Add a benchmark for the matmul slowness.

.github/workflows/ci_cuda.yaml

@@ -8,8 +8,6 @@ jobs:
   start-runner:
     name: Start self-hosted EC2 runner
     runs-on: ubuntu-latest
-    # Don't run on forks, they won't have access to secrets anyway.
-    if: ${{ github.event.pull_request.head.repo.full_name == github.event.pull_request.base.repo.full_name }}
     env:
       AWS_REGION: us-east-1
       EC2_AMI_ID: ami-03cfed9ea28f4b002
@@ -61,7 +59,7 @@ jobs:
       - name: Install Rust Stable
         run: curl https://sh.rustup.rs -sSf | sh -s -- -y
       - uses: Swatinem/rust-cache@v2
-      - run: apt-get update -y && apt-get install libssl-dev protobuf-compiler -y
+      - run: apt-get update -y && apt-get install libssl-dev -y
       - name: Test (cuda)
         run: PATH=$PATH:/usr/local/cuda-11.8/bin/ /root/.cargo/bin/cargo test --features cuda
   stop-runner:
@@ -72,7 +70,7 @@ jobs:
     runs-on: ubuntu-latest
     env:
       AWS_REGION: us-east-1
-    if: ${{ (success() || failure()) && github.event.pull_request.head.repo.full_name == github.event.pull_request.base.repo.full_name }} # required to stop the runner even if the error happened in the previous jobs
+    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

.github/workflows/python.yml

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

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 segment anything example
+- TinyViT backbone for the segemnt anything example
   [787](https://github.com/huggingface/candle/pull/787).
 - Shape with holes support
   [770](https://github.com/huggingface/candle/pull/770).

Cargo.toml

@@ -7,19 +7,20 @@ members = [
     "candle-nn",
     "candle-pyo3",
     "candle-transformers",
-    "candle-wasm-examples/*",
+    "candle-wasm-examples/llama2-c",
+    "candle-wasm-examples/segment-anything",
+    "candle-wasm-examples/whisper",
+    "candle-wasm-examples/yolo",
+    "candle-wasm-examples/bert",
+    "candle-wasm-examples/phi",
+    "candle-wasm-examples/t5",
     "candle-wasm-tests",
 ]
-exclude = [
-   "candle-flash-attn",
-   "candle-kernels",
-   "candle-metal-kernels",
-   "candle-onnx",
-]
+exclude = ["candle-flash-attn", "candle-kernels"]
 resolver = "2"
 
 [workspace.package]
-version = "0.3.3"
+version = "0.3.0"
 edition = "2021"
 description = "Minimalist ML framework."
 repository = "https://github.com/huggingface/candle"
@@ -31,18 +32,10 @@ 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" }
-candle-datasets = { path = "./candle-datasets" }
-candle-flash-attn = { path = "./candle-flash-attn" }
-candle-kernels = { path = "./candle-kernels" }
-candle-metal-kernels = { path = "./candle-metal-kernels" }
-candle-nn = { path = "./candle-nn" }
-candle-onnx = { path = "./candle-onnx" }
-candle-transformers = { path = "./candle-transformers" }
 clap = { version = "4.2.4", features = ["derive"] }
-criterion = { version = "0.5.1", default-features=false }
 cudarc = { version = "0.9.14", features = ["f16"] }
-gemm = { version = "0.16.6", features = ["wasm-simd128-enable"] }
+# TODO: Switch back to the official gemm implementation once it has caught up.
+gemm = { version = "0.16.0", package = "candle-gemm" }
 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"] }
@@ -60,7 +53,6 @@ rayon = "1.7.0"
 rusttype = { version = "0.9", default-features = false }
 safetensors = "0.3.1"
 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 }
@@ -70,7 +62,6 @@ 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 = "0.27.0", features = ["mps"]}
 
 [profile.release-with-debug]
 inherits = "release"

README.md

@@ -51,32 +51,22 @@ For more advanced examples, please have a look at the following section.
 These online demos run entirely in your browser:
 - [yolo](https://huggingface.co/spaces/lmz/candle-yolo): pose estimation and
   object recognition.
-- [whisper](https://huggingface.co/spaces/lmz/candle-whisper): speech recognition.
+- [whisper](https://huggingface.co/spaces/lmz/candle-whisper): text to speech.
 - [LLaMA2](https://huggingface.co/spaces/lmz/candle-llama2): text generation.
 - [T5](https://huggingface.co/spaces/radames/Candle-T5-Generation-Wasm): text generation.
-- [Phi-1.5, and Phi-2](https://huggingface.co/spaces/radames/Candle-Phi-1.5-Wasm): text generation.
+- [Phi-v1.5](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, includes
-  the SOLAR-10.7B variant.
+- [LLaMA and LLaMA-v2](./candle-examples/examples/llama/): general LLM.
 - [Falcon](./candle-examples/examples/falcon/): general LLM.
-- [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.
+- [Phi-v1.5](./candle-examples/examples/phi/): a 1.3b general LLM 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.
-- [Minimal Mamba](./candle-examples/examples/minimal-mamba/): a minimal
-  implementation of the Mamba state space model.
 - [Mistral7b-v0.1](./candle-examples/examples/mistral/): a 7b general LLM with
-  better performance than all publicly available 13b models as of 2023-09-28.
-- [Mixtral8x7b-v0.1](./candle-examples/examples/mixtral/): a sparse mixture of
-  experts 8x7b general LLM with better performance than a Llama 2 70B model with
-  much faster inference.
+  performance larger than all publicly available 13b models as of 2023-09-28.
 - [StarCoder](./candle-examples/examples/bigcode/): LLM specialized to code generation.
-- [Replit-code-v1.5](./candle-examples/examples/replit-code/): a 3.3b LLM specialized for code completion.
-- [Yi-6B / Yi-34B](./candle-examples/examples/yi/): two bilingual
-  (English/Chinese) general LLMs with 6b and 34b parameters.
 - [Quantized LLaMA](./candle-examples/examples/quantized/): quantized version of
   the LLaMA model using the same quantization techniques as
   [llama.cpp](https://github.com/ggerganov/llama.cpp).
@@ -84,7 +74,7 @@ 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, SDXL 1.0 and Turbo versions.
+  image generative model, support for the 1.5, 2.1, and SDXL 1.0 versions.
 
 <img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/stable-diffusion/assets/stable-diffusion-xl.jpg" width="200">
 
@@ -104,15 +94,10 @@ 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">
 
 - [Whisper](./candle-examples/examples/whisper/): speech recognition model.
-- [T5](./candle-examples/examples/t5), [Bert](./candle-examples/examples/bert/),
-  [JinaBert](./candle-examples/examples/jina-bert/) : useful for sentence embeddings.
+- [T5](./candle-examples/examples/t5), [Bert](./candle-examples/examples/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).
-- [BLIP](./candle-examples/examples/blip/): image to text model, can be used to
-  generate captions for an image.
-- [Marian-MT](./candle-examples/examples/marian-mt/): neural machine translation
-  model, generates the translated text from the input text.
 
 Run them using commands like:
 ```
@@ -128,7 +113,7 @@ 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-1.5, and Phi-2](https://huggingface.co/spaces/radames/Candle-Phi-1.5-Wasm),
+[Phi-v1.5](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
@@ -144,21 +129,8 @@ And then head over to
 
 <!--- ANCHOR: useful_libraries --->
 
-## Useful External Resources
-- [`candle-tutorial`](https://github.com/ToluClassics/candle-tutorial): A
-  very detailed tutorial showing how to convert a PyTorch model to Candle.
-- [`candle-lora`](https://github.com/EricLBuehler/candle-lora): Efficient and
-  ergonomic LoRA implementation for Candle. `candle-lora` has      
-  out-of-the-box LoRA support for many models from Candle, which can be found
-  [here](https://github.com/EricLBuehler/candle-lora/tree/master/candle-lora-transformers/examples).
-- [`optimisers`](https://github.com/KGrewal1/optimisers): A collection of optimisers
-  including SGD with momentum, AdaGrad, AdaDelta, AdaMax, NAdam, RAdam, and RMSprop.
-- [`candle-vllm`](https://github.com/EricLBuehler/candle-vllm): Efficient platform for inference and
-  serving local LLMs including an OpenAI compatible API server.
-- [`candle-ext`](https://github.com/mokeyish/candle-ext): An extension library to Candle that provides PyTorch functions not currently available in Candle.
-- [`kalosm`](https://github.com/floneum/floneum/tree/master/interfaces/kalosm): A multi-modal meta-framework in Rust for interfacing with local pre-trained models with support for controlled generation, custom samplers, in-memory vector databases, audio transcription, and more.
-- [`candle-sampling`](https://github.com/EricLBuehler/candle-sampling): Sampling techniques for Candle.
-- [`gpt-from-scratch-rs`](https://github.com/jeroenvlek/gpt-from-scratch-rs): A port of Andrej Karpathy's _Let's build GPT_ tutorial on YouTube showcasing the Candle API on a toy problem.
+## Useful Libraries
+- [`candle-lora`](https://github.com/EricLBuehler/candle-lora) provides a LoRA implementation that conforms to the official `peft` implementation.
 
 If you have an addition to this list, please submit a pull request.
 
@@ -177,35 +149,22 @@ If you have an addition to this list, please submit a pull request.
     - WASM support, run your models in a browser.
 - Included models.
     - Language Models.
-        - LLaMA v1 and v2 with variants such as SOLAR-10.7B.
+        - LLaMA v1 and v2.
         - Falcon.
         - StarCoder.
-        - Phi 1, 1.5, and 2.
-        - Minimal Mamba
+        - Phi v1.5.
         - Mistral 7b v0.1.
-        - Mixtral 8x7b v0.1.
         - StableLM-3B-4E1T.
-        - Replit-code-v1.5-3B.
+        - T5.
         - Bert.
-        - Yi-6B and Yi-34B.
-    - Quantized LLMs.
-        - Llama 7b, 13b, 70b, as well as the chat and code variants.
-        - Mistral 7b, and 7b instruct.
-        - Mixtral 8x7b.
-        - Zephyr 7b a and b (Mistral-7b based).
-        - OpenChat 3.5 (Mistral-7b based).
-    - Text to text.
-        - T5 and its variants: FlanT5, UL2, MADLAD400 (translation), CoEdit (Grammar correction).
-        - Marian MT (Machine Translation).
     - Whisper (multi-lingual support).
-    - Text to image.
-        - Stable Diffusion v1.5, v2.1, XL v1.0.
-        - Wurstchen v2.
-    - Image to text.
-        - BLIP.
+    - Stable Diffusion v1.5, v2.1, XL v1.0.
+    - Wurstchen v2.
     - Computer Vision Models.
-        - DINOv2, ConvMixer, EfficientNet, ResNet, ViT.
-        - yolo-v3, yolo-v8.
+        - DINOv2.
+        - EfficientNet.
+        - yolo-v3.
+        - yolo-v8.
         - Segment-Anything Model (SAM).
 - File formats: load models from safetensors, npz, ggml, or PyTorch files.
 - Serverless (on CPU), small and fast deployments.
@@ -243,7 +202,6 @@ Cheatsheet:
 - [candle-datasets](./candle-datasets/): Datasets and data loaders.
 - [candle-transformers](./candle-transformers): transformers-related utilities.
 - [candle-flash-attn](./candle-flash-attn): Flash attention v2 layer.
-- [candle-onnx](./candle-onnx/): ONNX model evaluation.
 
 ## FAQ
 

candle-book/Cargo.toml

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

candle-book/src/guide/installation.md

@@ -12,9 +12,6 @@ compute_cap
 8.9
 ```
 
-You can also compile the Cuda kernels for a specific compute cap using the 
-`CUDA_COMPUTE_CAP=<compute cap>` environment variable.
-
 If any of the above commands errors out, please make sure to update your Cuda version.
 
 2. Create a new app and add [`candle-core`](https://github.com/huggingface/candle/tree/main/candle-core) with Cuda support.

candle-book/src/lib.rs

@@ -28,7 +28,6 @@ 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;
@@ -46,10 +45,9 @@ 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;
@@ -104,7 +102,6 @@ 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

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

candle-core/benches/matmul.rs

@@ -1,42 +0,0 @@
-use candle_core::{DType, Device, Tensor};
-use criterion::{black_box, criterion_group, criterion_main, Criterion, Throughput};
-use std::time::Instant;
-
-fn run(a: &Tensor, b: &Tensor) {
-    a.matmul(&b.t().unwrap()).unwrap();
-}
-
-fn criterion_benchmark(c: &mut Criterion) {
-    let b = 1;
-    let m = 1;
-    let n = 2048;
-    let k = 2048;
-
-    let device = Device::new_metal(0).unwrap();
-    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("matmul_metal");
-    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));
-            }
-            if let Device::Metal(device) = &device {
-                device.wait_until_completed().unwrap();
-            } else {
-                panic!("Expected metal device");
-            }
-            start.elapsed()
-        })
-    });
-    group.finish();
-}
-
-criterion_group!(benches, criterion_benchmark);
-criterion_main!(benches);

candle-core/examples/basics.rs

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

candle-core/src/backend.rs

@@ -39,14 +39,6 @@ pub trait BackendStorage: Sized {
         _params: &crate::conv::ParamsConv1D,
     ) -> Result<Self>;
 
-    fn conv_transpose1d(
-        &self,
-        _l: &Layout,
-        _kernel: &Self,
-        _kernel_l: &Layout,
-        _params: &crate::conv::ParamsConvTranspose1D,
-    ) -> Result<Self>;
-
     fn conv2d(
         &self,
         _l: &Layout,

candle-core/src/backprop.rs

@@ -15,17 +15,6 @@ fn broadcast_back(arg: &Tensor, node: &Tensor, reduced_dims: &[usize]) -> Result
     }
 }
 
-thread_local! {
-    static CANDLE_GRAD_DO_NOT_DETACH: bool = {
-        match std::env::var("CANDLE_GRAD_DO_NOT_DETACH") {
-            Ok(s) => {
-                !s.is_empty() && s != "0"
-            },
-            Err(_) => false,
-        }
-    }
-}
-
 impl Tensor {
     /// Return all the nodes that lead to this value in a topologically sorted vec, the first
     /// elements having dependencies on the latter ones, e.g. the first element if any is the
@@ -47,8 +36,6 @@ impl Tensor {
                 // Do not call recursively on the "leaf" nodes.
                 track_grad = true;
                 nodes
-            } else if node.dtype().is_int() {
-                nodes
             } else if let Some(op) = node.op() {
                 match op {
                     Op::IndexAdd(t1, t2, t3, _)
@@ -68,11 +55,6 @@ impl Tensor {
                         kernel: rhs,
                         ..
                     }
-                    | Op::ConvTranspose1D {
-                        arg: lhs,
-                        kernel: rhs,
-                        ..
-                    }
                     | Op::Conv2D {
                         arg: lhs,
                         kernel: rhs,
@@ -114,13 +96,14 @@ impl Tensor {
                     | Op::Unary(_node, UnaryOp::Round) => nodes,
                     Op::Reshape(node)
                     | Op::UpsampleNearest1D(node)
-                    | Op::UpsampleNearest2D { arg: node, .. }
+                    | Op::UpsampleNearest2D(node)
                     | Op::AvgPool2D { arg: node, .. }
                     | Op::MaxPool2D { arg: node, .. }
                     | Op::Copy(node)
                     | Op::Broadcast(node)
                     | Op::Cmp(node, _)
                     | Op::Reduce(node, ReduceOp::Min | ReduceOp::Sum | ReduceOp::Max, _)
+                    | Op::ToDType(node)
                     | Op::ToDevice(node)
                     | Op::Transpose(node, _, _)
                     | Op::Permute(node, _)
@@ -133,15 +116,6 @@ impl Tensor {
                         track_grad |= tg;
                         nodes
                     }
-                    Op::ToDType(node) => {
-                        if node.dtype().is_float() {
-                            let (tg, nodes) = walk(node, nodes, already_seen);
-                            track_grad |= tg;
-                            nodes
-                        } else {
-                            nodes
-                        }
-                    }
                     Op::Reduce(_, ReduceOp::ArgMin | ReduceOp::ArgMax, _) => nodes,
                 }
             } else {
@@ -166,16 +140,10 @@ impl Tensor {
             if node.is_variable() {
                 continue;
             }
-            let grad = grads
-                .remove(node)
-                .expect("candle internal error - grad not populated");
-            // https://github.com/huggingface/candle/issues/1241
-            // Ideally, we would make these operations in place where possible to ensure that we
-            // do not have to allocate too often. Here we just call `.detach` to avoid computing
-            // the backprop graph of the backprop itself. This would be an issue for second order
-            // derivatives but these are out of scope at the moment.
-            let do_not_detach = CANDLE_GRAD_DO_NOT_DETACH.with(|b| *b);
-            let grad = if do_not_detach { grad } else { grad.detach()? };
+            let grad = grads.remove(node).unwrap();
+            // TODO: We should perform all these operations in place (or at least not track the
+            // whole graph). The only drawback would be if we wanted to support grad of grad but
+            // this is out of scope.
             if let Some(op) = node.op() {
                 match op {
                     Op::Binary(lhs, rhs, BinaryOp::Add) => {
@@ -230,44 +198,7 @@ impl Tensor {
                         let f_grad = pred.where_cond(&zeros, &grad)?;
                         *f_sum_grad = f_sum_grad.add(&f_grad)?;
                     }
-                    Op::Conv1D {
-                        arg,
-                        kernel,
-                        padding,
-                        stride,
-                        dilation,
-                    } => {
-                        // The output height for conv_transpose1d is:
-                        // (l_in - 1) * stride - 2 * padding + dilation * (k_size - 1) + out_padding + 1
-                        let grad_l_in = grad.dim(2)?;
-                        let k_size = kernel.dim(2)?;
-                        let out_size =
-                            (grad_l_in - 1) * stride + dilation * (k_size - 1) + 1 - 2 * padding;
-                        let out_padding = arg.dim(2)? - out_size;
-                        let grad_arg = grad.conv_transpose1d(
-                            kernel,
-                            *padding,
-                            out_padding,
-                            *stride,
-                            *dilation,
-                        )?;
-                        let sum_grad = grads.or_insert(arg)?;
-                        *sum_grad = sum_grad.add(&grad_arg)?;
-
-                        let grad_kernel = arg
-                            .transpose(0, 1)?
-                            .conv1d(&grad.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
-                            .transpose(0, 1)?;
-                        let sum_grad = grads.or_insert(kernel)?;
-                        let (_, _, k0) = kernel.dims3()?;
-                        let (_, _, g_k0) = grad_kernel.dims3()?;
-                        let grad_kernel = if g_k0 != k0 {
-                            grad_kernel.narrow(2, 0, k0)?
-                        } else {
-                            grad_kernel
-                        };
-                        *sum_grad = sum_grad.add(&grad_kernel)?;
-                    }
+                    Op::Conv1D { .. } => Err(Error::BackwardNotSupported { op: "conv1d" })?,
                     Op::Conv2D {
                         arg,
                         kernel,
@@ -297,18 +228,8 @@ impl Tensor {
                             .conv2d(&grad.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
                             .transpose(0, 1)?;
                         let sum_grad = grads.or_insert(kernel)?;
-                        let (_, _, k0, k1) = kernel.dims4()?;
-                        let (_, _, g_k0, g_k1) = grad_kernel.dims4()?;
-                        let grad_kernel = if g_k0 != k0 || g_k1 != k1 {
-                            grad_kernel.narrow(2, 0, k0)?.narrow(3, 0, k1)?
-                        } else {
-                            grad_kernel
-                        };
                         *sum_grad = sum_grad.add(&grad_kernel)?;
                     }
-                    Op::ConvTranspose1D { .. } => Err(Error::BackwardNotSupported {
-                        op: "conv-transpose1d",
-                    })?,
                     Op::ConvTranspose2D { .. } => Err(Error::BackwardNotSupported {
                         op: "conv-transpose2d",
                     })?,
@@ -350,27 +271,9 @@ impl Tensor {
                     Op::UpsampleNearest1D { .. } => Err(Error::BackwardNotSupported {
                         op: "upsample-nearest1d",
                     })?,
-                    Op::UpsampleNearest2D {
-                        arg,
-                        target_h,
-                        target_w,
-                    } => {
-                        let (_n, c, h, w) = arg.dims4()?;
-                        if target_h % h != 0 || target_w % w != 0 {
-                            crate::bail!("backward not supported for non integer upscaling factors")
-                        }
-                        let scale_h = target_h / h;
-                        let scale_w = target_w / w;
-
-                        if scale_h != scale_w {
-                            crate::bail!("backward not supported for non uniform upscaling factors")
-                        };
-                        let kernel =
-                            Tensor::ones((c, 1, scale_h, scale_w), arg.dtype(), arg.device())?;
-                        let conv_sum = grad.conv2d(&kernel, 0, scale_h, 1, c)?;
-                        let sum_grad = grads.or_insert(arg)?;
-                        *sum_grad = conv_sum;
-                    }
+                    Op::UpsampleNearest2D { .. } => Err(Error::BackwardNotSupported {
+                        op: "upsample-nearest2d",
+                    })?,
                     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)?)?;
@@ -471,7 +374,7 @@ impl Tensor {
                     }
                     Op::ToDType(arg) => {
                         let sum_grad = grads.or_insert(arg)?;
-                        *sum_grad = sum_grad.add(&grad.to_dtype(arg.dtype())?)?
+                        *sum_grad = sum_grad.add(&grad.to_dtype(node.dtype())?)?
                     }
                     Op::Copy(arg) => {
                         let sum_grad = grads.or_insert(arg)?;
@@ -558,47 +461,17 @@ impl Tensor {
                     Op::Unary(_, UnaryOp::Round) => {
                         Err(Error::BackwardNotSupported { op: "round" })?
                     }
-                    Op::Unary(arg, UnaryOp::Gelu) => {
-                        let sum_grad = grads.or_insert(arg)?;
-                        let cube = arg.powf(3.)?;
-                        let tanh = (0.0356774 * &cube + (0.797885 * arg)?)?.tanh()?;
-                        let gelu_grad = (((0.5 * &tanh)?
-                            + (0.0535161 * cube + (0.398942 * arg)?)? * (1. - tanh.powf(2.)?))?
-                            + 0.5)?;
-                        *sum_grad = sum_grad.add(&(&grad * gelu_grad)?)?
-                    }
-                    Op::Unary(arg, UnaryOp::Erf) => {
-                        let sum_grad = grads.or_insert(arg)?;
-                        // d/dx erf(x) = 2/sqrt(pi) * e^(-x^2)
-                        let erf_grad =
-                            (2. / std::f64::consts::PI.sqrt()) * (arg.sqr()?.neg()?).exp()?;
-                        *sum_grad = sum_grad.add(&(&grad * erf_grad)?)?
-                    }
-                    Op::Unary(arg, UnaryOp::GeluErf) => {
-                        let sum_grad = grads.or_insert(arg)?;
-                        // d/dx gelu_erf(x) = 0.5 + 0.398942 e^(-x^2/2) x + 0.5 erf(x/sqrt(2))
-                        let neg_half_square = (arg.sqr()?.neg()? / 2.)?;
-                        let scaled_exp_arg = (0.398942 * neg_half_square.exp()? * arg)?;
-                        let arg_scaled_sqrt = (arg / 2f64.sqrt())?;
-                        let erf_scaled_sqrt = (0.5 * arg_scaled_sqrt.erf()?)?;
-                        let gelu_erf_grad = (0.5 + scaled_exp_arg + erf_scaled_sqrt)?;
-                        *sum_grad = sum_grad.add(&(&grad * gelu_erf_grad)?)?;
+                    Op::Unary(_, UnaryOp::Gelu) => Err(Error::BackwardNotSupported { op: "gelu" })?,
+                    Op::Unary(_, UnaryOp::Erf) => Err(Error::BackwardNotSupported { op: "erf" })?,
+                    Op::Unary(_, UnaryOp::GeluErf) => {
+                        Err(Error::BackwardNotSupported { op: "gelu-erf" })?
                     }
                     Op::Unary(arg, UnaryOp::Relu) => {
                         let sum_grad = grads.or_insert(arg)?;
                         let relu_grad = arg.ge(&arg.zeros_like()?)?.to_dtype(arg.dtype())?;
                         *sum_grad = sum_grad.add(&(&grad * relu_grad)?)?
                     }
-                    Op::Elu(arg, alpha) => {
-                        // d/dx elu(x) = 1 for x > 0, alpha * e^x for x <= 0
-                        let sum_grad = grads.or_insert(arg)?;
-                        let zeros = arg.zeros_like()?;
-                        let positive_mask = arg.gt(&zeros)?.to_dtype(arg.dtype())?;
-                        let negative_mask = arg.le(&zeros)?.to_dtype(arg.dtype())?;
-                        let negative_exp_mask = ((negative_mask * arg.exp())? * *alpha)?;
-                        let combined_mask = (positive_mask + negative_exp_mask)?;
-                        *sum_grad = sum_grad.add(&(grad * combined_mask)?)?
-                    }
+                    Op::Elu(..) => Err(Error::BackwardNotSupported { op: "elu" })?,
                     Op::Powf(arg, e) => {
                         let arg_grad = (&(grad * arg.powf(e - 1.)?)? * *e)?;
                         let sum_grad = grads.or_insert(arg)?;

candle-core/src/conv.rs

@@ -25,33 +25,6 @@ impl ParamsConv1D {
     }
 }
 
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub struct ParamsConvTranspose1D {
-    pub(crate) b_size: usize,
-    pub(crate) l_in: usize,
-    pub(crate) c_out: usize,
-    pub(crate) c_in: usize,
-    pub(crate) k_size: usize,
-    pub(crate) padding: usize,
-    pub(crate) output_padding: usize,
-    pub(crate) stride: usize,
-    pub(crate) dilation: usize,
-}
-
-impl ParamsConvTranspose1D {
-    pub(crate) fn l_out(&self) -> usize {
-        (self.l_in - 1) * self.stride - 2 * self.padding
-            + self.dilation * (self.k_size - 1)
-            + self.output_padding
-            + 1
-    }
-
-    pub(crate) fn out_dims(&self) -> Vec<usize> {
-        let l_out = self.l_out();
-        vec![self.b_size, self.c_out, l_out]
-    }
-}
-
 #[derive(Debug, Clone, PartialEq, Eq, Hash)]
 pub enum CudnnFwdAlgo {
     ImplicitGemm,
@@ -187,49 +160,6 @@ impl Tensor {
         }
     }
 
-    /// Applies a 1D transposed convolution over the input tensor.
-    pub fn conv_transpose1d(
-        &self,
-        kernel: &Self,
-        padding: usize,
-        output_padding: usize,
-        stride: usize,
-        dilation: usize,
-    ) -> Result<Self> {
-        let (b_size, c_in, l_in) = self.dims3()?;
-        let (c_in_k, c_out, k_size) = kernel.dims3()?;
-        if c_in != c_in_k {
-            crate::bail!("in_channel mismatch between input ({c_in}) and kernel ({c_in_k})")
-        }
-        let params = ParamsConvTranspose1D {
-            b_size,
-            l_in,
-            k_size,
-            c_out,
-            c_in,
-            padding,
-            output_padding,
-            stride,
-            dilation,
-        };
-        let storage = self.storage().conv_transpose1d(
-            self.layout(),
-            &kernel.storage(),
-            kernel.layout(),
-            &params,
-        )?;
-        let op = BackpropOp::new2(self, kernel, |arg, kernel| Op::ConvTranspose1D {
-            arg,
-            kernel,
-            padding: params.padding,
-            output_padding: params.output_padding,
-            stride: params.stride,
-            dilation: params.dilation,
-        });
-        let out_dims = params.out_dims();
-        Ok(crate::tensor::from_storage(storage, out_dims, op, false))
-    }
-
     fn conv2d_single_group(&self, kernel: &Self, params: &ParamsConv2D) -> Result<Self> {
         let storage =
             self.storage()

candle-core/src/cpu_backend.rs

@@ -804,11 +804,11 @@ impl<'a, I: IntDType> Map1 for Gather<'a, I> {
     fn f<T: WithDType>(&self, src: &[T], src_l: &Layout) -> Result<Vec<T>> {
         let ids = match self.ids_l.contiguous_offsets() {
             Some((a, b)) => &self.ids[a..b],
-            None => Err(Error::RequiresContiguous { op: "gather" }.bt())?,
+            None => Err(Error::RequiresContiguous { op: "gather" })?,
         };
         let src = match src_l.contiguous_offsets() {
             Some((a, b)) => &src[a..b],
-            None => Err(Error::RequiresContiguous { op: "gather" }.bt())?,
+            None => Err(Error::RequiresContiguous { op: "gather" })?,
         };
         let dim = self.dim;
         let ids_dims = self.ids_l.dims();
@@ -857,7 +857,7 @@ impl<'a, I: IntDType> Map1 for IndexSelect<'a, I> {
     fn f<T: WithDType>(&self, src: &[T], layout: &Layout) -> Result<Vec<T>> {
         let src = match layout.contiguous_offsets() {
             Some((a, b)) => &src[a..b],
-            None => Err(Error::RequiresContiguous { op: "index-select" }.bt())?,
+            None => Err(Error::RequiresContiguous { op: "index-select" })?,
         };
         let dim = self.dim;
         let n_ids = match self.ids_l.dims() {
@@ -913,7 +913,7 @@ impl<'a, I: IntDType> Map2 for ScatterAdd<'a, I> {
         let mut dst = vec![T::zero(); dst_len];
         copy_strided_src_(v1, &mut dst, 0, l1);
         let src = match src_l.contiguous_offsets() {
-            None => Err(Error::RequiresContiguous { op: "scatter-add" }.bt())?,
+            None => Err(Error::RequiresContiguous { op: "scatter-add" })?,
             Some((o1, o2)) => &src[o1..o2],
         };
 
@@ -929,7 +929,7 @@ impl<'a, I: IntDType> Map2 for ScatterAdd<'a, I> {
 
         let ids = match self.ids_l.contiguous_offsets() {
             Some((a, b)) => &self.ids[a..b],
-            None => Err(Error::RequiresContiguous { op: "gather" }.bt())?,
+            None => Err(Error::RequiresContiguous { op: "gather" })?,
         };
         for left_i in 0..ids_left_len {
             let start_ids_idx = left_i * ids_right_len * ids_dim_len;
@@ -971,7 +971,7 @@ impl<'a, I: IntDType> Map2 for IndexAdd<'a, I> {
         let mut dst = vec![T::zero(); dst_len];
         copy_strided_src_(v1, &mut dst, 0, l1);
         let src = match src_l.contiguous_offsets() {
-            None => Err(Error::RequiresContiguous { op: "index-add" }.bt())?,
+            None => Err(Error::RequiresContiguous { op: "index-add" })?,
             Some((o1, o2)) => &src[o1..o2],
         };
         let dim = self.dim;
@@ -1256,74 +1256,6 @@ impl Map1 for Im2Col {
     }
 }
 
-struct ConvTranspose1D<'a>(&'a crate::conv::ParamsConvTranspose1D);
-
-impl<'a> Map2 for ConvTranspose1D<'a> {
-    const OP: &'static str = "conv_transpose1d";
-    fn f<T: WithDType>(&self, inp: &[T], inp_l: &Layout, k: &[T], k_l: &Layout) -> Result<Vec<T>> {
-        let p = self.0;
-        let inp = &inp[inp_l.start_offset()..];
-        let (inp_s0, inp_s1, inp_s2) = crate::shape::dims3(inp_l.stride())?;
-        let (k_s0, k_s1, k_s2) = crate::shape::dims3(k_l.stride())?;
-        let l_out = p.l_out();
-
-        // Output shape: [b_size, c_out, l_out].
-        let dst_elems = p.c_out * l_out * p.b_size;
-        let dst = vec![T::zero(); dst_elems];
-        let dst_s0 = p.c_out * l_out;
-        let dst_s1 = l_out;
-        let dst_s2 = 1;
-
-        // TODO: Avoid making this copy if `inp` already has the appropriate layout.
-        let mut inp_cont = vec![T::zero(); p.b_size * p.c_in * p.l_in];
-        let cont_s0 = p.l_in * p.c_in;
-        let cont_s1 = p.c_in;
-        for b_idx in 0..p.b_size {
-            for l_idx in 0..p.l_in {
-                for c_idx in 0..p.c_in {
-                    let src_idx = b_idx * inp_s0 + c_idx * inp_s1 + l_idx * inp_s2;
-                    let dst_idx = b_idx * cont_s0 + l_idx * cont_s1 + c_idx;
-                    inp_cont[dst_idx] = inp[src_idx]
-                }
-            }
-        }
-
-        for k_idx in 0..p.k_size {
-            (0..p.c_out).into_par_iter().for_each(|dst_c_idx| {
-                let k_cont = (0..p.c_in)
-                    .map(|c_in_idx| k[c_in_idx * k_s0 + dst_c_idx * k_s1 + k_idx * k_s2])
-                    .collect::<Vec<_>>();
-                for b_idx in 0..p.b_size {
-                    for l_idx in 0..p.l_in {
-                        let out_idx = l_idx * p.stride + k_idx * p.dilation;
-                        if out_idx < p.padding {
-                            continue;
-                        }
-                        let out_idx = out_idx - p.padding;
-                        if out_idx < l_out {
-                            let inp_cont = &inp_cont[b_idx * cont_s0 + l_idx * cont_s1..];
-                            let dst_idx = b_idx * dst_s0 + out_idx * dst_s2 + dst_c_idx * dst_s1;
-                            let mut d = T::zero();
-                            unsafe {
-                                T::vec_dot(inp_cont.as_ptr(), k_cont.as_ptr(), &mut d, p.c_in)
-                            }
-                            let dst_p = dst.as_ptr();
-                            // Safety: dst_idx are uniques per dst_c_idx which is used to
-                            // parallelise the different tasks so no two threads can try to
-                            // write at the same location.
-                            unsafe {
-                                let ptr = dst_p.add(dst_idx) as *mut T;
-                                *ptr += d
-                            }
-                        }
-                    }
-                }
-            })
-        }
-        Ok(dst)
-    }
-}
-
 struct Conv2D<'a>(&'a crate::conv::ParamsConv2D);
 
 impl<'a> Map2 for Conv2D<'a> {
@@ -2503,16 +2435,6 @@ impl BackendStorage for CpuStorage {
         Ok(res_t)
     }
 
-    fn conv_transpose1d(
-        &self,
-        l: &Layout,
-        kernel: &Self,
-        kernel_l: &Layout,
-        params: &crate::conv::ParamsConvTranspose1D,
-    ) -> Result<Self> {
-        ConvTranspose1D(params).map(self, l, kernel, kernel_l)
-    }
-
     fn conv2d(
         &self,
         l: &Layout,
@@ -2617,25 +2539,25 @@ impl BackendStorage for CpuStorage {
             Self::U8(ids) => {
                 let ids = match ids_l.contiguous_offsets() {
                     Some((a, b)) => &ids[a..b],
-                    None => Err(Error::RequiresContiguous { op: "index-add" }.bt())?,
+                    None => Err(Error::RequiresContiguous { op: "index-add" })?,
                 };
                 IndexAdd { ids, dim }.map(self, l, src, src_l)
             }
             Self::U32(ids) => {
                 let ids = match ids_l.contiguous_offsets() {
                     Some((a, b)) => &ids[a..b],
-                    None => Err(Error::RequiresContiguous { op: "index-add" }.bt())?,
+                    None => Err(Error::RequiresContiguous { op: "index-add" })?,
                 };
                 IndexAdd { ids, dim }.map(self, l, src, src_l)
             }
             Self::I64(ids) => {
                 let ids = match ids_l.contiguous_offsets() {
                     Some((a, b)) => &ids[a..b],
-                    None => Err(Error::RequiresContiguous { op: "index-add" }.bt())?,
+                    None => Err(Error::RequiresContiguous { op: "index-add" })?,
                 };
                 IndexAdd { ids, dim }.map(self, l, src, src_l)
             }
-            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "index-add").bt()),
+            _ => Err(Error::UnsupportedDTypeForOp(self.dtype(), "index-add")),
         }
     }
 

candle-core/src/cuda_backend.rs

@@ -224,10 +224,8 @@ impl BackendDevice for CudaDevice {
     }
 
     fn set_seed(&self, seed: u64) -> Result<()> {
-        // We do not call set_seed but instead create a new curand object. This ensures that the
-        // state will be identical and the same random numbers will be generated.
         let mut curand = self.curand.lock().unwrap();
-        curand.0 = cudarc::curand::CudaRng::new(seed, self.device.clone()).w()?;
+        curand.0.set_seed(seed).w()?;
         Ok(())
     }
 
@@ -1808,16 +1806,6 @@ impl BackendStorage for CudaStorage {
         Ok(res_t)
     }
 
-    fn conv_transpose1d(
-        &self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: &crate::conv::ParamsConvTranspose1D,
-    ) -> Result<Self> {
-        todo!()
-    }
-
     #[cfg(not(feature = "cudnn"))]
     fn conv2d(
         &self,
@@ -2181,7 +2169,7 @@ impl BackendStorage for CudaStorage {
                 if src_l.is_contiguous() {
                     dev.dtod_copy(&src, &mut dst).w()?
                 } else {
-                    let func = dev.get_or_load_func("ucopy_f64", kernels::UNARY)?;
+                    let func = dev.get_or_load_func("ucopy_64", kernels::UNARY)?;
                     // SAFETY: Set later by running the kernel.
                     let params = (el_count, dims.len(), &ds, &src, &mut dst);
                     // SAFETY: ffi.

candle-core/src/device.rs

@@ -8,14 +8,12 @@ use crate::{CpuStorage, DType, Result, Shape, Storage, WithDType};
 pub enum DeviceLocation {
     Cpu,
     Cuda { gpu_id: usize },
-    Metal { gpu_id: usize },
 }
 
 #[derive(Debug, Clone)]
 pub enum Device {
     Cpu,
     Cuda(crate::CudaDevice),
-    Metal(crate::MetalDevice),
 }
 
 pub trait NdArray {
@@ -130,23 +128,10 @@ impl Device {
         Ok(Self::Cuda(crate::CudaDevice::new(ordinal)?))
     }
 
-    pub fn new_metal(ordinal: usize) -> Result<Self> {
-        Ok(Self::Metal(crate::MetalDevice::new(ordinal)?))
-    }
-
-    pub fn set_seed(&self, seed: u64) -> Result<()> {
-        match self {
-            Self::Cpu => CpuDevice.set_seed(seed),
-            Self::Cuda(c) => c.set_seed(seed),
-            Self::Metal(m) => m.set_seed(seed),
-        }
-    }
-
     pub fn same_device(&self, rhs: &Self) -> bool {
         match (self, rhs) {
             (Self::Cpu, Self::Cpu) => true,
             (Self::Cuda(lhs), Self::Cuda(rhs)) => lhs.same_device(rhs),
-            (Self::Metal(lhs), Self::Metal(rhs)) => lhs.same_device(rhs),
             _ => false,
         }
     }
@@ -155,20 +140,21 @@ impl Device {
         match self {
             Self::Cpu => DeviceLocation::Cpu,
             Self::Cuda(device) => device.location(),
-            Device::Metal(device) => device.location(),
         }
     }
 
     pub fn is_cpu(&self) -> bool {
-        matches!(self, Self::Cpu)
+        match self {
+            Self::Cpu => true,
+            Self::Cuda(_) => false,
+        }
     }
 
     pub fn is_cuda(&self) -> bool {
-        matches!(self, Self::Cuda(_))
-    }
-
-    pub fn is_metal(&self) -> bool {
-        matches!(self, Self::Metal(_))
+        match self {
+            Self::Cpu => false,
+            Self::Cuda(_) => true,
+        }
     }
 
     pub fn cuda_if_available(ordinal: usize) -> Result<Self> {
@@ -192,18 +178,8 @@ impl Device {
                 Ok(Storage::Cpu(storage))
             }
             Device::Cuda(device) => {
-                // TODO: Remove the special case if we start supporting generating f16/bf16 directly.
-                if dtype == DType::F16 || dtype == DType::BF16 {
-                    let storage = device.rand_uniform(shape, DType::F32, lo, up)?;
-                    Storage::Cuda(storage).to_dtype(&crate::Layout::contiguous(shape), dtype)
-                } else {
-                    let storage = device.rand_uniform(shape, dtype, lo, up)?;
-                    Ok(Storage::Cuda(storage))
-                }
-            }
-            Device::Metal(device) => {
                 let storage = device.rand_uniform(shape, dtype, lo, up)?;
-                Ok(Storage::Metal(storage))
+                Ok(Storage::Cuda(storage))
             }
         }
     }
@@ -230,18 +206,8 @@ impl Device {
                 Ok(Storage::Cpu(storage))
             }
             Device::Cuda(device) => {
-                // TODO: Remove the special case if we start supporting generating f16/bf16 directly.
-                if dtype == DType::F16 || dtype == DType::BF16 {
-                    let storage = device.rand_normal(shape, DType::F32, mean, std)?;
-                    Storage::Cuda(storage).to_dtype(&crate::Layout::contiguous(shape), dtype)
-                } else {
-                    let storage = device.rand_normal(shape, dtype, mean, std)?;
-                    Ok(Storage::Cuda(storage))
-                }
-            }
-            Device::Metal(device) => {
                 let storage = device.rand_normal(shape, dtype, mean, std)?;
-                Ok(Storage::Metal(storage))
+                Ok(Storage::Cuda(storage))
             }
         }
     }
@@ -265,10 +231,6 @@ impl Device {
                 let storage = device.ones_impl(shape, dtype)?;
                 Ok(Storage::Cuda(storage))
             }
-            Device::Metal(device) => {
-                let storage = device.ones_impl(shape, dtype)?;
-                Ok(Storage::Metal(storage))
-            }
         }
     }
 
@@ -282,10 +244,6 @@ impl Device {
                 let storage = device.zeros_impl(shape, dtype)?;
                 Ok(Storage::Cuda(storage))
             }
-            Device::Metal(device) => {
-                let storage = device.zeros_impl(shape, dtype)?;
-                Ok(Storage::Metal(storage))
-            }
         }
     }
 
@@ -297,11 +255,6 @@ impl Device {
                 let storage = device.storage_from_cpu_storage(&storage)?;
                 Ok(Storage::Cuda(storage))
             }
-            Device::Metal(device) => {
-                let storage = array.to_cpu_storage();
-                let storage = device.storage_from_cpu_storage(&storage)?;
-                Ok(Storage::Metal(storage))
-            }
         }
     }
 
@@ -313,11 +266,6 @@ impl Device {
                 let storage = device.storage_from_cpu_storage(&storage)?;
                 Ok(Storage::Cuda(storage))
             }
-            Device::Metal(device) => {
-                let storage = S::to_cpu_storage_owned(data);
-                let storage = device.storage_from_cpu_storage(&storage)?;
-                Ok(Storage::Metal(storage))
-            }
         }
     }
 }

candle-core/src/display.rs

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

candle-core/src/dummy_cuda_backend.rs

@@ -79,16 +79,6 @@ impl crate::backend::BackendStorage for CudaStorage {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
-    fn conv_transpose1d(
-        &self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: &crate::conv::ParamsConvTranspose1D,
-    ) -> Result<Self> {
-        Err(Error::NotCompiledWithCudaSupport)
-    }
-
     fn conv2d(
         &self,
         _: &Layout,

candle-core/src/dummy_metal_backend.rs

@@ -1,223 +0,0 @@
-#![allow(dead_code)]
-use crate::op::{BinaryOpT, CmpOp, ReduceOp, UnaryOpT};
-use crate::{CpuStorage, DType, Error, Layout, Result, Shape};
-
-#[derive(Debug, Clone)]
-pub struct MetalDevice;
-
-#[derive(Debug)]
-pub struct MetalStorage;
-
-#[derive(thiserror::Error, Debug)]
-pub enum MetalError {
-    #[error("{0}")]
-    Message(String),
-}
-
-impl From<String> for MetalError {
-    fn from(e: String) -> Self {
-        MetalError::Message(e)
-    }
-}
-
-macro_rules! fail {
-    () => {
-        unimplemented!("metal support has not been enabled, add `metal` feature to enable.")
-    };
-}
-
-impl crate::backend::BackendStorage for MetalStorage {
-    type Device = MetalDevice;
-
-    fn try_clone(&self, _: &Layout) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn dtype(&self) -> DType {
-        fail!()
-    }
-
-    fn device(&self) -> &Self::Device {
-        fail!()
-    }
-
-    fn to_cpu_storage(&self) -> Result<CpuStorage> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn affine(&self, _: &Layout, _: f64, _: f64) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn powf(&self, _: &Layout, _: f64) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn elu(&self, _: &Layout, _: f64) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn reduce_op(&self, _: ReduceOp, _: &Layout, _: &[usize]) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn cmp(&self, _: CmpOp, _: &Self, _: &Layout, _: &Layout) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn to_dtype(&self, _: &Layout, _: DType) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn unary_impl<B: UnaryOpT>(&self, _: &Layout) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn binary_impl<B: BinaryOpT>(&self, _: &Self, _: &Layout, _: &Layout) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn where_cond(&self, _: &Layout, _: &Self, _: &Layout, _: &Self, _: &Layout) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn conv1d(
-        &self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: &crate::conv::ParamsConv1D,
-    ) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn conv_transpose1d(
-        &self,
-        _l: &Layout,
-        _kernel: &Self,
-        _kernel_l: &Layout,
-        _params: &crate::conv::ParamsConvTranspose1D,
-    ) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn conv2d(
-        &self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: &crate::conv::ParamsConv2D,
-    ) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn conv_transpose2d(
-        &self,
-        _l: &Layout,
-        _kernel: &Self,
-        _kernel_l: &Layout,
-        _params: &crate::conv::ParamsConvTranspose2D,
-    ) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn index_select(&self, _: &Self, _: &Layout, _: &Layout, _: usize) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-    fn gather(&self, _: &Layout, _: &Self, _: &Layout, _: usize) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn scatter_add(
-        &self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: usize,
-    ) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn index_add(
-        &self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: &Self,
-        _: &Layout,
-        _: usize,
-    ) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn matmul(
-        &self,
-        _: &Self,
-        _: (usize, usize, usize, usize),
-        _: &Layout,
-        _: &Layout,
-    ) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn copy_strided_src(&self, _: &mut Self, _: usize, _: &Layout) -> Result<()> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn avg_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn max_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn upsample_nearest1d(&self, _: &Layout, _: usize) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn upsample_nearest2d(&self, _: &Layout, _: usize, _: usize) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-}
-
-impl crate::backend::BackendDevice for MetalDevice {
-    type Storage = MetalStorage;
-    fn new(_: usize) -> Result<Self> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn set_seed(&self, _: u64) -> Result<()> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn location(&self) -> crate::DeviceLocation {
-        fail!()
-    }
-
-    fn same_device(&self, _: &Self) -> bool {
-        fail!()
-    }
-
-    fn zeros_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn ones_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-
-    fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
-        Err(Error::NotCompiledWithMetalSupport)
-    }
-}

candle-core/src/error.rs

@@ -1,4 +1,4 @@
-use crate::{DType, DeviceLocation, Layout, MetalError, Shape};
+use crate::{DType, DeviceLocation, Layout, Shape};
 
 #[derive(Debug, Clone)]
 pub struct MatMulUnexpectedStriding {
@@ -142,9 +142,6 @@ pub enum Error {
     #[error("{op} expects at least one tensor")]
     OpRequiresAtLeastOneTensor { op: &'static str },
 
-    #[error("{op} expects at least two tensors")]
-    OpRequiresAtLeastTwoTensors { op: &'static str },
-
     #[error("backward is not supported for {op}")]
     BackwardNotSupported { op: &'static str },
 
@@ -152,9 +149,6 @@ pub enum Error {
     #[error("the candle crate has not been built with cuda support")]
     NotCompiledWithCudaSupport,
 
-    #[error("the candle crate has not been built with metal support")]
-    NotCompiledWithMetalSupport,
-
     #[error("cannot find tensor {path}")]
     CannotFindTensor { path: String },
 
@@ -162,9 +156,6 @@ pub enum Error {
     #[error(transparent)]
     Cuda(Box<dyn std::error::Error + Send + Sync>),
 
-    #[error("Metal error {0}")]
-    Metal(#[from] MetalError),
-
     #[error(transparent)]
     TryFromIntError(#[from] core::num::TryFromIntError),
 

candle-core/src/indexer.rs

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

candle-core/src/lib.rs

@@ -49,12 +49,9 @@ mod device;
 pub mod display;
 mod dtype;
 mod dummy_cuda_backend;
-mod dummy_metal_backend;
 pub mod error;
 mod indexer;
 pub mod layout;
-#[cfg(feature = "metal")]
-pub mod metal_backend;
 #[cfg(feature = "mkl")]
 mod mkl;
 pub mod npy;
@@ -90,12 +87,6 @@ pub use cuda_backend::{CudaDevice, CudaStorage};
 #[cfg(not(feature = "cuda"))]
 pub use dummy_cuda_backend::{CudaDevice, CudaStorage};
 
-#[cfg(feature = "metal")]
-pub use metal_backend::{MetalDevice, MetalError, MetalStorage};
-
-#[cfg(not(feature = "metal"))]
-pub use dummy_metal_backend::{MetalDevice, MetalError, MetalStorage};
-
 #[cfg(feature = "mkl")]
 extern crate intel_mkl_src;
 
@@ -123,20 +114,14 @@ pub trait Module {
     fn forward(&self, xs: &Tensor) -> Result<Tensor>;
 }
 
+impl Module for quantized::QMatMul {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        self.forward(xs)
+    }
+}
+
 impl<T: Fn(&Tensor) -> Result<Tensor>> Module for T {
     fn forward(&self, xs: &Tensor) -> Result<Tensor> {
         self(xs)
     }
 }
-
-// A trait defining a module with forward method using a single tensor argument and a flag to
-// separate the training and evaluation behaviors.
-pub trait ModuleT {
-    fn forward_t(&self, xs: &Tensor, train: bool) -> Result<Tensor>;
-}
-
-impl<M: Module> ModuleT for M {
-    fn forward_t(&self, xs: &Tensor, _train: bool) -> Result<Tensor> {
-        self.forward(xs)
-    }
-}

candle-core/src/npy.rs

@@ -250,6 +250,8 @@ impl Tensor {
         if header.fortran_order {
             return Err(Error::Npy("fortran order not supported".to_string()));
         }
+        let mut data: Vec<u8> = vec![];
+        reader.read_to_end(&mut data)?;
         Self::from_reader(header.shape(), header.descr, &mut reader)
     }
 

candle-core/src/op.rs

@@ -1,5 +1,5 @@
 #![allow(clippy::redundant_closure_call)]
-use crate::{CpuStorage, CudaStorage, Layout, MetalStorage, Result, Shape, Tensor};
+use crate::{CpuStorage, CudaStorage, Layout, Result, Shape, Tensor};
 use half::{bf16, f16};
 use num_traits::float::Float;
 
@@ -90,16 +90,6 @@ pub enum Op {
         dilation: usize,
     },
 
-    #[allow(dead_code)]
-    ConvTranspose1D {
-        arg: Tensor,
-        kernel: Tensor,
-        padding: usize,
-        output_padding: usize,
-        stride: usize,
-        dilation: usize,
-    },
-
     #[allow(dead_code)]
     Conv2D {
         arg: Tensor,
@@ -132,11 +122,7 @@ pub enum Op {
     },
 
     UpsampleNearest1D(Tensor),
-    UpsampleNearest2D {
-        arg: Tensor,
-        target_h: usize,
-        target_w: usize,
-    },
+    UpsampleNearest2D(Tensor),
 
     Cat(Vec<Tensor>, usize),
 
@@ -188,18 +174,6 @@ pub trait CustomOp1 {
         ))
     }
 
-    /// The forward pass, as run on a metal gpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn metal_fwd(
-        &self,
-        _storage: &MetalStorage,
-        _layout: &Layout,
-    ) -> Result<(MetalStorage, Shape)> {
-        Err(crate::Error::Metal(
-            format!("no metal implementation for {}", self.name()).into(),
-        ))
-    }
-
     /// This function takes as argument the argument `arg` used in the forward pass, the result
     /// produced by the forward operation `res` and the gradient of the result `grad_res`.
     /// The function should return the gradient of the argument.
@@ -235,20 +209,6 @@ pub trait CustomOp2 {
         ))
     }
 
-    /// The forward pass, as run on a metal gpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn metal_fwd(
-        &self,
-        _: &MetalStorage,
-        _: &Layout,
-        _: &MetalStorage,
-        _: &Layout,
-    ) -> Result<(MetalStorage, Shape)> {
-        Err(crate::Error::Metal(
-            format!("no metal implementation for {}", self.name()).into(),
-        ))
-    }
-
     fn bwd(
         &self,
         _arg1: &Tensor,
@@ -291,22 +251,6 @@ pub trait CustomOp3 {
         ))
     }
 
-    /// The forward pass, as run on a metal gpu device. Note that the storage can use arbitrary strides,
-    /// offsets etc so the associated layout should be used to access it.
-    fn metal_fwd(
-        &self,
-        _: &MetalStorage,
-        _: &Layout,
-        _: &MetalStorage,
-        _: &Layout,
-        _: &MetalStorage,
-        _: &Layout,
-    ) -> Result<(MetalStorage, Shape)> {
-        Err(crate::Error::Metal(
-            format!("no metal implementation for {}", self.name()).into(),
-        ))
-    }
-
     fn bwd(
         &self,
         _arg1: &Tensor,
@@ -592,13 +536,13 @@ unary_op!(Log, "log", v, v.ln(), vs_ln, vd_ln);
 unary_op!(Sin, "sin", v, v.sin(), vs_sin, vd_sin);
 unary_op!(Cos, "cos", v, v.cos(), vs_cos, vd_cos);
 unary_op!(Tanh, "tanh", v, v.tanh(), vs_tanh, vd_tanh);
+unary_op!(Abs, "abs", v, v.abs());
 unary_op!(Neg, "neg", v, -v);
 unary_op!(Recip, "recip", v, v.recip());
 unary_op!(Sqr, "sqr", v, v * v, vs_sqr, vd_sqr);
 unary_op!(Sqrt, "sqrt", v, v.sqrt(), vs_sqrt, vd_sqrt);
 
-/// Tanh based approximation of the `gelu` operation
-/// GeluErf is the more precise one.
+/// `gelu` operation
 /// <https://en.wikipedia.org/wiki/Activation_function#Comparison_of_activation_functions>
 impl UnaryOpT for Gelu {
     const NAME: &'static str = "gelu";
@@ -688,8 +632,6 @@ impl UnaryOpT for Gelu {
     }
 }
 
-/// `erf` operation
-/// <https://en.wikipedia.org/wiki/Error_function>
 impl UnaryOpT for Erf {
     const NAME: &'static str = "erf";
     const KERNEL: &'static str = "uerf";
@@ -724,40 +666,6 @@ impl UnaryOpT for Erf {
     }
 }
 
-impl UnaryOpT for Abs {
-    const NAME: &'static str = "abs";
-    const KERNEL: &'static str = "uabs";
-    const V: Self = Abs;
-    #[inline(always)]
-    fn bf16(v: bf16) -> bf16 {
-        v.abs()
-    }
-    #[inline(always)]
-    fn f16(v: f16) -> f16 {
-        v.abs()
-    }
-    #[inline(always)]
-    fn f32(v: f32) -> f32 {
-        v.abs()
-    }
-    #[inline(always)]
-    fn f64(v: f64) -> f64 {
-        v.abs()
-    }
-    #[inline(always)]
-    fn u8(v: u8) -> u8 {
-        v
-    }
-    #[inline(always)]
-    fn u32(v: u32) -> u32 {
-        v
-    }
-    #[inline(always)]
-    fn i64(v: i64) -> i64 {
-        v.abs()
-    }
-}
-
 impl UnaryOpT for Ceil {
     const NAME: &'static str = "ceil";
     const KERNEL: &'static str = "uceil";
@@ -979,10 +887,6 @@ impl BackpropOp {
         };
         Self(op)
     }
-
-    pub(crate) fn is_none(&self) -> bool {
-        self.0.is_none()
-    }
 }
 
 impl std::ops::Deref for BackpropOp {

candle-core/src/pickle.rs

@@ -193,50 +193,6 @@ impl Object {
             _ => Err(self),
         }
     }
-
-    pub fn into_tensor_info(
-        self,
-        name: Self,
-        dir_name: &std::path::Path,
-    ) -> Result<Option<TensorInfo>> {
-        let name = match name.unicode() {
-            Ok(name) => name,
-            Err(_) => return Ok(None),
-        };
-        let (callable, args) = match self.reduce() {
-            Ok(callable_args) => callable_args,
-            _ => return Ok(None),
-        };
-        let (callable, args) = match callable {
-            Object::Class {
-                module_name,
-                class_name,
-            } if module_name == "torch._tensor" && class_name == "_rebuild_from_type_v2" => {
-                let mut args = args.tuple()?;
-                let callable = args.remove(0);
-                let args = args.remove(1);
-                (callable, args)
-            }
-            _ => (callable, args),
-        };
-        match callable {
-            Object::Class {
-                module_name,
-                class_name,
-            } if module_name == "torch._utils" && class_name == "_rebuild_tensor_v2" => {}
-            _ => return Ok(None),
-        };
-        let (layout, dtype, file_path, storage_size) = rebuild_args(args)?;
-        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(),
-            storage_size,
-        }))
-    }
 }
 
 impl TryFrom<Object> for String {
@@ -609,7 +565,6 @@ fn rebuild_args(args: Object) -> Result<(Layout, DType, String, usize)> {
         "HalfStorage" => DType::F16,
         "BFloat16Storage" => DType::BF16,
         "ByteStorage" => DType::U8,
-        "LongStorage" => DType::I64,
         other => {
             crate::bail!("unsupported storage type {other}")
         }
@@ -668,10 +623,50 @@ pub fn read_pth_tensor_info<P: AsRef<std::path::Path>>(
         };
         if let Object::Dict(key_values) = obj {
             for (name, value) in key_values.into_iter() {
-                match value.into_tensor_info(name, &dir_name) {
-                    Ok(Some(tensor_info)) => tensor_infos.push(tensor_info),
-                    Ok(None) => {}
-                    Err(err) => eprintln!("skipping: {err:?}"),
+                let name = match name.unicode() {
+                    Ok(name) => name,
+                    Err(_) => continue,
+                };
+                let (callable, args) = match value.reduce() {
+                    Ok(callable_args) => callable_args,
+                    _ => continue,
+                };
+                let (callable, args) = match callable {
+                    Object::Class {
+                        module_name,
+                        class_name,
+                    } if module_name == "torch._tensor"
+                        && class_name == "_rebuild_from_type_v2" =>
+                    {
+                        let mut args = args.tuple()?;
+                        let callable = args.remove(0);
+                        let args = args.remove(1);
+                        (callable, args)
+                    }
+                    _ => (callable, args),
+                };
+                match callable {
+                    Object::Class {
+                        module_name,
+                        class_name,
+                    } if module_name == "torch._utils" && class_name == "_rebuild_tensor_v2" => {}
+                    _ => continue,
+                };
+                match rebuild_args(args) {
+                    Ok((layout, dtype, file_path, storage_size)) => {
+                        let mut path = dir_name.clone();
+                        path.push(file_path);
+                        tensor_infos.push(TensorInfo {
+                            name,
+                            dtype,
+                            layout,
+                            path: path.to_string_lossy().into_owned(),
+                            storage_size,
+                        })
+                    }
+                    Err(err) => {
+                        eprintln!("skipping {name}: {err:?}")
+                    }
                 }
             }
         }
@@ -728,16 +723,3 @@ impl PthTensors {
         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)?;
-    let tensor_names = pth.tensor_infos.keys();
-    let mut tensors = Vec::with_capacity(tensor_names.len());
-    for name in tensor_names {
-        if let Some(tensor) = pth.get(name)? {
-            tensors.push((name.to_string(), tensor))
-        }
-    }
-    Ok(tensors)
-}

candle-core/src/quantized/avx.rs

@@ -50,9 +50,14 @@ pub(crate) unsafe fn mul_sum_i8_pairs_float(x: __m256i, y: __m256i) -> __m256 {
 #[inline(always)]
 pub(crate) fn vec_dot_q4_0_q8_0(n: usize, xs: &[BlockQ4_0], ys: &[BlockQ8_0]) -> Result<f32> {
     let qk = QK8_0;
+    let nb = n / qk;
     if n % QK8_0 != 0 {
         crate::bail!("vec_dot_q4_0_q8_0: {n} is not divisible by {qk}")
     }
+    if nb % 2 != 0 {
+        crate::bail!("vec_dot_q4_0_q8_0: {nb} is not even")
+    }
+
     unsafe {
         let mut acc = _mm256_setzero_ps();
         for (x, y) in xs.iter().zip(ys.iter()) {
@@ -353,7 +358,7 @@ 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 separate register
+                // We hardcode the shifts here to avoid loading them into a seperate 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 +591,7 @@ 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 separate register
+                //Similar to q3k we hardcode the shifts here to avoid loading them into a seperate 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/gguf_file.rs

@@ -29,7 +29,6 @@ impl TryFrom<u32> for Magic {
 pub enum VersionedMagic {
     GgufV1,
     GgufV2,
-    GgufV3,
 }
 
 impl VersionedMagic {
@@ -40,8 +39,7 @@ impl VersionedMagic {
         let versioned_magic = match (magic, version) {
             (Magic::Gguf, 1) => Self::GgufV1,
             (Magic::Gguf, 2) => Self::GgufV2,
-            (Magic::Gguf, 3) => Self::GgufV3,
-            _ => crate::bail!("gguf: unsupported magic/version {magic:?}/{version}"),
+            _ => crate::bail!("ggml: unsupported magic/version {magic:?}/{version}"),
         };
         Ok(versioned_magic)
     }
@@ -86,9 +84,7 @@ pub struct Content {
 fn read_string<R: std::io::Read>(reader: &mut R, magic: &VersionedMagic) -> Result<String> {
     let len = match magic {
         VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
-        VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
-            reader.read_u64::<LittleEndian>()? as usize
-        }
+        VersionedMagic::GgufV2 => reader.read_u64::<LittleEndian>()? as usize,
     };
     let mut v = vec![0u8; len];
     reader.read_exact(&mut v)?;
@@ -288,9 +284,7 @@ impl Value {
                 let value_type = ValueType::from_u32(value_type)?;
                 let len = match magic {
                     VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
-                    VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
-                        reader.read_u64::<LittleEndian>()? as usize
-                    }
+                    VersionedMagic::GgufV2 => reader.read_u64::<LittleEndian>()? as usize,
                 };
                 let mut vs = Vec::with_capacity(len);
                 for _ in 0..len {
@@ -387,15 +381,11 @@ impl Content {
 
         let tensor_count = match magic {
             VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
-            VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
-                reader.read_u64::<LittleEndian>()? as usize
-            }
+            VersionedMagic::GgufV2 => reader.read_u64::<LittleEndian>()? as usize,
         };
         let metadata_kv_count = match magic {
             VersionedMagic::GgufV1 => reader.read_u32::<LittleEndian>()? as usize,
-            VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
-                reader.read_u64::<LittleEndian>()? as usize
-            }
+            VersionedMagic::GgufV2 => reader.read_u64::<LittleEndian>()? as usize,
         };
 
         let mut metadata = HashMap::new();
@@ -417,7 +407,7 @@ impl Content {
                     reader.read_u32_into::<LittleEndian>(&mut dimensions)?;
                     dimensions.into_iter().map(|c| c as usize).collect()
                 }
-                VersionedMagic::GgufV2 | VersionedMagic::GgufV3 => {
+                VersionedMagic::GgufV2 => {
                     let mut dimensions = vec![0; n_dimensions as usize];
                     reader.read_u64_into::<LittleEndian>(&mut dimensions)?;
                     dimensions.into_iter().map(|c| c as usize).collect()
@@ -463,7 +453,7 @@ impl Content {
     ) -> Result<QTensor> {
         let tensor_info = match self.tensor_infos.get(name) {
             Some(tensor_info) => tensor_info,
-            None => crate::bail!("cannot find tensor info for {name}"),
+            None => crate::bail!("cannot find tensor-infor for {name}"),
         };
         tensor_info.read(reader, self.tensor_data_offset)
     }

candle-core/src/quantized/k_quants.rs

@@ -236,9 +236,14 @@ impl GgmlType for BlockQ4_0 {
 
     fn vec_dot_unopt(n: usize, xs: &[Self], ys: &[Self::VecDotType]) -> Result<f32> {
         let qk = QK8_0;
+        let nb = n / qk;
         if n % QK8_0 != 0 {
             crate::bail!("vec_dot_q4_0_q8_0: {n} is not divisible by {qk}")
         }
+        if nb % 2 != 0 {
+            crate::bail!("vec_dot_q4_0_q8_0: {nb} is not even")
+        }
+
         // Generic implementation.
         let mut sumf = 0f32;
         for (xs, ys) in xs.iter().zip(ys.iter()) {

candle-core/src/quantized/mod.rs

@@ -307,8 +307,8 @@ impl crate::CustomOp1 for QTensor {
     }
 }
 
-impl crate::Module for QMatMul {
-    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+impl QMatMul {
+    pub fn forward(&self, xs: &Tensor) -> Result<Tensor> {
         match self {
             Self::QTensor(t) => xs.apply_op1_no_bwd(t.as_ref()),
             Self::Tensor(w) => {

candle-core/src/quantized/neon.rs

@@ -12,14 +12,6 @@ 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;
@@ -27,39 +19,71 @@ pub(crate) fn vec_dot_q4_0_q8_0(n: usize, xs: &[BlockQ4_0], ys: &[BlockQ8_0]) ->
     if n % QK8_0 != 0 {
         crate::bail!("vec_dot_q4_0_q8_0: {n} is not divisible by {qk}")
     }
+    if nb % 2 != 0 {
+        crate::bail!("vec_dot_q4_0_q8_0: {nb} is not even")
+    }
 
     unsafe {
         let mut sumv0 = vdupq_n_f32(0.0f32);
-        for i in 0..nb {
+        let mut sumv1 = vdupq_n_f32(0.0f32);
+        for i in (0..nb).step_by(2) {
             let x0 = &xs[i];
+            let x1 = &xs[i + 1];
             let y0 = &ys[i];
+            let y1 = &ys[i + 1];
 
             let m4b = vdupq_n_u8(0x0F);
             let s8b = vdupq_n_s8(0x8);
 
             let v0_0 = vld1q_u8(x0.qs.as_ptr());
+            let v0_1 = vld1q_u8(x1.qs.as_ptr());
 
             // 4-bit -> 8-bit
             let v0_0l = vreinterpretq_s8_u8(vandq_u8(v0_0, m4b));
             let v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            let v0_1l = vreinterpretq_s8_u8(vandq_u8(v0_1, m4b));
+            let v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
 
             // sub 8
             let v0_0ls = vsubq_s8(v0_0l, s8b);
             let v0_0hs = vsubq_s8(v0_0h, s8b);
+            let v0_1ls = vsubq_s8(v0_1l, s8b);
+            let v0_1hs = vsubq_s8(v0_1h, s8b);
 
             // load y
             let v1_0l = vld1q_s8(y0.qs.as_ptr());
             let v1_0h = vld1q_s8(y0.qs.as_ptr().add(16));
+            let v1_1l = vld1q_s8(y1.qs.as_ptr());
+            let v1_1h = vld1q_s8(y1.qs.as_ptr().add(16));
+
+            // TODO: Support dotprod when it's available outside of nightly.
+            let pl0l = vmull_s8(vget_low_s8(v0_0ls), vget_low_s8(v1_0l));
+            let pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l));
+            let ph0l = vmull_s8(vget_low_s8(v0_0hs), vget_low_s8(v1_0h));
+            let ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h));
+
+            let pl1l = vmull_s8(vget_low_s8(v0_1ls), vget_low_s8(v1_1l));
+            let pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l));
+            let ph1l = vmull_s8(vget_low_s8(v0_1hs), vget_low_s8(v1_1h));
+            let ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h));
+
+            let pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
+            let ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
+            let pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
+            let ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
 
-            let pl0 = vdotq_s32(v0_0ls, v1_0l);
-            let ph0 = vdotq_s32(v0_0hs, v1_0h);
             sumv0 = vmlaq_n_f32(
                 sumv0,
                 vcvtq_f32_s32(vaddq_s32(pl0, ph0)),
                 x0.d.to_f32() * y0.d.to_f32(),
             );
+            sumv1 = vmlaq_n_f32(
+                sumv1,
+                vcvtq_f32_s32(vaddq_s32(pl1, ph1)),
+                x1.d.to_f32() * y1.d.to_f32(),
+            );
         }
-        Ok(vaddvq_f32(sumv0))
+        Ok(vaddvq_f32(sumv0) + vaddvq_f32(sumv1))
     }
 }
 
@@ -70,29 +94,57 @@ pub(crate) fn vec_dot_q8_0_q8_0(n: usize, xs: &[BlockQ8_0], ys: &[BlockQ8_0]) ->
         crate::bail!("vec_dot_q8_0_q8_0: {n} is not divisible by {qk}")
     }
     let nb = n / QK8_0;
+    if nb % 2 != 0 {
+        crate::bail!("vec_dot_q8_0_q8_0: {nb} is not even")
+    }
     unsafe {
         let mut sumv0 = vdupq_n_f32(0.0f32);
-        for i in 0..nb {
+        let mut sumv1 = vdupq_n_f32(0.0f32);
+        for i in (0..nb).step_by(2) {
             let x0 = &xs[i];
+            let x1 = &xs[i + 1];
             let y0 = &ys[i];
+            let y1 = &ys[i + 1];
 
             let x0_0 = vld1q_s8(x0.qs.as_ptr());
             let x0_1 = vld1q_s8(x0.qs.as_ptr().add(16));
+            let x1_0 = vld1q_s8(x1.qs.as_ptr());
+            let x1_1 = vld1q_s8(x1.qs.as_ptr().add(16));
 
             // load y
             let y0_0 = vld1q_s8(y0.qs.as_ptr());
             let y0_1 = vld1q_s8(y0.qs.as_ptr().add(16));
+            let y1_0 = vld1q_s8(y1.qs.as_ptr());
+            let y1_1 = vld1q_s8(y1.qs.as_ptr().add(16));
 
-            let p0 = vdotq_s32(x0_0, y0_0);
-            let p1 = vdotq_s32(x0_1, y0_1);
+            // TODO dotprod once this is the intrinsics are.
+            let p0_0 = vmull_s8(vget_low_s8(x0_0), vget_low_s8(y0_0));
+            let p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0));
+            let p0_2 = vmull_s8(vget_low_s8(x0_1), vget_low_s8(y0_1));
+            let p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1));
+
+            let p1_0 = vmull_s8(vget_low_s8(x1_0), vget_low_s8(y1_0));
+            let p1_1 = vmull_s8(vget_high_s8(x1_0), vget_high_s8(y1_0));
+            let p1_2 = vmull_s8(vget_low_s8(x1_1), vget_low_s8(y1_1));
+            let p1_3 = vmull_s8(vget_high_s8(x1_1), vget_high_s8(y1_1));
+
+            let p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1));
+            let p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3));
+            let p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1));
+            let p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3));
 
             sumv0 = vmlaq_n_f32(
                 sumv0,
                 vcvtq_f32_s32(vaddq_s32(p0, p1)),
                 x0.d.to_f32() * y0.d.to_f32(),
             );
+            sumv1 = vmlaq_n_f32(
+                sumv1,
+                vcvtq_f32_s32(vaddq_s32(p2, p3)),
+                x1.d.to_f32() * y1.d.to_f32(),
+            );
         }
-        Ok(vaddvq_f32(sumv0))
+        Ok(vaddvq_f32(sumv0) + vaddvq_f32(sumv1))
     }
 }
 
@@ -113,7 +165,10 @@ 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 = vdotq_s32(xs, ys);
+                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));
                 sum_i = vaddq_s32(sum_i, xy)
             }
             sumf += vaddvq_s32(sum_i) as f32 * scale
@@ -183,16 +238,30 @@ 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));
 
-                let p0 = vdotq_s32(q6bytes_0, q8bytes.0);
-                let p1 = vdotq_s32(q6bytes_1, q8bytes.1);
+                // 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 (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
-                isum += vaddvq_s32(p0) * scale0 + vaddvq_s32(p1) * scale1;
+                isum += vaddvq_s16(p0) as i32 * scale0 + vaddvq_s16(p1) as i32 * scale1;
                 scale = scale.add(2);
 
-                let p2 = vdotq_s32(q6bytes_2, q8bytes.2);
-                let p3 = vdotq_s32(q6bytes_3, q8bytes.3);
+                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 (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
-                isum += vaddvq_s32(p2) * scale0 + vaddvq_s32(p3) * scale1;
+                isum += vaddvq_s16(p2) as i32 * scale0 + vaddvq_s16(p3) as i32 * scale1;
                 scale = scale.add(2);
 
                 let q8bytes = vld1q_s8_x4(q8);
@@ -212,16 +281,29 @@ 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));
 
-                let p0 = vdotq_s32(q6bytes_0, q8bytes.0);
-                let p1 = vdotq_s32(q6bytes_1, q8bytes.1);
+                // 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 (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
-                isum += vaddvq_s32(p0) * scale0 + vaddvq_s32(p1) * scale1;
+                isum += vaddvq_s16(p0) as i32 * scale0 + vaddvq_s16(p1) as i32 * scale1;
                 scale = scale.add(2);
 
-                let p2 = vdotq_s32(q6bytes_2, q8bytes.2);
-                let p3 = vdotq_s32(q6bytes_3, q8bytes.3);
+                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 (scale0, scale1) = (*scale as i32, *scale.add(1) as i32);
-                isum += vaddvq_s32(p2) * scale0 + vaddvq_s32(p3) * scale1;
+                isum += vaddvq_s16(p2) as i32 * scale0 + vaddvq_s16(p3) as i32 * scale1;
                 scale = scale.add(2);
             }
             sum += d_all * y.d * ((isum - 32 * isum_mins) as f32);
@@ -298,14 +380,28 @@ 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));
 
-                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;
+                // 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;
                 scales = scales.add(1);
 
-                let p2 = vdotq_s32(q5bytes_2, q8bytes.2);
-                let p3 = vdotq_s32(q5bytes_3, q8bytes.3);
-                sumi += vaddvq_s32(vaddq_s32(p2, p3)) * *scales as i32;
+                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;
                 scales = scales.add(1);
             }
             sumf += d * sumi as f32 - dmin * sumi_mins as f32;
@@ -368,15 +464,22 @@ 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 = 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 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 q8bytes = vld1q_s8_x2(q8);
                 q8 = q8.add(32);
@@ -384,9 +487,15 @@ 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 = 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;
+                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;
             }
             sumf += d * (sumi1 + sumi2) as f32;
         }
@@ -464,14 +573,27 @@ pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Res
                     vreinterpretq_s8_u8(q3h_3),
                 );
 
-                let p0 = vdotq_s32(q3bytes_0, q8bytes_1.0);
-                let p1 = vdotq_s32(q3bytes_1, q8bytes_1.1);
-                let p2 = vdotq_s32(q3bytes_2, q8bytes_1.2);
-                let p3 = vdotq_s32(q3bytes_3, q8bytes_1.3);
-                isum += vaddvq_s32(p0) * *scale as i32
-                    + vaddvq_s32(p1) * *scale.add(1) as i32
-                    + vaddvq_s32(p2) * *scale.add(2) as i32
-                    + vaddvq_s32(p3) * *scale.add(3) as i32;
+                // 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;
                 scale = scale.add(4);
 
                 let q3h_0 = vbicq_u8(m2, qhbits.0);
@@ -496,14 +618,27 @@ pub(crate) fn vec_dot_q3k_q8k(n: usize, xs: &[BlockQ3K], ys: &[BlockQ8K]) -> Res
                     vreinterpretq_s8_u8(q3h_3),
                 );
 
-                let p0 = vdotq_s32(q3bytes_0, q8bytes_2.0);
-                let p1 = vdotq_s32(q3bytes_1, q8bytes_2.1);
-                let p2 = vdotq_s32(q3bytes_2, q8bytes_2.2);
-                let p3 = vdotq_s32(q3bytes_3, q8bytes_2.3);
-                isum += vaddvq_s32(p0) * *scale as i32
-                    + vaddvq_s32(p1) * *scale.add(1) as i32
-                    + vaddvq_s32(p2) * *scale.add(2) as i32
-                    + vaddvq_s32(p3) * *scale.add(3) as i32;
+                // 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;
                 scale = scale.add(4);
 
                 if j == 0 {
@@ -561,6 +696,7 @@ 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);
@@ -607,7 +743,14 @@ unsafe fn multiply_accum_with_scale(
     q2bytes: int8x16x2_t,
     q8bytes: int8x16x2_t,
 ) -> i32 {
-    let p1 = vdotq_s32(q2bytes.0, q8bytes.0);
-    let p2 = vdotq_s32(q2bytes.1, q8bytes.1);
-    vaddvq_s32(p1) * aux[is + index] as i32 + vaddvq_s32(p2) * aux[is + 1 + index] as i32
+    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
 }

candle-core/src/quantized/simd128.rs

@@ -11,6 +11,10 @@ pub(crate) fn vec_dot_q4_0_q8_0(n: usize, xs: &[BlockQ4_0], ys: &[BlockQ8_0]) ->
     if n % QK8_0 != 0 {
         crate::bail!("vec_dot_q4_0_q8_0: {n} is not divisible by {qk}")
     }
+    let nb = n / QK8_0;
+    if nb % 2 != 0 {
+        crate::bail!("vec_dot_q4_0_q8_0: {nb} is not even")
+    }
     unsafe {
         let mut acc = f32x4_splat(0.0f32);
         for (x, y) in xs.iter().zip(ys.iter()) {
@@ -57,6 +61,10 @@ pub(crate) fn vec_dot_q8_0_q8_0(n: usize, xs: &[BlockQ8_0], ys: &[BlockQ8_0]) ->
     if n % QK8_0 != 0 {
         crate::bail!("vec_dot_q8_0_q8_0: {n} is not divisible by {qk}")
     }
+    let nb = n / QK8_0;
+    if nb % 2 != 0 {
+        crate::bail!("vec_dot_q8_0_q8_0: {nb} is not even")
+    }
     unsafe {
         let mut acc = f32x4_splat(0.0f32);
         for (x, y) in xs.iter().zip(ys.iter()) {

candle-core/src/shape.rs

@@ -203,7 +203,7 @@ impl Shape {
 
     /// Check whether the two shapes are compatible for broadcast, and if it is the case return the
     /// broadcasted shape. This is to be used for binary pointwise ops.
-    pub fn broadcast_shape_binary_op(&self, rhs: &Self, op: &'static str) -> Result<Shape> {
+    pub(crate) fn broadcast_shape_binary_op(&self, rhs: &Self, op: &'static str) -> Result<Shape> {
         let lhs = self;
         let lhs_dims = lhs.dims();
         let rhs_dims = rhs.dims();
@@ -478,139 +478,6 @@ extract_dims!(
     (usize, usize, usize, usize, usize)
 );
 
-pub trait ShapeWithOneHole {
-    fn into_shape(self, el_count: usize) -> Result<Shape>;
-}
-
-impl<S: Into<Shape>> ShapeWithOneHole for S {
-    fn into_shape(self, _el_count: usize) -> Result<Shape> {
-        Ok(self.into())
-    }
-}
-
-impl ShapeWithOneHole for ((),) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        Ok(el_count.into())
-    }
-}
-
-fn hole_size(el_count: usize, prod_d: usize, s: &dyn std::fmt::Debug) -> Result<usize> {
-    if prod_d == 0 {
-        crate::bail!("cannot reshape tensor of {el_count} elements to {s:?}")
-    }
-    if el_count % prod_d != 0 {
-        crate::bail!("cannot reshape tensor with {el_count} elements to {s:?}")
-    }
-    Ok(el_count / prod_d)
-}
-
-impl ShapeWithOneHole for ((), usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let ((), d1) = self;
-        Ok((hole_size(el_count, d1, &self)?, d1).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, ()) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, ()) = self;
-        Ok((d1, hole_size(el_count, d1, &self)?).into())
-    }
-}
-
-impl ShapeWithOneHole for ((), usize, usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let ((), d1, d2) = self;
-        Ok((hole_size(el_count, d1 * d2, &self)?, d1, d2).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, (), usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, (), d2) = self;
-        Ok((d1, hole_size(el_count, d1 * d2, &self)?, d2).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, usize, ()) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, d2, ()) = self;
-        Ok((d1, d2, hole_size(el_count, d1 * d2, &self)?).into())
-    }
-}
-
-impl ShapeWithOneHole for ((), usize, usize, usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let ((), d1, d2, d3) = self;
-        let d = hole_size(el_count, d1 * d2 * d3, &self)?;
-        Ok((d, d1, d2, d3).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, (), usize, usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, (), d2, d3) = self;
-        let d = hole_size(el_count, d1 * d2 * d3, &self)?;
-        Ok((d1, d, d2, d3).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, usize, (), usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, d2, (), d3) = self;
-        let d = hole_size(el_count, d1 * d2 * d3, &self)?;
-        Ok((d1, d2, d, d3).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, usize, usize, ()) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, d2, d3, ()) = self;
-        let d = hole_size(el_count, d1 * d2 * d3, &self)?;
-        Ok((d1, d2, d3, d).into())
-    }
-}
-
-impl ShapeWithOneHole for ((), usize, usize, usize, usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let ((), d1, d2, d3, d4) = self;
-        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
-        Ok((d, d1, d2, d3, d4).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, (), usize, usize, usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, (), d2, d3, d4) = self;
-        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
-        Ok((d1, d, d2, d3, d4).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, usize, (), usize, usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, d2, (), d3, d4) = self;
-        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
-        Ok((d1, d2, d, d3, d4).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, usize, usize, (), usize) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, d2, d3, (), d4) = self;
-        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
-        Ok((d1, d2, d3, d, d4).into())
-    }
-}
-
-impl ShapeWithOneHole for (usize, usize, usize, usize, ()) {
-    fn into_shape(self, el_count: usize) -> Result<Shape> {
-        let (d1, d2, d3, d4, ()) = self;
-        let d = hole_size(el_count, d1 * d2 * d3 * d4, &self)?;
-        Ok((d1, d2, d3, d4, d).into())
-    }
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -627,3 +494,171 @@ mod tests {
         assert_eq!(shape.stride_contiguous(), [458 * 792, 458, 1]);
     }
 }
+
+pub trait ShapeWithOneHole {
+    fn into_shape(self, el_count: usize) -> Result<Shape>;
+}
+
+impl<S: Into<Shape>> ShapeWithOneHole for S {
+    fn into_shape(self, _el_count: usize) -> Result<Shape> {
+        Ok(self.into())
+    }
+}
+
+impl ShapeWithOneHole for ((),) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        Ok(el_count.into())
+    }
+}
+
+impl ShapeWithOneHole for ((), usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let ((), d1) = self;
+        if el_count % d1 != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d1}")
+        }
+        Ok((el_count / d1, d1).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, ()) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, ()) = self;
+        if el_count % d1 != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d1}")
+        }
+        Ok((d1, el_count / d1).into())
+    }
+}
+
+impl ShapeWithOneHole for ((), usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let ((), d1, d2) = self;
+        let d = d1 * d2;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((el_count / d, d1, d2).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, (), usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, (), d2) = self;
+        let d = d1 * d2;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((d1, el_count / d, d2).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, ()) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, ()) = self;
+        let d = d1 * d2;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((d1, d2, el_count / d).into())
+    }
+}
+
+impl ShapeWithOneHole for ((), usize, usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let ((), d1, d2, d3) = self;
+        let d = d1 * d2 * d3;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((el_count / d, d1, d2, d3).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, (), usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, (), d2, d3) = self;
+        let d = d1 * d2 * d3;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((d1, el_count / d, d2, d3).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, (), usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, (), d3) = self;
+        let d = d1 * d2 * d3;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((d1, d2, el_count / d, d3).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, usize, ()) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, d3, ()) = self;
+        let d = d1 * d2 * d3;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((d1, d2, d3, el_count / d).into())
+    }
+}
+
+impl ShapeWithOneHole for ((), usize, usize, usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let ((), d1, d2, d3, d4) = self;
+        let d = d1 * d2 * d3 * d4;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((el_count / d, d1, d2, d3, d4).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, (), usize, usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, (), d2, d3, d4) = self;
+        let d = d1 * d2 * d3 * d4;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((d1, el_count / d, d2, d3, d4).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, (), usize, usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, (), d3, d4) = self;
+        let d = d1 * d2 * d3 * d4;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((d1, d2, el_count / d, d3, d4).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, usize, (), usize) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, d3, (), d4) = self;
+        let d = d1 * d2 * d3 * d4;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((d1, d2, d3, el_count / d, d4).into())
+    }
+}
+
+impl ShapeWithOneHole for (usize, usize, usize, usize, ()) {
+    fn into_shape(self, el_count: usize) -> Result<Shape> {
+        let (d1, d2, d3, d4, ()) = self;
+        let d = d1 * d2 * d3 * d4;
+        if el_count % d != 0 {
+            crate::bail!("tensor number of elements {el_count} is not divisible by {d}")
+        }
+        Ok((d1, d2, d3, d4, el_count / d).into())
+    }
+}

candle-core/src/storage.rs

@@ -1,6 +1,6 @@
 use crate::backend::BackendStorage;
 use crate::op::{self, CmpOp, CustomOp1, CustomOp2, CustomOp3, ReduceOp};
-use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, MetalStorage, Result, Shape};
+use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, Result, Shape};
 
 // We do not want to implement Clone on Storage as cloning may fail because of
 // out of memory. Instead try_clone should be used.
@@ -8,7 +8,6 @@ use crate::{CpuStorage, CudaStorage, DType, Device, Error, Layout, MetalStorage,
 pub enum Storage {
     Cpu(CpuStorage),
     Cuda(CudaStorage),
-    Metal(MetalStorage),
 }
 
 impl Storage {
@@ -19,10 +18,6 @@ impl Storage {
                 let storage = storage.try_clone(layout)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.try_clone(layout)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -30,7 +25,6 @@ impl Storage {
         match self {
             Self::Cpu(_) => Device::Cpu,
             Self::Cuda(storage) => Device::Cuda(storage.device().clone()),
-            Self::Metal(storage) => Device::Metal(storage.device().clone()),
         }
     }
 
@@ -38,7 +32,6 @@ impl Storage {
         match self {
             Self::Cpu(storage) => storage.dtype(),
             Self::Cuda(storage) => storage.dtype(),
-            Self::Metal(storage) => storage.dtype(),
         }
     }
 
@@ -72,10 +65,6 @@ impl Storage {
                 let storage = storage.affine(layout, mul, add)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.affine(layout, mul, add)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -89,10 +78,6 @@ impl Storage {
                 let storage = storage.powf(layout, alpha)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.powf(layout, alpha)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -106,10 +91,6 @@ impl Storage {
                 let storage = storage.elu(layout, alpha)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.elu(layout, alpha)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -131,10 +112,6 @@ impl Storage {
                 let storage = lhs.cmp(op, rhs, lhs_layout, rhs_layout)?;
                 Ok(Self::Cuda(storage))
             }
-            (Self::Metal(lhs), Self::Metal(rhs)) => {
-                let storage = lhs.cmp(op, rhs, lhs_layout, rhs_layout)?;
-                Ok(Self::Metal(storage))
-            }
             (lhs, rhs) => {
                 // Should not happen because of the same device check above but we're defensive
                 // anyway.
@@ -158,10 +135,6 @@ impl Storage {
                 let storage = storage.reduce_op(op, layout, s)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.reduce_op(op, layout, s)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -175,10 +148,6 @@ impl Storage {
                 let storage = storage.to_dtype(layout, dtype)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.to_dtype(layout, dtype)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -192,10 +161,6 @@ impl Storage {
                 let (storage, shape) = c.cuda_fwd(storage, l)?;
                 Ok((Self::Cuda(storage), shape))
             }
-            Self::Metal(storage) => {
-                let (storage, shape) = c.metal_fwd(storage, l)?;
-                Ok((Self::Metal(storage), shape))
-            }
         }
     }
 
@@ -216,10 +181,6 @@ impl Storage {
                 let (s, shape) = c.cuda_fwd(s1, l1, s2, l2)?;
                 Ok((Self::Cuda(s), shape))
             }
-            (Self::Metal(s1), Self::Metal(s2)) => {
-                let (s, shape) = c.metal_fwd(s1, l1, s2, l2)?;
-                Ok((Self::Metal(s), shape))
-            }
             _ => unreachable!(),
         }
     }
@@ -244,10 +205,6 @@ impl Storage {
                 let (s, shape) = c.cuda_fwd(s1, l1, s2, l2, s3, l3)?;
                 Ok((Self::Cuda(s), shape))
             }
-            (Self::Metal(s1), Self::Metal(s2), Self::Metal(s3)) => {
-                let (s, shape) = c.metal_fwd(s1, l1, s2, l2, s3, l3)?;
-                Ok((Self::Metal(s), shape))
-            }
             _ => unreachable!(),
         }
     }
@@ -262,10 +219,6 @@ impl Storage {
                 let storage = storage.unary_impl::<B>(layout)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.unary_impl::<B>(layout)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -286,10 +239,6 @@ impl Storage {
                 let storage = lhs.binary_impl::<B>(rhs, lhs_layout, rhs_layout)?;
                 Ok(Self::Cuda(storage))
             }
-            (Self::Metal(lhs), Self::Metal(rhs)) => {
-                let storage = lhs.binary_impl::<B>(rhs, lhs_layout, rhs_layout)?;
-                Ok(Self::Metal(storage))
-            }
             (lhs, rhs) => {
                 // Should not happen because of the same device check above but we're defensive
                 // anyway.
@@ -321,10 +270,6 @@ impl Storage {
                 let s = inp.conv1d(l, kernel, kernel_l, params)?;
                 Ok(Self::Cuda(s))
             }
-            (Storage::Metal(inp), Storage::Metal(kernel)) => {
-                let s = inp.conv1d(l, kernel, kernel_l, params)?;
-                Ok(Self::Metal(s))
-            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -334,33 +279,6 @@ impl Storage {
         }
     }
 
-    pub(crate) fn conv_transpose1d(
-        &self,
-        l: &Layout,
-        kernel: &Self,
-        kernel_l: &Layout,
-        params: &crate::conv::ParamsConvTranspose1D,
-    ) -> Result<Self> {
-        self.same_device(kernel, "conv-transpose1d")?;
-        self.same_dtype(kernel, "conv-transpose1d")?;
-        match (self, &kernel) {
-            (Storage::Cpu(inp), Storage::Cpu(kernel)) => {
-                let s = inp.conv_transpose1d(l, kernel, kernel_l, params)?;
-                Ok(Self::Cpu(s))
-            }
-            (Storage::Cuda(inp), Storage::Cuda(kernel)) => {
-                let s = inp.conv_transpose1d(l, kernel, kernel_l, params)?;
-                Ok(Self::Cuda(s))
-            }
-            (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
-                lhs: lhs.device().location(),
-                rhs: rhs.device().location(),
-                op: "conv-transpose1d",
-            }
-            .bt()),
-        }
-    }
-
     pub(crate) fn conv2d(
         &self,
         l: &Layout,
@@ -379,10 +297,6 @@ impl Storage {
                 let s = inp.conv2d(l, kernel, kernel_l, params)?;
                 Ok(Self::Cuda(s))
             }
-            (Storage::Metal(inp), Storage::Metal(kernel)) => {
-                let s = inp.conv2d(l, kernel, kernel_l, params)?;
-                Ok(Self::Metal(s))
-            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -410,10 +324,6 @@ impl Storage {
                 let s = inp.conv_transpose2d(l, kernel, kernel_l, params)?;
                 Ok(Self::Cuda(s))
             }
-            (Storage::Metal(inp), Storage::Metal(kernel)) => {
-                let s = inp.conv_transpose2d(l, kernel, kernel_l, params)?;
-                Ok(Self::Metal(s))
-            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -438,10 +348,6 @@ impl Storage {
                 let storage = storage.avg_pool2d(layout, kernel_size, stride)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.avg_pool2d(layout, kernel_size, stride)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -460,10 +366,6 @@ impl Storage {
                 let storage = storage.max_pool2d(layout, kernel_size, stride)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.max_pool2d(layout, kernel_size, stride)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -477,10 +379,6 @@ impl Storage {
                 let storage = storage.upsample_nearest1d(layout, sz)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.upsample_nearest1d(layout, sz)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -494,10 +392,6 @@ impl Storage {
                 let storage = storage.upsample_nearest2d(layout, h, w)?;
                 Ok(Self::Cuda(storage))
             }
-            Self::Metal(storage) => {
-                let storage = storage.upsample_nearest2d(layout, h, w)?;
-                Ok(Self::Metal(storage))
-            }
         }
     }
 
@@ -521,10 +415,6 @@ impl Storage {
                 let storage = cond.where_cond(layout, t, layout_t, f, layout_f)?;
                 Ok(Self::Cuda(storage))
             }
-            (Self::Metal(cond), Self::Metal(t), Self::Metal(f)) => {
-                let storage = cond.where_cond(layout, t, layout_t, f, layout_f)?;
-                Ok(Self::Metal(storage))
-            }
             (_, lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -551,10 +441,6 @@ impl Storage {
                 let storage = s.gather(l, indexes, indexes_l, d)?;
                 Ok(Self::Cuda(storage))
             }
-            (Self::Metal(s), Self::Metal(indexes)) => {
-                let storage = s.gather(l, indexes, indexes_l, d)?;
-                Ok(Self::Metal(storage))
-            }
             _ => unreachable!(),
         }
     }
@@ -579,10 +465,6 @@ impl Storage {
                 let storage = s.scatter_add(l, indexes, indexes_l, source, source_l, d)?;
                 Ok(Self::Cuda(storage))
             }
-            (Self::Metal(s), Self::Metal(indexes), Self::Metal(source)) => {
-                let storage = s.scatter_add(l, indexes, indexes_l, source, source_l, d)?;
-                Ok(Self::Metal(storage))
-            }
             _ => unreachable!(),
         }
     }
@@ -607,10 +489,6 @@ impl Storage {
                 let storage = s.index_add(l, indexes, indexes_l, source, source_l, d)?;
                 Ok(Self::Cuda(storage))
             }
-            (Self::Metal(s), Self::Metal(indexes), Self::Metal(source)) => {
-                let storage = s.index_add(l, indexes, indexes_l, source, source_l, d)?;
-                Ok(Self::Metal(storage))
-            }
             _ => unreachable!(),
         }
     }
@@ -632,10 +510,6 @@ impl Storage {
                 let storage = lhs.index_select(rhs, lhs_l, rhs_l, d)?;
                 Ok(Self::Cuda(storage))
             }
-            (Self::Metal(lhs), Self::Metal(rhs)) => {
-                let storage = lhs.index_select(rhs, lhs_l, rhs_l, d)?;
-                Ok(Self::Metal(storage))
-            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -663,10 +537,6 @@ impl Storage {
                 let storage = lhs.matmul(rhs, bmnk, lhs_layout, rhs_layout)?;
                 Ok(Self::Cuda(storage))
             }
-            (Self::Metal(lhs), Self::Metal(rhs)) => {
-                let storage = lhs.matmul(rhs, bmnk, lhs_layout, rhs_layout)?;
-                Ok(Self::Metal(storage))
-            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),
@@ -686,9 +556,6 @@ impl Storage {
         match (self, dst) {
             (Self::Cpu(src), Self::Cpu(dst)) => src.copy_strided_src(dst, dst_offset, src_l),
             (Self::Cuda(src), Self::Cuda(dst)) => Ok(src.copy_strided_src(dst, dst_offset, src_l)?),
-            (Self::Metal(src), Self::Metal(dst)) => {
-                Ok(src.copy_strided_src(dst, dst_offset, src_l)?)
-            }
             (lhs, rhs) => Err(Error::DeviceMismatchBinaryOp {
                 lhs: lhs.device().location(),
                 rhs: rhs.device().location(),

candle-core/src/tensor.rs

@@ -1,4 +1,4 @@
-//! Tensors are N-dimensional matrixes of elements using a single data type.
+//! Tensors are N-dimenional matrixes of elements using a single data type.
 #![allow(clippy::redundant_closure_call)]
 use crate::backend::{BackendDevice, BackendStorage};
 use crate::op::{
@@ -6,7 +6,7 @@ use crate::op::{
 };
 use crate::scalar::TensorOrScalar;
 use crate::shape::{Dim, Dims};
-use crate::{bail, storage::Storage, DType, Device, Error, Layout, Result, Shape};
+use crate::{storage::Storage, DType, Device, Error, Layout, Result, Shape};
 use std::sync::{Arc, RwLock};
 
 /// Unique identifier for tensors.
@@ -361,16 +361,6 @@ 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>,
@@ -395,21 +385,11 @@ impl Tensor {
         step: D,
         device: &Device,
     ) -> Result<Self> {
-        if D::is_zero(&step) {
-            bail!("step cannot be zero")
-        }
         let mut data = vec![];
         let mut current = start;
-        if step >= D::zero() {
-            while current < end {
-                data.push(current);
-                current += step;
-            }
-        } else {
-            while current > end {
-                data.push(current);
-                current += step;
-            }
+        while current < end {
+            data.push(current);
+            current += step;
         }
         let len = data.len();
         Self::from_vec_impl(data, len, device, false)
@@ -469,7 +449,7 @@ impl Tensor {
 
     /// Returns true if the computation graph should track this op, that is if it is
     /// a variable or if it has some variable as dependencies.
-    pub fn track_op(&self) -> bool {
+    pub(crate) fn track_op(&self) -> bool {
         self.is_variable || self.op.is_some()
     }
 
@@ -487,12 +467,6 @@ impl Tensor {
     broadcast_binary_op!(broadcast_div, div);
     broadcast_binary_op!(broadcast_maximum, maximum);
     broadcast_binary_op!(broadcast_minimum, minimum);
-    broadcast_binary_op!(broadcast_eq, eq);
-    broadcast_binary_op!(broadcast_ne, ne);
-    broadcast_binary_op!(broadcast_lt, lt);
-    broadcast_binary_op!(broadcast_le, le);
-    broadcast_binary_op!(broadcast_gt, gt);
-    broadcast_binary_op!(broadcast_ge, ge);
 
     unary_op!(recip, Recip);
     unary_op!(neg, Neg);
@@ -539,7 +513,6 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(cpu_storage) => from_cpu_storage(cpu_storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
-            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -567,73 +540,6 @@ impl Tensor {
         Ok(inp)
     }
 
-    /// Creates grids of coordinates specified by the 1D inputs.
-    ///
-    /// # Arguments
-    ///
-    /// * `args` - A slice of 1D tensors.
-    /// * `xy_indexing` - Whether to use xy indexing or ij indexing. If xy is selected, the
-    /// first dimension corresponds to the cardinality of the second input and the second
-    /// dimension corresponds to the cardinality of the first input. If ij is selected, the
-    /// dimensions are in the same order as the cardinality of the inputs.
-    ///
-    /// # Examples
-    ///
-    /// ```rust
-    /// use candle_core::{Tensor, Device, Shape};
-    /// let x = Tensor::new(&[1f32, 2., 3.], &Device::Cpu)?;
-    /// let y = Tensor::new(&[4f32, 5., 6.], &Device::Cpu)?;
-    ///
-    /// let grids_xy = Tensor::meshgrid(&[&x, &y], true)?;
-    ///
-    /// assert_eq!(grids_xy.len(), 2);
-    /// assert_eq!(grids_xy[0].dims(), &[3, 3]);
-    ///
-    /// assert_eq!(grids_xy[0].to_vec2::<f32>()?, &[[1., 2., 3.], [1., 2., 3.], [1., 2., 3.]]);
-    /// assert_eq!(grids_xy[1].to_vec2::<f32>()?, &[[4., 4., 4.], [5., 5., 5.], [6., 6., 6.]]);
-    ///
-    /// let grids_ij = Tensor::meshgrid(&[&x, &y], false)?;
-    ///
-    /// assert_eq!(grids_ij[0].to_vec2::<f32>()?, &[[1., 1., 1.], [2., 2., 2.], [3., 3., 3.]]);
-    /// assert_eq!(grids_ij[1].to_vec2::<f32>()?, &[[4., 5., 6.], [4., 5., 6.], [4., 5., 6.]]);
-    /// # Ok::<(), candle_core::Error>(())
-    /// ```
-    ///
-    /// # Errors
-    ///
-    /// * Will return `Err` if `args` contains less than 2 tensors.
-    ///
-    pub fn meshgrid<A: AsRef<Tensor>>(args: &[A], xy_indexing: bool) -> Result<Vec<Self>> {
-        if args.len() <= 1 {
-            Err(Error::OpRequiresAtLeastTwoTensors { op: "meshgrid" }.bt())?
-        }
-        let args: Vec<_> = if xy_indexing {
-            args.iter().rev().collect()
-        } else {
-            args.iter().collect()
-        };
-
-        let mut shape = Vec::with_capacity(args.len());
-        for arg in args.iter() {
-            shape.push(arg.as_ref().dims1()?)
-        }
-
-        let mut grids = Vec::with_capacity(args.len());
-        for idx in 0..args.len() {
-            let mut ones = vec![1usize; args.len()];
-            ones[idx] = shape[idx];
-            let arg = args[idx].as_ref().reshape(ones)?;
-            let mut repeats = shape.clone();
-            repeats[idx] = 1;
-            let repeated_tensor = arg.repeat(repeats)?;
-            grids.push(repeated_tensor);
-        }
-        if xy_indexing {
-            grids.reverse();
-        }
-        Ok(grids)
-    }
-
     /// This operation multiplies the input tensor by `mul` then adds `add` and return the result.
     /// The input values `mul` and `add` are casted to the appropriate type so some rounding might
     /// be performed.
@@ -679,7 +585,7 @@ impl Tensor {
     }
 
     /// Split a tensor into the specified number of chunks, this may return less chunks than
-    /// specified.
+    /// specificed.
     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)?;
@@ -709,23 +615,15 @@ impl Tensor {
     pub fn narrow<D: Dim>(&self, dim: D, start: usize, len: usize) -> Result<Self> {
         let dims = self.dims();
         let dim = dim.to_index(self.shape(), "narrow")?;
-        let err = |msg| {
-            Err::<(), _>(
-                Error::NarrowInvalidArgs {
-                    shape: self.shape().clone(),
-                    dim,
-                    start,
-                    len,
-                    msg,
-                }
-                .bt(),
-            )
-        };
-        if start > dims[dim] {
-            err("start > dim_len")?
-        }
-        if start.saturating_add(len) > dims[dim] {
-            err("start + len > dim_len")?
+        if start + len > dims[dim] {
+            Err(Error::NarrowInvalidArgs {
+                shape: self.shape().clone(),
+                dim,
+                start,
+                len,
+                msg: "start + len > dim_len",
+            }
+            .bt())?
         }
         if start == 0 && dims[dim] == len {
             Ok(self.clone())
@@ -866,20 +764,6 @@ impl Tensor {
         self.sum_impl(mean_dims, false)? * scale
     }
 
-    /// Returns the unbiased variance over the selected dimension.
-    pub fn var_keepdim<D: Dim>(&self, dim: D) -> Result<Self> {
-        let dim = dim.to_index(self.shape(), "var")?;
-        let mean = self.mean_keepdim(dim)?;
-        let squares = self.broadcast_sub(&mean)?.sqr()?;
-        squares.sum_impl(dim, true)? / (self.dim(dim)? - 1) as f64
-    }
-
-    /// Returns the unbiased variance over the selected dimension.
-    pub fn var<D: Dim>(&self, dim: D) -> Result<Self> {
-        let dim = dim.to_index(self.shape(), "var")?;
-        self.var_keepdim(dim)?.squeeze(dim)
-    }
-
     /// Gathers the maximum value across the selected dimension. The resulting shape has the same
     /// number of dimensions as the original tensor and the select dimension has a single element.
     pub fn max_keepdim<D: Dim>(&self, dim: D) -> Result<Self> {
@@ -1004,11 +888,7 @@ 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, |arg| Op::UpsampleNearest2D {
-            arg,
-            target_h,
-            target_w,
-        });
+        let op = BackpropOp::new1(self, Op::UpsampleNearest2D);
         let storage = self
             .storage()
             .upsample_nearest2d(self.layout(), target_h, target_w)?;
@@ -1041,9 +921,6 @@ 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;
@@ -1079,9 +956,6 @@ 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;
@@ -1237,16 +1111,14 @@ impl Tensor {
                 op: "scatter-add (self, src)",
                 lhs: self.shape().clone(),
                 rhs: source.shape().clone(),
-            }
-            .bt())?
+            })?
         }
         if indexes.dims() != source.dims() {
             Err(Error::ShapeMismatchBinaryOp {
                 op: "scatter-add (indexes, src)",
                 lhs: indexes.shape().clone(),
                 rhs: source.shape().clone(),
-            }
-            .bt())?
+            })?
         }
         let storage = self.storage().scatter_add(
             self.layout(),
@@ -1318,8 +1190,7 @@ impl Tensor {
                 op: "slice-scatter (self, src)",
                 lhs: self.shape().clone(),
                 rhs: src.shape().clone(),
-            }
-            .bt())?
+            })?
         }
         let mut storage = self.device().zeros(self.shape(), self.dtype())?;
         self.storage()
@@ -1353,8 +1224,7 @@ impl Tensor {
                 op: "index-add (self, source)",
                 lhs: self.shape().clone(),
                 rhs: source.shape().clone(),
-            }
-            .bt())?
+            })?
         }
         // The number of element in indexes must match the dimension on which the add is
         // performed on the source tensor (and the index values from `indexes` are taken from
@@ -1365,8 +1235,7 @@ impl Tensor {
                 op: "index-add (ids, source))",
                 lhs: indexes.shape().clone(),
                 rhs: source.shape().clone(),
-            }
-            .bt())?
+            })?
         }
         let storage = self.storage().index_add(
             self.layout(),
@@ -1414,8 +1283,7 @@ impl Tensor {
                 op: "gather",
                 lhs: self.shape().clone(),
                 rhs: indexes.shape().clone(),
-            }
-            .bt())?
+            })?
         }
         let storage =
             self.storage()
@@ -1489,7 +1357,6 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(storage) => from_cpu_storage(storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
-            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -1520,7 +1387,6 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(storage) => from_cpu_storage(storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
-            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -1561,7 +1427,6 @@ impl Tensor {
         match &*self.storage() {
             Storage::Cpu(storage) => from_cpu_storage(storage),
             Storage::Cuda(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
-            Storage::Metal(storage) => from_cpu_storage(&storage.to_cpu_storage()?),
         }
     }
 
@@ -1725,24 +1590,6 @@ impl Tensor {
         }
     }
 
-    /// Returns the sub-tensor fixing the index at `index` on the dimension `dim`.
-    ///
-    /// ```rust
-    /// use candle_core::{Tensor, Device};
-    /// let tensor = Tensor::new(&[[0f32, 1.], [2., 3.], [4., 5.]], &Device::Cpu)?;
-    /// let t = tensor.get_on_dim(1, 0)?;
-    /// assert_eq!(t.to_vec1::<f32>()?, &[0., 2., 4.]);
-    /// let t = tensor.get_on_dim(1, 1)?;
-    /// assert_eq!(t.to_vec1::<f32>()?, &[1., 3., 5.]);
-    /// let t = tensor.get_on_dim(0, 1)?;
-    /// assert_eq!(t.to_vec1::<f32>()?, &[2., 3.]);
-    /// # Ok::<(), candle_core::Error>(())
-    /// ```
-    pub fn get_on_dim<D: Dim>(&self, dim: D, index: usize) -> Result<Tensor> {
-        let dim = dim.to_index(self.shape(), "get_on_dim")?;
-        self.narrow(dim, index, 1)?.squeeze(dim)
-    }
-
     /// Returns a tensor that is a transposed version of the input, the two last dimensions of the
     /// input are swapped.
     ///
@@ -1804,7 +1651,7 @@ impl Tensor {
         let is_permutation =
             dims.len() == self.rank() && (0..dims.len()).all(|i| dims.contains(&i));
         if !is_permutation {
-            bail!(
+            crate::bail!(
                 "dimension mismatch in permute, tensor {:?}, dims: {:?}",
                 self.dims(),
                 dims
@@ -1851,23 +1698,17 @@ impl Tensor {
 
     /// Returns a new tensor detached from the current graph, gradient are not propagated through
     /// this new node. The storage of this tensor is shared with the initial tensor.
-    ///
-    /// If the tensor is already detached from the computation graph, the same tensor is returned.
     pub fn detach(&self) -> Result<Tensor> {
-        if self.op.is_none() && !self.is_variable {
-            Ok(self.clone())
-        } else {
-            let tensor_ = Tensor_ {
-                id: TensorId::new(),
-                storage: self.storage.clone(),
-                layout: self.layout.clone(),
-                op: BackpropOp::none(),
-                is_variable: false,
-                dtype: self.dtype,
-                device: self.device.clone(),
-            };
-            Ok(Tensor(Arc::new(tensor_)))
-        }
+        let tensor_ = Tensor_ {
+            id: TensorId::new(),
+            storage: self.storage.clone(),
+            layout: self.layout.clone(),
+            op: BackpropOp::none(),
+            is_variable: false,
+            dtype: self.dtype,
+            device: self.device.clone(),
+        };
+        Ok(Tensor(Arc::new(tensor_)))
     }
 
     /// If the target device is the same as the tensor device, only a shallow copy is performed.
@@ -1879,11 +1720,7 @@ impl Tensor {
                 (Storage::Cpu(storage), Device::Cuda(cuda)) => {
                     Storage::Cuda(cuda.storage_from_cpu_storage(storage)?)
                 }
-                (Storage::Cpu(storage), Device::Metal(metal)) => {
-                    Storage::Metal(metal.storage_from_cpu_storage(storage)?)
-                }
                 (Storage::Cuda(storage), Device::Cpu) => Storage::Cpu(storage.to_cpu_storage()?),
-                (Storage::Metal(storage), Device::Cpu) => Storage::Cpu(storage.to_cpu_storage()?),
                 (Storage::Cuda(storage), Device::Cuda(cuda)) => {
                     // TODO: Avoid passing through the cpu storage here, especially if the gpu ids
                     // are the same.
@@ -1891,9 +1728,6 @@ impl Tensor {
                     Storage::Cuda(cuda.storage_from_cpu_storage(&cpu_storage)?)
                 }
                 (Storage::Cpu(storage), Device::Cpu) => Storage::Cpu(storage.clone()),
-                _ => {
-                    bail!("not implemented yet")
-                }
             };
             let op = BackpropOp::new1(self, Op::ToDevice);
             let tensor_ = Tensor_ {
@@ -2293,56 +2127,11 @@ impl Tensor {
         }
     }
 
-    /// Pad the input tensor using same values along dimension `dim`. This adds `left` elements before the
-    /// input tensor values and `right` elements after.
-    pub fn pad_with_same<D: Dim>(&self, dim: D, left: usize, right: usize) -> Result<Self> {
-        if left == 0 && right == 0 {
-            Ok(self.clone())
-        } else if self.elem_count() == 0 {
-            bail!("cannot use pad_with_same on an empty tensor")
-        } else if left == 0 {
-            let dim = dim.to_index(self.shape(), "pad_with_same")?;
-            let r = self.narrow(dim, self.dim(dim)? - 1, 1)?;
-            let mut v = vec![self];
-            for _ in 0..right {
-                v.push(&r)
-            }
-            Tensor::cat(&v, dim)
-        } else if right == 0 {
-            let dim = dim.to_index(self.shape(), "pad_with_same")?;
-            let l = self.narrow(dim, 0, 1)?;
-            let mut v = vec![];
-            for _ in 0..left {
-                v.push(&l)
-            }
-            v.push(self);
-            Tensor::cat(&v, dim)
-        } else {
-            let dim = dim.to_index(self.shape(), "pad_with_same")?;
-            let l = self.narrow(dim, 0, 1)?;
-            let r = self.narrow(dim, self.dim(dim)? - 1, 1)?;
-            let mut v = vec![];
-            for _ in 0..left {
-                v.push(&l)
-            }
-            v.push(self);
-            for _ in 0..right {
-                v.push(&r)
-            }
-            Tensor::cat(&v, dim)
-        }
-    }
-
     /// Run the `forward` method of `m` on `self`.
     pub fn apply<M: crate::Module>(&self, m: &M) -> Result<Self> {
         m.forward(self)
     }
 
-    /// Run the `forward` method of `m` on `self`.
-    pub fn apply_t<M: crate::ModuleT>(&self, m: &M, train: bool) -> Result<Self> {
-        m.forward_t(self, train)
-    }
-
     pub(crate) fn storage(&self) -> std::sync::RwLockReadGuard<'_, Storage> {
         self.storage.read().unwrap()
     }
@@ -2457,132 +2246,6 @@ impl Tensor {
     ) -> Result<Self> {
         self.apply_op3_arc(t2, t3, Arc::new(Box::new(c)))
     }
-
-    /// Normalize a 'relative' axis value: positive values are kept, negative
-    /// values means counting the dimensions from the back.
-    pub fn normalize_axis(&self, axis: i64) -> Result<usize> {
-        let rank = self.rank() as i64;
-        if rank <= axis {
-            bail!("axis {axis} is too large, tensor rank {rank}")
-        } else if 0 <= axis {
-            Ok(axis as usize)
-        } else {
-            let naxis = rank + axis;
-            if naxis < 0 {
-                bail!("axis {axis} is too small, tensor rank {rank}")
-            }
-            Ok(naxis as usize)
-        }
-    }
-
-    /// Returns a lower triangular matrix of ones of size n by n.
-    pub fn tril2(n: usize, dtype: DType, device: &Device) -> Result<Self> {
-        let t = Tensor::arange(0u32, n as u32, device)?;
-        let t1 = t.reshape((1, n))?.broadcast_as((n, n))?;
-        let t2 = t.reshape((n, 1))?.broadcast_as((n, n))?;
-        t1.le(&t2)?.to_dtype(dtype)
-    }
-
-    /// Returns an upper triangular matrix of ones of size n by n.
-    pub fn triu2(n: usize, dtype: DType, device: &Device) -> Result<Self> {
-        let t = Tensor::arange(0u32, n as u32, device)?;
-        let t1 = t.reshape((1, n))?.broadcast_as((n, n))?;
-        let t2 = t.reshape((n, 1))?.broadcast_as((n, n))?;
-        t1.ge(&t2)?.to_dtype(dtype)
-    }
-
-    /// Returns a matrix with a diagonal of ones of size n by n.
-    pub fn eye(n: usize, dtype: DType, device: &Device) -> Result<Self> {
-        let t = Tensor::arange(0u32, n as u32, device)?;
-        let t1 = t.reshape((1, n))?.broadcast_as((n, n))?;
-        let t2 = t.reshape((n, 1))?.broadcast_as((n, n))?;
-        t1.eq(&t2)?.to_dtype(dtype)
-    }
-
-    /// Returns the cumulative sum of elements of the input tensor summed over the specified
-    /// dimension.
-    ///
-    /// This operation is most efficient when dim is the last dimension of the tensor.
-    pub fn cumsum<D: Dim>(&self, dim: D) -> Result<Self> {
-        let dim = dim.to_index(self.shape(), "cumsum")?;
-        let rank = self.rank();
-        if rank == 0 {
-            return Ok(self.clone());
-        }
-        let n_axis = self.dim(dim)?;
-        let triu = Tensor::triu2(n_axis, self.dtype(), self.device())?;
-        if rank == 1 {
-            self.unsqueeze(0)?.matmul(&triu)?.squeeze(0)
-        } else {
-            let last = rank - 1;
-            let t = self.transpose(dim, last)?;
-            let t = t.broadcast_matmul(&triu)?;
-            t.transpose(dim, last)
-        }
-    }
-
-    /// Returns a copy of `self` where the values within `ranges` have been replaced with the
-    /// content of `src`.
-    pub fn slice_assign<D: std::ops::RangeBounds<usize>>(
-        &self,
-        ranges: &[D],
-        src: &Tensor,
-    ) -> Result<Self> {
-        let src_dims = src.dims();
-        let self_dims = self.dims();
-        if self_dims.len() != src_dims.len() {
-            bail!(
-                "slice-assign requires input with the same rank {} <> {}",
-                self_dims.len(),
-                src_dims.len()
-            )
-        }
-        if self_dims.len() != ranges.len() {
-            bail!(
-                "slice-assign requires input with the same rank as there are ranges {} <> {}",
-                self_dims.len(),
-                ranges.len()
-            )
-        }
-        let mut src = src.clone();
-        let mut mask = Self::ones(src.shape(), DType::U8, src.device())?;
-        for (i, range) in ranges.iter().enumerate() {
-            let start_included = match range.start_bound() {
-                std::ops::Bound::Unbounded => 0,
-                std::ops::Bound::Included(v) => *v,
-                std::ops::Bound::Excluded(v) => *v + 1,
-            };
-            let end_excluded = match range.end_bound() {
-                std::ops::Bound::Unbounded => self_dims[i],
-                std::ops::Bound::Included(v) => *v + 1,
-                std::ops::Bound::Excluded(v) => *v,
-            };
-            if end_excluded <= start_included {
-                bail!("slice-assign: empty range for dim {i}, {start_included} {end_excluded}")
-            }
-            if self_dims[i] < end_excluded {
-                bail!(
-                    "slice-assign: upper bound is out of range for dim {i}, {end_excluded} {}",
-                    self_dims[i]
-                )
-            }
-            if end_excluded - start_included != src_dims[i] {
-                bail!(
-                    "slice-assign: the range for dim {i} ({start_included}..{end_excluded}) does not match the size of src {}", src_dims[i]
-                )
-            }
-            src = src.pad_with_zeros(i, start_included, self_dims[i] - end_excluded)?;
-            mask = mask.pad_with_zeros(i, start_included, self_dims[i] - end_excluded)?
-        }
-        mask.where_cond(/* on_true= */ &src, /* on_false= */ self)
-    }
-
-    /// Returns log(sum(exp(tensor), dim)).
-    pub fn logsumexp<D: Dims>(&self, sum_dims: D) -> Result<Self> {
-        let exp = self.exp()?;
-        let sum = exp.sum(sum_dims)?;
-        sum.log()
-    }
 }
 
 macro_rules! bin_trait {

candle-core/src/test_utils.rs

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

candle-core/src/utils.rs

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

candle-core/tests/conv_tests.rs

@@ -13,11 +13,6 @@ res = torch.nn.functional.conv1d(t, w)
 print(res.flatten())
 res = torch.nn.functional.conv1d(t, w, padding=1)
 print(res.flatten())
-
-w_t = w.transpose(0, 1)
-res = torch.nn.functional.conv_transpose1d(t, w_t)
-print(res.shape)
-print(res)
 */
 fn conv1d(dev: &Device) -> Result<()> {
     let t = Tensor::new(
@@ -50,17 +45,6 @@ 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)?;
-        assert_eq!(res.dims(), [1, 2, 7]);
-        assert_eq!(
-            test_utils::to_vec1_round(&res.flatten_all()?, 4)?,
-            [
-                0.0699, -1.2899, 8.3018, 5.5873, 2.4572, -2.6143, -0.0706, 1.8765, 4.8318, 1.1538,
-                4.7076, -5.9745, -0.8276, 1.621
-            ],
-        );
-    }
     Ok(())
 }
 
@@ -495,103 +479,17 @@ fn conv2d_grad(dev: &Device) -> Result<()> {
             ]
         ]
     );
-
-    // Replicate the issue from https://github.com/huggingface/candle/issues/1212
-    let res = t.i((.., .., 0..4, 0..4))?.conv2d(&w, 0, 2, 1, 1)?;
-    let loss = res.sqr()?.sum_all()?;
-    assert_eq!(test_utils::to_vec0_round(&loss, 2)?, 21.12f32);
-    let grads = loss.backward()?;
-    let grad_t = grads.get(&t).unwrap();
-    let grad_w = grads.get(&w).unwrap();
-    assert_eq!(grad_t.dims(), [1, 4, 5, 5]);
-    assert_eq!(grad_w.dims(), [2, 4, 3, 3]);
-    assert_eq!(
-        test_utils::to_vec3_round(&grad_t.i(0)?, 2)?,
-        [
-            [
-                [9.29, -7.03, 7.87, 0.0, 0.0],
-                [-1.8, -7.82, 5.9, 0.0, 0.0],
-                [-3.12, 4.49, 5.52, 0.0, 0.0],
-                [0.0, 0.0, 0.0, 0.0, 0.0],
-                [0.0, 0.0, 0.0, 0.0, 0.0]
-            ],
-            [
-                [21.73, 3.39, 4.77, 0.0, 0.0],
-                [8.25, 3.73, 27.61, 0.0, 0.0],
-                [-20.55, -5.61, -2.77, 0.0, 0.0],
-                [0.0, 0.0, 0.0, 0.0, 0.0],
-                [0.0, 0.0, 0.0, 0.0, 0.0]
-            ],
-            [
-                [-8.98, 9.91, -7.15, 0.0, 0.0],
-                [4.93, -0.33, 4.56, 0.0, 0.0],
-                [-6.7, -5.76, -8.05, 0.0, 0.0],
-                [0.0, 0.0, 0.0, 0.0, 0.0],
-                [0.0, 0.0, 0.0, 0.0, 0.0]
-            ],
-            [
-                [23.54, 6.98, -10.0, 0.0, 0.0],
-                [9.65, 6.18, 18.72, 0.0, 0.0],
-                [3.29, -5.27, 0.79, 0.0, 0.0],
-                [0.0, 0.0, 0.0, 0.0, 0.0],
-                [0.0, 0.0, 0.0, 0.0, 0.0]
-            ]
-        ]
-    );
-    assert_eq!(
-        test_utils::to_vec3_round(&grad_w.i(0)?, 2)?,
-        [
-            [
-                [-3.47, 7.44, 0.66],
-                [12.89, -3.4, -9.29],
-                [-14.16, -0.83, 7.14]
-            ],
-            [
-                [-3.23, 5.37, -3.02],
-                [-2.12, -11.24, 1.94],
-                [6.97, 7.2, 2.99]
-            ],
-            [
-                [-4.04, -3.31, 4.87],
-                [-6.68, -5.68, 1.73],
-                [-5.54, 4.32, 0.52]
-            ],
-            [[-4.72, 1.5, 4.72], [3.79, 4.04, 6.76], [-4.6, 5.8, 6.93]]
-        ]
-    );
-
     Ok(())
 }
 
-test_device!(conv1d, conv1d_cpu, conv1d_gpu, conv1d_metal);
-test_device!(
-    conv1d_small,
-    conv1d_small_cpu,
-    conv1d_small_gpu,
-    conv1d_small_metal
-);
-test_device!(conv2d, conv2d_cpu, conv2d_gpu, conv2d_metal);
+test_device!(conv1d, conv1d_cpu, conv1d_gpu);
+test_device!(conv1d_small, conv1d_small_cpu, conv1d_small_gpu);
+test_device!(conv2d, conv2d_cpu, conv2d_gpu);
 test_device!(
     conv2d_non_square,
     conv2d_non_square_cpu,
-    conv2d_non_square_gpu,
-    conv2d_non_square_metal
-);
-test_device!(
-    conv2d_small,
-    conv2d_small_cpu,
-    conv2d_small_gpu,
-    conv2d_small_metal
-);
-test_device!(
-    conv2d_smaller,
-    conv2d_smaller_cpu,
-    conv2d_smaller_gpu,
-    conv2d_smaller_metal
-);
-test_device!(
-    conv2d_grad,
-    conv2d_grad_cpu,
-    conv2d_grad_gpu,
-    conv2_grad_metal
+    conv2d_non_square_gpu
 );
+test_device!(conv2d_small, conv2d_small_cpu, conv2d_small_gpu);
+test_device!(conv2d_smaller, conv2d_smaller_cpu, conv2d_smaller_gpu);
+test_device!(conv2d_grad, conv2d_grad_cpu, conv2d_grad_gpu);

candle-core/tests/grad_tests.rs

@@ -192,244 +192,6 @@ fn unary_grad(device: &Device) -> Result<()> {
         test_utils::to_vec1_round(grad_x, 2)?,
         [0.01, 0.42, 0.0, 0.98],
     );
-
-    // testing compared to pytorch nn.GELU(approximate = 'tanh')
-    let y = x.gelu()?;
-    let grads = y.backward()?;
-    let grad_x = grads.get(&x).context("no grad for x")?;
-    assert_eq!(
-        test_utils::to_vec1_round(&y, 4)?,
-        [2.9964, 0.8412, 3.9999, 0.0839]
-    );
-    assert_eq!(
-        test_utils::to_vec1_round(grad_x, 4)?,
-        [1.0116, 1.0830, 1.0003, 0.6188],
-    );
-
-    // Testing compared to pytorch torch.erf
-    //
-    // import torch
-    // x = torch.tensor([3.0, 1.0, 4.0, 0.15], requires_grad=True)
-    // y = x.erf()
-    // print(y)
-    // loss = y.sum()
-    // loss.backward()
-    // print(x.grad)
-    let y = x.erf()?;
-    let grads = y.backward()?;
-    let grad_x = grads.get(&x).context("no grad for x")?;
-    assert_eq!(test_utils::to_vec1_round(&y, 4)?, [1.0, 0.8427, 1.0, 0.168]);
-    assert_eq!(
-        test_utils::to_vec1_round(grad_x, 4)?,
-        [0.0001, 0.4151, 0.0, 1.1033],
-    );
-
-    // Testing compared to pytorch nn.GELU(approximate = 'none')
-    //
-    // import torch
-    // import torch.nn.functional as F
-    // x = torch.tensor([3.0, 1.0, 4.0, 0.15], requires_grad=True)
-    // y = F.gelu(x, approximate='none')
-    // print(y)
-    // loss = y.sum()
-    // loss.backward()
-    // print(x.grad)
-    let y = x.gelu_erf()?;
-    let grads = y.backward()?;
-    let grad_x = grads.get(&x).context("no grad for x")?;
-    assert_eq!(
-        test_utils::to_vec1_round(&y, 4)?,
-        [2.9960, 0.8413, 3.9999, 0.0839]
-    );
-    assert_eq!(
-        test_utils::to_vec1_round(grad_x, 4)?,
-        [1.0119, 1.0833, 1.0005, 0.6188],
-    );
-
-    // Testing compared to pytorch elu
-    //
-    // import torch
-    // import torch.nn.functional as F
-    // x = torch.tensor([-1.0, 0.0, -2.0, 3.0], requires_grad=True)
-    // y = F.elu(x, alpha=2.0)
-    // print(y)
-    // loss = y.min
-    // loss = y.sum()
-    // loss.backward()
-    // print(x.grad)
-    let elu_x = Var::new(&[-1.0f32, 0., -2., 3.], device)?;
-    let y = elu_x.elu(2.)?;
-    let grads = y.backward()?;
-    let grad_x = grads.get(&elu_x).context("no grad for x")?;
-    assert_eq!(
-        test_utils::to_vec1_round(&y, 4)?,
-        [-1.2642, 0.0000, -1.7293, 3.0000]
-    );
-    assert_eq!(
-        test_utils::to_vec1_round(grad_x, 4)?,
-        [0.7358, 2.0000, 0.2707, 1.0000]
-    );
-
-    // manually checked: see comments
-    let x = Var::new(&[[[[1f32, 2., 3.], [4., 5., 6.], [7., 8., 9.]]]], device)?;
-    let y = x.interpolate2d(6, 6)?.reshape(36)?;
-
-    #[rustfmt::skip]
-    let z = Tensor::new(
-        &[
-            1_f32, 02., 03., 04., 05., 06.,
-            07.,   08., 09., 10., 11., 12.,
-            13.,   14., 15., 16., 17., 18.,
-            19.,   20., 21., 22., 23., 24.,
-            25.,   26., 27., 28., 29., 30.,
-            31.,   32., 33., 34., 35., 36.,
-        ],
-        device,
-    )?;
-    // gradient should be
-    // row 1
-    // 1+2+7+8 = 18
-    // 3+4+9+10 = 26
-    // 5+6+11+12 = 34
-    // row 2
-    // 13+14+19+20 = 66
-    // 15+16+21+22 = 74
-    // 17+18+23+24 = 82
-    // row 3
-    // 25+26+31+32 = 114
-    // 27+28+33+34 = 122
-    // 29+30+35+36 = 130
-    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
-
-    let grads = loss.backward()?;
-
-    let grad_x = grads.get(&x).context("no grad for x")?;
-    assert_eq!(
-        test_utils::to_vec2_round(&grad_x.flatten(0, 2)?, 4)?,
-        [[18_f32, 26., 34.], [66., 74., 82.], [114., 122., 130.]]
-    );
-
-    // manually checked: see comments
-    let x = Var::new(&[[[[1f32, 2.], [4., 5.]]]], device)?;
-    let y = x.interpolate2d(6, 6)?.reshape(36)?;
-
-    #[rustfmt::skip]
-    let z = Tensor::new(
-        &[
-            1_f32, 02., 03., 04., 05., 06.,
-            07.,   08., 09., 10., 11., 12.,
-            13.,   14., 15., 16., 17., 18.,
-            19.,   20., 21., 22., 23., 24.,
-            25.,   26., 27., 28., 29., 30.,
-            31.,   32., 33., 34., 35., 36.,
-        ],
-        device,
-    )?;
-    // gradient should be
-    // row 1
-    // 1+2+3+7+8+9+13+14+15 = 72
-    // 4+5+6+10+11+12+16+17+18 = 99
-    // row 2
-    // 19+20+21+25+26+27+31+32+33 = 234
-    // 22+23+24+28+29+30+34+35+36 = 243
-    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
-
-    let grads = loss.backward()?;
-
-    let grad_x = grads.get(&x).context("no grad for x")?;
-    assert_eq!(
-        test_utils::to_vec2_round(&grad_x.flatten(0, 2)?, 4)?,
-        [[72_f32, 99.], [234., 261.]]
-    );
-
-    // manually checked: see comments
-    let x = Var::new(&[[[[1f32, 2.], [4., 5.]], [[6f32, 7.], [8., 9.]]]], device)?;
-
-    let y = x.interpolate2d(4, 4)?.reshape(32)?;
-
-    #[rustfmt::skip]
-    let z = Tensor::new(
-        &[
-            1_f32, 02., 03., 04.,
-            05.,   06., 07., 08.,
-            09.,   10., 11., 12.,
-            13.,   14., 15., 16.,
-            17.,   18., 19., 20.,
-            21.,   22., 23., 24.,
-            25.,   26., 27., 28.,
-            29.,   30., 31., 32.
-        ],
-        device,
-    )?;
-    // gradient should be
-    // m1r1
-    // 1+2+5+6=14
-    // 3+4+7+8=22
-    // m1r2
-    // 9+10+13+14=46
-    // 11+12+15+16=54
-    // m2r1
-    // 17+18+21+22=78
-    // 19+20+23+24=86
-    // m2r2
-    // 25+26+29+30=110
-    // 27+28+31+32=118
-    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
-
-    let grads = loss.backward()?;
-
-    let grad_x = grads.get(&x).context("no grad for x")?;
-
-    assert_eq!(
-        test_utils::to_vec3_round(&grad_x.flatten(0, 1)?, 4)?,
-        [[[14_f32, 22.], [46., 54.]], [[78., 86.], [110., 118.]]]
-    );
-
-    // manually checked: see comments
-    let x = Var::new(
-        &[[[[1f32, 2.], [4., 5.]]], [[[6f32, 7.], [8., 9.]]]],
-        device,
-    )?;
-
-    let y = x.interpolate2d(4, 4)?.reshape(32)?;
-
-    #[rustfmt::skip]
-       let z = Tensor::new(
-           &[
-               1_f32, 02., 03., 04.,
-               05.,   06., 07., 08.,
-               09.,   10., 11., 12.,
-               13.,   14., 15., 16.,
-               17.,   18., 19., 20.,
-               21.,   22., 23., 24.,
-               25.,   26., 27., 28.,
-               29.,   30., 31., 32.
-           ],
-           device,
-       )?;
-    // gradient should be
-    // m1r1
-    // 1+2+5+6=14
-    // 3+4+7+8=22
-    // m1r2
-    // 9+10+13+14=46
-    // 11+12+15+16=54
-    // m2r1
-    // 17+18+21+22=78
-    // 19+20+23+24=86
-    // m2r2
-    // 25+26+29+30=110
-    // 27+28+31+32=118
-    let loss = y.unsqueeze(1)?.transpose(0, 1)?.matmul(&z.unsqueeze(1)?)?;
-
-    let grads = loss.backward()?;
-
-    let grad_x = grads.get(&x).context("no grad for x")?;
-
-    assert_eq!(
-        test_utils::to_vec3_round(&grad_x.flatten(0, 1)?, 4)?,
-        [[[14_f32, 22.], [46., 54.]], [[78., 86.], [110., 118.]]]
-    );
     Ok(())
 }
 
@@ -475,29 +237,9 @@ fn binary_grad(device: &Device) -> Result<()> {
     Ok(())
 }
 
-test_device!(
-    simple_grad,
-    simple_grad_cpu,
-    simple_grad_gpu,
-    simple_grad_metal
-);
-test_device!(sum_grad, sum_grad_cpu, sum_grad_gpu, sum_grad_metal);
-test_device!(
-    matmul_grad,
-    matmul_grad_cpu,
-    matmul_grad_gpu,
-    matmul_grad_metal
-);
-test_device!(
-    grad_descent,
-    grad_descent_cpu,
-    grad_descent_gpu,
-    grad_descent_metal
-);
-test_device!(unary_grad, unary_grad_cpu, unary_grad_gpu, unary_grad_metal);
-test_device!(
-    binary_grad,
-    binary_grad_cpu,
-    binary_grad_gpu,
-    binary_grad_metal
-);
+test_device!(simple_grad, simple_grad_cpu, simple_grad_gpu);
+test_device!(sum_grad, sum_grad_cpu, sum_grad_gpu);
+test_device!(matmul_grad, matmul_grad_cpu, matmul_grad_gpu);
+test_device!(grad_descent, grad_descent_cpu, grad_descent_gpu);
+test_device!(unary_grad, unary_grad_cpu, unary_grad_gpu);
+test_device!(binary_grad, binary_grad_cpu, binary_grad_gpu);

candle-core/tests/indexing_tests.rs

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

candle-core/tests/layout_tests.rs

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

candle-core/tests/npy.py

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

candle-core/tests/pool_tests.rs

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

candle-core/tests/quantized_tests.rs

@@ -1,8 +1,7 @@
 use candle_core::{
-    bail,
     quantized::{self, GgmlDType},
     test_utils::to_vec2_round,
-    Device, Module, Result, Tensor,
+    Device, Result, Tensor,
 };
 use quantized::{k_quants, GgmlType};
 use rand::prelude::*;
@@ -266,8 +265,7 @@ fn compare_with_error(values: &[f32], expected: &[f32], tolerance: f32) {
     }
 }
 
-/// 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
+/// 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
 fn create_ggml_like_vector(offset: f32) -> Vec<f32> {
     (0..GGML_TEST_SIZE)
         .map(|i| 0.1 + 2.0 * (i as f32 + offset).cos())
@@ -286,15 +284,14 @@ fn calculate_rmse(a: &[f32], b: &[f32]) -> f32 {
     sum / a.len() as f32
 }
 
-/// Similar to the GGML quantization unit test:
-/// https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L43-L50
+/// 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<()> {
     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());
     if error > max_error {
-        bail!(
+        candle_core::bail!(
             "Quantization error {} exceeds max error {}",
             error,
             max_error
@@ -490,66 +487,54 @@ 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 => 0.008,
+        GgmlDType::Q4_1 => 0.007784,
         GgmlDType::Q5_0 => 0.001353,
-        GgmlDType::Q5_1 => 0.00149,
+        GgmlDType::Q5_1 => 0.001363,
         GgmlDType::Q8_0 => 0.000092,
 
         // Not from the ggml repo.
         GgmlDType::Q8K => 0.00065,
-        _ => bail!("No GGML results for quantization type {dtype:?}",),
+        _ => candle_core::bail!("No GGML results for quantization type {dtype:?}",),
     };
     Ok(err)
 }
 
-/// Similar to the GGML matmul unit test:
-/// https://github.com/ggerganov/llama.cpp/blob/master/tests/test-quantize-fns.cpp#L76-L91
+/// Mirrores 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 {
-        bail!(
+        candle_core::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)? * err_m;
+    let ggml_error = ggml_reference_matmul_error(T::DTYPE)?;
 
     if !error.is_finite() || error > GGML_MAX_DOT_PRODUCT_ERROR {
-        bail!("Dot product error {error} exceeds max error {GGML_MAX_DOT_PRODUCT_ERROR}",);
+        candle_core::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 {
-        bail!(
+        candle_core::bail!(
             "Dot product error {} exceeds ggml reference error {}",
             error,
             ggml_error
@@ -558,36 +543,6 @@ fn ggml_matmul_error_test_<T: GgmlType>(a: &[f32], b: &[f32], err_m: f32) -> Res
     Ok(())
 }
 
-fn get_small_tensors(
-    m: usize,
-    k: usize,
-    n: usize,
-    device: &Device,
-) -> Result<(Tensor, Tensor, Tensor)> {
-    let lhs = (0..m * k)
-        .map(|i| i as f32 / (m * k) as f32)
-        .collect::<Vec<_>>();
-    let rhs = (0..n * k)
-        .map(|i| i as f32 / (n * k) as f32)
-        .collect::<Vec<_>>();
-
-    let lhs = Tensor::from_vec(lhs, (m, k), device)?;
-    let rhs = Tensor::from_vec(rhs, (n, k), device)?;
-
-    let mm = lhs.matmul(&rhs.t()?)?;
-    Ok((lhs, rhs, mm))
-}
-
-#[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,
@@ -643,30 +598,20 @@ fn quantized_matmul_q3k() -> Result<()> {
 
     let cpu = &Device::Cpu;
     let (m, k, n) = (11, 512, 21);
-    let (lhs, rhs, mm) = get_small_tensors(m, k, n, cpu)?;
-    // assert_eq!(mm.dims(), [m, n]);
-    // let dst = mm.flatten_all()?.to_vec1::<f32>()?;
-    // let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
-    // assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
+    let (lhs, rhs, mm) = get_random_tensors(m, k, n, cpu)?;
+    assert_eq!(mm.dims(), [m, n]);
+    let dst = mm.flatten_all()?.to_vec1::<f32>()?;
+    let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
+    assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
     let rhs = quantized::QTensor::quantize::<BlockQ3K>(&rhs)?;
     let rhs = quantized::QMatMul::from_qtensor(rhs)?;
-    let qmm = rhs.forward(&lhs)?;
+    let mm = rhs.forward(&lhs)?;
 
-    let error: f32 = ((&mm - &qmm)?.abs()? / &mm.abs()?)?
-        .sum_all()?
-        .to_scalar()?;
-    let error = error / (m * n) as f32;
-
-    // assert_eq!(qmm.dims(), [m, n]);
-    // let dst = qmm.flatten_all()?.to_vec1::<f32>()?;
-    // let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
-    // assert_eq!(dst, [1.029, 1.418, -0.314, 1.495]);
-
-    assert!(
-        error < 0.01,
-        "{error} is too big, shouldn't exceed a few percent. \nGot:{qmm}\nExpected:\n{mm} "
-    );
+    assert_eq!(mm.dims(), [m, n]);
+    let dst = mm.flatten_all()?.to_vec1::<f32>()?;
+    let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
+    assert_eq!(dst, [1.029, 1.418, -0.314, 1.495]);
 
     ggml_matmul_error_test::<BlockQ3K>()?;
 
@@ -679,30 +624,20 @@ fn quantized_matmul_q4k() -> Result<()> {
 
     let cpu = &Device::Cpu;
     let (m, k, n) = (11, 512, 21);
-    let (lhs, rhs, mm) = get_small_tensors(m, k, n, cpu)?;
-    // assert_eq!(mm.dims(), [m, n]);
-    // let dst = mm.flatten_all()?.to_vec1::<f32>()?;
-    // let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
-    // assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
+    let (lhs, rhs, mm) = get_random_tensors(m, k, n, cpu)?;
+    assert_eq!(mm.dims(), [m, n]);
+    let dst = mm.flatten_all()?.to_vec1::<f32>()?;
+    let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
+    assert_eq!(dst, [1.262, 1.513, -0.208, 1.702]);
 
     let rhs = quantized::QTensor::quantize::<BlockQ4K>(&rhs)?;
     let rhs = quantized::QMatMul::from_qtensor(rhs)?;
-    let qmm = rhs.forward(&lhs)?;
+    let mm = rhs.forward(&lhs)?;
 
-    let error: f32 = ((&mm - &qmm)?.abs()? / &mm.abs()?)?
-        .sum_all()?
-        .to_scalar()?;
-    let error = error / (m * n) as f32;
-
-    assert!(
-        error < 0.01,
-        "{error} is too big, shouldn't exceed a few percent. \nGot:{qmm}\nExpected:\n{mm} "
-    );
-
-    // assert_eq!(mm.dims(), [m, n]);
-    // let dst = mm.flatten_all()?.to_vec1::<f32>()?;
-    // let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
-    // assert_eq!(dst, [1.125, 1.435, -0.201, 1.589]);
+    assert_eq!(mm.dims(), [m, n]);
+    let dst = mm.flatten_all()?.to_vec1::<f32>()?;
+    let dst = round_vector(&[dst[0], dst[m * n / 3], dst[m * n * 2 / 3], dst[m * n - 1]]);
+    assert_eq!(dst, [1.125, 1.435, -0.201, 1.589]);
 
     ggml_matmul_error_test::<BlockQ4K>()?;
 

candle-core/tests/serialization_tests.rs

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

candle-core/tests/tensor_tests.rs

@@ -1,4 +1,4 @@
-use candle_core::{test_device, test_utils, DType, Device, IndexOp, Result, Tensor, D};
+use candle_core::{test_device, test_utils, DType, Device, IndexOp, Result, Tensor};
 
 fn zeros(device: &Device) -> Result<()> {
     let tensor = Tensor::zeros((5, 2), DType::F32, device)?;
@@ -29,34 +29,7 @@ fn ones(device: &Device) -> Result<()> {
         Tensor::ones((2, 3), DType::F64, device)?.to_vec2::<f64>()?,
         [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
     );
-    Ok(())
-}
 
-fn full(device: &Device) -> Result<()> {
-    assert_eq!(
-        Tensor::full(42u32, (2, 3), device)?.to_vec2::<u32>()?,
-        [[42, 42, 42], [42, 42, 42]],
-    );
-    Ok(())
-}
-
-fn arange(device: &Device) -> Result<()> {
-    assert_eq!(
-        Tensor::arange(0u8, 5u8, device)?.to_vec1::<u8>()?,
-        [0, 1, 2, 3, 4],
-    );
-    assert_eq!(
-        Tensor::arange_step(0u8, 5u8, 2, device)?.to_vec1::<u8>()?,
-        [0, 2, 4],
-    );
-    assert_eq!(
-        Tensor::arange_step(0u8, 5u8, 3, device)?.to_vec1::<u8>()?,
-        [0, 3],
-    );
-    assert_eq!(
-        Tensor::arange_step(5i64, 0i64, -1, device)?.to_vec1::<i64>()?,
-        [5, 4, 3, 2, 1],
-    );
     Ok(())
 }
 
@@ -188,22 +161,6 @@ fn transpose(device: &Device) -> Result<()> {
     Ok(())
 }
 
-fn var(device: &Device) -> Result<()> {
-    // Values taken from https://pytorch.org/docs/stable/generated/torch.var.html
-    let data = &[
-        [0.2035f32, 1.2959, 1.8101, -0.4644],
-        [1.5027, -0.3270, 0.5905, 0.6538],
-        [-1.5745, 1.3330, -0.5596, -0.6548],
-        [0.1264, -0.5080, 1.6420, 0.1992],
-    ];
-    let tensor = Tensor::new(data, device)?;
-    assert_eq!(
-        test_utils::to_vec2_round(&tensor.var_keepdim(1)?, 4)?,
-        &[[1.0631], [0.559], [1.4893], [0.8258]]
-    );
-    Ok(())
-}
-
 fn sum(device: &Device) -> Result<()> {
     let data = &[[[3u32, 1, 4], [1, 5, 9]], [[2, 1, 7], [8, 2, 8]]];
     let tensor = Tensor::new(data, device)?;
@@ -1078,61 +1035,33 @@ fn randn(device: &Device) -> Result<()> {
     Ok(())
 }
 
-test_device!(zeros, zeros_cpu, zeros_gpu, zeros_metal);
-test_device!(ones, ones_cpu, ones_gpu, ones_metal);
-test_device!(full, full_cpu, full_gpu, full_metal);
-test_device!(arange, arange_cpu, arange_gpu, arange_metal);
-test_device!(add_mul, add_mul_cpu, add_mul_gpu, add_mul_metal);
-test_device!(tensor_2d, tensor_2d_cpu, tensor_2d_gpu, tensor_2d_metal);
-test_device!(narrow, narrow_cpu, narrow_gpu, narrow_metal);
-test_device!(broadcast, broadcast_cpu, broadcast_gpu, broadcast_metal);
-test_device!(cat, cat_cpu, cat_gpu, cat_metal);
-test_device!(sum, sum_cpu, sum_gpu, sum_metal);
-test_device!(min, min_cpu, min_gpu, min_metal);
-test_device!(max, max_cpu, max_gpu, max_metal);
-test_device!(argmax, argmax_cpu, argmax_gpu, argmax_metal);
-test_device!(argmin, argmin_cpu, argmin_gpu, argmin_metal);
-test_device!(transpose, transpose_cpu, transpose_gpu, transpose_metal);
-test_device!(unary_op, unary_op_cpu, unary_op_gpu, unary_op_metal);
-test_device!(binary_op, binary_op_cpu, binary_op_gpu, binary_op_metal);
-test_device!(embeddings, embeddings_cpu, embeddings_gpu, embeddings_metal);
-test_device!(cmp, cmp_cpu, cmp_gpu, cmp_metal);
-test_device!(matmul, matmul_cpu, matmul_gpu, matmul_metal);
-test_device!(
-    broadcast_matmul,
-    broadcast_matmul_cpu,
-    broadcast_matmul_gpu,
-    broadcast_matmul_metal
-);
-test_device!(
-    broadcasting,
-    broadcasting_cpu,
-    broadcasting_gpu,
-    broadcasting_metal
-);
-test_device!(
-    index_select,
-    index_select_cpu,
-    index_select_gpu,
-    index_select_metal
-);
-test_device!(index_add, index_add_cpu, index_add_gpu, index_add_metal);
-test_device!(gather, gather_cpu, gather_gpu, gather_metal);
-test_device!(
-    scatter_add,
-    scatter_add_cpu,
-    scatter_add_gpu,
-    scatter_add_metal
-);
-test_device!(
-    slice_scatter,
-    slice_scatter_cpu,
-    slice_scatter_gpu,
-    slice_scatter_metal
-);
-test_device!(randn, randn_cpu, randn_gpu, randn_metal);
-test_device!(clamp, clamp_cpu, clamp_gpu, clamp_metal);
-test_device!(var, var_cpu, var_gpu, var_metal);
+test_device!(zeros, zeros_cpu, zeros_gpu);
+test_device!(ones, ones_cpu, ones_gpu);
+test_device!(add_mul, add_mul_cpu, add_mul_gpu);
+test_device!(tensor_2d, tensor_2d_cpu, tensor_2d_gpu);
+test_device!(narrow, narrow_cpu, narrow_gpu);
+test_device!(broadcast, broadcast_cpu, broadcast_gpu);
+test_device!(cat, cat_cpu, cat_gpu);
+test_device!(sum, sum_cpu, sum_gpu);
+test_device!(min, min_cpu, min_gpu);
+test_device!(max, max_cpu, max_gpu);
+test_device!(argmax, argmax_cpu, argmax_gpu);
+test_device!(argmin, argmin_cpu, argmin_gpu);
+test_device!(transpose, transpose_cpu, transpose_gpu);
+test_device!(unary_op, unary_op_cpu, unary_op_gpu);
+test_device!(binary_op, binary_op_cpu, binary_op_gpu);
+test_device!(embeddings, embeddings_cpu, embeddings_gpu);
+test_device!(cmp, cmp_cpu, cmp_gpu);
+test_device!(matmul, matmul_cpu, matmul_gpu);
+test_device!(broadcast_matmul, broadcast_matmul_cpu, broadcast_matmul_gpu);
+test_device!(broadcasting, broadcasting_cpu, broadcasting_gpu);
+test_device!(index_select, index_select_cpu, index_select_gpu);
+test_device!(index_add, index_add_cpu, index_add_gpu);
+test_device!(gather, gather_cpu, gather_gpu);
+test_device!(scatter_add, scatter_add_cpu, scatter_add_gpu);
+test_device!(slice_scatter, slice_scatter_cpu, slice_scatter_gpu);
+test_device!(randn, randn_cpu, randn_gpu);
+test_device!(clamp, clamp_cpu, clamp_gpu);
 
 // There was originally a bug on the CPU implementation for randn
 // https://github.com/huggingface/candle/issues/381
@@ -1144,112 +1073,3 @@ fn randn_hasneg() -> Result<()> {
     }
     Ok(())
 }
-
-#[test]
-fn pad_with_same() -> Result<()> {
-    let t = Tensor::arange(1f32, 5f32, &Device::Cpu)?.reshape((2, 2))?;
-    let t0 = t.pad_with_same(0, 1, 2)?;
-    assert_eq!(
-        t0.to_vec2::<f32>()?,
-        [[1.0, 2.0], [1.0, 2.0], [3.0, 4.0], [3.0, 4.0], [3.0, 4.0]]
-    );
-    let t1 = t.pad_with_same(1, 1, 2)?;
-    assert_eq!(
-        t1.to_vec2::<f32>()?,
-        [[1.0, 1.0, 2.0, 2.0, 2.0], [3.0, 3.0, 4.0, 4.0, 4.0]]
-    );
-    Ok(())
-}
-
-#[test]
-fn i64_abs() -> Result<()> {
-    let t = Tensor::new(&[-42i64, 1337], &Device::Cpu)?;
-    let t = t.abs()?;
-    assert_eq!(t.to_vec1::<i64>()?, [42, 1337]);
-    Ok(())
-}
-
-#[test]
-fn tril_triu_eye() -> Result<()> {
-    let t = Tensor::tril2(4, DType::F32, &Device::Cpu)?;
-    assert_eq!(
-        t.to_vec2::<f32>()?,
-        [
-            [1.0, 0.0, 0.0, 0.0],
-            [1.0, 1.0, 0.0, 0.0],
-            [1.0, 1.0, 1.0, 0.0],
-            [1.0, 1.0, 1.0, 1.0]
-        ],
-    );
-    let t = Tensor::triu2(4, DType::F32, &Device::Cpu)?;
-    assert_eq!(
-        t.to_vec2::<f32>()?,
-        [
-            [1.0, 1.0, 1.0, 1.0],
-            [0.0, 1.0, 1.0, 1.0],
-            [0.0, 0.0, 1.0, 1.0],
-            [0.0, 0.0, 0.0, 1.0]
-        ]
-    );
-    let t = Tensor::eye(4, DType::F32, &Device::Cpu)?;
-    assert_eq!(
-        t.to_vec2::<f32>()?,
-        [
-            [1.0, 0.0, 0.0, 0.0],
-            [0.0, 1.0, 0.0, 0.0],
-            [0.0, 0.0, 1.0, 0.0],
-            [0.0, 0.0, 0.0, 1.0]
-        ]
-    );
-    Ok(())
-}
-
-#[test]
-fn cumsum() -> Result<()> {
-    let t = &[3f32, 1., 4., 1., 5.];
-    let t = Tensor::new(t, &Device::Cpu)?;
-    assert_eq!(t.cumsum(0)?.to_vec1::<f32>()?, [3., 4., 8., 9., 14.]);
-    let t = t.unsqueeze(1)?;
-    assert_eq!(
-        t.cumsum(0)?.to_vec2::<f32>()?,
-        [[3.0], [4.0], [8.0], [9.0], [14.0]]
-    );
-    assert_eq!(
-        t.cumsum(1)?.to_vec2::<f32>()?,
-        [[3.0], [1.0], [4.0], [1.0], [5.0]]
-    );
-    let t = &[[3f32, 1., 4., 1., 5.], [2., 1., 7., 8., 2.]];
-    let t = Tensor::new(t, &Device::Cpu)?;
-    assert_eq!(
-        t.cumsum(1)?.to_vec2::<f32>()?,
-        [[3.0, 4.0, 8.0, 9.0, 14.0], [2.0, 3.0, 10.0, 18.0, 20.0]],
-    );
-    assert_eq!(
-        t.cumsum(0)?.to_vec2::<f32>()?,
-        [[3.0, 1.0, 4.0, 1.0, 5.0], [5.0, 2.0, 11.0, 9.0, 7.0]]
-    );
-    Ok(())
-}
-
-/// A helper function for floating point comparison. Both a and b must be 1D Tensor and contains the same amount of data.
-/// Assertion passes if the difference of all pairs of a and b is smaller than epsilon.
-fn assert_close(a: &Tensor, b: &Tensor, epsilon: f64) -> Result<()> {
-    let a_vec: Vec<f64> = a.to_vec1()?;
-    let b_vec: Vec<f64> = b.to_vec1()?;
-
-    assert_eq!(a_vec.len(), b_vec.len());
-    for (a, b) in a_vec.iter().zip(b_vec.iter()) {
-        assert!((a - b).abs() < epsilon);
-    }
-    Ok(())
-}
-
-#[test]
-fn logsumexp() -> Result<()> {
-    let input = Tensor::new(&[[1f64, 2., 3.], [4., 5., 6.]], &Device::Cpu)?;
-    let output = input.logsumexp(D::Minus1)?;
-    // The expectations obtained from pytorch.
-    let expected = Tensor::new(&[3.4076, 6.4076], &Device::Cpu)?;
-    assert_close(&output, &expected, 0.00001)?;
-    Ok(())
-}

candle-datasets/Cargo.toml

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

candle-datasets/src/vision/cifar.rs

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

candle-examples/Cargo.toml

@@ -11,21 +11,16 @@ readme = "README.md"
 
 [dependencies]
 accelerate-src = { workspace = true, optional = true }
-candle = { workspace = true }
-candle-datasets = { workspace = true }
-candle-nn = { workspace = true }
-candle-transformers = { workspace = true }
-candle-flash-attn = { workspace = true, optional = true }
-candle-onnx = { workspace = true, optional = true }
-
-csv = "1.3.0"
+candle = { path = "../candle-core", version = "0.3.0", package = "candle-core" }
+candle-datasets = { path = "../candle-datasets", version = "0.3.0" }
+candle-nn = { path = "../candle-nn", version = "0.3.0" }
+candle-transformers = { path = "../candle-transformers", version = "0.3.0" }
+candle-flash-attn = { path = "../candle-flash-attn", version = "0.3.0", optional = true }
 cudarc = { workspace = true, optional = true }
 half = { workspace = true, optional = true }
-hf-hub = { workspace = true, features=["tokio"]}
 image = { workspace = true }
 intel-mkl-src = { workspace = true, optional = true }
 num-traits = { workspace = true }
-pyo3 = { version = "0.20.0", features = ["auto-initialize"], optional = true }
 rayon = { workspace = true }
 safetensors = { workspace = true }
 serde = { workspace = true }
@@ -36,6 +31,7 @@ tokenizers = { workspace = true, features = ["onig"] }
 anyhow = { workspace = true }
 byteorder = { workspace = true }
 clap = { workspace = true }
+hf-hub = { workspace = true, features=["tokio"]}
 imageproc = { workspace = true }
 memmap2 = { workspace = true }
 rand = { workspace = true }
@@ -49,31 +45,16 @@ 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", "dep:bindgen_cuda"]
+cuda = ["candle/cuda", "candle-nn/cuda", "candle-transformers/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"]
 
 [[example]]
 name = "llama_multiprocess"
 required-features = ["cuda", "nccl", "flash-attn"]
-
-[[example]]
-name = "reinforcement-learning"
-required-features = ["pyo3"]
-
-[[example]]
-name = "onnx"
-required-features = ["onnx"]
-
-[[example]]
-name = "onnx_basics"
-required-features = ["onnx"]

candle-examples/build.rs

@@ -4,34 +4,235 @@ use std::io::Write;
 use std::path::PathBuf;
 
 struct KernelDirectories {
-    kernel_glob: &'static str,
+    kernel_dir: &'static str,
     rust_target: &'static str,
     include_dirs: &'static [&'static str],
 }
 
-const KERNEL_DIRS: [KernelDirectories; 1] = [KernelDirectories {
-    kernel_glob: "examples/custom-ops/kernels/*.cu",
+const DIRS: [KernelDirectories; 1] = [KernelDirectories {
+    kernel_dir: "examples/custom-ops/kernels/",
     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")]
-    {
-        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()
-        }
-    }
+    set_cuda_include_dir()?;
+    #[cfg(feature = "cuda")]
+    let compute_cap = compute_cap()?;
     #[cfg(not(feature = "cuda"))]
-    {
-        for kdir in KERNEL_DIRS.iter() {
-            let _file = std::fs::File::create(kdir.rust_target)?;
-        }
+    let compute_cap = 0;
+    for d in DIRS {
+        d.process(&out_dir, compute_cap)?
     }
     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
@@ -24,48 +24,6 @@ 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

@@ -3,13 +3,13 @@ extern crate intel_mkl_src;
 
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
-use candle_transformers::models::bert::{BertModel, Config, HiddenAct, DTYPE};
+use candle_transformers::models::bert::{BertModel, Config, DTYPE};
 
-use anyhow::{Error as E, Result};
+use anyhow::{anyhow, Error as E, Result};
 use candle::Tensor;
 use candle_nn::VarBuilder;
 use clap::Parser;
-use hf_hub::{api::sync::Api, Repo, RepoType};
+use hf_hub::{api::sync::Api, Cache, Repo, RepoType};
 use tokenizers::{PaddingParams, Tokenizer};
 
 #[derive(Parser, Debug)]
@@ -19,6 +19,10 @@ struct Args {
     #[arg(long)]
     cpu: bool,
 
+    /// Run offline (you must have the files already cached)
+    #[arg(long)]
+    offline: bool,
+
     /// Enable tracing (generates a trace-timestamp.json file).
     #[arg(long)]
     tracing: bool,
@@ -34,10 +38,6 @@ struct Args {
     #[arg(long)]
     prompt: Option<String>,
 
-    /// Use the pytorch weights rather than the safetensors ones
-    #[arg(long)]
-    use_pth: bool,
-
     /// The number of times to run the prompt.
     #[arg(long, default_value = "1")]
     n: usize,
@@ -45,10 +45,6 @@ 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 {
@@ -64,30 +60,34 @@ impl Args {
         };
 
         let repo = Repo::with_revision(model_id, RepoType::Model, revision);
-        let (config_filename, tokenizer_filename, weights_filename) = {
+        let (config_filename, tokenizer_filename, weights_filename) = if self.offline {
+            let cache = Cache::default().repo(repo);
+            (
+                cache
+                    .get("config.json")
+                    .ok_or(anyhow!("Missing config file in cache"))?,
+                cache
+                    .get("tokenizer.json")
+                    .ok_or(anyhow!("Missing tokenizer file in cache"))?,
+                cache
+                    .get("model.safetensors")
+                    .ok_or(anyhow!("Missing weights file in cache"))?,
+            )
+        } else {
             let api = Api::new()?;
             let api = api.repo(repo);
-            let config = api.get("config.json")?;
-            let tokenizer = api.get("tokenizer.json")?;
-            let weights = if self.use_pth {
-                api.get("pytorch_model.bin")?
-            } else {
-                api.get("model.safetensors")?
-            };
-            (config, tokenizer, weights)
+            (
+                api.get("config.json")?,
+                api.get("tokenizer.json")?,
+                api.get("model.safetensors")?,
+            )
         };
         let config = std::fs::read_to_string(config_filename)?;
-        let mut config: Config = serde_json::from_str(&config)?;
+        let config: Config = serde_json::from_str(&config)?;
         let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
 
-        let vb = if self.use_pth {
-            VarBuilder::from_pth(&weights_filename, DTYPE, &device)?
-        } else {
-            unsafe { VarBuilder::from_mmaped_safetensors(&[weights_filename], DTYPE, &device)? }
-        };
-        if self.approximate_gelu {
-            config.hidden_act = HiddenAct::GeluApproximate;
-        }
+        let vb =
+            unsafe { VarBuilder::from_mmaped_safetensors(&[weights_filename], DTYPE, &device)? };
         let model = BertModel::load(vb, &config)?;
         Ok((model, tokenizer))
     }

candle-examples/examples/blip/README.md

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

candle-examples/examples/blip/main.rs

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

candle-examples/examples/convmixer/main.rs

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

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

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

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

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

candle-examples/examples/distilbert/README.md

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

candle-examples/examples/distilbert/main.rs

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

candle-examples/examples/falcon/main.rs

@@ -165,7 +165,14 @@ fn main() -> Result<()> {
         args.revision,
     ));
     let tokenizer_filename = repo.get("tokenizer.json")?;
-    let filenames = candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.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);
+    }
     println!("retrieved the files in {:?}", start.elapsed());
     let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
 

candle-examples/examples/jina-bert/README.md

@@ -1,45 +0,0 @@
-# candle-jina-bert
-
-Jina-Bert is a general large language model with a context size of 8192, [model
-card](https://huggingface.co/jinaai/jina-embeddings-v2-base-en). 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
-
-Jina-Bert is used to compute the sentence embeddings for a prompt. The model weights
-are downloaded from the hub on the first run.
-
-```bash
-cargo run --example jina-bert --release -- --prompt "Here is a test sentence"
-
-> [[[ 0.1595, -0.9885,  0.6494, ...,  0.3003, -0.6901, -1.2355],
->   [ 0.0374, -0.1798,  1.3359, ...,  0.6731,  0.2133, -1.6807],
->   [ 0.1700, -0.8534,  0.8924, ..., -0.1785, -0.0727, -1.5087],
->   ...
->   [-0.3113, -1.3665,  0.2027, ..., -0.2519,  0.1711, -1.5811],
->   [ 0.0907, -1.0492,  0.5382, ...,  0.0242, -0.7077, -1.0830],
->   [ 0.0369, -0.6343,  0.6105, ...,  0.0671,  0.3778, -1.1505]]]
-> Tensor[[1, 7, 768], f32]
-```
-
-## Similarities
-
-In this example, Jina-Bert is used to compute the sentence embeddings for a set of
-sentences (hardcoded in the examples). Then cosine similarities are computed for
-each sentence pair and they are reported by decreasing values, hence the first
-reported pair contains the two sentences that have the highest similarity score.
-The sentence embeddings are computed using average pooling through all the
-sentence tokens, including some potential padding.
-
-```bash
-cargo run --example jina-bert --release
-
-> score: 0.94 'The new movie is awesome' 'The new movie is so great'
-> score: 0.81 'The cat sits outside' 'The cat plays in the garden'
-> score: 0.78 'I love pasta' 'Do you like pizza?'
-> score: 0.68 'I love pasta' 'The new movie is awesome'
-> score: 0.67 'A man is playing guitar' 'A woman watches TV'
-```

candle-examples/examples/jina-bert/main.rs

@@ -1,180 +0,0 @@
-#[cfg(feature = "mkl")]
-extern crate intel_mkl_src;
-
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-
-use candle_transformers::models::jina_bert::{BertModel, Config};
-
-use anyhow::Error as E;
-use candle::{DType, Module, Tensor};
-use candle_nn::VarBuilder;
-use clap::Parser;
-
-#[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,
-
-    /// When set, compute embeddings for this prompt.
-    #[arg(long)]
-    prompt: Option<String>,
-
-    /// The number of times to run the prompt.
-    #[arg(long, default_value = "1")]
-    n: usize,
-
-    /// L2 normalization for embeddings.
-    #[arg(long, default_value = "true")]
-    normalize_embeddings: bool,
-
-    #[arg(long)]
-    tokenizer: Option<String>,
-
-    #[arg(long)]
-    model: Option<String>,
-}
-
-impl Args {
-    fn build_model_and_tokenizer(&self) -> anyhow::Result<(BertModel, tokenizers::Tokenizer)> {
-        use hf_hub::{api::sync::Api, Repo, RepoType};
-        let model = match &self.model {
-            Some(model_file) => std::path::PathBuf::from(model_file),
-            None => Api::new()?
-                .repo(Repo::new(
-                    "jinaai/jina-embeddings-v2-base-en".to_string(),
-                    RepoType::Model,
-                ))
-                .get("model.safetensors")?,
-        };
-        let tokenizer = match &self.tokenizer {
-            Some(file) => std::path::PathBuf::from(file),
-            None => Api::new()?
-                .repo(Repo::new(
-                    "sentence-transformers/all-MiniLM-L6-v2".to_string(),
-                    RepoType::Model,
-                ))
-                .get("tokenizer.json")?,
-        };
-        let device = candle_examples::device(self.cpu)?;
-        let config = Config::v2_base();
-        let tokenizer = tokenizers::Tokenizer::from_file(tokenizer).map_err(E::msg)?;
-        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model], DType::F32, &device)? };
-        let model = BertModel::new(vb, &config)?;
-        Ok((model, tokenizer))
-    }
-}
-
-fn main() -> anyhow::Result<()> {
-    use tracing_chrome::ChromeLayerBuilder;
-    use tracing_subscriber::prelude::*;
-
-    let args = Args::parse();
-    let _guard = if args.tracing {
-        println!("tracing...");
-        let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
-        tracing_subscriber::registry().with(chrome_layer).init();
-        Some(guard)
-    } else {
-        None
-    };
-    let start = std::time::Instant::now();
-
-    let (model, mut tokenizer) = args.build_model_and_tokenizer()?;
-    let device = &model.device;
-
-    if let Some(prompt) = args.prompt {
-        let tokenizer = tokenizer
-            .with_padding(None)
-            .with_truncation(None)
-            .map_err(E::msg)?;
-        let tokens = tokenizer
-            .encode(prompt, true)
-            .map_err(E::msg)?
-            .get_ids()
-            .to_vec();
-        let token_ids = Tensor::new(&tokens[..], device)?.unsqueeze(0)?;
-        println!("Loaded and encoded {:?}", start.elapsed());
-        for idx in 0..args.n {
-            let start = std::time::Instant::now();
-            let ys = model.forward(&token_ids)?;
-            if idx == 0 {
-                println!("{ys}");
-            }
-            println!("Took {:?}", start.elapsed());
-        }
-    } else {
-        let sentences = [
-            "The cat sits outside",
-            "A man is playing guitar",
-            "I love pasta",
-            "The new movie is awesome",
-            "The cat plays in the garden",
-            "A woman watches TV",
-            "The new movie is so great",
-            "Do you like pizza?",
-        ];
-        let n_sentences = sentences.len();
-        if let Some(pp) = tokenizer.get_padding_mut() {
-            pp.strategy = tokenizers::PaddingStrategy::BatchLongest
-        } else {
-            let pp = tokenizers::PaddingParams {
-                strategy: tokenizers::PaddingStrategy::BatchLongest,
-                ..Default::default()
-            };
-            tokenizer.with_padding(Some(pp));
-        }
-        let tokens = tokenizer
-            .encode_batch(sentences.to_vec(), true)
-            .map_err(E::msg)?;
-        let token_ids = tokens
-            .iter()
-            .map(|tokens| {
-                let tokens = tokens.get_ids().to_vec();
-                Tensor::new(tokens.as_slice(), device)
-            })
-            .collect::<candle::Result<Vec<_>>>()?;
-
-        let token_ids = Tensor::stack(&token_ids, 0)?;
-        println!("running inference on batch {:?}", token_ids.shape());
-        let embeddings = model.forward(&token_ids)?;
-        println!("generated embeddings {:?}", embeddings.shape());
-        // Apply some avg-pooling by taking the mean embedding value for all tokens (including padding)
-        let (_n_sentence, n_tokens, _hidden_size) = embeddings.dims3()?;
-        let embeddings = (embeddings.sum(1)? / (n_tokens as f64))?;
-        let embeddings = if args.normalize_embeddings {
-            normalize_l2(&embeddings)?
-        } else {
-            embeddings
-        };
-        println!("pooled embeddings {:?}", embeddings.shape());
-
-        let mut similarities = vec![];
-        for i in 0..n_sentences {
-            let e_i = embeddings.get(i)?;
-            for j in (i + 1)..n_sentences {
-                let e_j = embeddings.get(j)?;
-                let sum_ij = (&e_i * &e_j)?.sum_all()?.to_scalar::<f32>()?;
-                let sum_i2 = (&e_i * &e_i)?.sum_all()?.to_scalar::<f32>()?;
-                let sum_j2 = (&e_j * &e_j)?.sum_all()?.to_scalar::<f32>()?;
-                let cosine_similarity = sum_ij / (sum_i2 * sum_j2).sqrt();
-                similarities.push((cosine_similarity, i, j))
-            }
-        }
-        similarities.sort_by(|u, v| v.0.total_cmp(&u.0));
-        for &(score, i, j) in similarities[..5].iter() {
-            println!("score: {score:.2} '{}' '{}'", sentences[i], sentences[j])
-        }
-    }
-    Ok(())
-}
-
-pub fn normalize_l2(v: &Tensor) -> candle::Result<Tensor> {
-    v.broadcast_div(&v.sqr()?.sum_keepdim(1)?.sqrt()?)
-}

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, ValueEnum};
+use clap::Parser;
 
 use candle::{DType, Tensor};
 use candle_nn::VarBuilder;
@@ -22,21 +22,11 @@ use hf_hub::{api::sync::Api, Repo, RepoType};
 use std::io::Write;
 
 use candle_transformers::models::llama as model;
-use model::{Llama, LlamaConfig};
+use model::{Config, 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 {
@@ -44,6 +34,10 @@ 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>,
@@ -82,13 +76,17 @@ struct Args {
     #[arg(long)]
     revision: Option<String>,
 
-    /// The model size to use.
-    #[arg(long, default_value = "v2")]
-    which: Which,
+    #[arg(long)]
+    v1: bool,
 
     #[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,
@@ -120,34 +118,65 @@ fn main() -> Result<()> {
         Some(dtype) => bail!("Unsupported dtype {dtype}"),
         None => DType::F16,
     };
-    let (llama, tokenizer_filename, 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 (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 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 tokenizer_filename = match &args.local_weights {
+                Some(path) => (path.to_owned() + "tokenizer.json").into(),
+                _ => api.get("tokenizer.json")?,
+            };
 
-        let filenames = match args.which {
-            Which::V1 | Which::V2 | Which::Solar10_7B => {
-                candle_examples::hub_load_safetensors(&api, "model.safetensors.index.json")?
+            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);
+                    }
+                };
             }
-            Which::TinyLlama1_1BChat => vec![api.get("model.safetensors")?],
-        };
-        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)
+            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 tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
     let eos_token_id = tokenizer.token_to_id(EOS_TOKEN);
@@ -165,14 +194,14 @@ fn main() -> Result<()> {
     let mut index_pos = 0;
     let mut token_generated = 0;
     for index in 0..args.sample_len {
-        let (context_size, context_index) = if cache.use_kv_cache && index > 0 {
-            (1, index_pos)
+        let context_size = if cache.use_kv_cache && index > 0 {
+            1
         } else {
-            (tokens.len(), 0)
+            tokens.len()
         };
         let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
         let input = Tensor::new(ctxt, &device)?.unsqueeze(0)?;
-        let logits = llama.forward(&input, context_index)?;
+        let logits = llama.forward(&input, index_pos)?;
         let logits = logits.squeeze(0)?;
         let logits = if args.repeat_penalty == 1. {
             logits

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

@@ -6,10 +6,9 @@ extern crate accelerate_src;
 #[cfg(feature = "mkl")]
 extern crate intel_mkl_src;
 
-use candle_transformers::models::llama2_c as model;
-use candle_transformers::models::llama2_c_weights as weights;
-use candle_transformers::models::quantized_llama2_c as qmodel;
+mod model;
 mod training;
+mod weights;
 use clap::{Parser, Subcommand};
 
 use anyhow::{Error as E, Result};
@@ -20,7 +19,6 @@ use std::io::Write;
 use tokenizers::Tokenizer;
 
 use model::{Config, Llama};
-use qmodel::QLlama;
 use weights::TransformerWeights;
 
 #[derive(Parser, Debug, Clone)]
@@ -154,20 +152,6 @@ fn main() -> anyhow::Result<()> {
     Ok(())
 }
 
-enum Model {
-    Llama(Llama),
-    QLlama(QLlama),
-}
-
-impl Model {
-    fn forward(&self, xs: &Tensor, pos: usize) -> anyhow::Result<Tensor> {
-        match self {
-            Self::Llama(l) => Ok(l.forward(xs, pos)?),
-            Self::QLlama(l) => Ok(l.forward(xs, pos)?),
-        }
-    }
-}
-
 fn run_eval(args: &EvaluationCmd, common_args: &Args) -> Result<()> {
     use std::io::BufRead;
 
@@ -257,66 +241,24 @@ fn run_inference(args: &InferenceCmd, common_args: &Args) -> Result<()> {
 
     let device = candle_examples::device(common_args.cpu)?;
 
-    let is_gguf = config_path.extension().map_or(false, |v| v == "gguf");
     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 (_vocab_size, dim) = vb
-            .get_no_shape("model.embed_tokens.weight")?
-            .shape()
-            .dims2()?;
-        let config = match dim {
-            64 => Config::tiny_260k(),
-            288 => Config::tiny_15m(),
-            512 => Config::tiny_42m(),
-            768 => Config::tiny_110m(),
-            _ => anyhow::bail!("no config for dim {dim}"),
-        };
-        let freq_cis_real = vb
-            .get(
-                (config.seq_len, config.head_size() / 2),
-                "rot.freq_cis_real",
-            )?
-            .dequantize(&candle::Device::Cpu)?;
-        let freq_cis_imag = vb
-            .get(
-                (config.seq_len, config.head_size() / 2),
-                "rot.freq_cis_imag",
-            )?
-            .dequantize(&candle::Device::Cpu)?;
-
-        let fake_vb = candle_nn::VarBuilder::from_tensors(
-            [
-                ("freq_cis_real".to_string(), freq_cis_real),
-                ("freq_cis_imag".to_string(), freq_cis_imag),
-            ]
-            .into_iter()
-            .collect(),
-            candle::DType::F32,
-            &candle::Device::Cpu,
-        );
-        let cache = model::Cache::new(true, &config, fake_vb)?;
-        let model = Model::QLlama(QLlama::load(vb, &cache, config.clone())?);
-        (model, config)
-    } else if is_safetensors {
-        let config = Config::tiny_15m();
+    let (vb, config) = if is_safetensors {
+        let config = Config::tiny();
         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)
+        (vb, config)
     } else {
         let mut file = std::fs::File::open(config_path)?;
         let config = Config::from_reader(&mut file)?;
         println!("{config:?}");
         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)
+        (vb, config)
     };
+    let cache = model::Cache::new(true, &config, vb.pp("rot"))?;
+    let model = Llama::load(vb, &cache, config)?;
 
     println!("starting the inference loop");
     let mut logits_processor = LogitsProcessor::new(299792458, args.temperature, args.top_p);
@@ -331,7 +273,7 @@ fn run_inference(args: &InferenceCmd, common_args: &Args) -> Result<()> {
 
     let start_gen = std::time::Instant::now();
     for index in 0.. {
-        if tokens.len() >= config.seq_len {
+        if tokens.len() >= model.config.seq_len {
             break;
         }
         let context_size = if index > 0 { 1 } else { tokens.len() };

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

@@ -17,20 +17,7 @@ pub struct Config {
 }
 
 impl Config {
-    pub fn tiny_260k() -> Self {
-        Self {
-            dim: 64,
-            hidden_dim: 768,
-            n_layers: 5,
-            n_heads: 8,
-            n_kv_heads: 4,
-            vocab_size: 32000,
-            seq_len: 512,
-            norm_eps: 1e-5,
-        }
-    }
-
-    pub fn tiny_15m() -> Self {
+    pub fn tiny() -> Self {
         Self {
             dim: 288,
             hidden_dim: 768,
@@ -42,32 +29,6 @@ impl Config {
             norm_eps: 1e-5,
         }
     }
-
-    pub fn tiny_42m() -> Self {
-        Self {
-            dim: 512,
-            hidden_dim: 768,
-            n_layers: 8,
-            n_heads: 8,
-            n_kv_heads: 8,
-            vocab_size: 32000,
-            seq_len: 1024,
-            norm_eps: 1e-5,
-        }
-    }
-
-    pub fn tiny_110m() -> Self {
-        Self {
-            dim: 768,
-            hidden_dim: 768,
-            n_layers: 12,
-            n_heads: 12,
-            n_kv_heads: 12,
-            vocab_size: 32000,
-            seq_len: 1024,
-            norm_eps: 1e-5,
-        }
-    }
 }
 
 #[derive(Clone)]
@@ -75,9 +36,9 @@ pub struct Cache {
     masks: Arc<Mutex<HashMap<usize, Tensor>>>,
     pub use_kv_cache: bool,
     #[allow(clippy::type_complexity)]
-    pub kvs: Arc<Mutex<Vec<Option<(Tensor, Tensor)>>>>,
-    pub cos: Tensor,
-    pub sin: Tensor,
+    kvs: Arc<Mutex<Vec<Option<(Tensor, Tensor)>>>>,
+    cos: Tensor,
+    sin: Tensor,
     device: Device,
 }
 
@@ -114,7 +75,7 @@ impl Cache {
         })
     }
 
-    pub fn mask(&self, t: usize) -> Result<Tensor> {
+    fn mask(&self, t: usize) -> Result<Tensor> {
         let mut masks = self.masks.lock().unwrap();
         if let Some(mask) = masks.get(&t) {
             Ok(mask.clone())

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

@@ -33,7 +33,7 @@ pub fn run(args: &crate::TrainingCmd, common_args: &crate::Args) -> Result<()> {
     );
     let varmap = candle_nn::VarMap::new();
     let vb = candle_nn::VarBuilder::from_varmap(&varmap, DType::F32, &device);
-    let config = Config::tiny_15m();
+    let config = Config::tiny();
     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);
 

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

@@ -1,8 +1,9 @@
+use anyhow::Result;
 use byteorder::{LittleEndian, ReadBytesExt};
-use candle::{DType, Device, IndexOp, Result, Shape, Tensor};
+use candle::{DType, Device, IndexOp, Shape, Tensor};
 use candle_nn::VarBuilder;
 
-use super::llama2_c::Config;
+use crate::model::Config;
 
 pub struct TransformerWeights {
     // token embedding table

candle-examples/examples/llama_multiprocess/main.rs

@@ -143,7 +143,14 @@ 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 filenames = candle_examples::hub_load_safetensors(&api, "model.safetensors.index.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);
+    }
 
     if args.rank.is_none() {
         let children: Vec<_> = (0..args.num_shards)

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

@@ -1,12 +0,0 @@
-# candle-mamba-minimal: minimal implementation of Mamba
-
-This is based on [mamba-minimal](https://github.com/johnma2006/mamba-minimal).
-
-## 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

@@ -1,287 +0,0 @@
-#[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

@@ -1,204 +0,0 @@
-/// 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/marian-mt/README.md

@@ -1,38 +0,0 @@
-# candle-marian-mt
-
-`marian-mt` is a neural machine translation model. In this example it is used to
-translate text from French to English. See the associated [model
-card](https://huggingface.co/Helsinki-NLP/opus-mt-tc-big-fr-en) for details on
-the model itself.
-
-## Running an example
-
-```bash
-cargo run --example marian-mt --release -- \
-    --text "Demain, dès l'aube, à l'heure où blanchit la campagne, Je partirai. Vois-tu, je sais que tu m'attends. J'irai par la forêt, j'irai par la montagne. Je ne puis demeurer loin de toi plus longtemps."
-```
-
-```
-<NIL> Tomorrow, at dawn, at the time when the country is whitening, I will go. See,
-I know you are waiting for me. I will go through the forest, I will go through the
-mountain. I cannot stay far from you any longer.</s>
-```
-
-## Generating the tokenizer.json files
-
-You can use the following script to generate the `tokenizer.json` config files
-from the hf-hub repos. This requires the `tokenizers` and `sentencepiece`
-packages to be install and use the `convert_slow_tokenizer.py` script from this
-directory.
-
-```python
-from convert_slow_tokenizer import MarianConverter
-from transformers import AutoTokenizer
-
-
-tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-fr-en", use_fast=False)
-fast_tokenizer = MarianConverter(tokenizer, index=0).converted()
-fast_tokenizer.save(f"tokenizer-marian-base-fr.json")
-fast_tokenizer = MarianConverter(tokenizer, index=1).converted()
-fast_tokenizer.save(f"tokenizer-marian-base-en.json")
-```

candle-examples/examples/marian-mt/main.rs

@@ -1,152 +0,0 @@
-#[cfg(feature = "mkl")]
-extern crate intel_mkl_src;
-
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-
-use anyhow::Error as E;
-use clap::{Parser, ValueEnum};
-
-use candle::{DType, Tensor};
-use candle_examples::token_output_stream::TokenOutputStream;
-use candle_nn::VarBuilder;
-use candle_transformers::models::marian;
-
-use tokenizers::Tokenizer;
-
-#[derive(Clone, Debug, Copy, ValueEnum)]
-enum Which {
-    Base,
-    Big,
-}
-
-// TODO: Maybe add support for the conditional prompt.
-#[derive(Parser)]
-struct Args {
-    #[arg(long)]
-    model: Option<String>,
-
-    #[arg(long)]
-    tokenizer: Option<String>,
-
-    #[arg(long)]
-    tokenizer_dec: Option<String>,
-
-    /// Choose the variant of the model to run.
-    #[arg(long, default_value = "big")]
-    which: Which,
-
-    /// Run on CPU rather than on GPU.
-    #[arg(long)]
-    cpu: bool,
-
-    /// Use the quantized version of the model.
-    #[arg(long)]
-    quantized: bool,
-
-    /// Text to be translated
-    #[arg(long)]
-    text: String,
-}
-
-pub fn main() -> anyhow::Result<()> {
-    use hf_hub::api::sync::Api;
-    let args = Args::parse();
-
-    let config = match args.which {
-        Which::Base => marian::Config::opus_mt_fr_en(),
-        Which::Big => marian::Config::opus_mt_tc_big_fr_en(),
-    };
-    let tokenizer = {
-        let tokenizer = match args.tokenizer {
-            Some(tokenizer) => std::path::PathBuf::from(tokenizer),
-            None => {
-                let name = match args.which {
-                    Which::Base => "tokenizer-marian-base-fr.json",
-                    Which::Big => "tokenizer-marian-fr.json",
-                };
-                Api::new()?
-                    .model("lmz/candle-marian".to_string())
-                    .get(name)?
-            }
-        };
-        Tokenizer::from_file(&tokenizer).map_err(E::msg)?
-    };
-
-    let tokenizer_dec = {
-        let tokenizer = match args.tokenizer_dec {
-            Some(tokenizer) => std::path::PathBuf::from(tokenizer),
-            None => {
-                let name = match args.which {
-                    Which::Base => "tokenizer-marian-base-en.json",
-                    Which::Big => "tokenizer-marian-en.json",
-                };
-                Api::new()?
-                    .model("lmz/candle-marian".to_string())
-                    .get(name)?
-            }
-        };
-        Tokenizer::from_file(&tokenizer).map_err(E::msg)?
-    };
-    let mut tokenizer_dec = TokenOutputStream::new(tokenizer_dec);
-
-    let device = candle_examples::device(args.cpu)?;
-    let vb = {
-        let model = match args.model {
-            Some(model) => std::path::PathBuf::from(model),
-            None => match args.which {
-                Which::Base => Api::new()?
-                    .repo(hf_hub::Repo::with_revision(
-                        "Helsinki-NLP/opus-mt-fr-en".to_string(),
-                        hf_hub::RepoType::Model,
-                        "refs/pr/4".to_string(),
-                    ))
-                    .get("model.safetensors")?,
-                Which::Big => Api::new()?
-                    .model("Helsinki-NLP/opus-mt-tc-big-fr-en".to_string())
-                    .get("model.safetensors")?,
-            },
-        };
-        unsafe { VarBuilder::from_mmaped_safetensors(&[&model], DType::F32, &device)? }
-    };
-    let mut model = marian::MTModel::new(&config, vb)?;
-
-    let mut logits_processor =
-        candle_transformers::generation::LogitsProcessor::new(1337, None, None);
-
-    let encoder_xs = {
-        let mut tokens = tokenizer
-            .encode(args.text, true)
-            .map_err(E::msg)?
-            .get_ids()
-            .to_vec();
-        tokens.push(config.eos_token_id);
-        let tokens = Tensor::new(tokens.as_slice(), &device)?.unsqueeze(0)?;
-        model.encoder().forward(&tokens, 0)?
-    };
-
-    let mut token_ids = vec![config.decoder_start_token_id];
-    for index in 0..1000 {
-        let context_size = if index >= 1 { 1 } else { token_ids.len() };
-        let start_pos = token_ids.len().saturating_sub(context_size);
-        let input_ids = Tensor::new(&token_ids[start_pos..], &device)?.unsqueeze(0)?;
-        let logits = model.decode(&input_ids, &encoder_xs, start_pos)?;
-        let logits = logits.squeeze(0)?;
-        let logits = logits.get(logits.dim(0)? - 1)?;
-        let token = logits_processor.sample(&logits)?;
-        token_ids.push(token);
-        if let Some(t) = tokenizer_dec.next_token(token)? {
-            use std::io::Write;
-            print!("{t}");
-            std::io::stdout().flush()?;
-        }
-        if token == config.eos_token_id || token == config.forced_eos_token_id {
-            break;
-        }
-    }
-    if let Some(rest) = tokenizer_dec.decode_rest().map_err(E::msg)? {
-        print!("{rest}");
-    }
-    println!();
-    Ok(())
-}

candle-examples/examples/mistral/main.rs

@@ -155,8 +155,8 @@ struct Args {
     #[arg(long, short = 'n', default_value_t = 100)]
     sample_len: usize,
 
-    #[arg(long)]
-    model_id: Option<String>,
+    #[arg(long, default_value = "lmz/candle-mistral")]
+    model_id: String,
 
     #[arg(long, default_value = "main")]
     revision: String,
@@ -207,18 +207,8 @@ 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(
-        model_id,
+        args.model_id,
         RepoType::Model,
         args.revision,
     ));
@@ -235,7 +225,10 @@ fn main() -> Result<()> {
             if args.quantized {
                 vec![repo.get("model-q4k.gguf")?]
             } else {
-                candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.json")?
+                vec![
+                    repo.get("pytorch_model-00001-of-00002.safetensors")?,
+                    repo.get("pytorch_model-00002-of-00002.safetensors")?,
+                ]
             }
         }
     };

candle-examples/examples/mixtral/README.md

@@ -1,25 +0,0 @@
-# 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

@@ -1,241 +0,0 @@
-#[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 = 100)]
-    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/mnist-training/main.rs

@@ -9,7 +9,7 @@ use clap::{Parser, ValueEnum};
 use rand::prelude::*;
 
 use candle::{DType, Result, Tensor, D};
-use candle_nn::{loss, ops, Conv2d, Linear, Module, ModuleT, Optimizer, VarBuilder, VarMap};
+use candle_nn::{loss, ops, Conv2d, Linear, Module, Optimizer, VarBuilder, VarMap};
 
 const IMAGE_DIM: usize = 784;
 const LABELS: usize = 10;
@@ -95,7 +95,7 @@ impl ConvNet {
             .flatten_from(1)?
             .apply(&self.fc1)?
             .relu()?;
-        self.dropout.forward_t(&xs, train)?.apply(&self.fc2)
+        self.dropout.forward(&xs, train)?.apply(&self.fc2)
     }
 }
 

candle-examples/examples/musicgen/encodec_model.rs

@@ -8,7 +8,6 @@ use candle_nn::{conv1d, Conv1d, Conv1dConfig, VarBuilder};
 #[derive(Debug, Clone, PartialEq)]
 enum NormType {
     WeightNorm,
-    TimeGroupNorm,
     None,
 }
 
@@ -269,7 +268,6 @@ impl Module for EncodecConvTranspose1d {
 struct EncodecConv1d {
     causal: bool,
     conv: Conv1d,
-    norm: Option<candle_nn::GroupNorm>,
 }
 
 impl EncodecConv1d {
@@ -294,7 +292,7 @@ impl EncodecConv1d {
                 },
                 vb.pp("conv"),
             )?,
-            NormType::None | NormType::TimeGroupNorm => conv1d(
+            NormType::None => conv1d(
                 in_c,
                 out_c,
                 kernel_size,
@@ -307,17 +305,9 @@ impl EncodecConv1d {
                 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,
         })
     }
 }
@@ -326,10 +316,8 @@ 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),
-        }
+        // If we add support for NormType "time_group_norm", we should add some normalization here.
+        Ok(xs)
     }
 }
 

candle-examples/examples/musicgen/musicgen_model.rs

@@ -321,7 +321,7 @@ impl MusicgenDecoder {
         let positions = self.embed_positions.forward(&input)?.to_device(dev)?;
         let mut xs = inputs_embeds.broadcast_add(&positions)?;
         let attention_mask = self.prepare_decoder_attention_mask(b_sz, seq_len)?;
-        for decoder_layer in self.layers.iter_mut() {
+        for (_layer_idx, decoder_layer) in self.layers.iter_mut().enumerate() {
             xs = decoder_layer.forward(&xs, &attention_mask, None)?;
         }
         let xs = self.layer_norm.forward(&xs)?;

candle-examples/examples/onnx/README.md

@@ -1,10 +0,0 @@
-## Using ONNX models in Candle
-
-This example demonstrates how to run ONNX based models in Candle, the model
-being used here is a small sequeezenet variant.
-
-You can run the example with the following command:
-
-```bash
-cargo run --example squeezenet-onnx --release -- --image candle-examples/examples/yolo-v8/assets/bike.jpg
-```

candle-examples/examples/onnx/main.rs

@@ -1,78 +0,0 @@
-#[cfg(feature = "mkl")]
-extern crate intel_mkl_src;
-
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-
-use candle::{IndexOp, D};
-use clap::{Parser, ValueEnum};
-
-#[derive(Clone, Copy, Debug, ValueEnum)]
-enum Which {
-    SqueezeNet,
-    EfficientNet,
-}
-
-#[derive(Parser)]
-struct Args {
-    #[arg(long)]
-    image: String,
-
-    #[arg(long)]
-    model: Option<String>,
-
-    /// The model to be used.
-    #[arg(value_enum, long, default_value_t = Which::SqueezeNet)]
-    which: Which,
-}
-
-pub fn main() -> anyhow::Result<()> {
-    let args = Args::parse();
-    let image = candle_examples::imagenet::load_image224(args.image)?;
-    let image = match args.which {
-        Which::SqueezeNet => image,
-        Which::EfficientNet => image.permute((1, 2, 0))?,
-    };
-
-    println!("loaded image {image:?}");
-
-    let model = match args.model {
-        Some(model) => std::path::PathBuf::from(model),
-        None => match args.which {
-            Which::SqueezeNet => hf_hub::api::sync::Api::new()?
-                .model("lmz/candle-onnx".into())
-                .get("squeezenet1.1-7.onnx")?,
-            Which::EfficientNet => hf_hub::api::sync::Api::new()?
-                .model("onnx/EfficientNet-Lite4".into())
-                .get("efficientnet-lite4-11.onnx")?,
-        },
-    };
-
-    let model = candle_onnx::read_file(model)?;
-    let graph = model.graph.as_ref().unwrap();
-    let mut inputs = std::collections::HashMap::new();
-    inputs.insert(graph.input[0].name.to_string(), image.unsqueeze(0)?);
-    let mut outputs = candle_onnx::simple_eval(&model, inputs)?;
-    let output = outputs.remove(&graph.output[0].name).unwrap();
-    let prs = match args.which {
-        Which::SqueezeNet => candle_nn::ops::softmax(&output, D::Minus1)?,
-        Which::EfficientNet => output,
-    };
-    let prs = prs.i(0)?.to_vec1::<f32>()?;
-
-    // Sort the predictions and take the top 5
-    let mut top: Vec<_> = prs.iter().enumerate().collect();
-    top.sort_by(|a, b| b.1.partial_cmp(a.1).unwrap());
-    let top = top.into_iter().take(5).collect::<Vec<_>>();
-
-    // Print the top predictions
-    for &(i, p) in &top {
-        println!(
-            "{:50}: {:.2}%",
-            candle_examples::imagenet::CLASSES[i],
-            p * 100.0
-        );
-    }
-
-    Ok(())
-}

candle-examples/examples/onnx_basics.rs

@@ -1,87 +0,0 @@
-use anyhow::Result;
-use candle::{Device, Tensor};
-
-use clap::{Parser, Subcommand};
-
-#[derive(Subcommand, Debug, Clone)]
-enum Command {
-    Print {
-        #[arg(long)]
-        file: String,
-    },
-    SimpleEval {
-        #[arg(long)]
-        file: String,
-    },
-}
-
-#[derive(Parser, Debug)]
-#[command(author, version, about, long_about = None)]
-pub struct Args {
-    #[command(subcommand)]
-    command: Command,
-}
-
-pub fn main() -> Result<()> {
-    let args = Args::parse();
-    match args.command {
-        Command::Print { file } => {
-            let model = candle_onnx::read_file(file)?;
-            println!("{model:?}");
-            let graph = model.graph.unwrap();
-            for node in graph.node.iter() {
-                println!("{node:?}");
-            }
-        }
-        Command::SimpleEval { file } => {
-            let model = candle_onnx::read_file(file)?;
-            let graph = model.graph.as_ref().unwrap();
-            let constants: std::collections::HashSet<_> =
-                graph.initializer.iter().map(|i| i.name.as_str()).collect();
-            let mut inputs = std::collections::HashMap::new();
-            for input in graph.input.iter() {
-                use candle_onnx::onnx::tensor_proto::DataType;
-                if constants.contains(input.name.as_str()) {
-                    continue;
-                }
-
-                let type_ = input.r#type.as_ref().expect("no type for input");
-                let type_ = type_.value.as_ref().expect("no type.value for input");
-                let value = match type_ {
-                    candle_onnx::onnx::type_proto::Value::TensorType(tt) => {
-                        let dt = match DataType::try_from(tt.elem_type) {
-                            Ok(dt) => match candle_onnx::dtype(dt) {
-                                Some(dt) => dt,
-                                None => {
-                                    anyhow::bail!(
-                                        "unsupported 'value' data-type {dt:?} for {}",
-                                        input.name
-                                    )
-                                }
-                            },
-                            type_ => anyhow::bail!("unsupported input type {type_:?}"),
-                        };
-                        let shape = tt.shape.as_ref().expect("no tensortype.shape for input");
-                        let dims = shape
-                                .dim
-                                .iter()
-                                .map(|dim| match dim.value.as_ref().expect("no dim value") {
-                                    candle_onnx::onnx::tensor_shape_proto::dimension::Value::DimValue(v) => Ok(*v as usize),
-                                    candle_onnx::onnx::tensor_shape_proto::dimension::Value::DimParam(_) => Ok(42),
-                                })
-                                .collect::<Result<Vec<usize>>>()?;
-                        Tensor::zeros(dims, dt, &Device::Cpu)?
-                    }
-                    type_ => anyhow::bail!("unsupported input type {type_:?}"),
-                };
-                println!("input {}: {value:?}", input.name);
-                inputs.insert(input.name.clone(), value);
-            }
-            let outputs = candle_onnx::simple_eval(&model, inputs)?;
-            for (name, value) in outputs.iter() {
-                println!("output {name}: {value:?}")
-            }
-        }
-    }
-    Ok(())
-}

candle-examples/examples/phi/README.md

@@ -1,33 +1,14 @@
-# candle-phi: 1.3b and 2.7b LLM with state of the art performance for <10b models.
+# candle-phi: 1.3b LLM with state of the art performance for <10b models.
 
-[Phi-1.5](https://huggingface.co/microsoft/phi-1_5) and
-[Phi-2](https://huggingface.co/microsoft/phi-2) are language models using
-only 1.3 and 2.7 billion parameters but with state of the art performance compared to
+[Phi-1.5](https://huggingface.co/microsoft/phi-1_5) is a language model using
+only 1.3 billion parameters but with state of the art performance compared to
 models with up to 10 billion parameters.
 
 The candle implementation provides both the standard version as well as a
 quantized variant.
 
-## Running some examples
+## Running some example
 
-For the v2 version.
-```bash
-$ cargo run --example phi --release -- --model 2 \
-  --prompt "A skier slides down a frictionless slope of height 40m and length 80m. What's the skier speed at the bottom?"
-
-A skier slides down a frictionless slope of height 40m and length 80m. What's the skier speed at the bottom?
-
-Solution:
-The potential energy of the skier is converted into kinetic energy as it slides down the slope. The formula for potential energy is mgh, where m is mass, g is acceleration due to gravity (9.8 m/s^2), and h is height. Since there's no friction, all the potential energy is converted into kinetic energy at the bottom of the slope. The formula for kinetic energy is 1/2mv^2, where v is velocity. We can equate these two formulas:
-mgh = 1/2mv^2
-Solving for v, we get:
-v = sqrt(2gh)
-Substituting the given values, we get:
-v = sqrt(2*9.8*40) = 28 m/s
-Therefore, the skier speed at the bottom of the slope is 28 m/s.
-```
-
-For the v1.5 version.
 ```bash
 $ cargo run --example phi --release -- --prompt "def print_prime(n): "
 
@@ -60,16 +41,3 @@ def median(arr):
     else:
         return arr[n//2]
 ```
-
-This also supports the [Puffin Phi v2
-model](https://huggingface.co/teknium/Puffin-Phi-v2) for human interaction.
-```
-$ cargo run --example phi --release  -- \
-    --prompt "USER: What would you do on a sunny day in Paris?\nASSISTANT:" \
-    --sample-len 200 --model puffin-phi-v2 --quantized 
-USER: What would you do on a sunny day in Paris?
-ASSISTANT: On a sunny day in Paris, you could visit the Musée du Louvre to admire the famous
-painting "Mona Lisa" by Leonardo da Vinci. You might also want to stroll along the Champs-Élysées
-and enjoy the beautiful architecture of the buildings around you. Don't forget to stop by a café
-for a cup of coffee and to soak up the sun!"
-```

candle-examples/examples/phi/main.rs

@@ -5,7 +5,7 @@ extern crate intel_mkl_src;
 extern crate accelerate_src;
 
 use anyhow::{Error as E, Result};
-use clap::{Parser, ValueEnum};
+use clap::Parser;
 
 use candle_transformers::models::mixformer::{Config, MixFormerSequentialForCausalLM as MixFormer};
 use candle_transformers::models::quantized_mixformer::MixFormerSequentialForCausalLM as QMixFormer;
@@ -28,7 +28,6 @@ struct TextGeneration {
     logits_processor: LogitsProcessor,
     repeat_penalty: f32,
     repeat_last_n: usize,
-    verbose_prompt: bool,
 }
 
 impl TextGeneration {
@@ -41,7 +40,6 @@ impl TextGeneration {
         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);
@@ -51,7 +49,6 @@ impl TextGeneration {
             logits_processor,
             repeat_penalty,
             repeat_last_n,
-            verbose_prompt,
             device: device.clone(),
         }
     }
@@ -59,24 +56,20 @@ impl TextGeneration {
     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 phi 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();
+        print!("{prompt}");
+        std::io::stdout().flush()?;
+        let mut tokens = self
+            .tokenizer
+            .encode(prompt, true)
+            .map_err(E::msg)?
+            .get_ids()
+            .to_vec();
+
         let mut generated_tokens = 0usize;
         let eos_token = match self.tokenizer.get_vocab(true).get("<|endoftext|>") {
             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() };
@@ -117,18 +110,6 @@ impl TextGeneration {
     }
 }
 
-#[derive(Clone, Copy, Debug, ValueEnum)]
-enum WhichModel {
-    #[value(name = "1")]
-    V1,
-    #[value(name = "1.5")]
-    V1_5,
-    #[value(name = "2")]
-    V2,
-    PuffinPhiV2,
-    PhiHermes,
-}
-
 #[derive(Parser, Debug)]
 #[command(author, version, about, long_about = None)]
 struct Args {
@@ -140,15 +121,8 @@ struct Args {
     #[arg(long)]
     tracing: bool,
 
-    /// Display the token for the specified prompt.
     #[arg(long)]
-    verbose_prompt: bool,
-
-    #[arg(long)]
-    prompt: Option<String>,
-
-    #[arg(long)]
-    mmlu_dir: Option<String>,
+    prompt: String,
 
     /// The temperature used to generate samples.
     #[arg(long)]
@@ -163,24 +137,18 @@ struct Args {
     seed: u64,
 
     /// The length of the sample to generate (in tokens).
-    #[arg(long, short = 'n', default_value_t = 5000)]
+    #[arg(long, short = 'n', default_value_t = 100)]
     sample_len: usize,
 
-    #[arg(long)]
-    model_id: Option<String>,
+    #[arg(long, default_value = "microsoft/phi-1_5")]
+    model_id: String,
 
-    #[arg(long, default_value = "1.5")]
-    model: WhichModel,
-
-    #[arg(long)]
-    revision: Option<String>,
+    #[arg(long, default_value = "refs/pr/18")]
+    revision: String,
 
     #[arg(long)]
     weight_file: Option<String>,
 
-    #[arg(long)]
-    tokenizer: Option<String>,
-
     #[arg(long)]
     quantized: bool,
 
@@ -221,70 +189,20 @@ 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.to_string(),
+    let repo = api.repo(Repo::with_revision(
+        args.model_id,
+        RepoType::Model,
+        args.revision,
+    ));
+    let tokenizer_filename = repo.get("tokenizer.json")?;
+    let filename = match args.weight_file {
+        Some(weight_file) => std::path::PathBuf::from(weight_file),
         None => {
             if args.quantized {
-                "lmz/candle-quantized-phi".to_string()
+                api.model("lmz/candle-quantized-phi".to_string())
+                    .get("model-q4k.gguf")?
             } else {
-                match args.model {
-                    WhichModel::V1 => "microsoft/phi-1".to_string(),
-                    WhichModel::V1_5 => "microsoft/phi-1_5".to_string(),
-                    WhichModel::V2 => "microsoft/phi-2".to_string(),
-                    WhichModel::PuffinPhiV2 | WhichModel::PhiHermes => {
-                        "lmz/candle-quantized-phi".to_string()
-                    }
-                }
-            }
-        }
-    };
-    let revision = match args.revision {
-        Some(rev) => rev.to_string(),
-        None => {
-            if args.quantized {
-                "main".to_string()
-            } else {
-                match args.model {
-                    WhichModel::V1 => "refs/pr/2".to_string(),
-                    WhichModel::V1_5 => "refs/pr/18".to_string(),
-                    WhichModel::V2 | WhichModel::PuffinPhiV2 | WhichModel::PhiHermes => {
-                        "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 => match args.model {
-            WhichModel::V1 | WhichModel::V1_5 | WhichModel::V2 => repo.get("tokenizer.json")?,
-            WhichModel::PuffinPhiV2 | WhichModel::PhiHermes => {
-                repo.get("tokenizer-puffin-phi-v2.json")?
-            }
-        },
-    };
-    let filenames = match args.weight_file {
-        Some(weight_file) => vec![std::path::PathBuf::from(weight_file)],
-        None => {
-            if args.quantized {
-                match args.model {
-                    WhichModel::V1 => vec![repo.get("model-v1-q4k.gguf")?],
-                    WhichModel::V1_5 => vec![repo.get("model-q4k.gguf")?],
-                    WhichModel::V2 => vec![repo.get("model-v2-q4k.gguf")?],
-                    WhichModel::PuffinPhiV2 => vec![repo.get("model-puffin-phi-v2-q4k.gguf")?],
-                    WhichModel::PhiHermes => vec![repo.get("model-phi-hermes-1_3B-q4k.gguf")?],
-                }
-            } else {
-                match args.model {
-                    WhichModel::V1 | WhichModel::V1_5 => vec![repo.get("model.safetensors")?],
-                    WhichModel::V2 => candle_examples::hub_load_safetensors(
-                        &repo,
-                        "model.safetensors.index.json",
-                    )?,
-                    WhichModel::PuffinPhiV2 => vec![repo.get("model-puffin-phi-v2.safetensors")?],
-                    WhichModel::PhiHermes => vec![repo.get("model-phi-hermes-1_3B.safetensors")?],
-                }
+                repo.get("model.safetensors")?
             }
         }
     };
@@ -292,130 +210,29 @@ fn main() -> Result<()> {
     let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
 
     let start = std::time::Instant::now();
-    let config = match args.model {
-        WhichModel::V1 => Config::v1(),
-        WhichModel::V1_5 => Config::v1_5(),
-        WhichModel::V2 => Config::v2(),
-        WhichModel::PuffinPhiV2 => Config::puffin_phi_v2(),
-        WhichModel::PhiHermes => Config::phi_hermes_1_3b(),
-    };
+    let config = Config::v1_5();
     let (model, device) = if args.quantized {
-        let vb = candle_transformers::quantized_var_builder::VarBuilder::from_gguf(&filenames[0])?;
-        let model = match args.model {
-            WhichModel::V2 => QMixFormer::new_v2(&config, vb)?,
-            _ => QMixFormer::new(&config, vb)?,
-        };
+        let vb = candle_transformers::quantized_var_builder::VarBuilder::from_gguf(&filename)?;
+        let model = QMixFormer::new(&config, vb)?;
         (Model::Quantized(model), Device::Cpu)
     } else {
         let device = candle_examples::device(args.cpu)?;
-        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, DType::F32, &device)? };
-        let model = match args.model {
-            WhichModel::V2 => MixFormer::new_v2(&config, vb)?,
-            _ => MixFormer::new(&config, vb)?,
-        };
+        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[filename], DType::F32, &device)? };
+        let model = MixFormer::new(&config, vb)?;
         (Model::MixFormer(model), device)
     };
     println!("loaded the model in {:?}", start.elapsed());
 
-    match (args.prompt, args.mmlu_dir) {
-        (None, None) | (Some(_), Some(_)) => {
-            anyhow::bail!("exactly one of --prompt and --mmlu-dir must be specified")
-        }
-        (Some(prompt), None) => {
-            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(&prompt, args.sample_len)?;
-        }
-        (None, Some(mmlu_dir)) => mmlu(model, tokenizer, &device, mmlu_dir)?,
-    }
-    Ok(())
-}
-
-fn mmlu<P: AsRef<std::path::Path>>(
-    mut model: Model,
-    tokenizer: Tokenizer,
-    device: &Device,
-    mmlu_dir: P,
-) -> anyhow::Result<()> {
-    for dir_entry in mmlu_dir.as_ref().read_dir()?.flatten() {
-        let dir_entry = dir_entry.path();
-        let theme = match dir_entry.file_stem().and_then(|v| v.to_str()) {
-            None => "".to_string(),
-            Some(v) => match v.strip_suffix("_test") {
-                None => v.replace('_', " "),
-                Some(v) => v.replace('_', " "),
-            },
-        };
-        if dir_entry.extension().as_ref().and_then(|v| v.to_str()) != Some("csv") {
-            continue;
-        }
-        println!("reading {dir_entry:?}");
-        let dir_entry = std::fs::File::open(dir_entry)?;
-        let mut reader = csv::ReaderBuilder::new()
-            .has_headers(false)
-            .from_reader(dir_entry);
-        let token_a = tokenizer.token_to_id("A").unwrap();
-        let token_b = tokenizer.token_to_id("B").unwrap();
-        let token_c = tokenizer.token_to_id("C").unwrap();
-        let token_d = tokenizer.token_to_id("D").unwrap();
-        for row in reader.records() {
-            let row = match row {
-                Err(_) => continue,
-                Ok(row) => row,
-            };
-            if row.len() < 5 {
-                continue;
-            }
-            let question = row.get(0).unwrap();
-            let answer_a = row.get(1).unwrap();
-            let answer_b = row.get(2).unwrap();
-            let answer_c = row.get(3).unwrap();
-            let answer_d = row.get(4).unwrap();
-            let answer = row.get(5).unwrap();
-            let prompt = format!(
-                    "{} {theme}.\n{question}\nA. {answer_a}\nB. {answer_b}\nC. {answer_c}\nD. {answer_d}\nAnswer:\n",
-                    "The following are multiple choice questions (with answers) about"
-                );
-            let tokens = tokenizer.encode(prompt.as_str(), true).map_err(E::msg)?;
-            let tokens = tokens.get_ids().to_vec();
-            let input = Tensor::new(tokens, device)?.unsqueeze(0)?;
-            let logits = match &mut model {
-                Model::MixFormer(m) => {
-                    m.clear_kv_cache();
-                    m.forward(&input)?
-                }
-                Model::Quantized(m) => {
-                    m.clear_kv_cache();
-                    m.forward(&input)?
-                }
-            };
-            let logits = logits.squeeze(0)?.to_dtype(DType::F32)?;
-            let logits_v: Vec<f32> = logits.to_vec1()?;
-            let pr_a = logits_v[token_a as usize];
-            let pr_b = logits_v[token_b as usize];
-            let pr_c = logits_v[token_c as usize];
-            let pr_d = logits_v[token_d as usize];
-            let model_answer = if pr_a > pr_b && pr_a > pr_c && pr_a > pr_d {
-                "A"
-            } else if pr_b > pr_c && pr_b > pr_d {
-                "B"
-            } else if pr_c > pr_d {
-                "C"
-            } else {
-                "D"
-            };
-
-            println!("{prompt}\n -> {model_answer} vs {answer}");
-        }
-    }
+    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/quantized-t5/README.md

@@ -1,7 +1,5 @@
 # candle-quantized-t5
 
-## Seq2Seq example
-
 This example uses a quantized version of the t5 model.
 
 ```bash
@@ -10,8 +8,6 @@ $ cargo run --example quantized-t5 --release -- --prompt "translate to German: A
  Eine schöne Kerze.
 ```
 
-## Generating Quantized weight files
-
 The weight file is automatically retrieved from the hub. It is also possible to
 generate quantized weight files from the original safetensors file by using the
 `tensor-tools` command line utility via:
@@ -20,11 +16,8 @@ generate quantized weight files from the original safetensors file by using the
 $ cargo run --example tensor-tools --release -- quantize --quantization q6k PATH/TO/T5/model.safetensors /tmp/model.gguf
 ```
 
-## Using custom models
-
-To use a different model, specify the `model-id`.
-
-For example, for text editing, you can use quantized [CoEdit models](https://huggingface.co/jbochi/candle-coedit-quantized).
+To use a different model, specify the `model-id`. For example, you can use
+quantized [CoEdit models](https://huggingface.co/jbochi/candle-coedit-quantized).
 
 ```bash
 $ cargo run --example quantized-t5 --release  -- \
@@ -33,7 +26,6 @@ $ cargo run --example quantized-t5 --release  -- \
   --temperature 0
 ...
  Although their flight is weak, they run quickly through the tree canopy.
-```
 
 By default, it will look for `model.gguf` and `config.json`, but you can specify
 custom local or remote `weight-file` and `config-file`s:
@@ -48,16 +40,3 @@ cargo run --example quantized-t5 --release  -- \
 ...
  Note that a storm surge is what forecasters consider a hurricane's most dangerous part.
 ```
-
-### [MADLAD-400](https://arxiv.org/abs/2309.04662)
-
-MADLAD-400 is a series of multilingual machine translation T5 models trained on 250 billion tokens covering over 450 languages using publicly available data. These models are competitive with significantly larger models.
-
-```bash
-cargo run --example quantized-t5 --release  -- \
-  --model-id "jbochi/madlad400-3b-mt" --weight-file "model-q4k.gguf" \
-  --prompt "<2de> How are you, my friend?" \
-  --temperature 0
-...
- Wie geht es dir, mein Freund?
-```

candle-examples/examples/quantized-t5/main.rs

@@ -173,11 +173,7 @@ fn main() -> Result<()> {
         .to_vec();
     let input_token_ids = Tensor::new(&tokens[..], device)?.unsqueeze(0)?;
     let mut model = builder.build_model()?;
-    let mut output_token_ids = [builder
-        .config
-        .decoder_start_token_id
-        .unwrap_or(builder.config.pad_token_id) as u32]
-    .to_vec();
+    let mut output_token_ids = [builder.config.pad_token_id as u32].to_vec();
     let temperature = if args.temperature <= 0. {
         None
     } else {

candle-examples/examples/quantized/README.md

@@ -26,19 +26,6 @@ cargo run --example quantized --release -- --prompt "The best thing about coding
 > The best thing about coding in rust is 1.) that I don’t need to worry about memory leaks, 2.) speed and 3.) my program will compile even on old machines.
 ```
 
-Using the mixtral sparse mixture of expert model:
-```bash
-
-$ cargo run --example quantized --release -- --which mixtral --prompt "Lebesgue's integral is superior to Riemann's because "
-> avx: true, neon: false, simd128: false, f16c: true
-> temp: 0.80 repeat-penalty: 1.10 repeat-last-n: 64
-> loaded 995 tensors (26.44GB) in 0.03s
-Lebesgue's integral is superior to Riemann's because 1. it is defined for a wider class of functions, those which are absolutely integrable; 2. the definition does not involve limits in two variables---one being computed before the other (which makes some computations more difficult); and 3. interchange of order of integration is easier to establish than with Riemann's integral. On the other hand, Lebesgue's integral applies only for bounded functions defined on finite intervals; it does not provide numerical values for improper integrals. The latter are best evaluated using Cauchy's limit definition.
-
-The reason $f(x) = x^2$ is discontinuous at the ends of its interval of definition, and Riemann's integral requires continuity on the whole of an open interval containing it (see our earlier post), sine no such function exists with this property, is that the endpoints are infinite in measure for Lebesgue's integral.
- ```
-
-
 ## Command-line flags
 
 Run with `--help` to see all options.

candle-examples/examples/quantized/main.rs

@@ -12,7 +12,6 @@ use candle::quantized::{ggml_file, gguf_file};
 use candle::{Device, Tensor};
 use candle_transformers::generation::LogitsProcessor;
 
-use candle_examples::token_output_stream::TokenOutputStream;
 use candle_transformers::models::quantized_llama as model;
 use model::ModelWeights;
 
@@ -25,7 +24,7 @@ enum Prompt {
     One(String),
 }
 
-#[derive(Clone, Debug, Copy, PartialEq, Eq, ValueEnum)]
+#[derive(Clone, Debug, Copy, ValueEnum)]
 enum Which {
     #[value(name = "7b")]
     L7b,
@@ -45,28 +44,10 @@ enum Which {
     L13bCode,
     #[value(name = "32b-code")]
     L34bCode,
-    #[value(name = "7b-leo")]
-    Leo7b,
-    #[value(name = "13b-leo")]
-    Leo13b,
     #[value(name = "7b-mistral")]
     Mistral7b,
     #[value(name = "7b-mistral-instruct")]
     Mistral7bInstruct,
-    #[value(name = "7b-mistral-instruct-v0.2")]
-    Mistral7bInstructV02,
-    #[value(name = "7b-zephyr-a")]
-    Zephyr7bAlpha,
-    #[value(name = "7b-zephyr-b")]
-    Zephyr7bBeta,
-    #[value(name = "7b-open-chat-3.5")]
-    OpenChat35,
-    #[value(name = "7b-starling-a")]
-    Starling7bAlpha,
-    #[value(name = "mixtral")]
-    Mixtral,
-    #[value(name = "mixtral-instruct")]
-    MixtralInstruct,
 }
 
 impl Which {
@@ -80,93 +61,8 @@ impl Which {
             | Self::L70bChat
             | Self::L7bCode
             | Self::L13bCode
-            | Self::L34bCode
-            | Self::Leo7b
-            | Self::Leo13b => false,
-            // Zephyr and OpenChat are fine tuned versions of mistral and should be treated in the
-            // same way. Starling is a fine tuned version of OpenChat.
-            Self::OpenChat35
-            | Self::Starling7bAlpha
-            | Self::Zephyr7bAlpha
-            | Self::Zephyr7bBeta
-            | Self::Mixtral
-            | Self::MixtralInstruct
-            | Self::Mistral7b
-            | Self::Mistral7bInstruct
-            | Self::Mistral7bInstructV02 => true,
-        }
-    }
-
-    fn is_zephyr(&self) -> bool {
-        match self {
-            Self::L7b
-            | Self::L13b
-            | Self::L70b
-            | Self::L7bChat
-            | Self::L13bChat
-            | Self::L70bChat
-            | Self::L7bCode
-            | Self::L13bCode
-            | Self::L34bCode
-            | Self::Leo7b
-            | Self::Leo13b
-            | Self::Mixtral
-            | Self::MixtralInstruct
-            | Self::Mistral7b
-            | Self::Mistral7bInstruct
-            | Self::Mistral7bInstructV02
-            | Self::OpenChat35
-            | Self::Starling7bAlpha => false,
-            Self::Zephyr7bAlpha | Self::Zephyr7bBeta => true,
-        }
-    }
-
-    fn is_open_chat(&self) -> bool {
-        match self {
-            Self::L7b
-            | Self::L13b
-            | Self::L70b
-            | Self::L7bChat
-            | Self::L13bChat
-            | Self::L70bChat
-            | Self::L7bCode
-            | Self::L13bCode
-            | Self::L34bCode
-            | Self::Leo7b
-            | Self::Leo13b
-            | Self::Mixtral
-            | Self::MixtralInstruct
-            | Self::Mistral7b
-            | Self::Mistral7bInstruct
-            | Self::Mistral7bInstructV02
-            | Self::Zephyr7bAlpha
-            | Self::Zephyr7bBeta => false,
-            Self::OpenChat35 | Self::Starling7bAlpha => true,
-        }
-    }
-
-    fn tokenizer_repo(&self) -> &'static str {
-        match self {
-            Which::L7b
-            | Which::L13b
-            | Which::L70b
-            | Which::L7bChat
-            | Which::L13bChat
-            | Which::L70bChat
-            | Which::L7bCode
-            | Which::L13bCode
-            | Which::L34bCode => "hf-internal-testing/llama-tokenizer",
-            Which::Leo7b => "LeoLM/leo-hessianai-7b",
-            Which::Leo13b => "LeoLM/leo-hessianai-13b",
-            Which::Mixtral => "mistralai/Mixtral-8x7B-v0.1",
-            Which::MixtralInstruct => "mistralai/Mixtral-8x7B-Instruct-v0.1",
-            Which::Mistral7b
-            | Which::Mistral7bInstruct
-            | Which::Mistral7bInstructV02
-            | Which::Zephyr7bAlpha
-            | Which::Zephyr7bBeta => "mistralai/Mistral-7B-v0.1",
-            Which::OpenChat35 => "openchat/openchat_3.5",
-            Which::Starling7bAlpha => "berkeley-nest/Starling-LM-7B-alpha",
+            | Self::L34bCode => false,
+            Self::Mistral7b | Self::Mistral7bInstruct => true,
         }
     }
 }
@@ -174,7 +70,7 @@ impl Which {
 #[derive(Parser, Debug)]
 #[command(author, version, about, long_about = None)]
 struct Args {
-    /// GGML/GGUF file to load, typically a .bin/.gguf file generated by the quantize command from llama.cpp
+    /// GGML file to load, typically a .bin file generated by the quantize command from llama.cpp
     #[arg(long)]
     model: Option<String>,
 
@@ -185,7 +81,7 @@ struct Args {
     prompt: Option<String>,
 
     /// The length of the sample to generate (in tokens).
-    #[arg(short = 'n', long, default_value_t = 1000)]
+    #[arg(short = 'n', long, default_value_t = 100)]
     sample_len: usize,
 
     /// The tokenizer config in json format.
@@ -235,7 +131,11 @@ impl Args {
             Some(config) => std::path::PathBuf::from(config),
             None => {
                 let api = hf_hub::api::sync::Api::new()?;
-                let repo = self.which.tokenizer_repo();
+                let repo = if self.which.is_mistral() {
+                    "mistralai/Mistral-7B-v0.1"
+                } else {
+                    "hf-internal-testing/llama-tokenizer"
+                };
                 let api = api.model(repo.to_string());
                 api.get("tokenizer.json")?
             }
@@ -266,22 +166,6 @@ impl Args {
                     Which::L7bCode => ("TheBloke/CodeLlama-7B-GGUF", "codellama-7b.Q8_0.gguf"),
                     Which::L13bCode => ("TheBloke/CodeLlama-13B-GGUF", "codellama-13b.Q8_0.gguf"),
                     Which::L34bCode => ("TheBloke/CodeLlama-34B-GGUF", "codellama-34b.Q8_0.gguf"),
-                    Which::Leo7b => (
-                        "TheBloke/leo-hessianai-7B-GGUF",
-                        "leo-hessianai-7b.Q4_K_M.gguf",
-                    ),
-                    Which::Leo13b => (
-                        "TheBloke/leo-hessianai-13B-GGUF",
-                        "leo-hessianai-13b.Q4_K_M.gguf",
-                    ),
-                    Which::Mixtral => (
-                        "TheBloke/Mixtral-8x7B-v0.1-GGUF",
-                        "mixtral-8x7b-v0.1.Q4_K_M.gguf",
-                    ),
-                    Which::MixtralInstruct => (
-                        "TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF",
-                        "mixtral-8x7b-instruct-v0.1.Q4_K_M.gguf",
-                    ),
                     Which::Mistral7b => (
                         "TheBloke/Mistral-7B-v0.1-GGUF",
                         "mistral-7b-v0.1.Q4_K_S.gguf",
@@ -290,22 +174,6 @@ impl Args {
                         "TheBloke/Mistral-7B-Instruct-v0.1-GGUF",
                         "mistral-7b-instruct-v0.1.Q4_K_S.gguf",
                     ),
-                    Which::Mistral7bInstructV02 => (
-                        "TheBloke/Mistral-7B-Instruct-v0.2-GGUF",
-                        "mistral-7b-instruct-v0.2.Q4_K_S.gguf",
-                    ),
-                    Which::Zephyr7bAlpha => (
-                        "TheBloke/zephyr-7B-alpha-GGUF",
-                        "zephyr-7b-alpha.Q4_K_M.gguf",
-                    ),
-                    Which::Zephyr7bBeta => {
-                        ("TheBloke/zephyr-7B-beta-GGUF", "zephyr-7b-beta.Q4_K_M.gguf")
-                    }
-                    Which::OpenChat35 => ("TheBloke/openchat_3.5-GGUF", "openchat_3.5.Q4_K_M.gguf"),
-                    Which::Starling7bAlpha => (
-                        "TheBloke/Starling-LM-7B-alpha-GGUF",
-                        "starling-lm-7b-alpha.Q4_K_M.gguf",
-                    ),
                 };
                 let api = hf_hub::api::sync::Api::new()?;
                 let api = api.model(repo.to_string());
@@ -316,6 +184,31 @@ impl Args {
     }
 }
 
+fn print_token(next_token: u32, tokenizer: &Tokenizer) {
+    // 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('▁', " ");
+        let ascii = text
+            .strip_prefix("<0x")
+            .and_then(|t| t.strip_suffix('>'))
+            .and_then(|t| u8::from_str_radix(t, 16).ok());
+        match ascii {
+            None => print!("{text}"),
+            Some(ascii) => {
+                if let Some(chr) = char::from_u32(ascii as u32) {
+                    if chr.is_ascii() {
+                        print!("{chr}")
+                    }
+                }
+            }
+        }
+        let _ = std::io::stdout().flush();
+    }
+}
+
 fn format_size(size_in_bytes: usize) -> String {
     if size_in_bytes < 1_000 {
         format!("{}B", size_in_bytes)
@@ -364,7 +257,7 @@ fn main() -> anyhow::Result<()> {
 
     let mut model = match model_path.extension().and_then(|v| v.to_str()) {
         Some("gguf") => {
-            let model = gguf_file::Content::read(&mut file).map_err(|e| e.with_path(model_path))?;
+            let model = gguf_file::Content::read(&mut file)?;
             let mut total_size_in_bytes = 0;
             for (_, tensor) in model.tensor_infos.iter() {
                 let elem_count = tensor.shape.elem_count();
@@ -380,7 +273,7 @@ fn main() -> anyhow::Result<()> {
             ModelWeights::from_gguf(model, &mut file)?
         }
         Some("ggml" | "bin") | Some(_) | None => {
-            let model = ggml_file::Content::read(&mut file).map_err(|e| e.with_path(model_path))?;
+            let model = ggml_file::Content::read(&mut file)?;
             let mut total_size_in_bytes = 0;
             for (_, tensor) in model.tensors.iter() {
                 let elem_count = tensor.shape().elem_count();
@@ -401,20 +294,8 @@ fn main() -> anyhow::Result<()> {
                 | Which::L13bChat
                 | Which::L7bCode
                 | Which::L13bCode
-                | Which::L34bCode
-                | Which::Leo7b
-                | Which::Leo13b => 1,
-                Which::Mixtral
-                | Which::MixtralInstruct
-                | Which::Mistral7b
-                | Which::Mistral7bInstruct
-                | Which::Mistral7bInstructV02
-                | Which::Zephyr7bAlpha
-                | Which::Zephyr7bBeta
-                | Which::L70b
-                | Which::L70bChat
-                | Which::OpenChat35
-                | Which::Starling7bAlpha => 8,
+                | Which::L34bCode => 1,
+                Which::Mistral7b | Which::Mistral7bInstruct | Which::L70b | Which::L70bChat => 8,
             };
             ModelWeights::from_ggml(model, args.gqa.unwrap_or(default_gqa))?
         }
@@ -422,7 +303,6 @@ fn main() -> anyhow::Result<()> {
     println!("model built");
 
     let tokenizer = args.tokenizer()?;
-    let mut tos = TokenOutputStream::new(tokenizer);
     let prompt = match args.prompt.as_deref() {
         Some("chat") => Prompt::Chat,
         Some("interactive") => Prompt::Interactive,
@@ -431,11 +311,10 @@ fn main() -> anyhow::Result<()> {
     };
 
     let mut pre_prompt_tokens = vec![];
-    for prompt_index in 0.. {
+    loop {
         let prompt_str = match &prompt {
             Prompt::One(prompt) => prompt.clone(),
             Prompt::Interactive | Prompt::Chat => {
-                let is_interactive = matches!(prompt, Prompt::Interactive);
                 print!("> ");
                 std::io::stdout().flush()?;
                 let mut prompt = String::new();
@@ -446,15 +325,7 @@ fn main() -> anyhow::Result<()> {
                         prompt.pop();
                     }
                 }
-                if args.which.is_open_chat() {
-                    format!("GPT4 Correct User: {prompt}<|end_of_turn|>GPT4 Correct Assistant:")
-                } else if args.which.is_zephyr() {
-                    if prompt_index == 0 || is_interactive {
-                        format!("<|system|>\n</s>\n<|user|>\n{prompt}</s>\n<|assistant|>",)
-                    } else {
-                        format!("<|user|>\n{prompt}</s>\n<|assistant|>")
-                    }
-                } else if args.which.is_mistral() {
+                if args.which.is_mistral() {
                     format!("[INST] {prompt} [/INST]")
                 } else {
                     prompt
@@ -462,8 +333,7 @@ fn main() -> anyhow::Result<()> {
             }
         };
         print!("{}", &prompt_str);
-        let tokens = tos
-            .tokenizer()
+        let tokens = tokenizer
             .encode(prompt_str, true)
             .map_err(anyhow::Error::msg)?;
         if args.verbose_prompt {
@@ -493,19 +363,11 @@ fn main() -> anyhow::Result<()> {
         };
         let prompt_dt = start_prompt_processing.elapsed();
         all_tokens.push(next_token);
-        if let Some(t) = tos.next_token(next_token)? {
-            print!("{t}");
-            std::io::stdout().flush()?;
-        }
+        print_token(next_token, &tokenizer);
+
+        let eos_token = *tokenizer.get_vocab(true).get("</s>").unwrap();
 
-        let eos_token = if args.which.is_open_chat() {
-            "<|end_of_turn|>"
-        } else {
-            "</s>"
-        };
-        let eos_token = *tos.tokenizer().get_vocab(true).get(eos_token).unwrap();
         let start_post_prompt = std::time::Instant::now();
-        let mut sampled = 0;
         for index in 0..to_sample {
             let input = Tensor::new(&[next_token], &Device::Cpu)?.unsqueeze(0)?;
             let logits = model.forward(&input, prompt_tokens.len() + index)?;
@@ -522,19 +384,11 @@ fn main() -> anyhow::Result<()> {
             };
             next_token = logits_processor.sample(&logits)?;
             all_tokens.push(next_token);
-            if let Some(t) = tos.next_token(next_token)? {
-                print!("{t}");
-                std::io::stdout().flush()?;
-            }
-            sampled += 1;
+            print_token(next_token, &tokenizer);
             if next_token == eos_token {
                 break;
             };
         }
-        if let Some(rest) = tos.decode_rest().map_err(candle::Error::msg)? {
-            print!("{rest}");
-        }
-        std::io::stdout().flush()?;
         let dt = start_post_prompt.elapsed();
         println!(
             "\n\n{:4} prompt tokens processed: {:.2} token/s",
@@ -542,8 +396,9 @@ fn main() -> anyhow::Result<()> {
             prompt_tokens.len() as f64 / prompt_dt.as_secs_f64(),
         );
         println!(
-            "{sampled:4} tokens generated: {:.2} token/s",
-            sampled as f64 / dt.as_secs_f64(),
+            "{:4} tokens generated: {:.2} token/s",
+            to_sample,
+            to_sample as f64 / dt.as_secs_f64(),
         );
 
         match prompt {

candle-examples/examples/reinforcement-learning/README.md

@@ -1,23 +0,0 @@
-# candle-reinforcement-learning
-
-Reinforcement Learning examples for candle.
-
-This has been tested with `gymnasium` version `0.29.1`. You can install the
-Python package with:
-```bash
-pip install "gymnasium[accept-rom-license]"
-```
-
-In order to run the examples, use the following commands. Note the additional
-`--package` flag to ensure that there is no conflict with the `candle-pyo3`
-crate.
-
-For the Policy Gradient example:
-```bash
-cargo run --example reinforcement-learning --features=pyo3 --package candle-examples -- pg
-```
-
-For the Deep Deterministic Policy Gradient example:
-```bash
-cargo run --example reinforcement-learning --features=pyo3 --package candle-examples -- ddpg
-```

candle-examples/examples/reinforcement-learning/atari_wrappers.py

@@ -1,308 +0,0 @@
-import gymnasium as gym
-import numpy as np
-from collections import deque
-from PIL import Image
-from multiprocessing import Process, Pipe
-
-# atari_wrappers.py
-class NoopResetEnv(gym.Wrapper):
-    def __init__(self, env, noop_max=30):
-        """Sample initial states by taking random number of no-ops on reset.
-        No-op is assumed to be action 0.
-        """
-        gym.Wrapper.__init__(self, env)
-        self.noop_max = noop_max
-        self.override_num_noops = None
-        assert env.unwrapped.get_action_meanings()[0] == 'NOOP'
-
-    def reset(self):
-        """ Do no-op action for a number of steps in [1, noop_max]."""
-        self.env.reset()
-        if self.override_num_noops is not None:
-            noops = self.override_num_noops
-        else:
-            noops = self.unwrapped.np_random.integers(1, self.noop_max + 1) #pylint: disable=E1101
-        assert noops > 0
-        obs = None
-        for _ in range(noops):
-            obs, _, done, _ = self.env.step(0)
-            if done:
-                obs = self.env.reset()
-        return obs
-
-class FireResetEnv(gym.Wrapper):
-    def __init__(self, env):
-        """Take action on reset for environments that are fixed until firing."""
-        gym.Wrapper.__init__(self, env)
-        assert env.unwrapped.get_action_meanings()[1] == 'FIRE'
-        assert len(env.unwrapped.get_action_meanings()) >= 3
-
-    def reset(self):
-        self.env.reset()
-        obs, _, done, _ = self.env.step(1)
-        if done:
-            self.env.reset()
-        obs, _, done, _ = self.env.step(2)
-        if done:
-            self.env.reset()
-        return obs
-
-class ImageSaver(gym.Wrapper):
-    def __init__(self, env, img_path, rank):
-        gym.Wrapper.__init__(self, env)
-        self._cnt = 0
-        self._img_path = img_path
-        self._rank = rank
-
-    def step(self, action):
-        step_result = self.env.step(action)
-        obs, _, _, _ = step_result
-        img = Image.fromarray(obs, 'RGB')
-        img.save('%s/out%d-%05d.png' % (self._img_path, self._rank, self._cnt))
-        self._cnt += 1
-        return step_result
-
-class EpisodicLifeEnv(gym.Wrapper):
-    def __init__(self, env):
-        """Make end-of-life == end-of-episode, but only reset on true game over.
-        Done by DeepMind for the DQN and co. since it helps value estimation.
-        """
-        gym.Wrapper.__init__(self, env)
-        self.lives = 0
-        self.was_real_done  = True
-
-    def step(self, action):
-        obs, reward, done, info = self.env.step(action)
-        self.was_real_done = done
-        # check current lives, make loss of life terminal,
-        # then update lives to handle bonus lives
-        lives = self.env.unwrapped.ale.lives()
-        if lives < self.lives and lives > 0:
-            # for Qbert sometimes we stay in lives == 0 condition for a few frames
-            # so its important to keep lives > 0, so that we only reset once
-            # the environment advertises done.
-            done = True
-        self.lives = lives
-        return obs, reward, done, info
-
-    def reset(self):
-        """Reset only when lives are exhausted.
-        This way all states are still reachable even though lives are episodic,
-        and the learner need not know about any of this behind-the-scenes.
-        """
-        if self.was_real_done:
-            obs = self.env.reset()
-        else:
-            # no-op step to advance from terminal/lost life state
-            obs, _, _, _ = self.env.step(0)
-        self.lives = self.env.unwrapped.ale.lives()
-        return obs
-
-class MaxAndSkipEnv(gym.Wrapper):
-    def __init__(self, env, skip=4):
-        """Return only every `skip`-th frame"""
-        gym.Wrapper.__init__(self, env)
-        # most recent raw observations (for max pooling across time steps)
-        self._obs_buffer = deque(maxlen=2)
-        self._skip       = skip
-
-    def step(self, action):
-        """Repeat action, sum reward, and max over last observations."""
-        total_reward = 0.0
-        done = None
-        for _ in range(self._skip):
-            obs, reward, done, info = self.env.step(action)
-            self._obs_buffer.append(obs)
-            total_reward += reward
-            if done:
-                break
-        max_frame = np.max(np.stack(self._obs_buffer), axis=0)
-
-        return max_frame, total_reward, done, info
-
-    def reset(self):
-        """Clear past frame buffer and init. to first obs. from inner env."""
-        self._obs_buffer.clear()
-        obs = self.env.reset()
-        self._obs_buffer.append(obs)
-        return obs
-
-class ClipRewardEnv(gym.RewardWrapper):
-    def reward(self, reward):
-        """Bin reward to {+1, 0, -1} by its sign."""
-        return np.sign(reward)
-
-class WarpFrame(gym.ObservationWrapper):
-    def __init__(self, env):
-        """Warp frames to 84x84 as done in the Nature paper and later work."""
-        gym.ObservationWrapper.__init__(self, env)
-        self.res = 84
-        self.observation_space = gym.spaces.Box(low=0, high=255, shape=(self.res, self.res, 1), dtype='uint8')
-
-    def observation(self, obs):
-        frame = np.dot(obs.astype('float32'), np.array([0.299, 0.587, 0.114], 'float32'))
-        frame = np.array(Image.fromarray(frame).resize((self.res, self.res),
-            resample=Image.BILINEAR), dtype=np.uint8)
-        return frame.reshape((self.res, self.res, 1))
-
-class FrameStack(gym.Wrapper):
-    def __init__(self, env, k):
-        """Buffer observations and stack across channels (last axis)."""
-        gym.Wrapper.__init__(self, env)
-        self.k = k
-        self.frames = deque([], maxlen=k)
-        shp = env.observation_space.shape
-        assert shp[2] == 1  # can only stack 1-channel frames
-        self.observation_space = gym.spaces.Box(low=0, high=255, shape=(shp[0], shp[1], k), dtype='uint8')
-
-    def reset(self):
-        """Clear buffer and re-fill by duplicating the first observation."""
-        ob = self.env.reset()
-        for _ in range(self.k): self.frames.append(ob)
-        return self.observation()
-
-    def step(self, action):
-        ob, reward, done, info = self.env.step(action)
-        self.frames.append(ob)
-        return self.observation(), reward, done, info
-
-    def observation(self):
-        assert len(self.frames) == self.k
-        return np.concatenate(self.frames, axis=2)
-
-def wrap_deepmind(env, episode_life=True, clip_rewards=True):
-    """Configure environment for DeepMind-style Atari.
-
-    Note: this does not include frame stacking!"""
-    assert 'NoFrameskip' in env.spec.id  # required for DeepMind-style skip
-    if episode_life:
-        env = EpisodicLifeEnv(env)
-    env = NoopResetEnv(env, noop_max=30)
-    env = MaxAndSkipEnv(env, skip=4)
-    if 'FIRE' in env.unwrapped.get_action_meanings():
-        env = FireResetEnv(env)
-    env = WarpFrame(env)
-    if clip_rewards:
-        env = ClipRewardEnv(env)
-    return env
-
-# envs.py
-def make_env(env_id, img_dir, seed, rank):
-    def _thunk():
-        env = gym.make(env_id)
-        env.reset(seed=(seed + rank))
-        if img_dir is not None:
-            env = ImageSaver(env, img_dir, rank)
-        env = wrap_deepmind(env)
-        env = WrapPyTorch(env)
-        return env
-
-    return _thunk
-
-class WrapPyTorch(gym.ObservationWrapper):
-    def __init__(self, env=None):
-        super(WrapPyTorch, self).__init__(env)
-        self.observation_space = gym.spaces.Box(0.0, 1.0, [1, 84, 84], dtype='float32')
-
-    def observation(self, observation):
-        return observation.transpose(2, 0, 1)
-
-# vecenv.py
-class VecEnv(object):
-    """
-    Vectorized environment base class
-    """
-    def step(self, vac):
-        """
-        Apply sequence of actions to sequence of environments
-        actions -> (observations, rewards, news)
-
-        where 'news' is a boolean vector indicating whether each element is new.
-        """
-        raise NotImplementedError
-    def reset(self):
-        """
-        Reset all environments
-        """
-        raise NotImplementedError
-    def close(self):
-        pass
-
-# subproc_vec_env.py
-def worker(remote, env_fn_wrapper):
-    env = env_fn_wrapper.x()
-    while True:
-        cmd, data = remote.recv()
-        if cmd == 'step':
-            ob, reward, done, info = env.step(data)
-            if done:
-                ob = env.reset()
-            remote.send((ob, reward, done, info))
-        elif cmd == 'reset':
-            ob = env.reset()
-            remote.send(ob)
-        elif cmd == 'close':
-            remote.close()
-            break
-        elif cmd == 'get_spaces':
-            remote.send((env.action_space, env.observation_space))
-        else:
-            raise NotImplementedError
-
-class CloudpickleWrapper(object):
-    """
-    Uses cloudpickle to serialize contents (otherwise multiprocessing tries to use pickle)
-    """
-    def __init__(self, x):
-        self.x = x
-    def __getstate__(self):
-        import cloudpickle
-        return cloudpickle.dumps(self.x)
-    def __setstate__(self, ob):
-        import pickle
-        self.x = pickle.loads(ob)
-
-class SubprocVecEnv(VecEnv):
-    def __init__(self, env_fns):
-        """
-        envs: list of gym environments to run in subprocesses
-        """
-        nenvs = len(env_fns)
-        self.remotes, self.work_remotes = zip(*[Pipe() for _ in range(nenvs)])        
-        self.ps = [Process(target=worker, args=(work_remote, CloudpickleWrapper(env_fn))) 
-            for (work_remote, env_fn) in zip(self.work_remotes, env_fns)]
-        for p in self.ps:
-            p.start()
-
-        self.remotes[0].send(('get_spaces', None))
-        self.action_space, self.observation_space = self.remotes[0].recv()
-
-
-    def step(self, actions):
-        for remote, action in zip(self.remotes, actions):
-            remote.send(('step', action))
-        results = [remote.recv() for remote in self.remotes]
-        obs, rews, dones, infos = zip(*results)
-        return np.stack(obs), np.stack(rews), np.stack(dones), infos
-
-    def reset(self):
-        for remote in self.remotes:
-            remote.send(('reset', None))
-        return np.stack([remote.recv() for remote in self.remotes])
-
-    def close(self):
-        for remote in self.remotes:
-            remote.send(('close', None))
-        for p in self.ps:
-            p.join()
-
-    @property
-    def num_envs(self):
-        return len(self.remotes)
-
-# Create the environment.
-def make(env_name, img_dir, num_processes):
-    envs = SubprocVecEnv([
-        make_env(env_name, img_dir, 1337, i) for i in range(num_processes)
-    ])
-    return envs

candle-examples/examples/reinforcement-learning/ddpg.rs

@@ -1,556 +0,0 @@
-use std::collections::VecDeque;
-use std::fmt::Display;
-
-use candle::{DType, Device, Error, Module, Result, Tensor, Var};
-use candle_nn::{
-    func, linear, sequential::seq, Activation, AdamW, Optimizer, ParamsAdamW, Sequential,
-    VarBuilder, VarMap,
-};
-use rand::{distributions::Uniform, thread_rng, Rng};
-
-use super::gym_env::GymEnv;
-
-pub struct OuNoise {
-    mu: f64,
-    theta: f64,
-    sigma: f64,
-    state: Tensor,
-}
-impl OuNoise {
-    pub fn new(mu: f64, theta: f64, sigma: f64, size_action: usize) -> Result<Self> {
-        Ok(Self {
-            mu,
-            theta,
-            sigma,
-            state: Tensor::ones(size_action, DType::F32, &Device::Cpu)?,
-        })
-    }
-
-    pub fn sample(&mut self) -> Result<Tensor> {
-        let rand = Tensor::randn_like(&self.state, 0.0, 1.0)?;
-        let dx = ((self.theta * (self.mu - &self.state)?)? + (self.sigma * rand)?)?;
-        self.state = (&self.state + dx)?;
-        Ok(self.state.clone())
-    }
-}
-
-#[derive(Clone)]
-struct Transition {
-    state: Tensor,
-    action: Tensor,
-    reward: Tensor,
-    next_state: Tensor,
-    terminated: bool,
-    truncated: bool,
-}
-impl Transition {
-    fn new(
-        state: &Tensor,
-        action: &Tensor,
-        reward: &Tensor,
-        next_state: &Tensor,
-        terminated: bool,
-        truncated: bool,
-    ) -> Self {
-        Self {
-            state: state.clone(),
-            action: action.clone(),
-            reward: reward.clone(),
-            next_state: next_state.clone(),
-            terminated,
-            truncated,
-        }
-    }
-}
-
-pub struct ReplayBuffer {
-    buffer: VecDeque<Transition>,
-    capacity: usize,
-    size: usize,
-}
-impl ReplayBuffer {
-    pub fn new(capacity: usize) -> Self {
-        Self {
-            buffer: VecDeque::with_capacity(capacity),
-            capacity,
-            size: 0,
-        }
-    }
-
-    pub fn push(
-        &mut self,
-        state: &Tensor,
-        action: &Tensor,
-        reward: &Tensor,
-        next_state: &Tensor,
-        terminated: bool,
-        truncated: bool,
-    ) {
-        if self.size == self.capacity {
-            self.buffer.pop_front();
-        } else {
-            self.size += 1;
-        }
-        self.buffer.push_back(Transition::new(
-            state, action, reward, next_state, terminated, truncated,
-        ));
-    }
-
-    #[allow(clippy::type_complexity)]
-    pub fn random_batch(
-        &self,
-        batch_size: usize,
-    ) -> Result<Option<(Tensor, Tensor, Tensor, Tensor, Vec<bool>, Vec<bool>)>> {
-        if self.size < batch_size {
-            Ok(None)
-        } else {
-            let transitions: Vec<&Transition> = thread_rng()
-                .sample_iter(Uniform::from(0..self.size))
-                .take(batch_size)
-                .map(|i| self.buffer.get(i).unwrap())
-                .collect();
-
-            let states: Vec<Tensor> = transitions
-                .iter()
-                .map(|t| t.state.unsqueeze(0))
-                .collect::<Result<_>>()?;
-            let actions: Vec<Tensor> = transitions
-                .iter()
-                .map(|t| t.action.unsqueeze(0))
-                .collect::<Result<_>>()?;
-            let rewards: Vec<Tensor> = transitions
-                .iter()
-                .map(|t| t.reward.unsqueeze(0))
-                .collect::<Result<_>>()?;
-            let next_states: Vec<Tensor> = transitions
-                .iter()
-                .map(|t| t.next_state.unsqueeze(0))
-                .collect::<Result<_>>()?;
-            let terminateds: Vec<bool> = transitions.iter().map(|t| t.terminated).collect();
-            let truncateds: Vec<bool> = transitions.iter().map(|t| t.truncated).collect();
-
-            Ok(Some((
-                Tensor::cat(&states, 0)?,
-                Tensor::cat(&actions, 0)?,
-                Tensor::cat(&rewards, 0)?,
-                Tensor::cat(&next_states, 0)?,
-                terminateds,
-                truncateds,
-            )))
-        }
-    }
-}
-
-fn track(
-    varmap: &mut VarMap,
-    vb: &VarBuilder,
-    target_prefix: &str,
-    network_prefix: &str,
-    dims: &[(usize, usize)],
-    tau: f64,
-) -> Result<()> {
-    for (i, &(in_dim, out_dim)) in dims.iter().enumerate() {
-        let target_w = vb.get((out_dim, in_dim), &format!("{target_prefix}-fc{i}.weight"))?;
-        let network_w = vb.get((out_dim, in_dim), &format!("{network_prefix}-fc{i}.weight"))?;
-        varmap.set_one(
-            format!("{target_prefix}-fc{i}.weight"),
-            ((tau * network_w)? + ((1.0 - tau) * target_w)?)?,
-        )?;
-
-        let target_b = vb.get(out_dim, &format!("{target_prefix}-fc{i}.bias"))?;
-        let network_b = vb.get(out_dim, &format!("{network_prefix}-fc{i}.bias"))?;
-        varmap.set_one(
-            format!("{target_prefix}-fc{i}.bias"),
-            ((tau * network_b)? + ((1.0 - tau) * target_b)?)?,
-        )?;
-    }
-    Ok(())
-}
-
-struct Actor<'a> {
-    varmap: VarMap,
-    vb: VarBuilder<'a>,
-    network: Sequential,
-    target_network: Sequential,
-    size_state: usize,
-    size_action: usize,
-    dims: Vec<(usize, usize)>,
-}
-
-impl Actor<'_> {
-    fn new(device: &Device, dtype: DType, size_state: usize, size_action: usize) -> Result<Self> {
-        let mut varmap = VarMap::new();
-        let vb = VarBuilder::from_varmap(&varmap, dtype, device);
-
-        let dims = vec![(size_state, 400), (400, 300), (300, size_action)];
-
-        let make_network = |prefix: &str| {
-            let seq = seq()
-                .add(linear(
-                    dims[0].0,
-                    dims[0].1,
-                    vb.pp(format!("{prefix}-fc0")),
-                )?)
-                .add(Activation::Relu)
-                .add(linear(
-                    dims[1].0,
-                    dims[1].1,
-                    vb.pp(format!("{prefix}-fc1")),
-                )?)
-                .add(Activation::Relu)
-                .add(linear(
-                    dims[2].0,
-                    dims[2].1,
-                    vb.pp(format!("{prefix}-fc2")),
-                )?)
-                .add(func(|xs| xs.tanh()));
-            Ok::<Sequential, Error>(seq)
-        };
-
-        let network = make_network("actor")?;
-        let target_network = make_network("target-actor")?;
-
-        // this sets the two networks to be equal to each other using tau = 1.0
-        track(&mut varmap, &vb, "target-actor", "actor", &dims, 1.0);
-
-        Ok(Self {
-            varmap,
-            vb,
-            network,
-            target_network,
-            size_state,
-            size_action,
-            dims,
-        })
-    }
-
-    fn forward(&self, state: &Tensor) -> Result<Tensor> {
-        self.network.forward(state)
-    }
-
-    fn target_forward(&self, state: &Tensor) -> Result<Tensor> {
-        self.target_network.forward(state)
-    }
-
-    fn track(&mut self, tau: f64) -> Result<()> {
-        track(
-            &mut self.varmap,
-            &self.vb,
-            "target-actor",
-            "actor",
-            &self.dims,
-            tau,
-        )
-    }
-}
-
-struct Critic<'a> {
-    varmap: VarMap,
-    vb: VarBuilder<'a>,
-    network: Sequential,
-    target_network: Sequential,
-    size_state: usize,
-    size_action: usize,
-    dims: Vec<(usize, usize)>,
-}
-
-impl Critic<'_> {
-    fn new(device: &Device, dtype: DType, size_state: usize, size_action: usize) -> Result<Self> {
-        let mut varmap = VarMap::new();
-        let vb = VarBuilder::from_varmap(&varmap, dtype, device);
-
-        let dims: Vec<(usize, usize)> = vec![(size_state + size_action, 400), (400, 300), (300, 1)];
-
-        let make_network = |prefix: &str| {
-            let seq = seq()
-                .add(linear(
-                    dims[0].0,
-                    dims[0].1,
-                    vb.pp(format!("{prefix}-fc0")),
-                )?)
-                .add(Activation::Relu)
-                .add(linear(
-                    dims[1].0,
-                    dims[1].1,
-                    vb.pp(format!("{prefix}-fc1")),
-                )?)
-                .add(Activation::Relu)
-                .add(linear(
-                    dims[2].0,
-                    dims[2].1,
-                    vb.pp(format!("{prefix}-fc2")),
-                )?);
-            Ok::<Sequential, Error>(seq)
-        };
-
-        let network = make_network("critic")?;
-        let target_network = make_network("target-critic")?;
-
-        // this sets the two networks to be equal to each other using tau = 1.0
-        track(&mut varmap, &vb, "target-critic", "critic", &dims, 1.0);
-
-        Ok(Self {
-            varmap,
-            vb,
-            network,
-            target_network,
-            size_state,
-            size_action,
-            dims,
-        })
-    }
-
-    fn forward(&self, state: &Tensor, action: &Tensor) -> Result<Tensor> {
-        let xs = Tensor::cat(&[action, state], 1)?;
-        self.network.forward(&xs)
-    }
-
-    fn target_forward(&self, state: &Tensor, action: &Tensor) -> Result<Tensor> {
-        let xs = Tensor::cat(&[action, state], 1)?;
-        self.target_network.forward(&xs)
-    }
-
-    fn track(&mut self, tau: f64) -> Result<()> {
-        track(
-            &mut self.varmap,
-            &self.vb,
-            "target-critic",
-            "critic",
-            &self.dims,
-            tau,
-        )
-    }
-}
-
-#[allow(clippy::upper_case_acronyms)]
-pub struct DDPG<'a> {
-    actor: Actor<'a>,
-    actor_optim: AdamW,
-    critic: Critic<'a>,
-    critic_optim: AdamW,
-    gamma: f64,
-    tau: f64,
-    replay_buffer: ReplayBuffer,
-    ou_noise: OuNoise,
-
-    size_state: usize,
-    size_action: usize,
-    pub train: bool,
-}
-
-impl DDPG<'_> {
-    #[allow(clippy::too_many_arguments)]
-    pub fn new(
-        device: &Device,
-        size_state: usize,
-        size_action: usize,
-        train: bool,
-        actor_lr: f64,
-        critic_lr: f64,
-        gamma: f64,
-        tau: f64,
-        buffer_capacity: usize,
-        ou_noise: OuNoise,
-    ) -> Result<Self> {
-        let filter_by_prefix = |varmap: &VarMap, prefix: &str| {
-            varmap
-                .data()
-                .lock()
-                .unwrap()
-                .iter()
-                .filter_map(|(name, var)| name.starts_with(prefix).then_some(var.clone()))
-                .collect::<Vec<Var>>()
-        };
-
-        let actor = Actor::new(device, DType::F32, size_state, size_action)?;
-        let actor_optim = AdamW::new(
-            filter_by_prefix(&actor.varmap, "actor"),
-            ParamsAdamW {
-                lr: actor_lr,
-                ..Default::default()
-            },
-        )?;
-
-        let critic = Critic::new(device, DType::F32, size_state, size_action)?;
-        let critic_optim = AdamW::new(
-            filter_by_prefix(&critic.varmap, "critic"),
-            ParamsAdamW {
-                lr: critic_lr,
-                ..Default::default()
-            },
-        )?;
-
-        Ok(Self {
-            actor,
-            actor_optim,
-            critic,
-            critic_optim,
-            gamma,
-            tau,
-            replay_buffer: ReplayBuffer::new(buffer_capacity),
-            ou_noise,
-            size_state,
-            size_action,
-            train,
-        })
-    }
-
-    pub fn remember(
-        &mut self,
-        state: &Tensor,
-        action: &Tensor,
-        reward: &Tensor,
-        next_state: &Tensor,
-        terminated: bool,
-        truncated: bool,
-    ) {
-        self.replay_buffer
-            .push(state, action, reward, next_state, terminated, truncated)
-    }
-
-    pub fn actions(&mut self, state: &Tensor) -> Result<f32> {
-        let actions = self
-            .actor
-            .forward(&state.detach()?.unsqueeze(0)?)?
-            .squeeze(0)?;
-        let actions = if self.train {
-            (actions + self.ou_noise.sample()?)?
-        } else {
-            actions
-        };
-        actions.squeeze(0)?.to_scalar::<f32>()
-    }
-
-    pub fn train(&mut self, batch_size: usize) -> Result<()> {
-        let (states, actions, rewards, next_states, _, _) =
-            match self.replay_buffer.random_batch(batch_size)? {
-                Some(v) => v,
-                _ => return Ok(()),
-            };
-
-        let q_target = self
-            .critic
-            .target_forward(&next_states, &self.actor.target_forward(&next_states)?)?;
-        let q_target = (rewards + (self.gamma * q_target)?.detach())?;
-        let q = self.critic.forward(&states, &actions)?;
-        let diff = (q_target - q)?;
-
-        let critic_loss = diff.sqr()?.mean_all()?;
-        self.critic_optim.backward_step(&critic_loss)?;
-
-        let actor_loss = self
-            .critic
-            .forward(&states, &self.actor.forward(&states)?)?
-            .mean_all()?
-            .neg()?;
-        self.actor_optim.backward_step(&actor_loss)?;
-
-        self.critic.track(self.tau)?;
-        self.actor.track(self.tau)?;
-
-        Ok(())
-    }
-}
-
-// The impact of the q value of the next state on the current state's q value.
-const GAMMA: f64 = 0.99;
-// The weight for updating the target networks.
-const TAU: f64 = 0.005;
-// The capacity of the replay buffer used for sampling training data.
-const REPLAY_BUFFER_CAPACITY: usize = 100_000;
-// The training batch size for each training iteration.
-const TRAINING_BATCH_SIZE: usize = 100;
-// The total number of episodes.
-const MAX_EPISODES: usize = 100;
-// The maximum length of an episode.
-const EPISODE_LENGTH: usize = 200;
-// The number of training iterations after one episode finishes.
-const TRAINING_ITERATIONS: usize = 200;
-
-// Ornstein-Uhlenbeck process parameters.
-const MU: f64 = 0.0;
-const THETA: f64 = 0.15;
-const SIGMA: f64 = 0.1;
-
-const ACTOR_LEARNING_RATE: f64 = 1e-4;
-const CRITIC_LEARNING_RATE: f64 = 1e-3;
-
-pub fn run() -> Result<()> {
-    let env = GymEnv::new("Pendulum-v1")?;
-    println!("action space: {}", env.action_space());
-    println!("observation space: {:?}", env.observation_space());
-
-    let size_state = env.observation_space().iter().product::<usize>();
-    let size_action = env.action_space();
-
-    let mut agent = DDPG::new(
-        &Device::Cpu,
-        size_state,
-        size_action,
-        true,
-        ACTOR_LEARNING_RATE,
-        CRITIC_LEARNING_RATE,
-        GAMMA,
-        TAU,
-        REPLAY_BUFFER_CAPACITY,
-        OuNoise::new(MU, THETA, SIGMA, size_action)?,
-    )?;
-
-    let mut rng = rand::thread_rng();
-
-    for episode in 0..MAX_EPISODES {
-        // let mut state = env.reset(episode as u64)?;
-        let mut state = env.reset(rng.gen::<u64>())?;
-
-        let mut total_reward = 0.0;
-        for _ in 0..EPISODE_LENGTH {
-            let mut action = 2.0 * agent.actions(&state)?;
-            action = action.clamp(-2.0, 2.0);
-
-            let step = env.step(vec![action])?;
-            total_reward += step.reward;
-
-            agent.remember(
-                &state,
-                &Tensor::new(vec![action], &Device::Cpu)?,
-                &Tensor::new(vec![step.reward as f32], &Device::Cpu)?,
-                &step.state,
-                step.terminated,
-                step.truncated,
-            );
-
-            if step.terminated || step.truncated {
-                break;
-            }
-            state = step.state;
-        }
-
-        println!("episode {episode} with total reward of {total_reward}");
-
-        for _ in 0..TRAINING_ITERATIONS {
-            agent.train(TRAINING_BATCH_SIZE)?;
-        }
-    }
-
-    println!("Testing...");
-    agent.train = false;
-    for episode in 0..10 {
-        // let mut state = env.reset(episode as u64)?;
-        let mut state = env.reset(rng.gen::<u64>())?;
-        let mut total_reward = 0.0;
-        for _ in 0..EPISODE_LENGTH {
-            let mut action = 2.0 * agent.actions(&state)?;
-            action = action.clamp(-2.0, 2.0);
-
-            let step = env.step(vec![action])?;
-            total_reward += step.reward;
-
-            if step.terminated || step.truncated {
-                break;
-            }
-            state = step.state;
-        }
-        println!("episode {episode} with total reward of {total_reward}");
-    }
-    Ok(())
-}

candle-examples/examples/reinforcement-learning/gym_env.rs

@@ -1,112 +0,0 @@
-#![allow(unused)]
-//! Wrappers around the Python API of Gymnasium (the new version of OpenAI gym)
-use candle::{Device, Result, Tensor};
-use pyo3::prelude::*;
-use pyo3::types::PyDict;
-
-/// The return value for a step.
-#[derive(Debug)]
-pub struct Step<A> {
-    pub state: Tensor,
-    pub action: A,
-    pub reward: f64,
-    pub terminated: bool,
-    pub truncated: bool,
-}
-
-impl<A: Copy> Step<A> {
-    /// Returns a copy of this step changing the observation tensor.
-    pub fn copy_with_obs(&self, state: &Tensor) -> Step<A> {
-        Step {
-            state: state.clone(),
-            action: self.action,
-            reward: self.reward,
-            terminated: self.terminated,
-            truncated: self.truncated,
-        }
-    }
-}
-
-/// An OpenAI Gym session.
-pub struct GymEnv {
-    env: PyObject,
-    action_space: usize,
-    observation_space: Vec<usize>,
-}
-
-fn w(res: PyErr) -> candle::Error {
-    candle::Error::wrap(res)
-}
-
-impl GymEnv {
-    /// Creates a new session of the specified OpenAI Gym environment.
-    pub fn new(name: &str) -> Result<GymEnv> {
-        Python::with_gil(|py| {
-            let gym = py.import("gymnasium")?;
-            let make = gym.getattr("make")?;
-            let env = make.call1((name,))?;
-            let action_space = env.getattr("action_space")?;
-            let action_space = if let Ok(val) = action_space.getattr("n") {
-                val.extract()?
-            } else {
-                let action_space: Vec<usize> = action_space.getattr("shape")?.extract()?;
-                action_space[0]
-            };
-            let observation_space = env.getattr("observation_space")?;
-            let observation_space = observation_space.getattr("shape")?.extract()?;
-            Ok(GymEnv {
-                env: env.into(),
-                action_space,
-                observation_space,
-            })
-        })
-        .map_err(w)
-    }
-
-    /// Resets the environment, returning the observation tensor.
-    pub fn reset(&self, seed: u64) -> Result<Tensor> {
-        let state: Vec<f32> = Python::with_gil(|py| {
-            let kwargs = PyDict::new(py);
-            kwargs.set_item("seed", seed)?;
-            let state = self.env.call_method(py, "reset", (), Some(kwargs))?;
-            state.as_ref(py).get_item(0)?.extract()
-        })
-        .map_err(w)?;
-        Tensor::new(state, &Device::Cpu)
-    }
-
-    /// Applies an environment step using the specified action.
-    pub fn step<A: pyo3::IntoPy<pyo3::Py<pyo3::PyAny>> + Clone>(
-        &self,
-        action: A,
-    ) -> Result<Step<A>> {
-        let (state, reward, terminated, truncated) = Python::with_gil(|py| {
-            let step = self.env.call_method(py, "step", (action.clone(),), None)?;
-            let step = step.as_ref(py);
-            let state: Vec<f32> = step.get_item(0)?.extract()?;
-            let reward: f64 = step.get_item(1)?.extract()?;
-            let terminated: bool = step.get_item(2)?.extract()?;
-            let truncated: bool = step.get_item(3)?.extract()?;
-            Ok((state, reward, terminated, truncated))
-        })
-        .map_err(w)?;
-        let state = Tensor::new(state, &Device::Cpu)?;
-        Ok(Step {
-            state,
-            action,
-            reward,
-            terminated,
-            truncated,
-        })
-    }
-
-    /// Returns the number of allowed actions for this environment.
-    pub fn action_space(&self) -> usize {
-        self.action_space
-    }
-
-    /// Returns the shape of the observation tensors.
-    pub fn observation_space(&self) -> &[usize] {
-        &self.observation_space
-    }
-}

candle-examples/examples/reinforcement-learning/main.rs

@@ -1,37 +0,0 @@
-#![allow(unused)]
-
-#[cfg(feature = "mkl")]
-extern crate intel_mkl_src;
-
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-
-use candle::Result;
-use clap::{Parser, Subcommand};
-
-mod gym_env;
-mod vec_gym_env;
-
-mod ddpg;
-mod policy_gradient;
-
-#[derive(Parser)]
-struct Args {
-    #[command(subcommand)]
-    command: Command,
-}
-
-#[derive(Subcommand)]
-enum Command {
-    Pg,
-    Ddpg,
-}
-
-fn main() -> Result<()> {
-    let args = Args::parse();
-    match args.command {
-        Command::Pg => policy_gradient::run()?,
-        Command::Ddpg => ddpg::run()?,
-    }
-    Ok(())
-}

candle-examples/examples/reinforcement-learning/policy_gradient.rs

@@ -1,146 +0,0 @@
-use super::gym_env::{GymEnv, Step};
-use candle::{DType, Device, Error, Module, Result, Tensor};
-use candle_nn::{
-    linear, ops::log_softmax, ops::softmax, sequential::seq, Activation, AdamW, Optimizer,
-    ParamsAdamW, VarBuilder, VarMap,
-};
-use rand::{distributions::Distribution, rngs::ThreadRng, Rng};
-
-fn new_model(
-    input_shape: &[usize],
-    num_actions: usize,
-    dtype: DType,
-    device: &Device,
-) -> Result<(impl Module, VarMap)> {
-    let input_size = input_shape.iter().product();
-
-    let mut varmap = VarMap::new();
-    let var_builder = VarBuilder::from_varmap(&varmap, dtype, device);
-
-    let model = seq()
-        .add(linear(input_size, 32, var_builder.pp("lin1"))?)
-        .add(Activation::Relu)
-        .add(linear(32, num_actions, var_builder.pp("lin2"))?);
-
-    Ok((model, varmap))
-}
-
-fn accumulate_rewards(steps: &[Step<i64>]) -> Vec<f64> {
-    let mut rewards: Vec<f64> = steps.iter().map(|s| s.reward).collect();
-    let mut acc_reward = 0f64;
-    for (i, reward) in rewards.iter_mut().enumerate().rev() {
-        if steps[i].terminated {
-            acc_reward = 0.0;
-        }
-        acc_reward += *reward;
-        *reward = acc_reward;
-    }
-    rewards
-}
-
-fn weighted_sample(probs: Vec<f32>, rng: &mut ThreadRng) -> Result<usize> {
-    let distribution = rand::distributions::WeightedIndex::new(probs).map_err(Error::wrap)?;
-    let mut rng = rng;
-    Ok(distribution.sample(&mut rng))
-}
-
-pub fn run() -> Result<()> {
-    let env = GymEnv::new("CartPole-v1")?;
-
-    println!("action space: {:?}", env.action_space());
-    println!("observation space: {:?}", env.observation_space());
-
-    let (model, varmap) = new_model(
-        env.observation_space(),
-        env.action_space(),
-        DType::F32,
-        &Device::Cpu,
-    )?;
-
-    let optimizer_params = ParamsAdamW {
-        lr: 0.01,
-        weight_decay: 0.01,
-        ..Default::default()
-    };
-
-    let mut optimizer = AdamW::new(varmap.all_vars(), optimizer_params)?;
-
-    let mut rng = rand::thread_rng();
-
-    for epoch_idx in 0..100 {
-        let mut state = env.reset(rng.gen::<u64>())?;
-        let mut steps: Vec<Step<i64>> = vec![];
-
-        loop {
-            let action = {
-                let action_probs: Vec<f32> =
-                    softmax(&model.forward(&state.detach()?.unsqueeze(0)?)?, 1)?
-                        .squeeze(0)?
-                        .to_vec1()?;
-                weighted_sample(action_probs, &mut rng)? as i64
-            };
-
-            let step = env.step(action)?;
-            steps.push(step.copy_with_obs(&state));
-
-            if step.terminated || step.truncated {
-                state = env.reset(rng.gen::<u64>())?;
-                if steps.len() > 5000 {
-                    break;
-                }
-            } else {
-                state = step.state;
-            }
-        }
-
-        let total_reward: f64 = steps.iter().map(|s| s.reward).sum();
-        let episodes: i64 = steps
-            .iter()
-            .map(|s| (s.terminated || s.truncated) as i64)
-            .sum();
-        println!(
-            "epoch: {:<3} episodes: {:<5} avg reward per episode: {:.2}",
-            epoch_idx,
-            episodes,
-            total_reward / episodes as f64
-        );
-
-        let batch_size = steps.len();
-
-        let rewards = Tensor::from_vec(accumulate_rewards(&steps), batch_size, &Device::Cpu)?
-            .to_dtype(DType::F32)?
-            .detach()?;
-
-        let actions_mask = {
-            let actions: Vec<i64> = steps.iter().map(|s| s.action).collect();
-            let actions_mask: Vec<Tensor> = actions
-                .iter()
-                .map(|&action| {
-                    // One-hot encoding
-                    let mut action_mask = vec![0.0; env.action_space()];
-                    action_mask[action as usize] = 1.0;
-
-                    Tensor::from_vec(action_mask, env.action_space(), &Device::Cpu)
-                        .unwrap()
-                        .to_dtype(DType::F32)
-                        .unwrap()
-                })
-                .collect();
-            Tensor::stack(&actions_mask, 0)?.detach()?
-        };
-
-        let states = {
-            let states: Vec<Tensor> = steps.into_iter().map(|s| s.state).collect();
-            Tensor::stack(&states, 0)?.detach()?
-        };
-
-        let log_probs = actions_mask
-            .mul(&log_softmax(&model.forward(&states)?, 1)?)?
-            .sum(1)?;
-
-        let loss = rewards.mul(&log_probs)?.neg()?.mean_all()?;
-        optimizer.backward_step(&loss)?;
-    }
-
-    Ok(())
-}