_in, n), delta_a.dtype(), delta_a.device())?;
+    let mut ys = Vec::with_capacity(l);
+    for i in 0..l {
+        xs = ((delta_a.i((.., .., i))? * xs)? + delta_b_u.i((.., .., i))?)?;
+        let y = xs.matmul(&c.i((.., i, ..))?.unsqueeze(2)?)?.squeeze(2)?;
+        ys.push(y)
+    }
+    let ys = Tensor::stack(ys.as_slice(), 1)?;
+    ys + u.broadcast_mul(d)
+}
+
+impl Module for MambaBlock {
+    // https://github.com/johnma2006/mamba-minimal/blob/61f01953ca153f8c4a850d7111beecbf4be9cee1/model.py#L206
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let (_b_sz, seq_len, _dim) = xs.dims3()?;
+        let xs_and_res = xs.apply(&self.in_proj)?.chunk(2, D::Minus1)?;
+        let (xs, res) = (&xs_and_res[0], &xs_and_res[1]);
+        let xs = xs
+            .t()?
+            .apply(&self.conv1d)?
+            .narrow(D::Minus1, 0, seq_len)?
+            .t()?;
+        let xs = candle_nn::ops::silu(&xs)?;
+        let ys = (self.ssm(&xs)? * candle_nn::ops::silu(res))?;
+        ys.apply(&self.out_proj)
+    }
+}
+
+// https://github.com/johnma2006/mamba-minimal/blob/61f01953ca153f8c4a850d7111beecbf4be9cee1/model.py#L143
+#[derive(Clone, Debug)]
+pub struct ResidualBlock {
+    mixer: MambaBlock,
+    norm: RmsNorm,
+}
+
+impl ResidualBlock {
+    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let norm = candle_nn::rms_norm(cfg.d_model, 1e-5, vb.pp("norm"))?;
+        let mixer = MambaBlock::new(cfg, vb.pp("mixer"))?;
+        Ok(Self { mixer, norm })
+    }
+}
+
+impl Module for ResidualBlock {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        xs.apply(&self.norm)?.apply(&self.mixer)? + xs
+    }
+}
+
+// https://github.com/johnma2006/mamba-minimal/blob/61f01953ca153f8c4a850d7111beecbf4be9cee1/model.py#L56
+#[derive(Clone, Debug)]
+pub struct Model {
+    embedding: candle_nn::Embedding,
+    layers: Vec<ResidualBlock>,
+    norm_f: RmsNorm,
+    lm_head: Linear,
+}
+
+impl Model {
+    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let embedding = candle_nn::embedding(cfg.vocab_size(), cfg.d_model, vb.pp("embedding"))?;
+        let mut layers = Vec::with_capacity(cfg.n_layer);
+        let vb_l = vb.pp("layers");
+        for layer_idx in 0..cfg.n_layer {
+            let layer = ResidualBlock::new(cfg, vb_l.pp(layer_idx))?;
+            layers.push(layer)
+        }
+        let norm_f = candle_nn::rms_norm(cfg.d_model, 1e-5, vb.pp("norm_f"))?;
+        let lm_head = Linear::from_weights(embedding.embeddings().clone(), None);
+        Ok(Self {
+            embedding,
+            layers,
+            norm_f,
+            lm_head,
+        })
+    }
+}
+
+impl Module for Model {
+    fn forward(&self, input_ids: &Tensor) -> Result<Tensor> {
+        let (_b_size, seq_len) = input_ids.dims2()?;
+        let mut xs = self.embedding.forward(input_ids)?;
+        for layer in self.layers.iter() {
+            xs = layer.forward(&xs)?
+        }
+        xs.narrow(1, seq_len - 1, 1)?
+            .apply(&self.norm_f)?
+            .apply(&self.lm_head)
+    }
+}

candle-examples/examples/mamba/README.md

@@ -0,0 +1,17 @@
+# candle-mamba: Mamba implementation
+
+Candle implementation of *Mamba* [1] inference only. Mamba is an alternative to
+the transformer architecture. It leverages State Space Models (SSMs) with the
+goal of being computationally efficient on long sequences. The implementation is
+based on [mamba.rs](https://github.com/LaurentMazare/mamba.rs).
+
+- [1]. [Mamba: Linear-Time Sequence Modeling with Selective State Spaces](https://arxiv.org/abs/2312.00752).
+
+Compared to the mamba-minimal example, this version is far more efficient but
+would only work for inference.
+## Running the example
+
+```bash
+$ cargo run --example mamba-minimal --release -- --prompt "Mamba is the"
+```
+

candle-examples/examples/mamba/main.rs

@@ -0,0 +1,299 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::{Error as E, Result};
+use clap::{Parser, ValueEnum};
+
+use candle_transformers::models::mamba::{Config, Model, State};
+
+use candle::{DType, Device, Tensor};
+use candle_examples::token_output_stream::TokenOutputStream;
+use candle_nn::VarBuilder;
+use candle_transformers::generation::LogitsProcessor;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::Tokenizer;
+
+struct TextGeneration {
+    model: Model,
+    config: Config,
+    device: Device,
+    tokenizer: TokenOutputStream,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+}
+
+impl TextGeneration {
+    #[allow(clippy::too_many_arguments)]
+    fn new(
+        model: Model,
+        config: Config,
+        tokenizer: Tokenizer,
+        seed: u64,
+        temp: Option<f64>,
+        top_p: Option<f64>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        device: &Device,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            config,
+            tokenizer: TokenOutputStream::new(tokenizer),
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            device: device.clone(),
+        }
+    }
+
+    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
+        use std::io::Write;
+        self.tokenizer.clear();
+        let mut tokens = self
+            .tokenizer
+            .tokenizer()
+            .encode(prompt, true)
+            .map_err(E::msg)?
+            .get_ids()
+            .to_vec();
+        let mut generated_tokens = 0usize;
+        let eos_token = match self.tokenizer.get_token("<|endoftext|>") {
+            Some(token) => token,
+            None => anyhow::bail!("cannot find the </s> token"),
+        };
+        let mut state = State::new(1, &self.config, &self.device)?;
+        let mut next_logits = None;
+        for &t in tokens.iter() {
+            let input = Tensor::new(&[t], &self.device)?;
+            let logits = self.model.forward(&input, &mut state)?;
+            next_logits = Some(logits);
+            if let Some(t) = self.tokenizer.next_token(t)? {
+                print!("{t}")
+            }
+        }
+        std::io::stdout().flush()?;
+
+        let start_gen = std::time::Instant::now();
+        for _ in 0..sample_len {
+            let logits = match next_logits.as_ref() {
+                Some(logits) => logits,
+                None => anyhow::bail!("cannot work on an empty prompt"),
+            };
+            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;
+            }
+            if let Some(t) = self.tokenizer.next_token(next_token)? {
+                print!("{t}");
+                std::io::stdout().flush()?;
+            }
+
+            let input = Tensor::new(&[next_token], &self.device)?;
+            next_logits = Some(self.model.forward(&input, &mut state)?)
+        }
+        let dt = start_gen.elapsed();
+        if let Some(rest) = self.tokenizer.decode_rest().map_err(E::msg)? {
+            print!("{rest}");
+        }
+        std::io::stdout().flush()?;
+        println!(
+            "\n{generated_tokens} tokens generated ({:.2} token/s)",
+            generated_tokens as f64 / dt.as_secs_f64(),
+        );
+        Ok(())
+    }
+}
+
+#[derive(Parser, ValueEnum, Clone, Copy, PartialEq, Eq, Debug)]
+enum Which {
+    Mamba130m,
+    Mamba370m,
+    Mamba790m,
+    Mamba1_4b,
+    Mamba2_8b,
+    Mamba2_8bSlimPj,
+}
+
+impl std::fmt::Display for Which {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{:?}", self)
+    }
+}
+
+impl Which {
+    fn model_id(&self) -> &'static str {
+        match self {
+            Self::Mamba130m => "state-spaces/mamba-130m",
+            Self::Mamba370m => "state-spaces/mamba-370m",
+            Self::Mamba790m => "state-spaces/mamba-790m",
+            Self::Mamba1_4b => "state-spaces/mamba-1.4b",
+            Self::Mamba2_8b => "state-spaces/mamba-2.8b",
+            Self::Mamba2_8bSlimPj => "state-spaces/mamba-2.8b-slimpj'",
+        }
+    }
+
+    fn revision(&self) -> &'static str {
+        match self {
+            Self::Mamba130m
+            | Self::Mamba370m
+            | Self::Mamba790m
+            | Self::Mamba1_4b
+            | Self::Mamba2_8bSlimPj => "refs/pr/1",
+            Self::Mamba2_8b => "refs/pr/4",
+        }
+    }
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    #[arg(long)]
+    prompt: String,
+
+    /// The temperature used to generate samples.
+    #[arg(long)]
+    temperature: Option<f64>,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// The seed to use when generating random samples.
+    #[arg(long, default_value_t = 299792458)]
+    seed: u64,
+
+    /// The length of the sample to generate (in tokens).
+    #[arg(long, short = 'n', default_value_t = 5000)]
+    sample_len: usize,
+
+    #[arg(long, default_value = "mamba130m")]
+    which: Which,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long)]
+    revision: Option<String>,
+
+    #[arg(long)]
+    tokenizer_file: Option<String>,
+
+    #[arg(long)]
+    weight_files: Option<String>,
+
+    #[arg(long)]
+    config_file: 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 repo = api.repo(Repo::with_revision(
+        args.model_id
+            .unwrap_or_else(|| args.which.model_id().to_string()),
+        RepoType::Model,
+        args.revision
+            .unwrap_or_else(|| args.which.revision().to_string()),
+    ));
+    let tokenizer_filename = match args.tokenizer_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => api
+            .model("EleutherAI/gpt-neox-20b".to_string())
+            .get("tokenizer.json")?,
+    };
+    let config_filename = match args.config_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("config.json")?,
+    };
+    let filenames = match args.weight_files {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        None => {
+            vec![repo.get("model.safetensors")?]
+        }
+    };
+    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 = serde_json::from_slice(&std::fs::read(config_filename)?)?;
+    let device = candle_examples::device(args.cpu)?;
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, DType::F32, &device)? };
+    let model = Model::new(&config, vb.pp("backbone"))?;
+    println!("loaded the model in {:?}", start.elapsed());
+
+    let mut pipeline = TextGeneration::new(
+        model,
+        config,
+        tokenizer,
+        args.seed,
+        args.temperature,
+        args.top_p,
+        args.repeat_penalty,
+        args.repeat_last_n,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}

candle-examples/examples/metavoice/README.md

@@ -0,0 +1,18 @@
+# candle-metavoice
+
+MetaVoice-1B is a text-to-speech model trained on 100K hours of speech, more
+details on the [model
+card](https://huggingface.co/metavoiceio/metavoice-1B-v0.1).
+
+Note that the current candle implementation suffers from some limitations as of
+2024-03-02:
+- The speaker embeddings are hardcoded.
+- The generated audio file quality is weaker than the Python implementation,
+  probably because of some implementation discrepancies.
+
+## Run an example
+
+```bash
+cargo run --example metavoice --release -- \\
+  --prompt "This is a demo of text to speech by MetaVoice-1B, an open-source foundational audio model."
+```

candle-examples/examples/metavoice/main.rs

@@ -0,0 +1,277 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::Result;
+use clap::Parser;
+use std::io::Write;
+
+use candle_transformers::generation::LogitsProcessor;
+use candle_transformers::models::encodec;
+use candle_transformers::models::metavoice::{adapters, gpt, tokenizers, transformer};
+use candle_transformers::models::quantized_metavoice::transformer as qtransformer;
+
+use candle::{DType, IndexOp, Tensor};
+use candle_nn::VarBuilder;
+use hf_hub::api::sync::Api;
+use rand::{distributions::Distribution, SeedableRng};
+
+pub const ENCODEC_NTOKENS: u32 = 1024;
+
+#[derive(Clone, Debug, Copy, PartialEq, Eq, clap::ValueEnum)]
+enum ArgDType {
+    F32,
+    F16,
+    Bf16,
+}
+
+enum Transformer {
+    Normal(transformer::Model),
+    Quantized(qtransformer::Model),
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    #[arg(long)]
+    prompt: String,
+
+    /// Use the quantized version of the model.
+    #[arg(long)]
+    quantized: bool,
+
+    /// The guidance scale.
+    #[arg(long, default_value_t = 3.0)]
+    guidance_scale: f64,
+
+    /// The temperature used to generate samples.
+    #[arg(long, default_value_t = 1.0)]
+    temperature: f64,
+
+    /// The seed to use when generating random samples.
+    #[arg(long, default_value_t = 299792458)]
+    seed: u64,
+
+    /// The maximum number of tokens to generate for the first stage.
+    #[arg(long, default_value_t = 2000)]
+    max_tokens: u64,
+
+    /// The output file using the wav format.
+    #[arg(long, default_value = "out.wav")]
+    out_file: String,
+
+    #[arg(long)]
+    first_stage_meta: Option<String>,
+
+    #[arg(long)]
+    first_stage_weights: Option<String>,
+
+    #[arg(long)]
+    second_stage_weights: Option<String>,
+
+    #[arg(long)]
+    encodec_weights: Option<String>,
+
+    #[arg(long)]
+    spk_emb: Option<String>,
+
+    #[arg(long, default_value = "f32")]
+    dtype: ArgDType,
+}
+
+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()
+    );
+    let device = candle_examples::device(args.cpu)?;
+    let api = Api::new()?;
+    let repo = api.model("lmz/candle-metavoice".to_string());
+    let first_stage_meta = match &args.first_stage_meta {
+        Some(w) => std::path::PathBuf::from(w),
+        None => repo.get("first_stage.meta.json")?,
+    };
+    let first_stage_meta: serde_json::Value =
+        serde_json::from_reader(&std::fs::File::open(first_stage_meta)?)?;
+    let first_stage_tokenizer = match first_stage_meta.as_object() {
+        None => anyhow::bail!("not a json object"),
+        Some(j) => match j.get("tokenizer") {
+            None => anyhow::bail!("no tokenizer key"),
+            Some(j) => j,
+        },
+    };
+    let fs_tokenizer = tokenizers::BPE::from_json(first_stage_tokenizer, 512)?;
+
+    let second_stage_weights = match &args.second_stage_weights {
+        Some(w) => std::path::PathBuf::from(w),
+        None => repo.get("second_stage.safetensors")?,
+    };
+    let encodec_weights = match args.encodec_weights {
+        Some(w) => std::path::PathBuf::from(w),
+        None => Api::new()?
+            .model("facebook/encodec_24khz".to_string())
+            .get("model.safetensors")?,
+    };
+    let dtype = match args.dtype {
+        ArgDType::F32 => DType::F32,
+        ArgDType::F16 => DType::F16,
+        ArgDType::Bf16 => DType::BF16,
+    };
+
+    let first_stage_config = transformer::Config::cfg1b_v0_1();
+    let mut first_stage_model = if args.quantized {
+        let filename = match &args.first_stage_weights {
+            Some(w) => std::path::PathBuf::from(w),
+            None => repo.get("first_stage_q4k.gguf")?,
+        };
+        let vb =
+            candle_transformers::quantized_var_builder::VarBuilder::from_gguf(filename, &device)?;
+        let first_stage_model = qtransformer::Model::new(&first_stage_config, vb)?;
+        Transformer::Quantized(first_stage_model)
+    } else {
+        let first_stage_weights = match &args.first_stage_weights {
+            Some(w) => std::path::PathBuf::from(w),
+            None => repo.get("first_stage.safetensors")?,
+        };
+        let first_stage_vb =
+            unsafe { VarBuilder::from_mmaped_safetensors(&[first_stage_weights], dtype, &device)? };
+        let first_stage_model = transformer::Model::new(&first_stage_config, first_stage_vb)?;
+        Transformer::Normal(first_stage_model)
+    };
+
+    let second_stage_vb =
+        unsafe { VarBuilder::from_mmaped_safetensors(&[second_stage_weights], dtype, &device)? };
+    let second_stage_config = gpt::Config::cfg1b_v0_1();
+    let second_stage_model = gpt::Model::new(second_stage_config.clone(), second_stage_vb)?;
+
+    let encodec_device = if device.is_metal() {
+        &candle::Device::Cpu
+    } else {
+        &device
+    };
+    let encodec_vb =
+        unsafe { VarBuilder::from_mmaped_safetensors(&[encodec_weights], dtype, encodec_device)? };
+    let encodec_config = encodec::Config::default();
+    let encodec_model = encodec::Model::new(&encodec_config, encodec_vb)?;
+
+    println!("prompt: '{}'", args.prompt);
+    let prompt_tokens = fs_tokenizer.encode(&args.prompt)?;
+    let mut tokens = prompt_tokens.clone();
+    println!("{tokens:?}");
+    let spk_emb_file = match &args.spk_emb {
+        Some(w) => std::path::PathBuf::from(w),
+        None => repo.get("spk_emb.safetensors")?,
+    };
+    let spk_emb = candle::safetensors::load(&spk_emb_file, &candle::Device::Cpu)?;
+    let spk_emb = match spk_emb.get("spk_emb") {
+        None => anyhow::bail!("missing spk_emb tensor in {spk_emb_file:?}"),
+        Some(spk_emb) => spk_emb.to_dtype(dtype)?,
+    };
+    let spk_emb = spk_emb.to_device(&device)?;
+    let mut logits_processor = LogitsProcessor::new(args.seed, Some(args.temperature), Some(0.95));
+
+    // First stage generation.
+    for index in 0..args.max_tokens {
+        let context_size = if index > 0 { 1 } else { tokens.len() };
+        let start_pos = tokens.len().saturating_sub(context_size);
+        let ctxt = &tokens[start_pos..];
+        let input = Tensor::new(ctxt, &device)?;
+        let input = Tensor::stack(&[&input, &input], 0)?;
+        let logits = match &mut first_stage_model {
+            Transformer::Normal(m) => m.forward(&input, &spk_emb, tokens.len() - context_size)?,
+            Transformer::Quantized(m) => {
+                m.forward(&input, &spk_emb, tokens.len() - context_size)?
+            }
+        };
+        let logits0 = logits.i((0, 0))?;
+        let logits1 = logits.i((1, 0))?;
+        let logits = ((logits0 * args.guidance_scale)? + logits1 * (1. - args.guidance_scale))?;
+        let logits = logits.to_dtype(DType::F32)?;
+        let next_token = logits_processor.sample(&logits)?;
+        tokens.push(next_token);
+        print!(".");
+        std::io::stdout().flush()?;
+        if next_token == 2048 {
+            break;
+        }
+    }
+    println!();
+    let fie2c = adapters::FlattenedInterleavedEncodec2Codebook::new(ENCODEC_NTOKENS);
+    let (text_ids, ids1, ids2) = fie2c.decode(&tokens);
+    println!("text ids len: {}", text_ids.len());
+    let mut rng = rand::rngs::StdRng::seed_from_u64(args.seed + 1337);
+    // TODO: Use the config rather than hardcoding the offset here.
+    let encoded_text: Vec<_> = prompt_tokens.iter().map(|v| v - 1024).collect();
+    let mut hierarchies_in1 =
+        [encoded_text.as_slice(), ids1.as_slice(), &[ENCODEC_NTOKENS]].concat();
+    let mut hierarchies_in2 = [
+        vec![ENCODEC_NTOKENS; encoded_text.len()].as_slice(),
+        ids2.as_slice(),
+        &[ENCODEC_NTOKENS],
+    ]
+    .concat();
+    hierarchies_in1.resize(second_stage_config.block_size, ENCODEC_NTOKENS);
+    hierarchies_in2.resize(second_stage_config.block_size, ENCODEC_NTOKENS);
+    let in_x1 = Tensor::new(hierarchies_in1, &device)?;
+    let in_x2 = Tensor::new(hierarchies_in2, &device)?;
+    let in_x = Tensor::stack(&[in_x1, in_x2], 0)?.unsqueeze(0)?;
+    let logits = second_stage_model.forward(&in_x)?;
+    println!("sampling from logits...");
+    let mut codes = vec![];
+    for logits in logits.iter() {
+        let logits = logits.squeeze(0)?;
+        let (seq_len, _) = logits.dims2()?;
+        let mut codes_ = Vec::with_capacity(seq_len);
+        for step in 0..seq_len {
+            let logits = logits.i(step)?.to_dtype(DType::F32)?;
+            let logits = &(&logits / 1.0)?;
+            let prs = candle_nn::ops::softmax_last_dim(logits)?.to_vec1::<f32>()?;
+            let distr = rand::distributions::WeightedIndex::new(prs.as_slice())?;
+            let sample = distr.sample(&mut rng) as u32;
+            codes_.push(sample)
+        }
+        codes.push(codes_)
+    }
+
+    let codes = Tensor::new(codes, &device)?.unsqueeze(0)?;
+    let codes = Tensor::cat(&[in_x, codes], 1)?;
+    println!("codes: {codes}");
+    let tilted_encodec = adapters::TiltedEncodec::new(ENCODEC_NTOKENS);
+    let codes = codes.i(0)?.to_vec2::<u32>()?;
+    let (text_ids, audio_ids) = tilted_encodec.decode(&codes);
+    println!("text_ids len: {:?}", text_ids.len());
+    let audio_ids = Tensor::new(audio_ids, encodec_device)?.unsqueeze(0)?;
+    println!("audio_ids shape: {:?}", audio_ids.shape());
+    let pcm = encodec_model.decode(&audio_ids)?;
+    println!("output pcm shape: {:?}", pcm.shape());
+    let pcm = pcm.i(0)?.i(0)?.to_dtype(DType::F32)?;
+    let pcm = candle_examples::audio::normalize_loudness(&pcm, 24_000, true)?;
+    let pcm = pcm.to_vec1::<f32>()?;
+    let mut output = std::fs::File::create(&args.out_file)?;
+    candle_examples::wav::write_pcm_as_wav(&mut output, &pcm, 24_000)?;
+    Ok(())
+}

candle-examples/examples/mistral/main.rs

@@ -152,11 +152,11 @@ struct Args {
     seed: u64,
 
     /// The length of the sample to generate (in tokens).
-    #[arg(long, short = 'n', default_value_t = 100)]
+    #[arg(long, short = 'n', default_value_t = 10000)]
     sample_len: usize,
 
-    #[arg(long, default_value = "lmz/candle-mistral")]
-    model_id: String,
+    #[arg(long)]
+    model_id: Option<String>,
 
     #[arg(long, default_value = "main")]
     revision: String,
@@ -207,8 +207,18 @@ fn main() -> Result<()> {
 
     let start = std::time::Instant::now();
     let api = Api::new()?;
+    let model_id = match args.model_id {
+        Some(model_id) => model_id,
+        None => {
+            if args.quantized {
+                "lmz/candle-mistral".to_string()
+            } else {
+                "mistralai/Mistral-7B-v0.1".to_string()
+            }
+        }
+    };
     let repo = api.repo(Repo::with_revision(
-        args.model_id,
+        model_id,
         RepoType::Model,
         args.revision,
     ));
@@ -225,10 +235,7 @@ fn main() -> Result<()> {
             if args.quantized {
                 vec![repo.get("model-q4k.gguf")?]
             } else {
-                vec![
-                    repo.get("pytorch_model-00001-of-00002.safetensors")?,
-                    repo.get("pytorch_model-00002-of-00002.safetensors")?,
-                ]
+                candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.json")?
             }
         }
     };
@@ -237,13 +244,14 @@ fn main() -> Result<()> {
 
     let start = std::time::Instant::now();
     let config = Config::config_7b_v0_1(args.use_flash_attn);
+    let device = candle_examples::device(args.cpu)?;
     let (model, device) = if args.quantized {
         let filename = &filenames[0];
-        let vb = candle_transformers::quantized_var_builder::VarBuilder::from_gguf(filename)?;
+        let vb =
+            candle_transformers::quantized_var_builder::VarBuilder::from_gguf(filename, &device)?;
         let model = QMistral::new(&config, vb)?;
-        (Model::Quantized(model), Device::Cpu)
+        (Model::Quantized(model), device)
     } else {
-        let device = candle_examples::device(args.cpu)?;
         let dtype = if device.is_cuda() {
             DType::BF16
         } else {

candle-examples/examples/mixtral/README.md

@@ -0,0 +1,25 @@
+# candle-mixtral: 8x7b LLM using a sparse mixture of experts.
+
+Mixtral-8x7B-v0.1 is a pretrained generative LLM with 56 billion parameters. 
+
+- [Blog post](https://mistral.ai/news/mixtral-of-experts/) from Mistral announcing the model release.
+- [Model card](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1) on the HuggingFace Hub.
+
+## Running the example
+
+```bash
+$ cargo run --example mixtral --release  -- --prompt "def print_prime(n): "
+def print_prime(n):  # n is the number of prime numbers to be printed
+    i = 2
+    count = 0
+    while (count < n):
+        if (isPrime(i)):
+            print(i)
+            count += 1
+        i += 1
+
+def isPrime(n):
+    for x in range(2, int(n**0.5)+1):
+        if (n % x == 0):
+            ...
+```

candle-examples/examples/mixtral/main.rs

@@ -0,0 +1,241 @@
+#[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::mixtral::{Config, Model};
+
+use candle::{DType, Device, Tensor};
+use candle_examples::token_output_stream::TokenOutputStream;
+use candle_nn::VarBuilder;
+use candle_transformers::generation::LogitsProcessor;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::Tokenizer;
+
+struct TextGeneration {
+    model: Model,
+    device: Device,
+    tokenizer: TokenOutputStream,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+}
+
+impl TextGeneration {
+    #[allow(clippy::too_many_arguments)]
+    fn new(
+        model: Model,
+        tokenizer: Tokenizer,
+        seed: u64,
+        temp: Option<f64>,
+        top_p: Option<f64>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        device: &Device,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            tokenizer: TokenOutputStream::new(tokenizer),
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            device: device.clone(),
+        }
+    }
+
+    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
+        use std::io::Write;
+        self.tokenizer.clear();
+        let mut tokens = self
+            .tokenizer
+            .tokenizer()
+            .encode(prompt, true)
+            .map_err(E::msg)?
+            .get_ids()
+            .to_vec();
+        for &t in tokens.iter() {
+            if let Some(t) = self.tokenizer.next_token(t)? {
+                print!("{t}")
+            }
+        }
+        std::io::stdout().flush()?;
+
+        let mut generated_tokens = 0usize;
+        let eos_token = match self.tokenizer.get_token("</s>") {
+            Some(token) => token,
+            None => anyhow::bail!("cannot find the </s> token"),
+        };
+        let start_gen = std::time::Instant::now();
+        for index in 0..sample_len {
+            let context_size = if index > 0 { 1 } else { tokens.len() };
+            let start_pos = tokens.len().saturating_sub(context_size);
+            let ctxt = &tokens[start_pos..];
+            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+            let logits = self.model.forward(&input, start_pos)?;
+            let logits = logits.squeeze(0)?.squeeze(0)?.to_dtype(DType::F32)?;
+            let logits = if self.repeat_penalty == 1. {
+                logits
+            } else {
+                let start_at = tokens.len().saturating_sub(self.repeat_last_n);
+                candle_transformers::utils::apply_repeat_penalty(
+                    &logits,
+                    self.repeat_penalty,
+                    &tokens[start_at..],
+                )?
+            };
+
+            let next_token = self.logits_processor.sample(&logits)?;
+            tokens.push(next_token);
+            generated_tokens += 1;
+            if next_token == eos_token {
+                break;
+            }
+            if let Some(t) = self.tokenizer.next_token(next_token)? {
+                print!("{t}");
+                std::io::stdout().flush()?;
+            }
+        }
+        let dt = start_gen.elapsed();
+        if let Some(rest) = self.tokenizer.decode_rest().map_err(E::msg)? {
+            print!("{rest}");
+        }
+        std::io::stdout().flush()?;
+        println!(
+            "\n{generated_tokens} tokens generated ({:.2} token/s)",
+            generated_tokens as f64 / dt.as_secs_f64(),
+        );
+        Ok(())
+    }
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    #[arg(long)]
+    use_flash_attn: bool,
+
+    #[arg(long)]
+    prompt: String,
+
+    /// The temperature used to generate samples.
+    #[arg(long)]
+    temperature: Option<f64>,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// The seed to use when generating random samples.
+    #[arg(long, default_value_t = 299792458)]
+    seed: u64,
+
+    /// The length of the sample to generate (in tokens).
+    #[arg(long, short = 'n', default_value_t = 10000)]
+    sample_len: usize,
+
+    #[arg(long, default_value = "mistralai/Mixtral-8x7B-v0.1")]
+    model_id: String,
+
+    #[arg(long, default_value = "main")]
+    revision: String,
+
+    #[arg(long)]
+    tokenizer_file: Option<String>,
+
+    #[arg(long)]
+    weight_files: Option<String>,
+
+    /// Penalty to be applied for repeating tokens, 1. means no penalty.
+    #[arg(long, default_value_t = 1.1)]
+    repeat_penalty: f32,
+
+    /// The context size to consider for the repeat penalty.
+    #[arg(long, default_value_t = 64)]
+    repeat_last_n: usize,
+}
+
+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 repo = api.repo(Repo::with_revision(
+        args.model_id,
+        RepoType::Model,
+        args.revision,
+    ));
+    let tokenizer_filename = match args.tokenizer_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("tokenizer.json")?,
+    };
+    let filenames = match args.weight_files {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        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::v0_1_8x7b(args.use_flash_attn);
+    let device = candle_examples::device(args.cpu)?;
+    let dtype = if device.is_cuda() {
+        DType::BF16
+    } else {
+        DType::F32
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &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,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}

candle-examples/examples/mobileone/README.md

@@ -0,0 +1,22 @@
+# candle-mobileone
+
+[MobileOne: An Improved One millisecond Mobile Backbone](https://arxiv.org/abs/2206.04040).
+
+This candle implementation uses a pre-trained MobileOne 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 mobileone --release -- --image candle-examples/examples/yolo-v8/assets/bike.jpg --which s2
+
+loaded image Tensor[dims 3, 224, 224; f32]
+model built
+mountain bike, all-terrain bike, off-roader: 79.33%
+bicycle-built-for-two, tandem bicycle, tandem: 15.32%
+crash helmet            : 2.58%
+unicycle, monocycle     : 1.70%
+alp                     : 0.21%
+
+```

candle-examples/examples/mobileone/main.rs

@@ -0,0 +1,96 @@
+#[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::mobileone;
+
+#[derive(Clone, Copy, Debug, ValueEnum)]
+enum Which {
+    S0,
+    S1,
+    S2,
+    S3,
+    S4,
+}
+
+impl Which {
+    fn model_filename(&self) -> String {
+        let name = match self {
+            Self::S0 => "s0",
+            Self::S1 => "s1",
+            Self::S2 => "s2",
+            Self::S3 => "s3",
+            Self::S4 => "s4",
+        };
+        format!("timm/mobileone_{}.apple_in1k", name)
+    }
+
+    fn config(&self) -> mobileone::Config {
+        match self {
+            Self::S0 => mobileone::Config::s0(),
+            Self::S1 => mobileone::Config::s1(),
+            Self::S2 => mobileone::Config::s2(),
+            Self::S3 => mobileone::Config::s3(),
+            Self::S4 => mobileone::Config::s4(),
+        }
+    }
+}
+
+#[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::S0)]
+    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 = mobileone::mobileone(&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/musicgen/encodec_model.rs

@@ -1,580 +0,0 @@
-use crate::nn::conv1d_weight_norm;
-use candle::{DType, IndexOp, Module, Result, Tensor};
-use candle_nn::{conv1d, Conv1d, Conv1dConfig, VarBuilder};
-
-// Encodec Model
-// https://github.com/huggingface/transformers/blob/main/src/transformers/models/encodec/modeling_encodec.py
-
-#[derive(Debug, Clone, PartialEq)]
-enum NormType {
-    WeightNorm,
-    TimeGroupNorm,
-    None,
-}
-
-#[derive(Debug, Clone, PartialEq)]
-pub struct Config {
-    target_bandwidths: Vec<f64>,
-    sampling_rate: usize,
-    audio_channels: usize,
-    normalize: bool,
-    chunk_length_s: Option<usize>,
-    overlap: Option<usize>,
-    hidden_size: usize,
-    num_filters: usize,
-    num_residual_layers: usize,
-    upsampling_ratios: Vec<usize>,
-    norm_type: NormType,
-    kernel_size: usize,
-    last_kernel_size: usize,
-    residual_kernel_size: usize,
-    dilation_growth_rate: usize,
-    use_causal_conv: bool,
-    pad_mode: &'static str,
-    compress: usize,
-    num_lstm_layers: usize,
-    trim_right_ratio: f64,
-    codebook_size: usize,
-    codebook_dim: Option<usize>,
-    use_conv_shortcut: bool,
-}
-
-impl Default for Config {
-    fn default() -> Self {
-        Self {
-            target_bandwidths: vec![1.5, 3.0, 6.0, 12.0, 24.0],
-            sampling_rate: 24_000,
-            audio_channels: 1,
-            normalize: false,
-            chunk_length_s: None,
-            overlap: None,
-            hidden_size: 128,
-            num_filters: 32,
-            num_residual_layers: 1,
-            upsampling_ratios: vec![8, 5, 4, 2],
-            norm_type: NormType::WeightNorm,
-            kernel_size: 7,
-            last_kernel_size: 7,
-            residual_kernel_size: 3,
-            dilation_growth_rate: 2,
-            use_causal_conv: true,
-            pad_mode: "reflect",
-            compress: 2,
-            num_lstm_layers: 2,
-            trim_right_ratio: 1.0,
-            codebook_size: 1024,
-            codebook_dim: None,
-            use_conv_shortcut: true,
-        }
-    }
-}
-
-impl Config {
-    // https://huggingface.co/facebook/musicgen-small/blob/495da4ad086b3416a27c6187f9239f9fd96f3962/config.json#L6
-    pub fn musicgen_small() -> Self {
-        Self {
-            audio_channels: 1,
-            chunk_length_s: None,
-            codebook_dim: Some(128),
-            codebook_size: 2048,
-            compress: 2,
-            dilation_growth_rate: 2,
-            hidden_size: 128,
-            kernel_size: 7,
-            last_kernel_size: 7,
-            norm_type: NormType::WeightNorm,
-            normalize: false,
-            num_filters: 64,
-            num_lstm_layers: 2,
-            num_residual_layers: 1,
-            overlap: None,
-            pad_mode: "reflect",
-            residual_kernel_size: 3,
-            sampling_rate: 32_000,
-            target_bandwidths: vec![2.2],
-            trim_right_ratio: 1.0,
-            upsampling_ratios: vec![8, 5, 4, 4],
-            use_causal_conv: false,
-            use_conv_shortcut: false,
-        }
-    }
-
-    fn codebook_dim(&self) -> usize {
-        self.codebook_dim.unwrap_or(self.codebook_size)
-    }
-
-    fn frame_rate(&self) -> usize {
-        let hop_length: usize = self.upsampling_ratios.iter().product();
-        (self.sampling_rate + hop_length - 1) / hop_length
-    }
-
-    fn num_quantizers(&self) -> usize {
-        let num = 1000f64
-            * self
-                .target_bandwidths
-                .last()
-                .expect("empty target_bandwidths");
-        (num as usize) / (self.frame_rate() * 10)
-    }
-}
-
-// https://github.com/huggingface/transformers/blob/abaca9f9432a84cfaa95531de4c72334f38a42f2/src/transformers/models/encodec/modeling_encodec.py#L340
-#[derive(Debug)]
-struct EncodecEuclideanCodebook {
-    inited: Tensor,
-    cluster_size: Tensor,
-    embed: Tensor,
-    embed_avg: Tensor,
-}
-
-impl EncodecEuclideanCodebook {
-    fn load(vb: VarBuilder, cfg: &Config) -> Result<Self> {
-        let inited = vb.get(1, "inited")?;
-        let cluster_size = vb.get(cfg.codebook_size, "cluster_size")?;
-        let e_shape = (cfg.codebook_size, cfg.codebook_dim());
-        let embed = vb.get(e_shape, "embed")?;
-        let embed_avg = vb.get(e_shape, "embed_avg")?;
-        Ok(Self {
-            inited,
-            cluster_size,
-            embed,
-            embed_avg,
-        })
-    }
-
-    fn decode(&self, embed_ind: &Tensor) -> Result<Tensor> {
-        let quantize = self.embed.embedding(embed_ind)?;
-        Ok(quantize)
-    }
-}
-
-#[derive(Debug)]
-struct EncodecVectorQuantization {
-    codebook: EncodecEuclideanCodebook,
-}
-
-impl EncodecVectorQuantization {
-    fn load(vb: VarBuilder, cfg: &Config) -> Result<Self> {
-        let codebook = EncodecEuclideanCodebook::load(vb.pp("codebook"), cfg)?;
-        Ok(Self { codebook })
-    }
-
-    fn decode(&self, embed_ind: &Tensor) -> Result<Tensor> {
-        let quantize = self.codebook.decode(embed_ind)?;
-        let quantize = quantize.transpose(1, 2)?;
-        Ok(quantize)
-    }
-}
-
-#[derive(Debug)]
-struct EncodecResidualVectorQuantizer {
-    layers: Vec<EncodecVectorQuantization>,
-}
-
-impl EncodecResidualVectorQuantizer {
-    fn load(vb: VarBuilder, cfg: &Config) -> Result<Self> {
-        let vb = &vb.pp("layers");
-        let layers = (0..cfg.num_quantizers())
-            .map(|i| EncodecVectorQuantization::load(vb.pp(&i.to_string()), cfg))
-            .collect::<Result<Vec<_>>>()?;
-        Ok(Self { layers })
-    }
-
-    fn decode(&self, codes: &Tensor) -> Result<Tensor> {
-        let mut quantized_out = Tensor::zeros((), DType::F32, codes.device())?;
-        if codes.dim(0)? != self.layers.len() {
-            candle::bail!(
-                "codes shape {:?} does not match the number of quantization layers {}",
-                codes.shape(),
-                self.layers.len()
-            )
-        }
-        for (i, layer) in self.layers.iter().enumerate() {
-            let quantized = layer.decode(&codes.i(i)?)?;
-            quantized_out = quantized.broadcast_add(&quantized_out)?;
-        }
-        Ok(quantized_out)
-    }
-}
-
-// https://github.com/huggingface/transformers/blob/abaca9f9432a84cfaa95531de4c72334f38a42f2/src/transformers/models/encodec/modeling_encodec.py#L226
-#[derive(Debug)]
-struct EncodecLSTM {
-    layers: Vec<candle_nn::LSTM>,
-}
-
-impl EncodecLSTM {
-    fn load(dim: usize, vb: VarBuilder, cfg: &Config) -> Result<Self> {
-        let vb = &vb.pp("lstm");
-        let mut layers = vec![];
-        for layer_idx in 0..cfg.num_lstm_layers {
-            let config = candle_nn::LSTMConfig {
-                layer_idx,
-                ..Default::default()
-            };
-            let lstm = candle_nn::lstm(dim, dim, config, vb.clone())?;
-            layers.push(lstm)
-        }
-        Ok(Self { layers })
-    }
-}
-
-impl Module for EncodecLSTM {
-    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
-        use candle_nn::RNN;
-        let mut xs = xs.clone();
-        for layer in self.layers.iter() {
-            let states = layer.seq(&xs)?;
-            xs = layer.states_to_tensor(&states)?;
-        }
-        Ok(xs)
-    }
-}
-
-#[derive(Debug)]
-struct EncodecConvTranspose1d {
-    weight_g: Tensor,
-    weight_v: Tensor,
-    bias: Tensor,
-}
-
-impl EncodecConvTranspose1d {
-    fn load(
-        in_c: usize,
-        out_c: usize,
-        k: usize,
-        _stride: usize,
-        vb: VarBuilder,
-        _cfg: &Config,
-    ) -> Result<Self> {
-        let vb = &vb.pp("conv");
-        let weight_g = vb.get((in_c, 1, 1), "weight_g")?;
-        let weight_v = vb.get((in_c, out_c, k), "weight_v")?;
-        let bias = vb.get(out_c, "bias")?;
-        Ok(Self {
-            weight_g,
-            weight_v,
-            bias,
-        })
-    }
-}
-
-impl Module for EncodecConvTranspose1d {
-    fn forward(&self, _xs: &Tensor) -> Result<Tensor> {
-        todo!()
-    }
-}
-
-#[derive(Debug)]
-struct EncodecConv1d {
-    causal: bool,
-    conv: Conv1d,
-    norm: Option<candle_nn::GroupNorm>,
-}
-
-impl EncodecConv1d {
-    fn load(
-        in_c: usize,
-        out_c: usize,
-        kernel_size: usize,
-        stride: usize,
-        vb: VarBuilder,
-        cfg: &Config,
-    ) -> Result<Self> {
-        let conv = match cfg.norm_type {
-            NormType::WeightNorm => conv1d_weight_norm(
-                in_c,
-                out_c,
-                kernel_size,
-                Conv1dConfig {
-                    padding: 0,
-                    stride,
-                    groups: 1,
-                    dilation: 1,
-                },
-                vb.pp("conv"),
-            )?,
-            NormType::None | NormType::TimeGroupNorm => conv1d(
-                in_c,
-                out_c,
-                kernel_size,
-                Conv1dConfig {
-                    padding: 0,
-                    stride,
-                    groups: 1,
-                    dilation: 1,
-                },
-                vb.pp("conv"),
-            )?,
-        };
-        let norm = match cfg.norm_type {
-            NormType::None | NormType::WeightNorm => None,
-            NormType::TimeGroupNorm => {
-                let gn = candle_nn::group_norm(1, out_c, 1e-5, vb.pp("norm"))?;
-                Some(gn)
-            }
-        };
-        Ok(Self {
-            causal: cfg.use_causal_conv,
-            conv,
-            norm,
-        })
-    }
-}
-
-impl Module for EncodecConv1d {
-    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
-        // TODO: padding, depending on causal.
-        let xs = self.conv.forward(xs)?;
-        match &self.norm {
-            None => Ok(xs),
-            Some(norm) => xs.apply(norm),
-        }
-    }
-}
-
-#[derive(Debug)]
-struct EncodecResnetBlock {
-    block_conv1: EncodecConv1d,
-    block_conv2: EncodecConv1d,
-    shortcut: Option<EncodecConv1d>,
-}
-
-impl EncodecResnetBlock {
-    fn load(dim: usize, dilations: &[usize], vb: VarBuilder, cfg: &Config) -> Result<Self> {
-        let h = dim / cfg.compress;
-        let mut layer = Layer::new(vb.pp("block"));
-        if dilations.len() != 2 {
-            candle::bail!("expected dilations of size 2")
-        }
-        // TODO: Apply dilations!
-        layer.inc();
-        let block_conv1 =
-            EncodecConv1d::load(dim, h, cfg.residual_kernel_size, 1, layer.next(), cfg)?;
-        layer.inc();
-        let block_conv2 = EncodecConv1d::load(h, dim, 1, 1, layer.next(), cfg)?;
-        let shortcut = if cfg.use_conv_shortcut {
-            let conv = EncodecConv1d::load(dim, dim, 1, 1, vb.pp("shortcut"), cfg)?;
-            Some(conv)
-        } else {
-            None
-        };
-        Ok(Self {
-            block_conv1,
-            block_conv2,
-            shortcut,
-        })
-    }
-}
-
-impl Module for EncodecResnetBlock {
-    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
-        let residual = xs.clone();
-        let xs = xs.elu(1.)?;
-        let xs = self.block_conv1.forward(&xs)?;
-        let xs = xs.elu(1.)?;
-        let xs = self.block_conv2.forward(&xs)?;
-        let xs = match &self.shortcut {
-            None => (xs + residual)?,
-            Some(shortcut) => xs.add(&shortcut.forward(&residual)?)?,
-        };
-        Ok(xs)
-    }
-}
-
-struct Layer<'a> {
-    vb: VarBuilder<'a>,
-    cnt: usize,
-}
-
-impl<'a> Layer<'a> {
-    fn new(vb: VarBuilder<'a>) -> Self {
-        Self { vb, cnt: 0 }
-    }
-
-    fn inc(&mut self) {
-        self.cnt += 1;
-    }
-
-    fn next(&mut self) -> VarBuilder {
-        let vb = self.vb.pp(&self.cnt.to_string());
-        self.cnt += 1;
-        vb
-    }
-}
-
-#[derive(Debug)]
-struct EncodecEncoder {
-    init_conv: EncodecConv1d,
-    sampling_layers: Vec<(Vec<EncodecResnetBlock>, EncodecConv1d)>,
-    final_lstm: EncodecLSTM,
-    final_conv: EncodecConv1d,
-}
-
-impl EncodecEncoder {
-    fn load(vb: VarBuilder, cfg: &Config) -> Result<Self> {
-        let mut layer = Layer::new(vb.pp("layers"));
-        let init_conv = EncodecConv1d::load(
-            cfg.audio_channels,
-            cfg.num_filters,
-            cfg.kernel_size,
-            1,
-            layer.next(),
-            cfg,
-        )?;
-        let mut sampling_layers = vec![];
-        let mut scaling = 1;
-        for &ratio in cfg.upsampling_ratios.iter().rev() {
-            let current_scale = scaling * cfg.num_filters;
-            let mut resnets = vec![];
-            for j in 0..(cfg.num_residual_layers as u32) {
-                let resnet = EncodecResnetBlock::load(
-                    current_scale,
-                    &[cfg.dilation_growth_rate.pow(j), 1],
-                    layer.next(),
-                    cfg,
-                )?;
-                resnets.push(resnet)
-            }
-            layer.inc(); // ELU
-            let conv1d = EncodecConv1d::load(
-                current_scale,
-                current_scale * 2,
-                ratio * 2,
-                ratio,
-                layer.next(),
-                cfg,
-            )?;
-            sampling_layers.push((resnets, conv1d));
-            scaling *= 2;
-        }
-        let final_lstm = EncodecLSTM::load(cfg.num_filters * scaling, layer.next(), cfg)?;
-        layer.inc(); // ELU
-        let final_conv = EncodecConv1d::load(
-            cfg.num_filters * scaling,
-            cfg.hidden_size,
-            cfg.last_kernel_size,
-            1,
-            layer.next(),
-            cfg,
-        )?;
-        Ok(Self {
-            init_conv,
-            sampling_layers,
-            final_conv,
-            final_lstm,
-        })
-    }
-
-    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
-        let mut xs = xs.apply(&self.init_conv)?;
-        for (resnets, conv) in self.sampling_layers.iter() {
-            for resnet in resnets.iter() {
-                xs = xs.apply(resnet)?;
-            }
-            xs = xs.elu(1.0)?.apply(conv)?;
-        }
-        xs.apply(&self.final_lstm)?
-            .elu(1.0)?
-            .apply(&self.final_conv)
-    }
-}
-
-#[derive(Debug)]
-struct EncodecDecoder {
-    init_conv: EncodecConv1d,
-    init_lstm: EncodecLSTM,
-    sampling_layers: Vec<(EncodecConvTranspose1d, Vec<EncodecResnetBlock>)>,
-    final_conv: EncodecConv1d,
-}
-
-impl EncodecDecoder {
-    fn load(vb: VarBuilder, cfg: &Config) -> Result<Self> {
-        let mut layer = Layer::new(vb.pp("layers"));
-        let mut scaling = usize::pow(2, cfg.upsampling_ratios.len() as u32);
-        let init_conv = EncodecConv1d::load(
-            cfg.hidden_size,
-            cfg.num_filters * scaling,
-            cfg.last_kernel_size,
-            1,
-            layer.next(),
-            cfg,
-        )?;
-        let init_lstm = EncodecLSTM::load(cfg.num_filters * scaling, layer.next(), cfg)?;
-        let mut sampling_layers = vec![];
-        for &ratio in cfg.upsampling_ratios.iter() {
-            let current_scale = scaling * cfg.num_filters;
-            layer.inc(); // ELU
-            let conv1d = EncodecConvTranspose1d::load(
-                current_scale,
-                current_scale / 2,
-                ratio * 2,
-                ratio,
-                layer.next(),
-                cfg,
-            )?;
-            let mut resnets = vec![];
-            for j in 0..(cfg.num_residual_layers as u32) {
-                let resnet = EncodecResnetBlock::load(
-                    current_scale / 2,
-                    &[cfg.dilation_growth_rate.pow(j), 1],
-                    layer.next(),
-                    cfg,
-                )?;
-                resnets.push(resnet)
-            }
-            sampling_layers.push((conv1d, resnets));
-            scaling /= 2;
-        }
-        layer.inc(); // ELU
-        let final_conv = EncodecConv1d::load(
-            cfg.num_filters,
-            cfg.audio_channels,
-            cfg.last_kernel_size,
-            1,
-            layer.next(),
-            cfg,
-        )?;
-        Ok(Self {
-            init_conv,
-            init_lstm,
-            sampling_layers,
-            final_conv,
-        })
-    }
-
-    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
-        let mut xs = xs.apply(&self.init_conv)?.apply(&self.init_lstm)?;
-        for (conv, resnets) in self.sampling_layers.iter() {
-            xs = xs.elu(1.)?.apply(conv)?;
-            for resnet in resnets.iter() {
-                xs = xs.apply(resnet)?
-            }
-        }
-        xs.elu(1.)?.apply(&self.final_conv)
-    }
-}
-
-#[derive(Debug)]
-pub struct EncodecModel {
-    encoder: EncodecEncoder,
-    decoder: EncodecDecoder,
-    quantizer: EncodecResidualVectorQuantizer,
-}
-
-impl EncodecModel {
-    pub fn load(vb: VarBuilder, cfg: &Config) -> Result<Self> {
-        let encoder = EncodecEncoder::load(vb.pp("encoder"), cfg)?;
-        let decoder = EncodecDecoder::load(vb.pp("decoder"), cfg)?;
-        let quantizer = EncodecResidualVectorQuantizer::load(vb.pp("quantizer"), cfg)?;
-        Ok(Self {
-            encoder,
-            decoder,
-            quantizer,
-        })
-    }
-
-    pub fn forward(&self, _xs: &Tensor) -> Result<Tensor> {
-        todo!()
-    }
-}

candle-examples/examples/musicgen/main.rs

@@ -10,9 +10,7 @@ extern crate intel_mkl_src;
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
 
-mod encodec_model;
 mod musicgen_model;
-mod nn;
 
 use musicgen_model::{GenConfig, MusicgenForConditionalGeneration};