Bump the crate version to 0.8.0. (#2612)

* add module docs for candle-core

* doc each of the candle-nn modules and add the links to the doc page

.github/workflows/ci_cuda.yaml

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

.github/workflows/python.yml

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

.github/workflows/rust-ci.yml

@@ -1,6 +1,6 @@
-on: 
+on:
   push:
-    branches: 
+    branches:
       - main
   pull_request:
 
@@ -15,7 +15,7 @@ jobs:
         os: [ubuntu-latest, windows-latest, macOS-latest]
         rust: [stable]
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal
@@ -34,7 +34,7 @@ jobs:
         os: [ubuntu-latest, windows-latest, macOS-latest]
         rust: [stable]
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal
@@ -49,7 +49,7 @@ jobs:
     name: Rustfmt
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal
@@ -65,7 +65,7 @@ jobs:
     name: Clippy
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
           profile: minimal

.github/workflows/trufflehog.yml

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

.gitignore

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

Cargo.toml

@@ -20,7 +20,7 @@ exclude = [
 resolver = "2"
 
 [workspace.package]
-version = "0.5.0"
+version = "0.8.0"
 edition = "2021"
 description = "Minimalist ML framework."
 repository = "https://github.com/huggingface/candle"
@@ -33,22 +33,23 @@ ab_glyph = "0.2.23"
 accelerate-src = { version = "0.3.2" }
 anyhow = { version = "1", features = ["backtrace"] }
 byteorder = "1.4.3"
-candle = { path = "./candle-core", package = "candle-core", version = "0.5.0" }
-candle-datasets = { path = "./candle-datasets", version = "0.5.0" }
-candle-flash-attn = { path = "./candle-flash-attn", version = "0.5.0" }
-candle-kernels = { path = "./candle-kernels", version = "0.5.0" }
-candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.5.0" }
-candle-nn = { path = "./candle-nn", version = "0.5.0" }
-candle-onnx = { path = "./candle-onnx", version = "0.5.0" }
-candle-transformers = { path = "./candle-transformers", version = "0.5.0" }
+candle = { path = "./candle-core", package = "candle-core", version = "0.8.0" }
+candle-datasets = { path = "./candle-datasets", version = "0.8.0" }
+candle-flash-attn = { path = "./candle-flash-attn", version = "0.8.0" }
+candle-kernels = { path = "./candle-kernels", version = "0.8.0" }
+candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.8.0" }
+candle-nn = { path = "./candle-nn", version = "0.8.0" }
+candle-onnx = { path = "./candle-onnx", version = "0.8.0" }
+candle-transformers = { path = "./candle-transformers", version = "0.8.0" }
 clap = { version = "4.2.4", features = ["derive"] }
 criterion = { version = "0.5.1", default-features=false }
-cudarc = { version = "0.10.0", features = ["f16"] }
+cudarc = { version = "0.12.1", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16", "cuda-version-from-build-system", "dynamic-linking"], default-features=false }
 fancy-regex = "0.13.0"
 gemm = { version = "0.17.0", features = ["wasm-simd128-enable"] }
-hf-hub = "0.3.0"
+hf-hub = { version = "0.3.3", package = "candle-hf-hub" }
 half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
-image = { version = "0.25.0", default-features = false, features = ["jpeg", "png"] }
+hound = "3.5.1"
+image = { version = "0.25.2", default-features = false, features = ["jpeg", "png"] }
 imageproc = { version = "0.24.0", default-features = false }
 intel-mkl-src = { version = "0.8.1", features = ["mkl-static-lp64-iomp"] }
 libc = { version = "0.2.147" }
@@ -65,13 +66,15 @@ serde = { version = "1.0.171", features = ["derive"] }
 serde_plain = "1.0.2"
 serde_json = "1.0.99"
 thiserror = "1"
-tokenizers = { version = "0.15.0", default-features = false }
+tokenizers = { version = "0.19.1", default-features = false }
 tracing = "0.1.37"
 tracing-chrome = "0.7.1"
 tracing-subscriber = "0.3.7"
-wav = "1.0.0"
+ug = "0.0.2"
+ug-cuda = "0.0.2"
+ug-metal = "0.0.2"
 yoke = { version = "0.7.2", features = ["derive"] }
-zip = { version = "0.6.6", default-features = false }
+zip = { version = "1.1.1", default-features = false }
 metal = { version = "0.27.0", features = ["mps"]}
 
 [profile.release-with-debug]

README.md

@@ -2,7 +2,8 @@
 [![discord server](https://dcbadge.vercel.app/api/server/hugging-face-879548962464493619)](https://discord.gg/hugging-face-879548962464493619)
 [![Latest version](https://img.shields.io/crates/v/candle-core.svg)](https://crates.io/crates/candle-core)
 [![Documentation](https://docs.rs/candle-core/badge.svg)](https://docs.rs/candle-core)
-![License](https://img.shields.io/crates/l/candle-core.svg)
+[![License](https://img.shields.io/github/license/base-org/node?color=blue)](https://github.com/huggingface/candle/blob/main/LICENSE-MIT)
+[![License](https://img.shields.io/badge/license-Apache%202.0-blue?style=flat-square)](https://github.com/huggingface/candle/blob/main/LICENSE-APACHE)
 
 Candle is a minimalist ML framework for Rust with a focus on performance (including GPU support) 
 and ease of use. Try our online demos: 
@@ -60,12 +61,16 @@ These online demos run entirely in your browser:
 
 We also provide a some command line based examples using state of the art models:
 
-- [LLaMA and LLaMA-v2](./candle-examples/examples/llama/): general LLM, includes
+- [LLaMA v1, v2, and v3](./candle-examples/examples/llama/): general LLM, includes
   the SOLAR-10.7B variant.
 - [Falcon](./candle-examples/examples/falcon/): general LLM.
-- [Gemma](./candle-examples/examples/gemma/): 2b and 7b general LLMs from Google
-  Deepmind.
-- [Phi-1, Phi-1.5, and Phi-2](./candle-examples/examples/phi/): 1.3b and 2.7b general LLMs with performance on par with LLaMA-v2 7b.
+- [Codegeex4](./candle-examples/examples/codegeex4-9b/): Code completion,code interpreter,web search,fuction calling,repository-level
+- [GLM4](./candle-examples/examples/glm4/): Open Multilingual Multimodal Chat LMs by THUDM
+- [Gemma v1 and v2](./candle-examples/examples/gemma/): 2b and 7b+/9b general LLMs from Google Deepmind.
+- [RecurrentGemma](./candle-examples/examples/recurrent-gemma/): 2b and 7b
+  Griffin based models from Google that mix attention with a RNN like state.
+- [Phi-1, Phi-1.5, Phi-2, and Phi-3](./candle-examples/examples/phi/): 1.3b,
+  2.7b, and 3.8b general LLMs with performance on par with 7b models.
 - [StableLM-3B-4E1T](./candle-examples/examples/stable-lm/): a 3b general LLM
   pre-trained on 1T tokens of English and code datasets. Also supports
   StableLM-2, a 1.6b LLM trained on 2T tokens, as well as the code variants.
@@ -110,12 +115,14 @@ We also provide a some command line based examples using state of the art models
 
 <img src="https://github.com/huggingface/candle/raw/main/candle-examples/examples/segment-anything/assets/sam_merged.jpg" width="200">
 
-- [SegFormer](./candle-examples/examples/segformer/): transformer based semantic segmantation model.
+- [SegFormer](./candle-examples/examples/segformer/): transformer based semantic segmentation model.
 - [Whisper](./candle-examples/examples/whisper/): speech recognition model.
 - [EnCodec](./candle-examples/examples/encodec/): high-quality audio compression
   model using residual vector quantization.
 - [MetaVoice](./candle-examples/examples/metavoice/): foundational model for
   text-to-speech.
+- [Parler-TTS](./candle-examples/examples/parler-tts/): large text-to-speech
+  model.
 - [T5](./candle-examples/examples/t5), [Bert](./candle-examples/examples/bert/),
   [JinaBert](./candle-examples/examples/jina-bert/) : useful for sentence embeddings.
 - [DINOv2](./candle-examples/examples/dinov2/): computer vision model trained
@@ -181,6 +188,7 @@ And then head over to
 - [`candle-sampling`](https://github.com/EricLBuehler/candle-sampling): Sampling techniques for Candle.
 - [`gpt-from-scratch-rs`](https://github.com/jeroenvlek/gpt-from-scratch-rs): A port of Andrej Karpathy's _Let's build GPT_ tutorial on YouTube showcasing the Candle API on a toy problem.
 - [`candle-einops`](https://github.com/tomsanbear/candle-einops): A pure rust implementation of the python [einops](https://github.com/arogozhnikov/einops) library.
+- [`atoma-infer`](https://github.com/atoma-network/atoma-infer): A Rust library for fast inference at scale, leveraging FlashAttention2 for efficient attention computation, PagedAttention for efficient KV-cache memory management, and multi-GPU support. It is OpenAI api compatible.
 
 If you have an addition to this list, please submit a pull request.
 
@@ -199,12 +207,12 @@ If you have an addition to this list, please submit a pull request.
     - WASM support, run your models in a browser.
 - Included models.
     - Language Models.
-        - LLaMA v1 and v2 with variants such as SOLAR-10.7B.
+        - LLaMA v1, v2, and v3 with variants such as SOLAR-10.7B.
         - Falcon.
         - StarCoder, StarCoder2.
-        - Phi 1, 1.5, and 2.
+        - Phi 1, 1.5, 2, and 3.
         - Mamba, Minimal Mamba
-        - Gemma 2b and 7b.
+        - Gemma v1 2b and 7b+, v2 2b and 9b.
         - Mistral 7b v0.1.
         - Mixtral 8x7b v0.1.
         - StableLM-3B-4E1T, StableLM-2-1.6B, Stable-Code-3B.
@@ -232,9 +240,10 @@ If you have an addition to this list, please submit a pull request.
         - Whisper, multi-lingual speech-to-text.
         - EnCodec, audio compression model.
         - MetaVoice-1B, text-to-speech model.
+        - Parler-TTS, text-to-speech model.
     - Computer Vision Models.
         - DINOv2, ConvMixer, EfficientNet, ResNet, ViT, VGG, RepVGG, ConvNeXT,
-          ConvNeXTv2, MobileOne, EfficientVit (MSRA).
+          ConvNeXTv2, MobileOne, EfficientVit (MSRA), MobileNetv4, Hiera, FastViT.
         - yolo-v3, yolo-v8.
         - Segment-Anything Model (SAM).
         - SegFormer.
@@ -374,9 +383,9 @@ git submodule update --init
 /usr/include/c++/11/bits/std_function.h:530:146: error: parameter packs not expanded with ‘...’:
 ```
 
-This is a bug in gcc-11 triggered by the Cuda compiler. To fix this, install a different, supported gcc version - for example gcc-10, and specify the path to the compiler in the CANDLE_NVCC_CCBIN environment variable.
+This is a bug in gcc-11 triggered by the Cuda compiler. To fix this, install a different, supported gcc version - for example gcc-10, and specify the path to the compiler in the NVCC_CCBIN environment variable.
 ```
-env CANDLE_NVCC_CCBIN=/usr/lib/gcc/x86_64-linux-gnu/10 cargo ...
+env NVCC_CCBIN=/usr/lib/gcc/x86_64-linux-gnu/10 cargo ...
 ```
 
 #### Linking error on windows when running rustdoc or mdbook tests
@@ -406,3 +415,10 @@ This may be caused by the models being loaded from `/mnt/c`, more details on
 
 You can set `RUST_BACKTRACE=1` to be provided with backtraces when a candle
 error is generated.
+
+#### CudaRC error
+
+If you encounter an error like this one `called `Result::unwrap()` on an `Err` value: LoadLibraryExW { source: Os { code: 126, kind: Uncategorized, message: "The specified module could not be found." } }` on windows. To fix copy and rename these 3 files (make sure they are in path). The paths depend on your cuda version.
+`c:\Windows\System32\nvcuda.dll` -> `cuda.dll`
+`c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin\cublas64_12.dll` -> `cublas.dll`
+`c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin\curand64_10.dll` -> `curand.dll`

candle-book/Cargo.toml

@@ -37,7 +37,6 @@ tokenizers = { workspace = true, features = ["onig"] }
 tracing = { workspace = true }
 tracing-chrome = { workspace = true }
 tracing-subscriber = { workspace = true }
-wav = { workspace = true }
 # Necessary to disambiguate with tokio in wasm examples which are 1.28.1
 parquet = { workspace = true }
 image = { workspace = true }

candle-book/src/inference/hub.md

@@ -11,8 +11,8 @@ Then let's start by downloading the [model file](https://huggingface.co/bert-bas
 
 ```rust
 # extern crate candle_core;
-# extern crate hf_hub;
-use hf_hub::api::sync::Api;
+# extern crate candle_hf_hub;
+use candle_hf_hub::api::sync::Api;
 use candle_core::Device;
 
 let api = Api::new().unwrap();
@@ -50,8 +50,8 @@ Now that we have our weights, we can use them in our bert architecture:
 ```rust
 # extern crate candle_core;
 # extern crate candle_nn;
-# extern crate hf_hub;
-# use hf_hub::api::sync::Api;
+# extern crate candle_hf_hub;
+# use candle_hf_hub::api::sync::Api;
 # 
 # let api = Api::new().unwrap();
 # let repo = api.model("bert-base-uncased".to_string());

candle-book/src/lib.rs

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

candle-core/Cargo.toml

@@ -28,6 +28,9 @@ rand_distr = { workspace = true }
 rayon = { workspace = true }
 safetensors = { workspace = true }
 thiserror = { workspace = true }
+ug = { workspace = true }
+ug-cuda = { workspace = true, optional = true }
+ug-metal = { workspace = true, optional = true }
 yoke = { workspace = true }
 zip = { workspace = true }
 
@@ -39,12 +42,16 @@ criterion = { workspace = true }
 
 [features]
 default = []
-cuda = ["cudarc", "dep:candle-kernels"]
+cuda = ["cudarc", "dep:candle-kernels", "dep:ug-cuda"]
 cudnn = ["cuda", "cudarc/cudnn"]
 mkl = ["dep:libc", "dep:intel-mkl-src"]
 accelerate = ["dep:libc", "dep:accelerate-src"]
-metal = ["dep:metal", "dep:candle-metal-kernels"]
+metal = ["dep:metal", "dep:candle-metal-kernels", "dep:ug-metal"]
 
 [[bench]]
 name = "bench_main"
 harness = false
+
+[[example]]
+name = "metal_basics"
+required-features = ["metal"]

candle-core/benches/bench_main.rs

@@ -7,4 +7,6 @@ criterion_main!(
     benchmarks::random::benches,
     benchmarks::where_cond::benches,
     benchmarks::conv_transpose2d::benches,
+    benchmarks::qmatmul::benches,
+    benchmarks::unary::benches
 );

candle-core/benches/benchmarks/affine.rs

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

candle-core/benches/benchmarks/mod.rs

@@ -1,7 +1,9 @@
 pub(crate) mod affine;
 pub(crate) mod conv_transpose2d;
 pub(crate) mod matmul;
+pub(crate) mod qmatmul;
 pub(crate) mod random;
+pub(crate) mod unary;
 pub(crate) mod where_cond;
 
 use candle_core::{Device, Result};

candle-core/benches/benchmarks/qmatmul.rs

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

candle-core/benches/benchmarks/unary.rs

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

candle-core/benches/benchmarks/where_cond.rs

@@ -25,9 +25,9 @@ const SIZE: usize = B * M * K;
 const DATA: [u8; SIZE] = create_cond_arr::<SIZE>();
 
 fn run_where_cond_benchmark(c: &mut Criterion, device: &Device, dtype: DType, name: &str) {
-    let tensor = Tensor::from_slice(DATA.as_slice(), (B, M, K), &device).unwrap();
-    let on_true = Tensor::ones((B, M, K), dtype, &device).unwrap();
-    let on_false = Tensor::zeros((B, M, K), dtype, &device).unwrap();
+    let tensor = Tensor::from_slice(DATA.as_slice(), (B, M, K), device).unwrap();
+    let on_true = Tensor::ones((B, M, K), dtype, device).unwrap();
+    let on_false = Tensor::zeros((B, M, K), dtype, device).unwrap();
 
     let elements = B * M * K;
     // E.g. 2 f32 tensors + 1 u8 tensor

candle-core/examples/cuda_basics.rs

@@ -5,32 +5,29 @@ extern crate accelerate_src;
 extern crate intel_mkl_src;
 
 use anyhow::Result;
-use candle_core::{Device, Module, Tensor};
-
-use candle_core::quantized::{QMatMul, QTensor};
+use candle_core::{Device, Tensor};
 
 fn main() -> Result<()> {
     let device = Device::new_cuda(0)?;
-    let q = Tensor::randn(0f32, 1.0, (72, 256), &device)?;
-    let q_cpu = q.to_device(&Device::Cpu)?;
-    let q = QTensor::quantize(&q, candle_core::quantized::GgmlDType::Q8K)?;
-    let q = QMatMul::from_qtensor(q)?;
-    let x = Tensor::randn(0f32, 1.0, (5, 256), &device)?;
-    let res_q_cuda = q.forward(&x)?;
-    println!("{res_q_cuda}");
-
-    let q_cpu = QTensor::quantize(&q_cpu, candle_core::quantized::GgmlDType::Q8K)?;
-    let q_cpu_tensor = q_cpu.dequantize(&Device::Cpu)?;
-    let q_cpu = QMatMul::from_qtensor(q_cpu)?;
-    let x_cpu = x.to_device(&Device::Cpu)?;
-    let res_q_cpu = q_cpu.forward(&x_cpu)?;
-    println!("{res_q_cpu}");
-
-    let res_mm = x_cpu.matmul(&q_cpu_tensor.t()?)?;
-    let diff = (res_mm - res_q_cuda.to_device(&Device::Cpu))?
-        .abs()?
-        .flatten_all()?
-        .max(0)?;
-    println!("{diff}");
+    let x = Tensor::randn(0f32, 1.0, (8 * 4096, 8 * 4096), &device)?
+        .to_dtype(candle_core::DType::BF16)?;
+    candle_core::cuda::set_gemm_reduced_precision_f32(false);
+    candle_core::cuda::set_gemm_reduced_precision_bf16(false);
+    let _x1 = x.matmul(&x)?;
+    drop(_x1);
+    let start_time = std::time::Instant::now();
+    let _x1 = x.matmul(&x)?;
+    device.synchronize()?;
+    println!("fp32: {:?}", start_time.elapsed());
+    drop(_x1);
+    candle_core::cuda::set_gemm_reduced_precision_f32(true);
+    candle_core::cuda::set_gemm_reduced_precision_bf16(true);
+    let _x1 = x.matmul(&x)?;
+    drop(_x1);
+    let start_time = std::time::Instant::now();
+    let _x1 = x.matmul(&x)?;
+    device.synchronize()?;
+    println!("tf32: {:?}", start_time.elapsed());
+    drop(_x1);
     Ok(())
 }

candle-core/examples/metal_basics.rs

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

candle-core/src/backend.rs

@@ -133,6 +133,8 @@ pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
     /// after this call.
     unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;
 
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage>;
+
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage>;
 
     fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage>;
@@ -142,4 +144,7 @@ pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;
 
     fn set_seed(&self, _: u64) -> Result<()>;
+
+    /// Synchronize should block until all the operations on the device are completed.
+    fn synchronize(&self) -> Result<()>;
 }

candle-core/src/backprop.rs

@@ -320,13 +320,13 @@ impl Tensor {
                         dilation,
                         output_padding: _output_padding,
                     } => {
-                        let grad_arg = grad.conv2d(kernel, *padding, *dilation, *stride, 1)?;
+                        let grad_arg = grad.conv2d(kernel, *padding, *stride, *dilation, 1)?;
                         let sum_grad = grads.or_insert(arg)?;
                         *sum_grad = sum_grad.add(&grad_arg)?;
 
                         let grad_kernel = grad
                             .transpose(0, 1)?
-                            .conv2d(&arg.transpose(0, 1)?, *padding, *stride, *dilation, 1)?
+                            .conv2d(&arg.transpose(0, 1)?, *padding, *dilation, *stride, 1)?
                             .transpose(0, 1)?;
                         let sum_grad = grads.or_insert(kernel)?;
                         let (_, _, k0, k1) = kernel.dims4()?;
@@ -623,9 +623,9 @@ impl Tensor {
                     }
                     Op::Unary(arg, UnaryOp::Silu) => {
                         let sum_grad = grads.or_insert(arg)?;
-                        // d/dx silu = sigmoid(x) * (1 + x * (1 - sigmoid(x)))
-                        let sigmoid_arg = (*node / arg)?;
-                        let silu_grad = (&sigmoid_arg * (1. + (arg * (1. - &sigmoid_arg)?)?)?)?;
+                        // d/dx silu = sigmoid(x) * (1 + x * (1 - sigmoid(x))) = sigmoid(x) * (1 - node) + node
+                        let sigmoid_arg = (arg.neg()?.exp()? + 1.)?.recip()?;
+                        let silu_grad = &sigmoid_arg * (1. - *node) + *node;
                         *sum_grad = sum_grad.add(&(&grad * silu_grad)?)?
                     }
                     Op::Elu(arg, alpha) => {
@@ -634,7 +634,8 @@ impl Tensor {
                         let zeros = arg.zeros_like()?;
                         let positive_mask = arg.gt(&zeros)?.to_dtype(arg.dtype())?;
                         let negative_mask = arg.le(&zeros)?.to_dtype(arg.dtype())?;
-                        let negative_exp_mask = ((negative_mask * arg.exp())? * *alpha)?;
+                        // node == alpha * (e^x - 1) for x <= 0, reuse it
+                        let negative_exp_mask = (negative_mask * (*node + *alpha))?;
                         let combined_mask = (positive_mask + negative_exp_mask)?;
                         *sum_grad = sum_grad.add(&(grad * combined_mask)?)?
                     }
@@ -755,4 +756,9 @@ impl GradStore {
         };
         Ok(grad)
     }
+
+    /// Get the tensor ids of the stored gradient tensors
+    pub fn get_ids(&self) -> impl Iterator<Item = &TensorId> {
+        self.0.keys()
+    }
 }

candle-core/src/cpu/mod.rs

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

candle-core/src/cpu_backend/mod.rs

@@ -10,7 +10,7 @@ pub use utils::{
 };
 
 const USE_IM2COL_CONV1D: bool = true;
-const USE_IM2COL_CONV1D_TR: bool = true;
+const USE_COL2IM_CONV1D_TR: bool = true;
 const USE_IM2COL_CONV2D: bool = true;
 
 // TODO: Maybe we should not implement [Clone] here and instead have an explicit allocator +
@@ -26,6 +26,17 @@ pub enum CpuStorage {
     F64(Vec<f64>),
 }
 
+#[derive(Debug, Clone)]
+pub enum CpuStorageRef<'a> {
+    U8(&'a [u8]),
+    U32(&'a [u32]),
+    I64(&'a [i64]),
+    BF16(&'a [bf16]),
+    F16(&'a [f16]),
+    F32(&'a [f32]),
+    F64(&'a [f64]),
+}
+
 #[derive(Debug, Clone)]
 pub struct CpuDevice;
 
@@ -110,7 +121,8 @@ impl ReduceIndex {
         let dst_len = src_l.shape().elem_count() / reduce_dim_size;
         let mut dst: Vec<U> = Vec::with_capacity(dst_len);
         let dst_to_set = dst.spare_capacity_mut();
-        let dst_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(dst_to_set) };
+        let dst_to_set =
+            unsafe { std::mem::transmute::<&mut [std::mem::MaybeUninit<U>], &mut [U]>(dst_to_set) };
         match src_l.contiguous_offsets() {
             Some((o1, o2)) => {
                 let src = &src[o1..o2];
@@ -2238,7 +2250,7 @@ impl BackendStorage for CpuStorage {
             && params.dilation == 1
             && params.padding == 0
             && params.output_padding == 0;
-        if USE_IM2COL_CONV1D_TR && can_use_col2im {
+        if USE_COL2IM_CONV1D_TR && can_use_col2im {
             let (b_size, c_in, l_in) = l.shape().dims3()?;
             let (c_in2, c_out, k_size) = kernel_l.shape().dims3()?;
             if !kernel_l.is_contiguous() {
@@ -2445,6 +2457,10 @@ impl BackendDevice for CpuDevice {
         true
     }
 
+    fn storage_from_slice<T: crate::WithDType>(&self, s: &[T]) -> Result<Self::Storage> {
+        Ok(T::to_cpu_storage(s))
+    }
+
     fn storage_from_cpu_storage(&self, s: &CpuStorage) -> Result<Self::Storage> {
         Ok(s.clone())
     }
@@ -2628,6 +2644,10 @@ impl BackendDevice for CpuDevice {
         };
         Ok(storage)
     }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
 }
 
 #[macro_export]

candle-core/src/cpu_backend/utils.rs

@@ -174,7 +174,9 @@ pub fn binary_map_vec<T: Copy, F: FnMut(T, T) -> T, FV: FnMut(&[T], &[T], &mut [
         (Some((o_l1, o_l2)), Some((o_r1, o_r2))) => {
             let mut ys: Vec<T> = Vec::with_capacity(el_count);
             let ys_to_set = ys.spare_capacity_mut();
-            let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
+            let ys_to_set = unsafe {
+                std::mem::transmute::<&mut [std::mem::MaybeUninit<T>], &mut [T]>(ys_to_set)
+            };
             f_vec(&lhs[o_l1..o_l2], &rhs[o_r1..o_r2], ys_to_set);
             // SAFETY: values are all set by f_vec.
             unsafe { ys.set_len(el_count) };
@@ -185,7 +187,9 @@ pub fn binary_map_vec<T: Copy, F: FnMut(T, T) -> T, FV: FnMut(&[T], &[T], &mut [
                 let rhs = &rhs[ob.start..ob.start + ob.len];
                 let mut ys: Vec<T> = Vec::with_capacity(el_count);
                 let ys_to_set = ys.spare_capacity_mut();
-                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
+                let ys_to_set = unsafe {
+                    std::mem::transmute::<&mut [std::mem::MaybeUninit<T>], &mut [T]>(ys_to_set)
+                };
                 let mut dst_i = 0;
                 for src_i in (o_l1..o_l2).step_by(ob.len) {
                     f_vec(
@@ -224,7 +228,9 @@ pub fn binary_map_vec<T: Copy, F: FnMut(T, T) -> T, FV: FnMut(&[T], &[T], &mut [
                 let lhs = &lhs[ob.start..ob.start + ob.len];
                 let mut ys: Vec<T> = Vec::with_capacity(el_count);
                 let ys_to_set = ys.spare_capacity_mut();
-                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [T]>(ys_to_set) };
+                let ys_to_set = unsafe {
+                    std::mem::transmute::<&mut [std::mem::MaybeUninit<T>], &mut [T]>(ys_to_set)
+                };
                 let mut dst_i = 0;
                 for src_i in (o_r1..o_r2).step_by(ob.len) {
                     f_vec(
@@ -311,7 +317,9 @@ pub fn unary_map_vec<T: Copy, U: Copy, F: FnMut(T) -> U, FV: FnMut(&[T], &mut [U
         crate::StridedBlocks::SingleBlock { start_offset, len } => {
             let mut ys: Vec<U> = Vec::with_capacity(len);
             let ys_to_set = ys.spare_capacity_mut();
-            let ys_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(ys_to_set) };
+            let ys_to_set = unsafe {
+                std::mem::transmute::<&mut [std::mem::MaybeUninit<U>], &mut [U]>(ys_to_set)
+            };
             f_vec(&vs[start_offset..start_offset + len], ys_to_set);
             // SAFETY: values are all set by f_vec.
             unsafe { ys.set_len(len) };
@@ -333,7 +341,9 @@ pub fn unary_map_vec<T: Copy, U: Copy, F: FnMut(T) -> U, FV: FnMut(&[T], &mut [U
             } else {
                 let mut ys: Vec<U> = Vec::with_capacity(el_count);
                 let ys_to_set = ys.spare_capacity_mut();
-                let ys_to_set = unsafe { std::mem::transmute::<_, &mut [U]>(ys_to_set) };
+                let ys_to_set = unsafe {
+                    std::mem::transmute::<&mut [std::mem::MaybeUninit<U>], &mut [U]>(ys_to_set)
+                };
                 let mut dst_index = 0;
                 for src_index in block_start_index {
                     let vs = &vs[src_index..src_index + block_len];

candle-core/src/cuda_backend/cudnn.rs

@@ -1,6 +1,6 @@
 use crate::WithDType;
 use cudarc;
-use cudarc::cudnn::safe::{Conv2dForward, Cudnn};
+use cudarc::cudnn::safe::{ConvForward, Cudnn};
 use cudarc::driver::{CudaSlice, CudaView, DeviceRepr, ValidAsZeroBits};
 use std::cell::RefCell;
 use std::collections::HashMap;
@@ -26,6 +26,7 @@ impl From<cudarc::driver::DriverError> for crate::Error {
 
 pub(crate) fn launch_conv2d<
     T: DeviceRepr + WithDType + ValidAsZeroBits + cudarc::cudnn::CudnnDataType,
+    Y: cudarc::cudnn::CudnnDataType,
 >(
     src: &CudaView<T>,
     src_l: &crate::Layout,
@@ -48,7 +49,7 @@ pub(crate) fn launch_conv2d<
         }
         c
     })?;
-    let conv = cudnn.create_conv2d::<T>(
+    let conv = cudnn.create_conv2d::<Y>(
         /* pad */ [params.padding as i32, params.padding as i32],
         /* stride */ [params.stride as i32, params.stride as i32],
         /* dilation */ [params.dilation as i32, params.dilation as i32],
@@ -62,18 +63,18 @@ pub(crate) fn launch_conv2d<
     ];
     // Note that `src` already starts at the proper offset.
     let x = if src_l.is_contiguous() {
-        cudnn.create_4d_tensor(
+        cudnn.create_4d_tensor::<T>(
             cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
             x_shape,
         )?
     } else {
         let s = src_l.stride();
-        cudnn.create_4d_tensor_ex(
+        cudnn.create_4d_tensor_ex::<T>(
             x_shape,
             [s[0] as i32, s[1] as i32, s[2] as i32, s[3] as i32],
         )?
     };
-    let w = cudnn.create_4d_filter(
+    let w = cudnn.create_4d_filter::<T>(
         cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
         [
             params.c_out as i32,
@@ -83,11 +84,11 @@ pub(crate) fn launch_conv2d<
         ],
     )?;
     let (w_out, h_out) = (params.out_w() as i32, params.out_h() as i32);
-    let y = cudnn.create_4d_tensor(
+    let y = cudnn.create_4d_tensor::<T>(
         cudarc::cudnn::sys::cudnnTensorFormat_t::CUDNN_TENSOR_NCHW,
         [params.b_size as i32, params.c_out as i32, h_out, w_out],
     )?;
-    let conv2d = Conv2dForward {
+    let conv2d = ConvForward {
         conv: &conv,
         x: &x,
         w: &w,

candle-core/src/cuda_backend/device.rs

@@ -1,5 +1,5 @@
 use crate::backend::BackendDevice;
-use crate::{CpuStorage, DType, Layout, Result, Shape};
+use crate::{CpuStorage, CpuStorageRef, DType, Layout, Result, Shape};
 pub use candle_kernels as kernels;
 pub use cudarc;
 use cudarc::driver::{CudaFunction, LaunchAsync, LaunchConfig};
@@ -51,6 +51,27 @@ impl CudaDevice {
         self.device.clone()
     }
 
+    pub fn compile(
+        &self,
+        func_name: &'static str,
+        kernel: ug::lang::ssa::Kernel,
+    ) -> Result<CudaFunction> {
+        let mut buf = vec![];
+        ug_cuda::code_gen::gen(&mut buf, func_name, &kernel)?;
+        let cuda_code = String::from_utf8(buf)?;
+        let opts = cudarc::nvrtc::CompileOptions {
+            use_fast_math: Some(true),
+            ..Default::default()
+        };
+        let ptx = cudarc::nvrtc::safe::compile_ptx_with_opts(cuda_code, opts).w()?;
+        self.device.load_ptx(ptx, "ug", &[func_name]).w()?;
+        let func = match self.device.get_func("ug", func_name) {
+            Some(func) => func,
+            None => crate::bail!("unknown function ug::{func_name}"),
+        };
+        Ok(func)
+    }
+
     pub fn id(&self) -> DeviceId {
         self.id
     }
@@ -144,6 +165,20 @@ impl CudaDevice {
     }
 }
 
+impl CudaDevice {
+    pub fn new_with_stream(ordinal: usize) -> Result<Self> {
+        let device = cudarc::driver::CudaDevice::new_with_stream(ordinal).w()?;
+        let blas = cudarc::cublas::CudaBlas::new(device.clone()).w()?;
+        let curand = cudarc::curand::CudaRng::new(299792458, device.clone()).w()?;
+        Ok(Self {
+            id: DeviceId::new(),
+            device,
+            blas: Arc::new(blas),
+            curand: Arc::new(Mutex::new(CudaRng(curand))),
+        })
+    }
+}
+
 impl BackendDevice for CudaDevice {
     type Storage = CudaStorage;
 
@@ -334,6 +369,43 @@ impl BackendDevice for CudaDevice {
         })
     }
 
+    fn storage_from_slice<T: crate::WithDType>(&self, s: &[T]) -> Result<Self::Storage> {
+        let slice = match T::cpu_storage_ref(s) {
+            CpuStorageRef::U8(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::U8(data)
+            }
+            CpuStorageRef::U32(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::U32(data)
+            }
+            CpuStorageRef::I64(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::I64(data)
+            }
+            CpuStorageRef::BF16(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::BF16(data)
+            }
+            CpuStorageRef::F16(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::F16(data)
+            }
+            CpuStorageRef::F32(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::F32(data)
+            }
+            CpuStorageRef::F64(storage) => {
+                let data = self.htod_sync_copy(storage).w()?;
+                CudaStorageSlice::F64(data)
+            }
+        };
+        Ok(CudaStorage {
+            slice,
+            device: self.clone(),
+        })
+    }
+
     fn storage_from_cpu_storage(&self, storage: &CpuStorage) -> Result<CudaStorage> {
         let slice = match storage {
             CpuStorage::U8(storage) => {
@@ -407,4 +479,9 @@ impl BackendDevice for CudaDevice {
             device: self.clone(),
         })
     }
+
+    fn synchronize(&self) -> Result<()> {
+        self.device.synchronize().map_err(crate::Error::wrap)?;
+        Ok(())
+    }
 }

candle-core/src/cuda_backend/mod.rs

@@ -16,9 +16,9 @@ mod error;
 mod utils;
 pub use device::{CudaDevice, DeviceId};
 pub use error::{CudaError, WrapErr};
-pub use utils::{Map1, Map1Any, Map2, Map2Any, Map2InPlace, S};
+pub use utils::{Map1, Map1Any, Map2, Map2Any, Map2InPlace, Map3, S};
 
-enum SlicePtrOrNull<T> {
+pub enum SlicePtrOrNull<T> {
     Ptr(CudaSlice<T>),
     Null,
 }
@@ -33,7 +33,7 @@ unsafe impl<T: DeviceRepr> DeviceRepr for &SlicePtrOrNull<T> {
 }
 
 impl SlicePtrOrNull<usize> {
-    fn params_from_layout(dev: &CudaDevice, l: &Layout) -> Result<Self> {
+    pub fn params_from_layout(dev: &CudaDevice, l: &Layout) -> Result<Self> {
         let ds = if l.is_contiguous() {
             SlicePtrOrNull::Null
         } else {
@@ -174,6 +174,7 @@ impl Map1 for Im2Col1D {
     }
 }
 
+#[allow(unused)]
 struct Im2Col {
     h_k: usize,
     w_k: usize,
@@ -183,6 +184,7 @@ struct Im2Col {
 }
 
 impl Im2Col {
+    #[allow(unused)]
     fn hw_out(&self, h: usize, w: usize) -> (usize, usize) {
         let h_out = (h + 2 * self.padding - self.dilation * (self.h_k - 1) - 1) / self.stride + 1;
         let w_out = (w + 2 * self.padding - self.dilation * (self.w_k - 1) - 1) / self.stride + 1;
@@ -250,44 +252,6 @@ impl Map1 for Powf {
     }
 }
 
-struct Sum<'a>(&'a [usize]);
-impl<'a> Map1 for Sum<'a> {
-    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
-        &self,
-        src: &CudaSlice<T>,
-        dev: &CudaDevice,
-        layout: &Layout,
-    ) -> Result<CudaSlice<T>> {
-        let shape = layout.shape();
-        let src_dims = shape.dims();
-        let el = shape.elem_count();
-        let mut dst_el = el;
-        for &sum_dim in self.0.iter() {
-            dst_el /= src_dims[sum_dim];
-        }
-        let mut sum_dims = self.0.to_vec();
-        // Sort the sum_dims as they have to be processed from left to right when converting the
-        // indexes.
-        sum_dims.sort();
-        let sum_dims_l: Vec<usize> = sum_dims.iter().map(|&d| src_dims[d]).collect();
-        let sum_dims_s: Vec<usize> = sum_dims
-            .iter()
-            .map(|&d| src_dims[d + 1..].iter().product::<usize>())
-            .collect();
-        let cfg = LaunchConfig::for_num_elems(el as u32);
-        let ds = dev
-            .htod_copy([src_dims, layout.stride(), &sum_dims_l, &sum_dims_s].concat())
-            .w()?;
-        let src = &src.slice(layout.start_offset()..);
-        let func = dev.get_or_load_func(&kernel_name::<T>("sum"), kernels::REDUCE)?;
-        let out = dev.alloc_zeros::<T>(dst_el).w()?;
-        let params = (el, src_dims.len(), sum_dims.len(), &ds, src, &out);
-        // SAFETY: ffi.
-        unsafe { func.launch(cfg, params) }.w()?;
-        Ok(out)
-    }
-}
-
 struct FastReduce<'a>(&'a [usize], ReduceOp);
 impl<'a> Map1Any for FastReduce<'a> {
     fn f<T: DeviceRepr + WithDType + ValidAsZeroBits, W: Fn(CudaSlice<T>) -> S>(
@@ -668,6 +632,31 @@ impl<'a> Map2 for Conv2D<'a> {
     }
 }
 
+struct Col2Im1D {
+    stride: usize,
+}
+
+impl Map1 for Col2Im1D {
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        col: &CudaSlice<T>,
+        dev: &CudaDevice,
+        l: &Layout,
+    ) -> Result<CudaSlice<T>> {
+        let (b_size, l_in, c_out, k_size) = l.shape().dims4()?;
+        let stride = self.stride;
+        let l_out = (l_in - 1) * stride + k_size;
+        let dst_el = b_size * c_out * l_out;
+        let mut im = unsafe { dev.alloc::<T>(dst_el) }.w()?;
+
+        let cfg = LaunchConfig::for_num_elems(dst_el as u32);
+        let params = (dst_el, l_out, l_in, c_out, k_size, stride, col, &mut im);
+        let func = dev.get_or_load_func(&kernel_name::<T>("col2im1d"), kernels::CONV)?;
+        unsafe { func.launch(cfg, params) }.w()?;
+        Ok(im)
+    }
+}
+
 struct ConvTranspose1D<'a>(&'a crate::conv::ParamsConvTranspose1D);
 impl<'a> Map2 for ConvTranspose1D<'a> {
     fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
@@ -1404,9 +1393,55 @@ impl BackendStorage for CudaStorage {
         kernel_l: &Layout,
         params: &crate::conv::ParamsConvTranspose1D,
     ) -> Result<Self> {
+        const USE_COL2IM_CONV1D_TR: bool = true;
+
         let device = self.device().clone();
-        let slice =
-            ConvTranspose1D(params).map(&self.slice, l, &kernel.slice, kernel_l, &device)?;
+        let can_use_col2im = kernel_l.is_contiguous()
+            && params.dilation == 1
+            && params.padding == 0
+            && params.output_padding == 0;
+        let slice = if USE_COL2IM_CONV1D_TR && can_use_col2im {
+            let (b_size, c_in, l_in) = l.shape().dims3()?;
+            let (c_in2, c_out, k_size) = kernel_l.shape().dims3()?;
+            if !kernel_l.is_contiguous() {
+                crate::bail!(
+                    "convtr1d: the second argument (kernel) has to be contiguous {kernel_l:?}"
+                )
+            }
+            if c_in != c_in2 {
+                crate::bail!(
+                    "convtr1d: shape mismatch on c_in {:?} {:?}",
+                    l.shape(),
+                    kernel_l.shape()
+                )
+            }
+            let col = {
+                // This merges the last two dimensions of the kernel together.
+                let kernel_l_mm = Layout::new(
+                    (b_size, c_in, k_size * c_out).into(),
+                    vec![0, k_size * c_out, 1],
+                    kernel_l.start_offset(),
+                );
+                self.matmul(
+                    kernel,
+                    (
+                        b_size,
+                        /* m */ l_in,
+                        /* n */ c_out * k_size,
+                        /* k */ c_in,
+                    ),
+                    &l.transpose(1, 2)?,
+                    &kernel_l_mm,
+                )?
+            };
+            let col_l = Layout::contiguous((b_size, l_in, c_out, k_size));
+            Col2Im1D {
+                stride: params.stride,
+            }
+            .map(&col.slice, &device, &col_l)?
+        } else {
+            ConvTranspose1D(params).map(&self.slice, l, &kernel.slice, kernel_l, &device)?
+        };
         Ok(Self { slice, device })
     }
 
@@ -1487,7 +1522,7 @@ impl BackendStorage for CudaStorage {
                 let inp = &inp.slice(inp_l.start_offset()..);
                 let k = &k.slice(kernel_l.start_offset()..);
                 let mut out = unsafe { device.alloc::<u8>(dst_el) }.w()?;
-                crate::cudnn::launch_conv2d::<u8>(inp, inp_l, k, &mut out, params, &device)
+                crate::cudnn::launch_conv2d::<u8, u8>(inp, inp_l, k, &mut out, params, &device)
                     .map_err(crate::Error::wrap)?;
                 S::U8(out)
             }
@@ -1495,7 +1530,10 @@ impl BackendStorage for CudaStorage {
                 let inp = &inp.slice(inp_l.start_offset()..);
                 let k = &k.slice(kernel_l.start_offset()..);
                 let mut out = unsafe { device.alloc::<bf16>(dst_el) }.w()?;
-                crate::cudnn::launch_conv2d::<bf16>(inp, inp_l, k, &mut out, params, &device)
+                // Only PSEUDO_BFLOAT16_CONFIG is supported in cudnn, there is no "true bfloat16"
+                // version.
+                // https://docs.nvidia.com/deeplearning/cudnn/latest/api/cudnn-cnn-library.html#id88
+                crate::cudnn::launch_conv2d::<bf16, f32>(inp, inp_l, k, &mut out, params, &device)
                     .map_err(crate::Error::wrap)?;
                 S::BF16(out)
             }
@@ -1503,7 +1541,7 @@ impl BackendStorage for CudaStorage {
                 let inp = &inp.slice(inp_l.start_offset()..);
                 let k = &k.slice(kernel_l.start_offset()..);
                 let mut out = unsafe { device.alloc::<f16>(dst_el) }.w()?;
-                crate::cudnn::launch_conv2d::<f16>(inp, inp_l, k, &mut out, params, &device)
+                crate::cudnn::launch_conv2d::<f16, f16>(inp, inp_l, k, &mut out, params, &device)
                     .map_err(crate::Error::wrap)?;
                 S::F16(out)
             }
@@ -1511,7 +1549,7 @@ impl BackendStorage for CudaStorage {
                 let inp = &inp.slice(inp_l.start_offset()..);
                 let k = &k.slice(kernel_l.start_offset()..);
                 let mut out = unsafe { device.alloc::<f32>(dst_el) }.w()?;
-                crate::cudnn::launch_conv2d::<f32>(inp, inp_l, k, &mut out, params, &device)
+                crate::cudnn::launch_conv2d::<f32, f32>(inp, inp_l, k, &mut out, params, &device)
                     .map_err(crate::Error::wrap)?;
                 S::F32(out)
             }
@@ -1519,7 +1557,7 @@ impl BackendStorage for CudaStorage {
                 let inp = &inp.slice(inp_l.start_offset()..);
                 let k = &k.slice(kernel_l.start_offset()..);
                 let mut out = unsafe { device.alloc::<f64>(dst_el) }.w()?;
-                crate::cudnn::launch_conv2d::<f64>(inp, inp_l, k, &mut out, params, &device)
+                crate::cudnn::launch_conv2d::<f64, f64>(inp, inp_l, k, &mut out, params, &device)
                     .map_err(crate::Error::wrap)?;
                 S::F64(out)
             }
@@ -1635,12 +1673,8 @@ impl BackendStorage for CudaStorage {
                 let rhs = &rhs.slice(rhs_l.start_offset()..);
                 let cfg = gemm_config(bf16::ONE, bf16::ZERO, (b, m, n, k), lhs_l, rhs_l)?;
                 let mut out = unsafe { dev.alloc::<bf16>(elem_count) }.w()?;
-                unsafe {
-                    self.device
-                        .blas
-                        .gemm_strided_batched(cfg, rhs, lhs, &mut out)
-                }
-                .w()?;
+                unsafe { gemm_strided_batched_bf16(&self.device.blas, cfg, rhs, lhs, &mut out) }
+                    .w()?;
                 CudaStorageSlice::BF16(out)
             }
             (CudaStorageSlice::F16(lhs), CudaStorageSlice::F16(rhs)) => {
@@ -1648,12 +1682,8 @@ impl BackendStorage for CudaStorage {
                 let rhs = &rhs.slice(rhs_l.start_offset()..);
                 let cfg = gemm_config(f16::ONE, f16::ZERO, (b, m, n, k), lhs_l, rhs_l)?;
                 let mut out = unsafe { dev.alloc::<f16>(elem_count) }.w()?;
-                unsafe {
-                    self.device
-                        .blas
-                        .gemm_strided_batched(cfg, rhs, lhs, &mut out)
-                }
-                .w()?;
+                unsafe { gemm_strided_batched_f16(&self.device.blas, cfg, rhs, lhs, &mut out) }
+                    .w()?;
                 CudaStorageSlice::F16(out)
             }
             (CudaStorageSlice::F32(lhs), CudaStorageSlice::F32(rhs)) => {
@@ -1661,12 +1691,8 @@ impl BackendStorage for CudaStorage {
                 let rhs = &rhs.slice(rhs_l.start_offset()..);
                 let cfg = gemm_config(1., 0., (b, m, n, k), lhs_l, rhs_l)?;
                 let mut out = unsafe { dev.alloc::<f32>(elem_count) }.w()?;
-                unsafe {
-                    self.device
-                        .blas
-                        .gemm_strided_batched(cfg, rhs, lhs, &mut out)
-                }
-                .w()?;
+                unsafe { gemm_strided_batched_f32(&self.device.blas, cfg, rhs, lhs, &mut out) }
+                    .w()?;
                 CudaStorageSlice::F32(out)
             }
             (CudaStorageSlice::F64(lhs), CudaStorageSlice::F64(rhs)) => {
@@ -1856,3 +1882,203 @@ impl BackendStorage for CudaStorage {
         Ok(())
     }
 }
+
+// Default for the reduced precision setting is false, similar to pytorch.
+// https://github.com/pytorch/pytorch/issues/123157
+static MM_F16_REDUCED_PRECISION: std::sync::atomic::AtomicBool =
+    std::sync::atomic::AtomicBool::new(false);
+static MM_BF16_REDUCED_PRECISION: std::sync::atomic::AtomicBool =
+    std::sync::atomic::AtomicBool::new(false);
+static MM_F32_REDUCED_PRECISION: std::sync::atomic::AtomicBool =
+    std::sync::atomic::AtomicBool::new(false);
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn gemm_reduced_precision_f32() -> bool {
+    MM_F32_REDUCED_PRECISION.load(std::sync::atomic::Ordering::Relaxed)
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn set_gemm_reduced_precision_f32(b: bool) {
+    MM_F32_REDUCED_PRECISION.store(b, std::sync::atomic::Ordering::Relaxed)
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn gemm_reduced_precision_f16() -> bool {
+    MM_F16_REDUCED_PRECISION.load(std::sync::atomic::Ordering::Relaxed)
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn set_gemm_reduced_precision_f16(b: bool) {
+    MM_F16_REDUCED_PRECISION.store(b, std::sync::atomic::Ordering::Relaxed)
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn gemm_reduced_precision_bf16() -> bool {
+    MM_BF16_REDUCED_PRECISION.load(std::sync::atomic::Ordering::Relaxed)
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn set_gemm_reduced_precision_bf16(b: bool) {
+    MM_BF16_REDUCED_PRECISION.store(b, std::sync::atomic::Ordering::Relaxed)
+}
+
+unsafe fn gemm_strided_batched_f32(
+    cublas: &cudarc::cublas::CudaBlas,
+    cfg: StridedBatchedConfig<f32>,
+    a: &cudarc::driver::CudaView<f32>,
+    b: &cudarc::driver::CudaView<f32>,
+    c: &mut CudaSlice<f32>,
+) -> std::result::Result<(), cudarc::cublas::result::CublasError> {
+    use cudarc::cublas::sys;
+    use cudarc::driver::DevicePtrMut;
+
+    let compute_type = if gemm_reduced_precision_f32() {
+        sys::cublasComputeType_t::CUBLAS_COMPUTE_32F_FAST_TF32
+    } else {
+        sys::cublasComputeType_t::CUBLAS_COMPUTE_32F
+    };
+    let alpha = &cfg.gemm.alpha as *const f32 as *const _;
+    let beta = &cfg.gemm.beta as *const f32 as *const _;
+
+    cudarc::cublas::result::gemm_strided_batched_ex(
+        *cublas.handle(),
+        cfg.gemm.transa,
+        cfg.gemm.transb,
+        cfg.gemm.m,
+        cfg.gemm.n,
+        cfg.gemm.k,
+        alpha,
+        *a.device_ptr() as *const _,
+        sys::cudaDataType_t::CUDA_R_32F,
+        cfg.gemm.lda,
+        cfg.stride_a,
+        *b.device_ptr() as *const _,
+        sys::cudaDataType_t::CUDA_R_32F,
+        cfg.gemm.ldb,
+        cfg.stride_b,
+        beta,
+        *c.device_ptr_mut() as *mut _,
+        sys::cudaDataType_t::CUDA_R_32F,
+        cfg.gemm.ldc,
+        cfg.stride_c,
+        cfg.batch_size,
+        compute_type,
+        sys::cublasGemmAlgo_t::CUBLAS_GEMM_DEFAULT_TENSOR_OP,
+    )
+}
+
+unsafe fn gemm_strided_batched_f16(
+    cublas: &cudarc::cublas::CudaBlas,
+    cfg: StridedBatchedConfig<f16>,
+    a: &cudarc::driver::CudaView<f16>,
+    b: &cudarc::driver::CudaView<f16>,
+    c: &mut CudaSlice<f16>,
+) -> std::result::Result<(), cudarc::cublas::result::CublasError> {
+    use cudarc::cublas::sys;
+    use cudarc::driver::DevicePtrMut;
+
+    let alpha = cfg.gemm.alpha;
+    let beta = cfg.gemm.beta;
+    let alpha_f32: f32 = cfg.gemm.alpha.to_f32();
+    let beta_f32: f32 = cfg.gemm.beta.to_f32();
+    let (compute_type, alpha, beta) = if gemm_reduced_precision_f16() {
+        (
+            sys::cublasComputeType_t::CUBLAS_COMPUTE_16F,
+            (&alpha) as *const f16 as *const _,
+            (&beta) as *const f16 as *const _,
+        )
+    } else {
+        (
+            sys::cublasComputeType_t::CUBLAS_COMPUTE_32F,
+            (&alpha_f32) as *const f32 as *const _,
+            (&beta_f32) as *const f32 as *const _,
+        )
+    };
+
+    cudarc::cublas::result::gemm_strided_batched_ex(
+        *cublas.handle(),
+        cfg.gemm.transa,
+        cfg.gemm.transb,
+        cfg.gemm.m,
+        cfg.gemm.n,
+        cfg.gemm.k,
+        alpha,
+        *a.device_ptr() as *const _,
+        sys::cudaDataType_t::CUDA_R_16F,
+        cfg.gemm.lda,
+        cfg.stride_a,
+        *b.device_ptr() as *const _,
+        sys::cudaDataType_t::CUDA_R_16F,
+        cfg.gemm.ldb,
+        cfg.stride_b,
+        beta,
+        *c.device_ptr_mut() as *mut _,
+        sys::cudaDataType_t::CUDA_R_16F,
+        cfg.gemm.ldc,
+        cfg.stride_c,
+        cfg.batch_size,
+        compute_type,
+        sys::cublasGemmAlgo_t::CUBLAS_GEMM_DEFAULT_TENSOR_OP,
+    )
+}
+
+unsafe fn gemm_strided_batched_bf16(
+    cublas: &cudarc::cublas::CudaBlas,
+    cfg: StridedBatchedConfig<bf16>,
+    a: &cudarc::driver::CudaView<bf16>,
+    b: &cudarc::driver::CudaView<bf16>,
+    c: &mut CudaSlice<bf16>,
+) -> std::result::Result<(), cudarc::cublas::result::CublasError> {
+    use cudarc::cublas::sys;
+    use cudarc::driver::DevicePtrMut;
+
+    let alpha_f32: f32 = cfg.gemm.alpha.to_f32();
+    let beta_f32: f32 = cfg.gemm.beta.to_f32();
+    // The type for alpha and beta depends on the computeType.
+    // https://docs.nvidia.com/cuda/cublas/index.html#cublasgemmstridedbatchedex
+    let (compute_type, alpha, beta) = if gemm_reduced_precision_bf16() {
+        (
+            sys::cublasComputeType_t::CUBLAS_COMPUTE_32F_FAST_16BF,
+            (&alpha_f32) as *const f32 as *const _,
+            (&beta_f32) as *const f32 as *const _,
+        )
+    } else {
+        (
+            sys::cublasComputeType_t::CUBLAS_COMPUTE_32F,
+            (&alpha_f32) as *const f32 as *const _,
+            (&beta_f32) as *const f32 as *const _,
+        )
+    };
+
+    cudarc::cublas::result::gemm_strided_batched_ex(
+        *cublas.handle(),
+        cfg.gemm.transa,
+        cfg.gemm.transb,
+        cfg.gemm.m,
+        cfg.gemm.n,
+        cfg.gemm.k,
+        alpha,
+        *a.device_ptr() as *const _,
+        sys::cudaDataType_t::CUDA_R_16BF,
+        cfg.gemm.lda,
+        cfg.stride_a,
+        *b.device_ptr() as *const _,
+        sys::cudaDataType_t::CUDA_R_16BF,
+        cfg.gemm.ldb,
+        cfg.stride_b,
+        beta,
+        *c.device_ptr_mut() as *mut _,
+        sys::cudaDataType_t::CUDA_R_16BF,
+        cfg.gemm.ldc,
+        cfg.stride_c,
+        cfg.batch_size,
+        compute_type,
+        sys::cublasGemmAlgo_t::CUBLAS_GEMM_DEFAULT_TENSOR_OP,
+    )
+}

candle-core/src/cuda_backend/utils.rs

@@ -54,6 +54,44 @@ pub trait Map2 {
     }
 }
 
+pub trait Map3 {
+    #[allow(clippy::too_many_arguments)]
+    fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
+        &self,
+        src1: &CudaSlice<T>,
+        layout1: &Layout,
+        src2: &CudaSlice<T>,
+        layout2: &Layout,
+        src3: &CudaSlice<T>,
+        layout3: &Layout,
+        dev: &CudaDevice,
+    ) -> Result<CudaSlice<T>>;
+
+    #[allow(clippy::too_many_arguments)]
+    fn map(
+        &self,
+        s1: &S,
+        l1: &Layout,
+        s2: &S,
+        l2: &Layout,
+        s3: &S,
+        l3: &Layout,
+        d: &CudaDevice,
+    ) -> Result<S> {
+        let out = match (s1, s2, s3) {
+            (S::U8(s1), S::U8(s2), S::U8(s3)) => S::U8(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::U32(s1), S::U32(s2), S::U32(s3)) => S::U32(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::I64(s1), S::I64(s2), S::I64(s3)) => S::I64(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::BF16(s1), S::BF16(s2), S::BF16(s3)) => S::BF16(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::F16(s1), S::F16(s2), S::F16(s3)) => S::F16(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::F32(s1), S::F32(s2), S::F32(s3)) => S::F32(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            (S::F64(s1), S::F64(s2), S::F64(s3)) => S::F64(self.f(s1, l1, s2, l2, s3, l3, d)?),
+            _ => Err(CudaError::InternalError("dtype mismatch in ternary op"))?,
+        };
+        Ok(out)
+    }
+}
+
 pub trait Map2InPlace {
     fn f<T: DeviceRepr + WithDType + ValidAsZeroBits>(
         &self,

candle-core/src/custom_op.rs

@@ -375,3 +375,110 @@ impl Tensor {
         )
     }
 }
+
+pub struct UgIOp1 {
+    name: &'static str,
+    #[cfg(feature = "cuda")]
+    func: cudarc::driver::CudaFunction,
+    #[cfg(feature = "metal")]
+    func: metal::ComputePipelineState,
+}
+
+impl UgIOp1 {
+    #[allow(unused)]
+    pub fn new(
+        name: &'static str,
+        kernel: ug::lang::ssa::Kernel,
+        device: &crate::Device,
+    ) -> Result<Self> {
+        #[cfg(feature = "cuda")]
+        {
+            let device = device.as_cuda_device()?;
+            let func = device.compile(name, kernel)?;
+            Ok(Self { name, func })
+        }
+        #[cfg(feature = "metal")]
+        {
+            let device = device.as_metal_device()?;
+            let func = device.compile(name, kernel)?;
+            Ok(Self { name, func })
+        }
+        #[cfg(not(any(feature = "cuda", feature = "metal")))]
+        {
+            Ok(Self { name })
+        }
+    }
+}
+
+impl InplaceOp1 for UgIOp1 {
+    fn name(&self) -> &'static str {
+        self.name
+    }
+
+    fn cpu_fwd(&self, _: &mut CpuStorage, _: &Layout) -> Result<()> {
+        crate::bail!("ug ops are only supported on metal/cuda at the moment")
+    }
+
+    #[cfg(feature = "metal")]
+    fn metal_fwd(&self, sto: &mut MetalStorage, layout: &Layout) -> Result<()> {
+        use crate::backend::BackendStorage;
+        use candle_metal_kernels::utils::EncoderProvider;
+
+        let elem_count = layout.shape().elem_count();
+        if sto.dtype() != crate::DType::F32 {
+            // TODO: support more dtypes.
+            crate::bail!("input is not a f32 tensor")
+        }
+        let device = sto.device();
+        println!("here");
+        let command_buffer = device.command_buffer()?;
+        let command_buffer = &command_buffer;
+        let encoder = command_buffer.encoder();
+        let encoder = encoder.as_ref();
+        encoder.set_compute_pipeline_state(&self.func);
+        let (g, b) = if elem_count % 32 == 0 {
+            (elem_count / 32, 32)
+        } else {
+            (elem_count, 1)
+        };
+        let grid_dims = metal::MTLSize {
+            width: g as u64,
+            height: 1,
+            depth: 1,
+        };
+        let group_dims = candle_metal_kernels::utils::get_block_dims(b as u64, 1, 1);
+        candle_metal_kernels::utils::set_param(encoder, 0, (sto.buffer(), 0usize));
+
+        encoder.use_resource(sto.buffer(), metal::MTLResourceUsage::Write);
+        encoder.dispatch_threads(grid_dims, group_dims);
+
+        Ok(())
+    }
+
+    #[cfg(feature = "cuda")]
+    fn cuda_fwd(&self, sto: &mut CudaStorage, layout: &Layout) -> Result<()> {
+        use crate::cuda_backend::WrapErr;
+        use cudarc::driver::LaunchAsync;
+
+        let elem_count = layout.shape().elem_count();
+        // TODO: support more dtypes.
+        let sto = sto.as_cuda_slice::<f32>()?;
+        let sto = match layout.contiguous_offsets() {
+            None => crate::bail!("input has to be contiguous"),
+            Some((o1, o2)) => sto.slice(o1..o2),
+        };
+        let params = (&sto,);
+        let (g, b) = if elem_count % 32 == 0 {
+            (elem_count / 32, 32)
+        } else {
+            (elem_count, 1)
+        };
+        let cfg = cudarc::driver::LaunchConfig {
+            grid_dim: (g as u32, 1, 1),
+            block_dim: (b as u32, 1, 1),
+            shared_mem_bytes: 0,
+        };
+        unsafe { self.func.clone().launch(cfg, params) }.w()?;
+        Ok(())
+    }
+}

candle-core/src/device.rs

@@ -130,6 +130,26 @@ impl Device {
         Ok(Self::Cuda(crate::CudaDevice::new(ordinal)?))
     }
 
+    pub fn as_cuda_device(&self) -> Result<&crate::CudaDevice> {
+        match self {
+            Self::Cuda(d) => Ok(d),
+            Self::Cpu => crate::bail!("expected a cuda device, got cpu"),
+            Self::Metal(_) => crate::bail!("expected a cuda device, got Metal"),
+        }
+    }
+
+    pub fn as_metal_device(&self) -> Result<&crate::MetalDevice> {
+        match self {
+            Self::Cuda(_) => crate::bail!("expected a metal device, got cuda"),
+            Self::Cpu => crate::bail!("expected a metal device, got cpu"),
+            Self::Metal(d) => Ok(d),
+        }
+    }
+
+    pub fn new_cuda_with_stream(ordinal: usize) -> Result<Self> {
+        Ok(Self::Cuda(crate::CudaDevice::new_with_stream(ordinal)?))
+    }
+
     pub fn new_metal(ordinal: usize) -> Result<Self> {
         Ok(Self::Metal(crate::MetalDevice::new(ordinal)?))
     }
@@ -171,6 +191,22 @@ impl Device {
         matches!(self, Self::Metal(_))
     }
 
+    pub fn supports_bf16(&self) -> bool {
+        match self {
+            Self::Cuda(_) | Self::Metal(_) => true,
+            Self::Cpu => false,
+        }
+    }
+
+    /// Return `BF16` for devices that support it, otherwise default to `F32`.
+    pub fn bf16_default_to_f32(&self) -> DType {
+        if self.supports_bf16() {
+            DType::BF16
+        } else {
+            DType::F32
+        }
+    }
+
     pub fn cuda_if_available(ordinal: usize) -> Result<Self> {
         if crate::utils::cuda_is_available() {
             Self::new_cuda(ordinal)
@@ -306,6 +342,20 @@ impl Device {
         }
     }
 
+    pub(crate) fn storage_from_slice<D: WithDType>(&self, data: &[D]) -> Result<Storage> {
+        match self {
+            Device::Cpu => Ok(Storage::Cpu(data.to_cpu_storage())),
+            Device::Cuda(device) => {
+                let storage = device.storage_from_slice(data)?;
+                Ok(Storage::Cuda(storage))
+            }
+            Device::Metal(device) => {
+                let storage = device.storage_from_slice(data)?;
+                Ok(Storage::Metal(storage))
+            }
+        }
+    }
+
     pub(crate) fn storage<A: NdArray>(&self, array: A) -> Result<Storage> {
         match self {
             Device::Cpu => Ok(Storage::Cpu(array.to_cpu_storage())),
@@ -337,4 +387,12 @@ impl Device {
             }
         }
     }
+
+    pub fn synchronize(&self) -> Result<()> {
+        match self {
+            Self::Cpu => Ok(()),
+            Self::Cuda(d) => d.synchronize(),
+            Self::Metal(d) => d.synchronize(),
+        }
+    }
 }

candle-core/src/dtype.rs

@@ -1,7 +1,7 @@
 //! Types for elements that can be stored and manipulated using tensors.
 #![allow(clippy::redundant_closure_call)]
 use crate::backend::BackendStorage;
-use crate::{CpuStorage, Error, Result};
+use crate::{CpuStorage, CpuStorageRef, Error, Result};
 
 /// The different types of elements allowed in tensors.
 #[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
@@ -100,12 +100,14 @@ pub trait WithDType:
     + 'static
     + Send
     + Sync
+    + std::any::Any
     + crate::cpu::kernels::VecOps
 {
     const DTYPE: DType;
 
     fn from_f64(v: f64) -> Self;
     fn to_f64(self) -> f64;
+    fn cpu_storage_ref(data: &[Self]) -> CpuStorageRef<'_>;
     fn to_cpu_storage_owned(data: Vec<Self>) -> CpuStorage;
 
     fn to_cpu_storage(data: &[Self]) -> CpuStorage {
@@ -129,6 +131,10 @@ macro_rules! with_dtype {
                 $to_f64(self)
             }
 
+            fn cpu_storage_ref(data: &[Self]) -> CpuStorageRef<'_> {
+                CpuStorageRef::$dtype(data)
+            }
+
             fn to_cpu_storage_owned(data: Vec<Self>) -> CpuStorage {
                 CpuStorage::$dtype(data)
             }

candle-core/src/dummy_cuda_backend.rs

@@ -14,6 +14,12 @@ macro_rules! fail {
     };
 }
 
+impl CudaDevice {
+    pub fn new_with_stream(_: usize) -> Result<Self> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+}
+
 impl crate::backend::BackendStorage for CudaStorage {
     type Device = CudaDevice;
 
@@ -214,6 +220,10 @@ impl crate::backend::BackendDevice for CudaDevice {
         Err(Error::NotCompiledWithCudaSupport)
     }
 
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithCudaSupport)
+    }
+
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
@@ -229,4 +239,38 @@ impl crate::backend::BackendDevice for CudaDevice {
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithCudaSupport)
     }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
 }
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn gemm_reduced_precision_f16() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with f16 GEMMs.
+pub fn set_gemm_reduced_precision_f16(_: bool) {}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn gemm_reduced_precision_bf16() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with fp16 accumulation type) are
+/// allowed with bf16 GEMMs.
+pub fn set_gemm_reduced_precision_bf16(_: bool) {}
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn gemm_reduced_precision_f32() -> bool {
+    true
+}
+
+/// This bool controls whether reduced precision reductions (e.g., with tf32 accumulation type) are
+/// allowed with f32 GEMMs.
+pub fn set_gemm_reduced_precision_f32(_b: bool) {}

candle-core/src/dummy_metal_backend.rs

@@ -226,6 +226,10 @@ impl crate::backend::BackendDevice for MetalDevice {
         Err(Error::NotCompiledWithMetalSupport)
     }
 
+    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage> {
+        Err(Error::NotCompiledWithMetalSupport)
+    }
+
     fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithMetalSupport)
     }
@@ -241,4 +245,8 @@ impl crate::backend::BackendDevice for MetalDevice {
     fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage> {
         Err(Error::NotCompiledWithMetalSupport)
     }
+
+    fn synchronize(&self) -> Result<()> {
+        Ok(())
+    }
 }

candle-core/src/error.rs

@@ -165,6 +165,9 @@ pub enum Error {
     #[error("Metal error {0}")]
     Metal(#[from] MetalError),
 
+    #[error(transparent)]
+    Ug(#[from] ug::Error),
+
     #[error(transparent)]
     TryFromIntError(#[from] core::num::TryFromIntError),
 
@@ -179,6 +182,10 @@ pub enum Error {
     #[error(transparent)]
     ParseInt(#[from] std::num::ParseIntError),
 
+    /// Utf8 parse error.
+    #[error(transparent)]
+    FromUtf8(#[from] std::string::FromUtf8Error),
+
     /// I/O error.
     #[error(transparent)]
     Io(#[from] std::io::Error),
@@ -219,10 +226,14 @@ impl Error {
         Self::Wrapped(Box::new(err)).bt()
     }
 
-    pub fn msg(err: impl std::error::Error + Send + Sync + 'static) -> Self {
+    pub fn msg(err: impl std::error::Error) -> Self {
         Self::Msg(err.to_string()).bt()
     }
 
+    pub fn debug(err: impl std::fmt::Debug) -> Self {
+        Self::Msg(format!("{err:?}")).bt()
+    }
+
     pub fn bt(self) -> Self {
         let backtrace = std::backtrace::Backtrace::capture();
         match backtrace.status() {

candle-core/src/indexer.rs

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

candle-core/src/layout.rs

@@ -35,6 +35,12 @@ impl Layout {
         self.shape.dims()
     }
 
+    /// The dimension size for a specified dimension index.
+    pub fn dim<D: crate::shape::Dim>(&self, dim: D) -> Result<usize> {
+        let dim = dim.to_index(&self.shape, "dim")?;
+        Ok(self.dims()[dim])
+    }
+
     pub fn shape(&self) -> &Shape {
         &self.shape
     }

candle-core/src/lib.rs

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

candle-core/src/metal_backend/device.rs

@@ -4,7 +4,7 @@ use metal::{Buffer, CommandBuffer, CommandQueue, MTLResourceOptions, NSUInteger}
 use std::collections::HashMap;
 use std::ffi::c_void;
 use std::path::Path;
-use std::sync::{Arc, Mutex, RwLock, RwLockWriteGuard};
+use std::sync::{Arc, Mutex, RwLock};
 
 use super::MetalError;
 
@@ -22,7 +22,73 @@ impl DeviceId {
 }
 
 type BufferMap = HashMap<(NSUInteger, MTLResourceOptions), Vec<Arc<Buffer>>>;
-type AllocatedBuffers = Arc<RwLock<BufferMap>>;
+pub(crate) struct Commands {
+    /// Single command queue for the entire device.
+    command_queue: CommandQueue,
+    /// One command buffer at a time.
+    /// The scheduler works by allowing multiple
+    /// [ComputeCommandEncoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc)
+    /// on a single command buffer. Using a single command buffer would be fastest on the GPU but
+    /// prevents overlapping of CPU and GPU commands (because command buffer needs to be committed
+    /// to start to work).
+    /// Despite what the documentation says, command buffers are NOT ordered. They are ordered
+    /// for their START time, but there's no guarantee that command buffer1 will finish before
+    /// command buffer2 starts (or there are metal bugs there)
+    command_buffer: CommandBuffer,
+    /// Keeps track of the current amount of compute command encoders on the current
+    /// command buffer
+    /// Arc, RwLock because of the interior mutability.
+    command_buffer_index: usize,
+    /// The maximum amount of [compute command encoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc) per [command buffer](https://developer.apple.com/documentation/metal/mtlcommandbuffer?language=objc)
+    compute_per_buffer: usize,
+}
+
+impl Commands {
+    pub(crate) fn new(command_queue: CommandQueue) -> Result<Self> {
+        let command_buffer = command_queue.new_command_buffer().to_owned();
+        command_buffer.enqueue();
+        let compute_per_buffer = match std::env::var("CANDLE_METAL_COMPUTE_PER_BUFFER") {
+            Ok(val) => val.parse()?,
+            _ => 50,
+        };
+        Ok(Self {
+            command_queue,
+            command_buffer,
+            command_buffer_index: 0,
+            compute_per_buffer,
+        })
+    }
+
+    pub fn command_buffer(&mut self) -> Result<(bool, CommandBuffer)> {
+        let mut command_buffer = self.command_buffer.to_owned();
+        let mut flushed = false;
+        if self.command_buffer_index > self.compute_per_buffer {
+            self.command_buffer.commit();
+            command_buffer = self.command_queue.new_command_buffer().to_owned();
+            self.command_buffer = command_buffer.clone();
+            self.command_buffer_index = 0;
+            flushed = true;
+        }
+        self.command_buffer_index += 1;
+        Ok((flushed, command_buffer))
+    }
+
+    pub fn wait_until_completed(&mut self) -> Result<()> {
+        match self.command_buffer.status() {
+            metal::MTLCommandBufferStatus::Committed
+            | metal::MTLCommandBufferStatus::Scheduled
+            | metal::MTLCommandBufferStatus::Completed => {
+                panic!("Already committed");
+            }
+            _ => {}
+        }
+        self.command_buffer.commit();
+        self.command_buffer.wait_until_completed();
+        self.command_buffer = self.command_queue.new_command_buffer().to_owned();
+
+        Ok(())
+    }
+}
 
 #[derive(Clone)]
 pub struct MetalDevice {
@@ -33,27 +99,8 @@ pub struct MetalDevice {
     /// Raw metal device: <https://developer.apple.com/documentation/metal/mtldevice?language=objc>
     pub(crate) device: metal::Device,
 
-    /// Single command queue for the entire device.
-    pub(crate) command_queue: CommandQueue,
-    /// One command buffer at a time.
-    /// The scheduler works by allowing multiple
-    /// [ComputeCommandEncoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc)
-    /// on a single command buffer. Using a single command buffer would be fastest on the GPU but
-    /// prevents overlapping of CPU and GPU commands (because command buffer needs to be committed
-    /// to start to work).
-    /// Despite what the documentation says, command buffers are NOT ordered. They are ordered
-    /// for their START time, but there's no guarantee that command buffer1 will finish before
-    /// command buffer2 starts (or there are metal bugs there)
-    pub(crate) command_buffer: Arc<RwLock<CommandBuffer>>,
-    /// Keeps track of the current amount of compute command encoders on the current
-    /// command buffer
-    /// Arc, RwLock because of the interior mutability.
-    pub(crate) command_buffer_index: Arc<RwLock<usize>>,
-    /// The maximum amount of [compute command encoder](https://developer.apple.com/documentation/metal/mtlcomputecommandencoder?language=objc) per [command buffer](https://developer.apple.com/documentation/metal/mtlcommandbuffer?language=objc)
-    pub(crate) compute_per_buffer: usize,
-    /// Simple keeper struct to keep track of the already compiled kernels so we can reuse them.
-    /// Heavily used by [`candle_metal_kernels`]
-    pub(crate) kernels: Arc<Kernels>,
+    pub(crate) commands: Arc<RwLock<Commands>>,
+
     /// Simple allocator struct.
     /// The buffers are stored in size buckets since ML tends to use similar shapes over and over.
     /// We store the buffers in [`Arc`] because it's much faster than Obj-c internal ref counting
@@ -67,9 +114,15 @@ pub struct MetalDevice {
     ///
     /// Whenever we actually allocate a new buffer, we make a full sweep to clean up unused buffers
     /// (strong_count = 1).
-    pub(crate) buffers: AllocatedBuffers,
+    pub(crate) buffers: Arc<RwLock<BufferMap>>,
+
+    /// Simple keeper struct to keep track of the already compiled kernels so we can reuse them.
+    /// Heavily used by [`candle_metal_kernels`]
+    pub(crate) kernels: Arc<Kernels>,
     /// Seed for random number generation.
     pub(crate) seed: Arc<Mutex<Buffer>>,
+    /// Whether to use the MLX matmul kernels instead of the MFA ones.
+    pub(crate) use_mlx_mm: bool,
 }
 
 impl std::fmt::Debug for MetalDevice {
@@ -87,6 +140,32 @@ impl std::ops::Deref for MetalDevice {
 }
 
 impl MetalDevice {
+    pub fn set_use_mlx_mm(&mut self, use_mlx_mm: bool) {
+        self.use_mlx_mm = use_mlx_mm
+    }
+
+    pub fn compile(
+        &self,
+        func_name: &'static str,
+        kernel: ug::lang::ssa::Kernel,
+    ) -> Result<metal::ComputePipelineState> {
+        let mut buf = vec![];
+        ug_metal::code_gen::gen(&mut buf, func_name, &kernel)?;
+        let metal_code = String::from_utf8(buf)?;
+        let lib = self
+            .device
+            .new_library_with_source(&metal_code, &metal::CompileOptions::new())
+            .map_err(MetalError::from)?;
+        let func = lib
+            .get_function(func_name, None)
+            .map_err(MetalError::from)?;
+        let pl = self
+            .device
+            .new_compute_pipeline_state_with_function(&func)
+            .map_err(MetalError::from)?;
+        Ok(pl)
+    }
+
     pub fn id(&self) -> DeviceId {
         self.id
     }
@@ -95,44 +174,31 @@ impl MetalDevice {
         &self.device
     }
 
-    pub fn command_queue(&self) -> &CommandQueue {
-        &self.command_queue
+    fn drop_unused_buffers(&self) -> Result<()> {
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+        for subbuffers in buffers.values_mut() {
+            let newbuffers = subbuffers
+                .iter()
+                .filter(|s| Arc::strong_count(*s) > 1)
+                .map(Arc::clone)
+                .collect();
+            *subbuffers = newbuffers;
+        }
+        Ok(())
     }
 
     pub fn command_buffer(&self) -> Result<CommandBuffer> {
-        let mut command_buffer_lock = self.command_buffer.try_write().map_err(MetalError::from)?;
-        let mut command_buffer = command_buffer_lock.to_owned();
-        let mut index = self
-            .command_buffer_index
-            .try_write()
-            .map_err(MetalError::from)?;
-        if *index > self.compute_per_buffer {
-            command_buffer.commit();
-            command_buffer = self.command_queue.new_command_buffer().to_owned();
-            *command_buffer_lock = command_buffer.clone();
-            *index = 0;
-
-            self.drop_unused_buffers()?;
+        let mut commands = self.commands.write().map_err(MetalError::from)?;
+        let (flushed, command_buffer) = commands.command_buffer()?;
+        if flushed {
+            self.drop_unused_buffers()?
         }
-        *index += 1;
         Ok(command_buffer)
     }
 
     pub fn wait_until_completed(&self) -> Result<()> {
-        let mut command_buffer = self.command_buffer.try_write().map_err(MetalError::from)?;
-        match command_buffer.status() {
-            metal::MTLCommandBufferStatus::Committed
-            | metal::MTLCommandBufferStatus::Scheduled
-            | metal::MTLCommandBufferStatus::Completed => {
-                panic!("Already committed");
-            }
-            _ => {}
-        }
-        command_buffer.commit();
-        command_buffer.wait_until_completed();
-        *command_buffer = self.command_queue.new_command_buffer().to_owned();
-
-        Ok(())
+        let mut commands = self.commands.write().map_err(MetalError::from)?;
+        commands.wait_until_completed()
     }
 
     pub fn kernels(&self) -> &Kernels {
@@ -179,7 +245,8 @@ impl MetalDevice {
             size,
             MTLResourceOptions::StorageModeManaged,
         );
-        let mut buffers = self.buffers.try_write().map_err(MetalError::from)?;
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+
         let subbuffers = buffers
             .entry((size, MTLResourceOptions::StorageModeManaged))
             .or_insert(vec![]);
@@ -210,40 +277,6 @@ impl MetalDevice {
         Ok(buffer)
     }
 
-    fn find_available_buffer(
-        &self,
-        size: NSUInteger,
-        option: MTLResourceOptions,
-        buffers: &RwLockWriteGuard<BufferMap>,
-    ) -> Option<Arc<Buffer>> {
-        let mut best_buffer: Option<&Arc<Buffer>> = None;
-        let mut best_buffer_size: NSUInteger = NSUInteger::MAX;
-        for ((buffer_size, buffer_option), subbuffers) in buffers.iter() {
-            if buffer_size >= &size && buffer_size < &best_buffer_size && buffer_option == &option {
-                for sub in subbuffers {
-                    if Arc::strong_count(sub) == 1 {
-                        best_buffer = Some(sub);
-                        best_buffer_size = *buffer_size;
-                    }
-                }
-            }
-        }
-        best_buffer.cloned()
-    }
-
-    fn drop_unused_buffers(&self) -> Result<()> {
-        let mut buffers = self.buffers.try_write().map_err(MetalError::from)?;
-        for subbuffers in buffers.values_mut() {
-            let newbuffers = subbuffers
-                .iter()
-                .filter(|s| Arc::strong_count(*s) > 1)
-                .map(Arc::clone)
-                .collect();
-            *subbuffers = newbuffers;
-        }
-        Ok(())
-    }
-
     /// The critical allocator algorithm
     fn allocate_buffer(
         &self,
@@ -251,8 +284,8 @@ impl MetalDevice {
         option: MTLResourceOptions,
         _name: &str,
     ) -> Result<Arc<Buffer>> {
-        let mut buffers = self.buffers.try_write().map_err(MetalError::from)?;
-        if let Some(b) = self.find_available_buffer(size, option, &buffers) {
+        let mut buffers = self.buffers.write().map_err(MetalError::from)?;
+        if let Some(b) = find_available_buffer(size, option, &buffers) {
             // Cloning also ensures we increment the strong count
             return Ok(b.clone());
         }
@@ -273,7 +306,13 @@ impl MetalDevice {
         let descriptor = metal::CaptureDescriptor::new();
         descriptor.set_destination(metal::MTLCaptureDestination::GpuTraceDocument);
         descriptor.set_capture_device(self);
-        descriptor.set_output_url(path);
+        // The [set_output_url] call requires an absolute path so we convert it if needed.
+        if path.as_ref().is_absolute() {
+            descriptor.set_output_url(path);
+        } else {
+            let path = std::env::current_dir()?.join(path);
+            descriptor.set_output_url(path);
+        }
 
         capture
             .start_capture(&descriptor)
@@ -283,5 +322,25 @@ impl MetalDevice {
 }
 
 fn buf_size(size: NSUInteger) -> NSUInteger {
-    (size - 1).next_power_of_two() as NSUInteger
+    size.saturating_sub(1).next_power_of_two() as NSUInteger
+}
+
+fn find_available_buffer(
+    size: NSUInteger,
+    option: MTLResourceOptions,
+    buffers: &BufferMap,
+) -> Option<Arc<Buffer>> {
+    let mut best_buffer: Option<&Arc<Buffer>> = None;
+    let mut best_buffer_size: NSUInteger = NSUInteger::MAX;
+    for ((buffer_size, buffer_option), subbuffers) in buffers.iter() {
+        if buffer_size >= &size && buffer_size < &best_buffer_size && buffer_option == &option {
+            for sub in subbuffers {
+                if Arc::strong_count(sub) == 1 {
+                    best_buffer = Some(sub);
+                    best_buffer_size = *buffer_size;
+                }
+            }
+        }
+    }
+    best_buffer.cloned()
 }

candle-core/src/npy.rs

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

candle-core/src/quantized/cuda.rs

@@ -2,12 +2,19 @@ use super::{GgmlDType, QStorage};
 use crate::quantized::k_quants::GgmlType;
 use crate::{backend::BackendDevice, cuda_backend::WrapErr};
 use crate::{CudaDevice, CudaStorage, Result};
+use half::f16;
 
 use cudarc::driver::{CudaSlice, CudaView, DeviceSlice};
 
+#[derive(Clone, Debug)]
+struct PaddedCudaSlice {
+    inner: CudaSlice<u8>,
+    len: usize,
+}
+
 #[derive(Clone, Debug)]
 pub struct QCudaStorage {
-    data: CudaSlice<u8>,
+    data: PaddedCudaSlice,
     dtype: GgmlDType,
     device: CudaDevice,
 }
@@ -40,6 +47,7 @@ fn quantize_q8_1(
     src: &CudaView<f32>,
     dst: &mut CudaSlice<u8>,
     elem_count: usize,
+    ky: usize,
     dev: &CudaDevice,
 ) -> Result<()> {
     use cudarc::driver::LaunchAsync;
@@ -49,7 +57,7 @@ fn quantize_q8_1(
     let num_blocks = ceil_div(kx_padded, CUDA_QUANTIZE_BLOCK_SIZE);
     let func = dev.get_or_load_func("quantize_q8_1", candle_kernels::QUANTIZED)?;
     let cfg = cudarc::driver::LaunchConfig {
-        grid_dim: (num_blocks as u32, 1, 1),
+        grid_dim: (num_blocks as u32, ky as u32, 1),
         block_dim: (CUDA_QUANTIZE_BLOCK_SIZE as u32, 1, 1),
         shared_mem_bytes: 0,
     };
@@ -58,8 +66,8 @@ fn quantize_q8_1(
     Ok(())
 }
 
-fn dequantize(
-    data: &CudaSlice<u8>,
+fn dequantize_f32(
+    data: &PaddedCudaSlice,
     dtype: GgmlDType,
     elem_count: usize,
     dev: &CudaDevice,
@@ -68,27 +76,27 @@ fn dequantize(
 
     let nb = (elem_count + 255) / 256;
     let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
-        GgmlDType::Q4_0 => ("dequantize_block_q4_0", false, 32, nb),
-        GgmlDType::Q4_1 => ("dequantize_block_q4_1", false, 32, nb),
+        GgmlDType::Q4_0 => ("dequantize_block_q4_0_f32", false, 32, nb),
+        GgmlDType::Q4_1 => ("dequantize_block_q4_1_f32", false, 32, nb),
         GgmlDType::Q5_0 => (
-            "dequantize_block_q5_0",
+            "dequantize_block_q5_0_f32",
             false,
             CUDA_DEQUANTIZE_BLOCK_SIZE,
             ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
         ),
         GgmlDType::Q5_1 => (
-            "dequantize_block_q5_1",
+            "dequantize_block_q5_1_f32",
             false,
             CUDA_DEQUANTIZE_BLOCK_SIZE,
             ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
         ),
-        GgmlDType::Q8_0 => ("dequantize_block_q8_0", false, 32, nb),
-        GgmlDType::Q2K => ("dequantize_block_q2_K", true, 64, nb),
-        GgmlDType::Q3K => ("dequantize_block_q3_K", true, 64, nb),
-        GgmlDType::Q4K => ("dequantize_block_q4_K", true, 32, nb),
-        GgmlDType::Q5K => ("dequantize_block_q5_K", true, 64, nb),
-        GgmlDType::Q6K => ("dequantize_block_q6_K", true, 64, nb),
-        GgmlDType::Q8K => ("dequantize_block_q8_K", true, 32, nb),
+        GgmlDType::Q8_0 => ("dequantize_block_q8_0_f32", false, 32, nb),
+        GgmlDType::Q2K => ("dequantize_block_q2_K_f32", true, 64, nb),
+        GgmlDType::Q3K => ("dequantize_block_q3_K_f32", true, 64, nb),
+        GgmlDType::Q4K => ("dequantize_block_q4_K_f32", true, 32, nb),
+        GgmlDType::Q5K => ("dequantize_block_q5_K_f32", true, 64, nb),
+        GgmlDType::Q6K => ("dequantize_block_q6_K_f32", true, 64, nb),
+        GgmlDType::Q8K => ("dequantize_block_q8_K_f32", true, 32, nb),
         _ => crate::bail!("unsupported dtype for dequantize {dtype:?}"),
     };
     let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
@@ -102,21 +110,78 @@ fn dequantize(
     };
 
     if is_k {
-        let params = (data, &dst);
+        let params = (&data.inner, &dst);
         unsafe { func.launch(cfg, params) }.w()?;
     } else {
         let nb32 = match dtype {
             GgmlDType::Q5_0 | GgmlDType::Q5_1 => elem_count,
             _ => elem_count / 32,
         };
-        let params = (data, &dst, nb32 as i32);
+        let params = (&data.inner, &dst, nb32 as i32);
+        unsafe { func.launch(cfg, params) }.w()?;
+    }
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
+fn dequantize_f16(
+    data: &PaddedCudaSlice,
+    dtype: GgmlDType,
+    elem_count: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    use cudarc::driver::LaunchAsync;
+
+    let nb = (elem_count + 255) / 256;
+    let (kernel_name, is_k, block_dim, num_blocks) = match dtype {
+        GgmlDType::Q4_0 => ("dequantize_block_q4_0_f16", false, 32, nb),
+        GgmlDType::Q4_1 => ("dequantize_block_q4_1_f16", false, 32, nb),
+        GgmlDType::Q5_0 => (
+            "dequantize_block_q5_0_f16",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q5_1 => (
+            "dequantize_block_q5_1_f16",
+            false,
+            CUDA_DEQUANTIZE_BLOCK_SIZE,
+            ceil_div(elem_count, 2 * CUDA_DEQUANTIZE_BLOCK_SIZE),
+        ),
+        GgmlDType::Q8_0 => ("dequantize_block_q8_0_f16", false, 32, nb),
+        GgmlDType::Q2K => ("dequantize_block_q2_K_f16", true, 64, nb),
+        GgmlDType::Q3K => ("dequantize_block_q3_K_f16", true, 64, nb),
+        GgmlDType::Q4K => ("dequantize_block_q4_K_f16", true, 32, nb),
+        GgmlDType::Q5K => ("dequantize_block_q5_K_f16", true, 64, nb),
+        GgmlDType::Q6K => ("dequantize_block_q6_K_f16", true, 64, nb),
+        GgmlDType::Q8K => ("dequantize_block_q8_K_f16", true, 32, nb),
+        _ => crate::bail!("unsupported dtype for dequantize {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f16>(elem_count).w()? };
+    // See e.g.
+    // https://github.com/ggerganov/llama.cpp/blob/cbbd1efa06f8c09f9dff58ff9d9af509cc4c152b/ggml-cuda.cu#L7270
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (num_blocks as u32, 1, 1),
+        block_dim: (block_dim as u32, 1, 1),
+        shared_mem_bytes: 0,
+    };
+
+    if is_k {
+        let params = (&data.inner, &dst);
+        unsafe { func.launch(cfg, params) }.w()?;
+    } else {
+        let nb32 = match dtype {
+            GgmlDType::Q5_0 | GgmlDType::Q5_1 => elem_count,
+            _ => elem_count / 32,
+        };
+        let params = (&data.inner, &dst, nb32 as i32);
         unsafe { func.launch(cfg, params) }.w()?;
     }
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
 }
 
 fn dequantize_mul_mat_vec(
-    data: &CudaSlice<u8>,
+    data: &PaddedCudaSlice,
     y: &CudaView<f32>,
     dtype: GgmlDType,
     ncols: usize,
@@ -125,7 +190,7 @@ fn dequantize_mul_mat_vec(
 ) -> Result<CudaStorage> {
     use cudarc::driver::LaunchAsync;
 
-    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    let data_elems = data.len / dtype.type_size() * dtype.block_size();
     if data_elems < ncols * nrows {
         crate::bail!("unexpected data size {}, ncols {ncols} {nrows}", data_elems)
     }
@@ -154,33 +219,38 @@ fn dequantize_mul_mat_vec(
         shared_mem_bytes: 0,
     };
 
-    let params = (data, y, &dst, ncols as i32, nrows as i32);
+    let params = (&data.inner, y, &dst, ncols as i32, nrows as i32);
     unsafe { func.launch(cfg, params) }.w()?;
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
 }
 
 fn mul_mat_vec_via_q8_1(
-    data: &CudaSlice<u8>,
+    data: &PaddedCudaSlice,
     y: &CudaView<f32>,
     dtype: GgmlDType,
     ncols: usize,
     nrows: usize,
+    b_size: usize,
     dev: &CudaDevice,
 ) -> Result<CudaStorage> {
     use cudarc::driver::LaunchAsync;
 
-    let data_elems = data.len() / dtype.type_size() * dtype.block_size();
+    let data_elems = data.len / dtype.type_size() * dtype.block_size();
     if data_elems < ncols * nrows {
         crate::bail!("unexpected data size {}, ncols {ncols} {nrows}", data_elems)
     }
-    if y.len() != ncols {
+    if y.len() != ncols * b_size {
         crate::bail!("unexpected y size {}, ncols {ncols} {nrows}", y.len())
     }
+    if b_size == 0 || b_size > 8 {
+        crate::bail!("only bsize between 1 and 8 are supported, got {b_size}")
+    }
     // Start by quantizing y
     let ncols_padded = pad(ncols, MATRIX_ROW_PADDING);
-    let y_size_in_bytes = ncols_padded * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
+    let y_size_in_bytes =
+        b_size * ncols_padded * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
     let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes).w()? };
-    quantize_q8_1(y, &mut y_q8_1, ncols, dev)?;
+    quantize_q8_1(y, &mut y_q8_1, ncols, b_size, dev)?;
 
     let kernel_name = match dtype {
         GgmlDType::Q4_0 => "mul_mat_vec_q4_0_q8_1_cuda",
@@ -195,22 +265,100 @@ fn mul_mat_vec_via_q8_1(
         GgmlDType::Q6K => "mul_mat_vec_q6_K_q8_1_cuda",
         _ => crate::bail!("unsupported dtype for quantized matmul {dtype:?}"),
     };
-    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
-    let dst = unsafe { dev.alloc::<f32>(nrows).w()? };
+    let kernel_name = format!("{kernel_name}{b_size}");
+    let func = dev.get_or_load_func(&kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(nrows * b_size).w()? };
+    // https://github.com/ggerganov/llama.cpp/blob/facb8b56f8fd3bb10a693bf0943ae9d69d0828ef/ggml-cuda/mmvq.cu#L98
+    let (nblocks, nwarps) = match b_size {
+        1 => (nrows as u32, 4),
+        2..=4 => ((nrows as u32 + 1) / 2, 4),
+        5..=8 => ((nrows as u32 + 1) / 2, 2),
+        _ => crate::bail!("unexpected bsize {b_size}"),
+    };
     let cfg = cudarc::driver::LaunchConfig {
-        grid_dim: (nrows as u32, 1, 1),
+        grid_dim: (nblocks, 1, 1),
+        block_dim: (WARP_SIZE as u32, nwarps, 1),
+        shared_mem_bytes: 0,
+    };
+
+    let params = (
+        &data.inner,
+        &y_q8_1,
+        &dst,
+        /* ncols_x */ ncols as i32,
+        /* nrows_x */ nrows as i32,
+        /* nrows_y */ ncols_padded as i32,
+        /* nrows_dst */ nrows as i32,
+    );
+    unsafe { func.launch(cfg, params) }.w()?;
+    Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
+}
+
+#[allow(clippy::too_many_arguments)]
+fn mul_mat_via_q8_1(
+    data: &PaddedCudaSlice,
+    y: &CudaView<f32>,
+    dtype: GgmlDType,
+    x_rows: usize,
+    x_cols: usize,
+    y_rows: usize,
+    y_cols: usize,
+    dev: &CudaDevice,
+) -> Result<CudaStorage> {
+    use cudarc::driver::LaunchAsync;
+
+    let data_elems = data.len / dtype.type_size() * dtype.block_size();
+    if data_elems < x_rows * x_cols {
+        crate::bail!("unexpected lhs size {}, {x_rows} {x_cols}", data_elems)
+    }
+    if y.len() != y_rows * y_cols {
+        crate::bail!("unexpected y size {}, {y_rows} {y_cols}", y.len())
+    }
+    if x_cols != y_rows {
+        crate::bail!("unexpected x/y size {x_rows} {x_cols} {y_rows} {y_cols}")
+    }
+    let k = x_cols;
+    // Start by quantizing y
+    let k_padded = pad(k, MATRIX_ROW_PADDING);
+    let y_size_in_bytes =
+        k_padded * y_cols * GgmlDType::Q8_1.type_size() / GgmlDType::Q8_1.block_size();
+    let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes).w()? };
+    quantize_q8_1(y, &mut y_q8_1, k, y_cols, dev)?;
+
+    let (kernel_name, mmq_x, mmq_y) = match dtype {
+        GgmlDType::Q4_0 => ("mul_mat_q4_0", 64, 128),
+        GgmlDType::Q4_1 => ("mul_mat_q4_1", 64, 128),
+        GgmlDType::Q5_0 => ("mul_mat_q5_0", 128, 64),
+        GgmlDType::Q5_1 => ("mul_mat_q5_1", 128, 64),
+        GgmlDType::Q8_0 => ("mul_mat_q8_0", 128, 64),
+        GgmlDType::Q2K => ("mul_mat_q2_K", 64, 128),
+        GgmlDType::Q3K => ("mul_mat_q3_K", 128, 128),
+        GgmlDType::Q4K => ("mul_mat_q4_K", 64, 128),
+        GgmlDType::Q5K => ("mul_mat_q5_K", 64, 128),
+        GgmlDType::Q6K => ("mul_mat_q6_K", 64, 64),
+        _ => crate::bail!("unsupported dtype for quantized matmul {dtype:?}"),
+    };
+    let func = dev.get_or_load_func(kernel_name, candle_kernels::QUANTIZED)?;
+    let dst = unsafe { dev.alloc::<f32>(x_rows * y_cols).w()? };
+    let cfg = cudarc::driver::LaunchConfig {
+        grid_dim: (
+            ceil_div(x_rows, mmq_y) as u32,
+            ceil_div(y_cols, mmq_x) as u32,
+            1,
+        ),
         block_dim: (WARP_SIZE as u32, 4, 1),
         shared_mem_bytes: 0,
     };
 
     let params = (
-        data,
-        &y_q8_1,
-        &dst,
-        /* ncols_x */ ncols as i32,
-        /* nrows_x */ nrows as i32,
-        /* nrows_y */ ncols as i32,
-        /* nrows_dst */ nrows as i32,
+        /* vx */ &data.inner,
+        /* vy */ &y_q8_1,
+        /* dst */ &dst,
+        /* ncols_x */ x_cols as i32,
+        /* nrows_x */ x_rows as i32,
+        /* ncols_y */ y_cols as i32,
+        /* nrows_y */ k_padded as i32,
+        /* nrows_dst */ x_rows as i32,
     );
     unsafe { func.launch(cfg, params) }.w()?;
     Ok(CudaStorage::wrap_cuda_slice(dst, dev.clone()))
@@ -219,9 +367,14 @@ fn mul_mat_vec_via_q8_1(
 impl QCudaStorage {
     pub fn zeros(device: &CudaDevice, el_count: usize, dtype: GgmlDType) -> Result<Self> {
         let size_in_bytes = ceil_div(el_count, dtype.block_size()) * dtype.type_size();
-        let data = device.alloc_zeros::<u8>(size_in_bytes).w()?;
+        let padded_size_in_bytes =
+            ceil_div(el_count + MATRIX_ROW_PADDING, dtype.block_size()) * dtype.type_size();
+        let inner = device.alloc_zeros::<u8>(padded_size_in_bytes).w()?;
         Ok(QCudaStorage {
-            data,
+            data: PaddedCudaSlice {
+                inner,
+                len: size_in_bytes,
+            },
             device: device.clone(),
             dtype,
         })
@@ -257,11 +410,14 @@ impl QCudaStorage {
                 | GgmlDType::Q8K
         );
         if fast_kernel {
-            return dequantize(&self.data, self.dtype, elem_count, self.device());
+            return dequantize_f32(&self.data, self.dtype, elem_count, self.device());
         }
         // Run the dequantization on cpu.
 
-        let buffer = self.device.dtoh_sync_copy(&self.data).w()?;
+        let buffer = self
+            .device
+            .dtoh_sync_copy(&self.data.inner.slice(..self.data.len))
+            .w()?;
         let mut out = vec![0.0; elem_count];
         let block_len = elem_count / self.dtype.block_size();
         match self.dtype {
@@ -285,6 +441,10 @@ impl QCudaStorage {
             .storage_from_cpu_storage(&crate::CpuStorage::F32(out))
     }
 
+    pub fn dequantize_f16(&self, elem_count: usize) -> Result<CudaStorage> {
+        dequantize_f16(&self.data, self.dtype, elem_count, self.device())
+    }
+
     pub fn quantize(&mut self, src: &CudaStorage) -> Result<()> {
         // Run the quantization on cpu.
         let src = match &src.slice {
@@ -298,13 +458,21 @@ impl QCudaStorage {
         let mut qcpu_storage = crate::Device::Cpu.qzeros(src_len, self.dtype)?;
         qcpu_storage.quantize(&src)?;
         let data = qcpu_storage.data()?;
-        let data = self.device.htod_sync_copy(data.as_ref()).w()?;
-        self.data = data;
+        let padded_len =
+            data.len() + MATRIX_ROW_PADDING * self.dtype.type_size() / self.dtype.block_size();
+        let mut inner = unsafe { self.device.alloc::<u8>(padded_len).w()? };
+        self.device
+            .htod_sync_copy_into(data.as_ref(), &mut inner.slice_mut(..data.len()))
+            .w()?;
+        self.data = PaddedCudaSlice {
+            inner,
+            len: data.len(),
+        };
         Ok(())
     }
 
     pub fn storage_size_in_bytes(&self) -> usize {
-        self.data.len()
+        self.data.len
     }
 
     pub fn fwd(
@@ -313,7 +481,17 @@ impl QCudaStorage {
         storage: &CudaStorage,
         layout: &crate::Layout,
     ) -> Result<(CudaStorage, crate::Shape)> {
-        if matches!(layout.shape().dims(), [1, 1, _] | [1, _]) {
+        let max_bm = if FORCE_DMMV.load(std::sync::atomic::Ordering::Relaxed) {
+            1
+        } else {
+            8
+        };
+        let use_vec_kernel = match layout.shape().dims() {
+            [b, m, _k] => b * m <= max_bm,
+            [b, _k] => *b <= max_bm,
+            _ => false,
+        };
+        if use_vec_kernel {
             self.dequantize_matmul_vec(self_shape, storage, layout)
         } else {
             self.dequantize_matmul(self_shape, storage, layout)
@@ -334,25 +512,31 @@ impl QCudaStorage {
             Some((o1, o2)) => rhs.slice(o1..o2),
             None => Err(crate::Error::RequiresContiguous { op: "dmmv" }.bt())?,
         };
-        let (with_batch, k) = match rhs_l.shape().dims() {
-            [1, 1, k] => (true, k),
-            [1, k] => (false, k),
+        let (b_size, k) = match rhs_l.shape().dims() {
+            [b, m, k] => (b * m, *k),
+            [b, k] => (*b, *k),
             _ => crate::bail!("unexpected rhs shape in dmmv {:?}", rhs_l.shape()),
         };
-        if ncols != *k {
+        if ncols != k {
             crate::bail!("mismatch on matmul dim {self_shape:?} {:?}", rhs_l.shape())
         }
 
         let out = if FORCE_DMMV.load(std::sync::atomic::Ordering::Relaxed) {
             dequantize_mul_mat_vec(&self.data, &rhs, self.dtype, ncols, nrows, self.device())?
         } else {
-            mul_mat_vec_via_q8_1(&self.data, &rhs, self.dtype, ncols, nrows, self.device())?
-        };
-        let out_shape = if with_batch {
-            vec![1, 1, nrows]
-        } else {
-            vec![1, nrows]
+            mul_mat_vec_via_q8_1(
+                &self.data,
+                &rhs,
+                self.dtype,
+                ncols,
+                nrows,
+                b_size,
+                self.device(),
+            )?
         };
+        let mut out_shape = rhs_l.shape().dims().to_vec();
+        out_shape.pop();
+        out_shape.push(nrows);
         Ok((out, out_shape.into()))
     }
 
@@ -373,9 +557,30 @@ impl QCudaStorage {
             crate::bail!("mismatch on matmul dim {self_shape:?} {:?}", layout.shape())
         }
 
-        let data_f32 = self.dequantize(n * k)?;
-        let rhs_l = crate::Layout::new((k, n).into(), vec![1, k], 0).broadcast_as((b, k, n))?;
-        let out = storage.matmul(&data_f32, (b, m, n, k), layout, &rhs_l)?;
+        let out = if FORCE_DMMV.load(std::sync::atomic::Ordering::Relaxed) {
+            let data_f32 = self.dequantize(n * k)?;
+            let rhs_l = crate::Layout::new((k, n).into(), vec![1, k], 0).broadcast_as((b, k, n))?;
+            storage.matmul(&data_f32, (b, m, n, k), layout, &rhs_l)?
+        } else {
+            let storage = storage.as_cuda_slice::<f32>()?;
+            let storage = match layout.contiguous_offsets() {
+                Some((o1, o2)) => storage.slice(o1..o2),
+                None => Err(crate::Error::RequiresContiguous {
+                    op: "quantized-matmul",
+                }
+                .bt())?,
+            };
+            mul_mat_via_q8_1(
+                &self.data,
+                &storage,
+                self.dtype,
+                /* x_rows */ n,
+                /* x_cols */ k,
+                /* y_rows */ k,
+                /* y_cols */ b * m,
+                self.device(),
+            )?
+        };
         let mut out_shape = layout.shape().dims().to_vec();
         out_shape.pop();
         out_shape.push(n);
@@ -390,11 +595,19 @@ pub fn load_quantized<T: super::GgmlType + Send + Sync + 'static>(
     let data = unsafe {
         std::slice::from_raw_parts(data.as_ptr() as *const u8, core::mem::size_of_val(data))
     };
-    let data = device.htod_sync_copy(data).w()?;
+    let dtype = T::DTYPE;
+    let padded_len = data.len() + MATRIX_ROW_PADDING * dtype.type_size() / dtype.block_size();
+    let mut inner = unsafe { device.alloc::<u8>(padded_len).w()? };
+    device
+        .htod_sync_copy_into(data, &mut inner.slice_mut(..data.len()))
+        .w()?;
     Ok(QStorage::Cuda(QCudaStorage {
-        data,
+        data: PaddedCudaSlice {
+            inner,
+            len: data.len(),
+        },
         device: device.clone(),
-        dtype: T::DTYPE,
+        dtype,
     }))
 }
 
@@ -412,7 +625,7 @@ mod test {
         let mut y_q8_1 = unsafe { dev.alloc::<u8>(y_size_in_bytes).w()? };
         let vs: Vec<f32> = (0..el).map(|v| v as f32).collect();
         let y = dev.htod_sync_copy(&vs).w()?;
-        quantize_q8_1(&y.slice(..), &mut y_q8_1, el, &dev)?;
+        quantize_q8_1(&y.slice(..), &mut y_q8_1, el, 1, &dev)?;
         Ok(())
     }
 
@@ -430,6 +643,7 @@ mod test {
             /* dtype */ GgmlDType::Q4_0,
             /* ncols */ ncols,
             /* nrows */ 1,
+            /* b_size */ 1,
             &dev,
         )?;
         let vs = cuda_storage.as_cuda_slice::<f32>()?;
@@ -453,4 +667,68 @@ mod test {
         assert_eq!(vs[0], 5561851.0);
         Ok(())
     }
+
+    #[test]
+    fn cuda_mm_q8_1() -> Result<()> {
+        let dev = CudaDevice::new(0)?;
+        let ncols = 256;
+        let vs: Vec<f32> = (0..ncols * 4).map(|v| v as f32 / 4.).collect();
+        let y = dev.htod_sync_copy(&vs).w()?;
+        let mut xs = QCudaStorage::zeros(&dev, ncols * 4, GgmlDType::Q4_0)?;
+        xs.quantize(&CudaStorage::wrap_cuda_slice(y.clone(), dev.clone()))?;
+        let cuda_storage = mul_mat_via_q8_1(
+            &xs.data,
+            &y.slice(..),
+            /* dtype */ GgmlDType::Q4_0,
+            /* x_rows */ 4,
+            /* x_cols */ ncols,
+            /* y_rows */ ncols,
+            /* y_cols */ 4,
+            &dev,
+        )?;
+        let vs = cuda_storage.as_cuda_slice::<f32>()?;
+        let vs = dev.dtoh_sync_copy(&vs.slice(..)).unwrap();
+
+        /*
+           x = torch.tensor([float(v) for v in range(1024)]).reshape(4, 256)
+           x @ x.t() / 16
+        tensor([[  347480.0000,   869720.0000,  1391960.0000,  1914200.0000],
+                [  869720.0000,  2440536.0000,  4011352.0000,  5582166.5000],
+                [ 1391960.0000,  4011352.0000,  6630742.0000,  9250132.0000],
+                [ 1914200.0000,  5582166.5000,  9250132.0000, 12918099.0000]])
+                */
+        assert_eq!(vs.len(), 16);
+        assert_eq!(vs[0], 347604.0);
+        assert_eq!(vs[1], 888153.06);
+        assert_eq!(vs[4], 869780.7);
+        assert_eq!(vs[5], 2483145.0);
+        assert_eq!(vs[11], 9407368.0);
+        assert_eq!(vs[14], 9470856.0);
+        assert_eq!(vs[15], 13138824.0);
+        Ok(())
+    }
+
+    // The following test used to fail under compute-sanitizer until #2526.
+    #[test]
+    fn cuda_mm_q8_1_pad() -> Result<()> {
+        let dev = CudaDevice::new(0)?;
+        let (x_rows, ncols, y_cols) = (4, 16, 2048);
+        let vs: Vec<f32> = (0..ncols * y_cols).map(|v| v as f32 / 256.).collect();
+        let y = dev.htod_sync_copy(&vs).w()?;
+        let mut xs = QCudaStorage::zeros(&dev, ncols * x_rows, GgmlDType::Q4_0)?;
+        xs.quantize(&CudaStorage::wrap_cuda_slice(y.clone(), dev.clone()))?;
+        let cuda_storage = mul_mat_via_q8_1(
+            &xs.data,
+            &y.slice(..),
+            /* dtype */ GgmlDType::Q4_0,
+            /* x_rows */ x_rows,
+            /* x_cols */ ncols,
+            /* y_rows */ ncols,
+            /* y_cols */ y_cols,
+            &dev,
+        )?;
+        let vs = cuda_storage.as_cuda_slice::<f32>()?;
+        let _vs = dev.dtoh_sync_copy(&vs.slice(..)).unwrap();
+        Ok(())
+    }
 }

candle-core/src/quantized/dummy_cuda.rs

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

candle-core/src/quantized/gguf_file.rs

@@ -135,7 +135,6 @@ pub enum ValueType {
     // The value is a UTF-8 non-null-terminated string, with length prepended.
     String,
     // The value is an array of other values, with the length and type prepended.
-    ///
     // Arrays can be nested, and the length of the array is the number of elements in the array, not the number of bytes.
     Array,
 }
@@ -218,10 +217,16 @@ impl Value {
         }
     }
 
+    /// This will also automatically upcast any integral types which will not truncate.
     pub fn to_u64(&self) -> Result<u64> {
         match self {
             Self::U64(v) => Ok(*v),
-            v => crate::bail!("not a u64 {v:?}"),
+            // Autoupcast cases here
+            Self::U8(v) => Ok(*v as u64),
+            Self::U16(v) => Ok(*v as u64),
+            Self::U32(v) => Ok(*v as u64),
+            Self::Bool(v) => Ok(*v as u64),
+            v => crate::bail!("not a u64 or upcastable to u64 {v:?}"),
         }
     }
 

candle-core/src/quantized/metal.rs

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

candle-core/src/quantized/mod.rs

@@ -1,4 +1,4 @@
-use crate::{CpuStorage, Device, Result, Shape, Storage, Tensor};
+use crate::{CpuStorage, DType, Device, Result, Shape, Storage, Tensor};
 use k_quants::*;
 use std::borrow::Cow;
 
@@ -360,9 +360,24 @@ impl QTensor {
     pub fn dequantize(&self, device: &Device) -> Result<Tensor> {
         let storage = self.storage.dequantize(self.shape.elem_count())?;
         let none = crate::op::BackpropOp::none();
-        let is_variable = false;
-        crate::tensor::from_storage(storage, self.shape.clone(), none, is_variable)
-            .to_device(device)
+        crate::tensor::from_storage(storage, self.shape.clone(), none, false).to_device(device)
+    }
+
+    pub fn dequantize_f16(&self, device: &Device) -> Result<Tensor> {
+        // In the CUDA case, we have a specialized kernel as this can be useful for volta
+        // architectures. https://github.com/huggingface/candle/issues/2136
+        match &self.storage {
+            QStorage::Cuda(s) => {
+                let s = s.dequantize_f16(self.shape.elem_count())?;
+                let none = crate::op::BackpropOp::none();
+                crate::tensor::from_storage(Storage::Cuda(s), self.shape.clone(), none, false)
+                    .to_device(device)
+            }
+            _ => {
+                let s = self.dequantize(device)?.to_dtype(crate::DType::F16)?;
+                Ok(s)
+            }
+        }
     }
 
     pub fn storage_size_in_bytes(&self) -> usize {
@@ -378,6 +393,7 @@ impl QTensor {
 pub enum QMatMul {
     QTensor(std::sync::Arc<QTensor>),
     Tensor(Tensor),
+    TensorF16(Tensor),
 }
 
 thread_local! {
@@ -391,6 +407,17 @@ thread_local! {
     }
 }
 
+thread_local! {
+    static DEQUANTIZE_ALL_F16: bool = {
+        match std::env::var("CANDLE_DEQUANTIZE_ALL_F16") {
+            Ok(s) => {
+                !s.is_empty() && s != "0"
+            },
+            Err(_) => false,
+        }
+    }
+}
+
 impl QMatMul {
     pub fn from_arc(qtensor: std::sync::Arc<QTensor>) -> Result<Self> {
         let dequantize = match qtensor.dtype() {
@@ -400,6 +427,9 @@ impl QMatMul {
         let t = if dequantize {
             let tensor = qtensor.dequantize(&qtensor.device())?;
             Self::Tensor(tensor)
+        } else if DEQUANTIZE_ALL_F16.with(|b| *b) {
+            let tensor = qtensor.dequantize_f16(&qtensor.device())?;
+            Self::TensorF16(tensor)
         } else {
             Self::QTensor(qtensor)
         };
@@ -409,6 +439,25 @@ impl QMatMul {
     pub fn from_qtensor(qtensor: QTensor) -> Result<Self> {
         Self::from_arc(std::sync::Arc::new(qtensor))
     }
+
+    pub fn dequantize_f16(&self) -> Result<Tensor> {
+        match self {
+            Self::QTensor(t) => t.dequantize_f16(&t.device()),
+            Self::Tensor(t) => t.to_dtype(DType::F16),
+            Self::TensorF16(t) => Ok(t.clone()),
+        }
+    }
+
+    pub fn forward_via_f16(&self, xs: &Tensor) -> Result<Tensor> {
+        let w = self.dequantize_f16()?;
+        let in_dtype = xs.dtype();
+        let w = match *xs.dims() {
+            [b1, b2, _, _] => w.broadcast_left((b1, b2))?.t()?,
+            [bsize, _, _] => w.broadcast_left(bsize)?.t()?,
+            _ => w.t()?,
+        };
+        xs.to_dtype(DType::F16)?.matmul(&w)?.to_dtype(in_dtype)
+    }
 }
 
 impl crate::CustomOp1 for QTensor {
@@ -486,6 +535,15 @@ impl crate::Module for QMatMul {
                 };
                 xs.matmul(&w)
             }
+            Self::TensorF16(w) => {
+                let in_dtype = xs.dtype();
+                let w = match *xs.dims() {
+                    [b1, b2, _, _] => w.broadcast_left((b1, b2))?.t()?,
+                    [bsize, _, _] => w.broadcast_left(bsize)?.t()?,
+                    _ => w.t()?,
+                };
+                xs.to_dtype(DType::F16)?.matmul(&w)?.to_dtype(in_dtype)
+            }
         }
     }
 }

candle-core/src/safetensors.rs

@@ -349,6 +349,30 @@ impl MmapedSafetensors {
     }
 }
 
+pub struct SliceSafetensors<'a> {
+    safetensors: SafeTensors<'a>,
+}
+
+impl<'a> SliceSafetensors<'a> {
+    /// Creates a wrapper around a binary buffer and deserialize the safetensors header.
+    pub fn new(buffer: &'a [u8]) -> Result<Self> {
+        let safetensors = safetensors::SafeTensors::deserialize(buffer)?;
+        Ok(Self { safetensors })
+    }
+
+    pub fn load(&self, name: &str, dev: &Device) -> Result<Tensor> {
+        self.safetensors.tensor(name)?.load(dev)
+    }
+
+    pub fn tensors(&self) -> Vec<(String, st::TensorView<'_>)> {
+        self.safetensors.tensors()
+    }
+
+    pub fn get(&self, name: &str) -> Result<st::TensorView<'_>> {
+        Ok(self.safetensors.tensor(name)?)
+    }
+}
+
 pub struct BufferedSafetensors {
     safetensors: yoke::Yoke<SafeTensors_<'static>, Vec<u8>>,
 }

candle-core/src/shape.rs

@@ -142,6 +142,12 @@ impl Shape {
         &self.0
     }
 
+    /// The dimension size for a specified dimension index.
+    pub fn dim<D: Dim>(&self, dim: D) -> Result<usize> {
+        let dim = dim.to_index(self, "dim")?;
+        Ok(self.dims()[dim])
+    }
+
     /// The total number of elements, this is the product of all dimension sizes.
     pub fn elem_count(&self) -> usize {
         self.0.iter().product()
@@ -304,6 +310,7 @@ impl Dim for usize {
 pub enum D {
     Minus1,
     Minus2,
+    Minus(usize),
 }
 
 impl D {
@@ -311,6 +318,7 @@ impl D {
         let dim = match self {
             Self::Minus1 => -1,
             Self::Minus2 => -2,
+            Self::Minus(u) => -(*u as i32),
         };
         Error::DimOutOfRange {
             shape: shape.clone(),
@@ -327,6 +335,7 @@ impl Dim for D {
         match self {
             Self::Minus1 if rank >= 1 => Ok(rank - 1),
             Self::Minus2 if rank >= 2 => Ok(rank - 2),
+            Self::Minus(u) if *u > 0 && rank >= *u => Ok(rank - *u),
             _ => Err(self.out_of_range(shape, op)),
         }
     }
@@ -336,6 +345,7 @@ impl Dim for D {
         match self {
             Self::Minus1 => Ok(rank),
             Self::Minus2 if rank >= 1 => Ok(rank - 1),
+            Self::Minus(u) if *u > 0 && rank + 1 >= *u => Ok(rank + 1 - *u),
             _ => Err(self.out_of_range(shape, op)),
         }
     }

candle-core/src/sort.rs

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

candle-core/src/streaming.rs

@@ -0,0 +1,206 @@
+use crate::{Result, Shape, Tensor};
+
+pub trait Dim: crate::shape::Dim + Copy {}
+impl<T: crate::shape::Dim + Copy> Dim for T {}
+
+/// A stream tensor is used in streaming module. It can either contain an actual tensor or be
+/// empty.
+#[derive(Clone)]
+pub struct StreamTensor(Option<Tensor>);
+
+impl std::fmt::Debug for StreamTensor {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match &self.0 {
+            Some(t) => write!(f, "{:?}", t.shape()),
+            None => write!(f, "Empty"),
+        }
+    }
+}
+
+impl std::convert::From<Option<Tensor>> for StreamTensor {
+    fn from(value: Option<Tensor>) -> Self {
+        Self(value)
+    }
+}
+
+impl std::convert::From<Tensor> for StreamTensor {
+    fn from(value: Tensor) -> Self {
+        Self(Some(value))
+    }
+}
+
+impl std::convert::From<()> for StreamTensor {
+    fn from(_value: ()) -> Self {
+        Self(None)
+    }
+}
+
+impl StreamTensor {
+    pub fn empty() -> Self {
+        Self(None)
+    }
+
+    pub fn from_tensor(tensor: Tensor) -> Self {
+        Self(Some(tensor))
+    }
+
+    pub fn shape(&self) -> Option<&Shape> {
+        self.0.as_ref().map(|t| t.shape())
+    }
+
+    pub fn cat2<D: Dim>(&self, rhs: &Self, dim: D) -> Result<Self> {
+        let xs = match (&self.0, &rhs.0) {
+            (Some(lhs), Some(rhs)) => {
+                let xs = Tensor::cat(&[lhs, rhs], dim)?;
+                Some(xs)
+            }
+            (Some(xs), None) | (None, Some(xs)) => Some(xs.clone()),
+            (None, None) => None,
+        };
+        Ok(Self(xs))
+    }
+
+    pub fn seq_len<D: Dim>(&self, dim: D) -> Result<usize> {
+        match &self.0 {
+            None => Ok(0),
+            Some(v) => v.dim(dim),
+        }
+    }
+
+    pub fn reset(&mut self) {
+        self.0 = None
+    }
+
+    pub fn narrow<D: Dim>(&self, dim: D, offset: usize, len: usize) -> Result<StreamTensor> {
+        let t = match &self.0 {
+            None => None,
+            Some(t) => {
+                let seq_len = t.dim(dim)?;
+                if seq_len <= offset {
+                    None
+                } else {
+                    let t = t.narrow(dim, offset, usize::min(len, seq_len - offset))?;
+                    Some(t)
+                }
+            }
+        };
+        Ok(Self(t))
+    }
+
+    /// Splits the Streaming Tensor on the time axis `dim` with the first `lhs_len` elements
+    /// returned in the first output and the remaining in the second output.
+    pub fn split<D: Dim>(&self, dim: D, lhs_len: usize) -> Result<(Self, Self)> {
+        match &self.0 {
+            None => Ok((Self::empty(), Self::empty())),
+            Some(t) => {
+                let seq_len = t.dim(dim)?;
+                let lhs_len = usize::min(seq_len, lhs_len);
+                if lhs_len == 0 {
+                    Ok((Self::empty(), t.clone().into()))
+                } else {
+                    let lhs = Self::from_tensor(t.narrow(dim, 0, lhs_len)?);
+                    let rhs_len = seq_len - lhs_len;
+                    let rhs = if rhs_len == 0 {
+                        Self::empty()
+                    } else {
+                        Self::from_tensor(t.narrow(dim, lhs_len, rhs_len)?)
+                    };
+                    Ok((lhs, rhs))
+                }
+            }
+        }
+    }
+
+    pub fn as_option(&self) -> Option<&Tensor> {
+        self.0.as_ref()
+    }
+
+    pub fn apply<M: crate::Module>(&self, m: &M) -> Result<Self> {
+        match &self.0 {
+            None => Ok(Self::empty()),
+            Some(t) => Ok(Self::from_tensor(t.apply(m)?)),
+        }
+    }
+}
+
+/// Streaming modules take as input a stream tensor and return a stream tensor. They may perform
+/// some internal buffering so that enough data has been received for the module to be able to
+/// perform some operations.
+pub trait StreamingModule {
+    // TODO: Should we also have a flush method?
+    fn step(&mut self, xs: &StreamTensor) -> Result<StreamTensor>;
+    fn reset_state(&mut self);
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum BinOp {
+    Add,
+    Mul,
+    Sub,
+    Div,
+}
+
+#[derive(Debug, Clone)]
+pub struct StreamingBinOp {
+    prev_lhs: StreamTensor,
+    prev_rhs: StreamTensor,
+    pub op: BinOp,
+    pub dim: crate::D,
+}
+
+impl StreamingBinOp {
+    pub fn new(op: BinOp, dim: crate::D) -> Self {
+        Self {
+            prev_lhs: StreamTensor::empty(),
+            prev_rhs: StreamTensor::empty(),
+            op,
+            dim,
+        }
+    }
+
+    pub fn reset_state(&mut self) {
+        self.prev_lhs.reset();
+        self.prev_rhs.reset();
+    }
+
+    pub fn forward(&self, lhs: &Tensor, rhs: &Tensor) -> Result<Tensor> {
+        match self.op {
+            BinOp::Add => Tensor::add(lhs, rhs),
+            BinOp::Mul => Tensor::mul(lhs, rhs),
+            BinOp::Sub => Tensor::sub(lhs, rhs),
+            BinOp::Div => Tensor::div(lhs, rhs),
+        }
+    }
+
+    pub fn step(&mut self, lhs: &StreamTensor, rhs: &StreamTensor) -> Result<StreamTensor> {
+        let lhs = StreamTensor::cat2(&self.prev_lhs, lhs, self.dim)?;
+        let rhs = StreamTensor::cat2(&self.prev_rhs, rhs, self.dim)?;
+        let lhs_len = lhs.seq_len(self.dim)?;
+        let rhs_len = rhs.seq_len(self.dim)?;
+        let common_len = usize::min(lhs_len, rhs_len);
+        let (lhs, prev_lhs) = lhs.split(self.dim, common_len)?;
+        let (rhs, prev_rhs) = rhs.split(self.dim, common_len)?;
+        let ys = match (lhs.0, rhs.0) {
+            (Some(lhs), Some(rhs)) => {
+                let ys = self.forward(&lhs, &rhs)?;
+                StreamTensor::from_tensor(ys)
+            }
+            (None, None) => StreamTensor::empty(),
+            (lhs, rhs) => crate::bail!("INTERNAL ERROR inconsistent lhs and rhs {lhs:?} {rhs:?}"),
+        };
+        self.prev_lhs = prev_lhs;
+        self.prev_rhs = prev_rhs;
+        Ok(ys)
+    }
+}
+
+/// Simple wrapper that doesn't do any buffering.
+pub struct Map<T: crate::Module>(T);
+
+impl<T: crate::Module> StreamingModule for Map<T> {
+    fn reset_state(&mut self) {}
+
+    fn step(&mut self, xs: &StreamTensor) -> Result<StreamTensor> {
+        xs.apply(&self.0)
+    }
+}

candle-core/src/tensor.rs

@@ -79,6 +79,9 @@ macro_rules! unary_op {
     ($fn_name:ident, $op_name:ident) => {
         pub fn $fn_name(&self) -> Result<Self> {
             let shape = self.shape();
+            if shape.elem_count() == 0 {
+                return Ok(self.clone());
+            }
             let storage = self
                 .storage()
                 .unary_impl::<crate::op::$op_name>(self.layout())?;
@@ -92,6 +95,9 @@ macro_rules! binary_op {
     ($fn_name:ident, $op_name:ident) => {
         pub fn $fn_name(&self, rhs: &Self) -> Result<Self> {
             let shape = self.same_shape_binary_op(rhs, stringify!($fn_name))?;
+            if shape.elem_count() == 0 {
+                return Ok(self.clone());
+            }
             let storage = self.storage().binary_impl::<crate::op::$op_name>(
                 &*rhs.storage(),
                 self.layout(),
@@ -114,6 +120,9 @@ macro_rules! binary_op_scalar {
                     .broadcast_as(self.shape())?,
             };
             let shape = self.same_shape_binary_op(&rhs, stringify!($fn_name))?;
+            if self.elem_count() == 0 {
+                return Ok(self.clone());
+            }
             let storage = self.storage().binary_impl::<crate::op::$op_name>(
                 &*rhs.storage(),
                 self.layout(),
@@ -361,6 +370,15 @@ impl Tensor {
 
     /// 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.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::full(3.5, (2, 4), &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec2::<f64>()?, &[
+    ///     [3.5, 3.5, 3.5, 3.5],
+    ///     [3.5, 3.5, 3.5, 3.5],
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
     pub fn full<D: crate::WithDType, S: Into<Shape>>(
         value: D,
         shape: S,
@@ -370,6 +388,13 @@ impl Tensor {
     }
 
     /// Creates a new 1D tensor from an iterator.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::from_iter( [1.0, 2.0, 3.0, 4.0].into_iter(), &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec1::<f64>()?, &[1.0, 2.0, 3.0, 4.0]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     pub fn from_iter<D: crate::WithDType>(
         iter: impl IntoIterator<Item = D>,
         device: &Device,
@@ -381,12 +406,26 @@ impl Tensor {
 
     /// Creates a new 1D tensor with values from the interval `[start, end)` taken with a common
     /// difference `1` from `start`.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::arange(2., 5., &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec1::<f64>()?, &[2., 3., 4.]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     pub fn arange<D: crate::WithDType>(start: D, end: D, device: &Device) -> Result<Self> {
         Self::arange_step(start, end, D::one(), device)
     }
 
     /// Creates a new 1D tensor with values from the interval `[start, end)` taken with a common
     /// difference `step` from `start`.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::arange_step(2.0, 4.0, 0.5, &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec1::<f64>()?, &[2.0, 2.5, 3.0, 3.5]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     pub fn arange_step<D: crate::WithDType>(
         start: D,
         end: D,
@@ -432,6 +471,16 @@ impl Tensor {
     /// Creates a new tensor initialized with values from the input vector. The number of elements
     /// in this vector must be the same as the number of elements defined by the shape.
     /// If the device is cpu, no data copy is made.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::from_vec(vec!{1., 2., 3., 4., 5., 6.}, (2, 3), &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec2::<f64>()?, &[
+    ///     [1., 2., 3.],
+    ///     [4., 5., 6.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     pub fn from_vec<S: Into<Shape>, D: crate::WithDType>(
         data: Vec<D>,
         shape: S,
@@ -442,12 +491,31 @@ impl Tensor {
 
     /// Creates a new tensor initialized with values from the input slice. The number of elements
     /// in this vector must be the same as the number of elements defined by the shape.
+    ///```rust
+    /// use candle_core::{Tensor, Device};
+    /// let values = vec![1., 2., 3., 4., 5., 6., 7., 8.];
+    /// let a = Tensor::from_slice(&values[1..7], (2, 3), &Device::Cpu)?;
+    ///
+    /// assert_eq!(a.to_vec2::<f64>()?, &[
+    ///     [2., 3., 4.],
+    ///     [5., 6., 7.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     pub fn from_slice<S: Into<Shape>, D: crate::WithDType>(
         array: &[D],
         shape: S,
         device: &Device,
     ) -> Result<Self> {
-        Self::new_impl(array, shape.into(), device, false)
+        let shape = shape.into();
+        let n: usize = shape.elem_count();
+        let buffer_size: usize = array.len();
+        if buffer_size != n {
+            return Err(Error::ShapeMismatch { buffer_size, shape }.bt());
+        }
+        let storage = device.storage_from_slice(array)?;
+        let none = BackpropOp::none();
+        Ok(from_storage(storage, shape, none, false))
     }
 
     pub(crate) fn same_shape_binary_op(&self, rhs: &Self, op: &'static str) -> Result<&Shape> {
@@ -573,9 +641,9 @@ impl Tensor {
     ///
     /// * `args` - A slice of 1D tensors.
     /// * `xy_indexing` - Whether to use xy indexing or ij indexing. If xy is selected, the
-    /// first dimension corresponds to the cardinality of the second input and the second
-    /// dimension corresponds to the cardinality of the first input. If ij is selected, the
-    /// dimensions are in the same order as the cardinality of the inputs.
+    ///   first dimension corresponds to the cardinality of the second input and the second
+    ///   dimension corresponds to the cardinality of the first input. If ij is selected, the
+    ///   dimensions are in the same order as the cardinality of the inputs.
     ///
     /// # Examples
     ///
@@ -646,6 +714,9 @@ impl Tensor {
     /// # Ok::<(), candle_core::Error>(())
     /// ```
     pub fn affine(&self, mul: f64, add: f64) -> Result<Self> {
+        if self.elem_count() == 0 {
+            return Ok(self.clone());
+        }
         let storage = self.storage().affine(self.layout(), mul, add)?;
         let op = BackpropOp::new1(self, |arg| Op::Affine { arg, mul, add });
         Ok(from_storage(storage, self.shape(), op, false))
@@ -653,6 +724,9 @@ impl Tensor {
 
     /// Applies the Exponential Linear Unit (ELU) function on each element of the input tensor.
     pub fn elu(&self, alpha: f64) -> Result<Self> {
+        if self.elem_count() == 0 {
+            return Ok(self.clone());
+        }
         let storage = self.storage().elu(self.layout(), alpha)?;
         let op = BackpropOp::new1(self, |t| Op::Elu(t, alpha));
         Ok(from_storage(storage, self.shape(), op, false))
@@ -660,6 +734,9 @@ impl Tensor {
 
     /// Raise the tensor to some float exponent `e`.
     pub fn powf(&self, e: f64) -> Result<Self> {
+        if self.elem_count() == 0 {
+            return Ok(self.clone());
+        }
         let storage = self.storage().powf(self.layout(), e)?;
         let op = BackpropOp::new1(self, |t| Op::Powf(t, e));
         Ok(from_storage(storage, self.shape(), op, false))
@@ -706,6 +783,30 @@ impl Tensor {
 
     /// Returns a new tensor that is a narrowed version of the input, the dimension `dim`
     /// ranges from `start` to `start + len`.
+    /// ```
+    /// use candle_core::{Tensor, Device};
+    /// let a = Tensor::new(&[
+    ///     [0f32, 1., 2.],
+    ///     [3.  , 4., 5.],
+    ///     [6.  , 7., 8.]
+    /// ], &Device::Cpu)?;
+    ///
+    /// let b = a.narrow(0, 1, 2)?;
+    /// assert_eq!(b.shape().dims(), &[2, 3]);
+    /// assert_eq!(b.to_vec2::<f32>()?, &[
+    ///     [3., 4., 5.],
+    ///     [6., 7., 8.]
+    /// ]);
+    ///
+    /// let c = a.narrow(1, 1, 1)?;
+    /// assert_eq!(c.shape().dims(), &[3, 1]);
+    /// assert_eq!(c.to_vec2::<f32>()?, &[
+    ///     [1.],
+    ///     [4.],
+    ///     [7.]
+    /// ]);
+    /// # Ok::<(), candle_core::Error>(())
+    /// ```
     pub fn narrow<D: Dim>(&self, dim: D, start: usize, len: usize) -> Result<Self> {
         let dims = self.dims();
         let dim = dim.to_index(self.shape(), "narrow")?;
@@ -1154,6 +1255,9 @@ impl Tensor {
         let n = b_dims[dim - 1];
 
         let c_shape = Shape::from(&a_dims[..dim - 2]).extend(&[m, n]);
+        if c_shape.elem_count() == 0 || k == 0 {
+            return Tensor::zeros(c_shape, self.dtype(), self.device());
+        }
         let batching: usize = a_dims[..dim - 2].iter().product();
         let batching_b: usize = b_dims[..dim - 2].iter().product();
         if k != k2 || batching != batching_b {
@@ -1416,14 +1520,15 @@ impl Tensor {
     /// # Arguments
     ///
     /// * `self` - The input tensor.
-    /// * `indexes` - The indices of elements to gather, this should have the same shape as `self`
-    ///   but can have a different number of elements on the target dimension.
+    /// * `indexes` - The indices of elements to gather, this should have same number of dimensions as `self`
+    ///   and indexes.dims()[d] <= self.dims()[d] for all dimensions d != dim
     /// * `dim` - the target dimension.
     ///
     /// The resulting tensor has the same shape as `indexes` and use values from `self` indexed on
     /// dimension `dim` by the values in `indexes`.
     pub fn gather<D: Dim>(&self, indexes: &Self, dim: D) -> Result<Self> {
         let dim = dim.to_index(self.shape(), "gather")?;
+
         let self_dims = self.dims();
         let indexes_dims = indexes.dims();
         let mismatch = if indexes_dims.len() != self_dims.len() {
@@ -1431,7 +1536,7 @@ impl Tensor {
         } else {
             let mut mismatch = false;
             for (i, (&d1, &d2)) in self_dims.iter().zip(indexes_dims.iter()).enumerate() {
-                if i != dim && d1 != d2 {
+                if i != dim && d1 < d2 {
                     mismatch = true;
                     break;
                 }
@@ -1921,7 +2026,11 @@ impl Tensor {
                 }
                 (Storage::Cpu(storage), Device::Cpu) => Storage::Cpu(storage.clone()),
                 _ => {
-                    bail!("not implemented yet")
+                    bail!(
+                        "not implemented yet, self.device: {:?}, device: {:?}",
+                        self.device(),
+                        device
+                    )
                 }
             };
             let op = BackpropOp::new1(self, Op::ToDevice);
@@ -2411,9 +2520,19 @@ impl Tensor {
 
     /// Returns log(sum(exp(tensor), dim)).
     pub fn log_sum_exp<D: Dims>(&self, sum_dims: D) -> Result<Self> {
-        let exp = self.exp()?;
-        let sum = exp.sum(sum_dims)?;
-        sum.log()
+        let sum_dims = sum_dims.to_indexes(self.shape(), "log-sum-exp")?;
+        if sum_dims.is_empty() {
+            return Ok(self.clone());
+        }
+        let max = sum_dims[1..]
+            .iter()
+            .try_fold(self.max_keepdim(sum_dims[0])?, |max, &dim| {
+                max.max_keepdim(dim)
+            })?;
+        let exp = self.broadcast_sub(&max)?.exp()?;
+        let sum = exp.sum(sum_dims.clone())?;
+
+        sum.log()? + max.squeeze_dims(&sum_dims)
     }
 
     /// Pointwise pow operation.

candle-core/src/tensor_cat.rs

@@ -235,4 +235,66 @@ impl Tensor {
         }
         Ok(crate::tensor::from_storage(storage, shape, op, false))
     }
+
+    /// Set the values on `self` using values from `src`. The copy starts at the specified
+    /// `offset` for the target dimension `dim` on `self`.
+    /// `self` and `src` must have the same shape except on dimension `dim` where the `self` size
+    /// has to be greater than or equal to `offset` plus the `src` size.
+    ///
+    /// Note that this modifies `self` in place and as such is not compatibel with
+    /// back-propagation.  
+    pub fn slice_set<D: Dim>(&self, src: &Self, dim: D, offset: usize) -> Result<()> {
+        let dim = dim.to_index(self.shape(), "slice-set")?;
+        if !self.is_contiguous() || !src.is_contiguous() {
+            Err(Error::RequiresContiguous { op: "slice-set" }.bt())?
+        }
+        if self.dtype() != src.dtype() {
+            Err(Error::DTypeMismatchBinaryOp {
+                lhs: self.dtype(),
+                rhs: src.dtype(),
+                op: "slice-set",
+            }
+            .bt())?
+        }
+        if self.device().location() != src.device().location() {
+            Err(Error::DeviceMismatchBinaryOp {
+                lhs: self.device().location(),
+                rhs: src.device().location(),
+                op: "slice-set",
+            }
+            .bt())?
+        }
+        if self.rank() != src.rank() {
+            Err(Error::UnexpectedNumberOfDims {
+                expected: self.rank(),
+                got: src.rank(),
+                shape: self.shape().clone(),
+            }
+            .bt())?
+        }
+        for (dim_idx, (v1, v2)) in self.dims().iter().zip(src.dims().iter()).enumerate() {
+            if dim_idx == dim && *v2 + offset > *v1 {
+                crate::bail!("shape mismatch on target dim, dst: {v1}, src: {v2} + {offset}")
+            }
+            if dim_idx != dim && v1 != v2 {
+                crate::bail!("shape mismatch on dim {dim_idx}, {v1} <> {v2}")
+            }
+        }
+        let block_size: usize = src.dims().iter().skip(1 + dim).product();
+        let d1: usize = src.dims().iter().take(dim).product();
+        let d2 = block_size * src.dims()[dim];
+        let dst_o = self.layout().start_offset() + offset * block_size;
+        let src_o = src.layout().start_offset();
+        src.storage().copy2d(
+            &mut self.storage_mut(),
+            d1,
+            d2,
+            /* src_s */ d2,
+            /* dst_s */ block_size * self.dims()[dim],
+            src_o,
+            dst_o,
+        )?;
+
+        Ok(())
+    }
 }

candle-core/src/variable.rs

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

candle-core/tests/conv_tests.rs

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

candle-core/tests/custom_op_tests.rs

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

candle-core/tests/matmul_tests.rs

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

candle-core/tests/quantized_tests.rs

@@ -3,7 +3,7 @@ use candle_core::{
     quantized::{self, GgmlDType},
     test_device,
     test_utils::to_vec2_round,
-    Device, Module, Result, Tensor,
+    DType, Device, IndexOp, Module, Result, Tensor,
 };
 use quantized::{k_quants, GgmlType};
 use rand::prelude::*;
@@ -47,18 +47,14 @@ fn test_matmul(
 }
 
 fn quantized_matmul(device: &Device) -> Result<()> {
-    // TODO Enable this later when we enable cuda.
-    if device.is_cuda() {
-        return Ok(());
-    }
     let (m, k, n) = (3, 64, 4);
-    let lhs = (0..(m * k)).map(|v| v as f32).collect::<Vec<_>>();
-    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), device)?;
+    let lhs_s = (0..(m * k)).map(|v| v as f32).collect::<Vec<_>>();
+    let lhs = Tensor::from_slice(&lhs_s, (m, k), device)?;
     let mut dst = vec![42.; 3 * 4];
     let mut rhs_t = vec![k_quants::BlockQ4_0::zeros(); 8];
     let rhs = (0..(k * n)).map(|v| v as f32).collect::<Vec<_>>();
     k_quants::BlockQ4_0::from_float(&rhs, &mut rhs_t)?;
-    k_quants::matmul((m, k, n), &lhs, &rhs_t, &mut dst)?;
+    k_quants::matmul((m, k, n), &lhs_s, &rhs_t, &mut dst)?;
     assert_eq!(
         dst.iter().map(|x| x.round()).collect::<Vec<_>>(),
         &[
@@ -67,7 +63,7 @@ fn quantized_matmul(device: &Device) -> Result<()> {
         ]
     );
     let tensor_rhs = Tensor::from_slice(&rhs, (n, k), device)?.t()?;
-    let mm = tensor_lhs.matmul(&tensor_rhs)?;
+    let mm = lhs.matmul(&tensor_rhs)?;
     assert_eq!(
         mm.to_vec2::<f32>()?,
         &[
@@ -79,7 +75,7 @@ fn quantized_matmul(device: &Device) -> Result<()> {
 
     let qtensor = quantized::QTensor::quantize(&tensor_rhs.t()?, GgmlDType::Q4_0)?;
     let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
-    let res = matmul.forward(&tensor_lhs)?;
+    let res = matmul.forward(&lhs)?;
     match device {
         Device::Metal(_) => assert_eq!(
             to_vec2_round(&res, 0)?,
@@ -89,7 +85,15 @@ fn quantized_matmul(device: &Device) -> Result<()> {
                 [341970.0, 994574.0, 1656181.0, 2302182.0]
             ]
         ),
-        _ => assert_eq!(
+        Device::Cuda(_) => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [84866.0, 214045.0, 344676.0, 473707.0],
+                [213425.0, 604313.0, 1000431.0, 1387960.0],
+                [342030.0, 994630.0, 1656248.0, 2302250.0]
+            ]
+        ),
+        Device::Cpu => assert_eq!(
             to_vec2_round(&res, 0)?,
             &[
                 [85120.0, 214562.0, 345455.0, 474748.0],
@@ -98,22 +102,16 @@ fn quantized_matmul(device: &Device) -> Result<()> {
             ]
         ),
     }
-
     test_matmul(device, (1, 3, 4, 256), GgmlDType::Q4_0)?;
-
     Ok(())
 }
 
 fn quantized_matmul_neg(device: &Device) -> Result<()> {
-    // TODO Enable this later when we enable cuda.
-    if device.is_cuda() {
-        return Ok(());
-    }
     let (m, k, n) = (3, 64, 4);
-    let lhs = (0..(m * k))
+    let lhs_s = (0..(m * k))
         .map(|v| v as f32 - (m * k) as f32 / 2.0)
         .collect::<Vec<_>>();
-    let tensor_lhs = Tensor::from_slice(&lhs, (m, k), device)?;
+    let lhs = Tensor::from_slice(&lhs_s, (m, k), device)?;
     let mut dst = vec![42.; 3 * 4];
     let mut rhs_t = vec![k_quants::BlockQ4_0::zeros(); 8];
     let rhs = (0..k * n)
@@ -121,7 +119,7 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
         .collect::<Vec<_>>();
     let tensor_rhs = Tensor::from_slice(&rhs, (n, k), device)?.t()?;
     k_quants::BlockQ4_0::from_float(&rhs, &mut rhs_t)?;
-    k_quants::matmul((m, k, n), &lhs, &rhs_t, &mut dst)?;
+    k_quants::matmul((m, k, n), &lhs_s, &rhs_t, &mut dst)?;
     assert_eq!(
         dst.iter().map(|x| x.round()).collect::<Vec<_>>(),
         &[
@@ -129,7 +127,7 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
             -196472.0, 63012.0, 324585.0, 587902.0
         ]
     );
-    let mm = tensor_lhs.matmul(&tensor_rhs)?;
+    let mm = lhs.matmul(&tensor_rhs)?;
     assert_eq!(
         to_vec2_round(&mm, 0)?,
         &[
@@ -141,7 +139,7 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
 
     let qtensor = quantized::QTensor::quantize(&tensor_rhs.t()?, GgmlDType::Q4_0)?;
     let matmul = quantized::QMatMul::from_qtensor(qtensor)?;
-    let res = matmul.forward(&tensor_lhs)?;
+    let res = matmul.forward(&lhs)?;
     match device {
         Device::Metal(_) => assert_eq!(
             to_vec2_round(&res, 0)?,
@@ -151,7 +149,15 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
                 [-196102.0, 63022.0, 324233.0, 587191.0]
             ]
         ),
-        _ => assert_eq!(
+        Device::Cuda(_) => assert_eq!(
+            to_vec2_round(&res, 0)?,
+            &[
+                [243740.0, -19762.0, -285476.0, -550498.0],
+                [23774.0, 21645.0, 19395.0, 18364.0],
+                [-196045.0, 63030.0, 324120.0, 587079.0]
+            ]
+        ),
+        Device::Cpu => assert_eq!(
             to_vec2_round(&res, 0)?,
             &[
                 [243524.0, -19596.0, -285051.0, -549815.0],
@@ -160,22 +166,58 @@ fn quantized_matmul_neg(device: &Device) -> Result<()> {
             ]
         ),
     }
-
+    let lhs2 = Tensor::stack(&[&lhs, &lhs], 0)?;
+    let res2 = matmul.forward(&lhs2)?;
+    let res2 = res2.i(1)?;
+    let diff = (res - res2)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    if device.is_cuda() {
+        assert!(diff < 0.1);
+    } else {
+        assert_eq!(diff, 0.);
+    }
     Ok(())
 }
 
-test_device!(
-    quantized_matmul,
-    quantized_matmul_cpu,
-    quantized_matmul_cuda,
-    quantized_matmul_metal
-);
-test_device!(
-    quantized_matmul_neg,
-    quantized_matmul_neg_cpu,
-    quantized_matmul_neg_cuda,
-    quantized_matmul_neg_metal
-);
+fn qmm_batch(dev: &Device) -> Result<()> {
+    let (lhs, rhs, _mm) = get_random_tensors(2, 256, 6, dev)?;
+    let rhs = quantized::QTensor::quantize(&rhs, GgmlDType::Q2K)?;
+    let rhs = quantized::QMatMul::from_qtensor(rhs)?;
+    let mm = rhs.forward(&lhs)?;
+    assert_eq!(mm.shape().dims(), [2, 6]);
+    let lhs2 = Tensor::cat(&[&lhs, &lhs], 0)?;
+    let mm2 = rhs.forward(&lhs2)?;
+    assert_eq!(mm2.shape().dims(), [4, 6]);
+    let diff2 = (mm2.i(2..)? - &mm)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff2, 0.0);
+    let lhs3 = Tensor::cat(&[&lhs2, &lhs], 0)?;
+    let mm3 = rhs.forward(&lhs3)?;
+    assert_eq!(mm3.shape().dims(), [6, 6]);
+    let diff3 = (mm3.i(2..4)? - &mm)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff3, 0.0);
+    let diff3 = (mm3.i(4..)? - &mm)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff3, 0.0);
+    let lhs4 = Tensor::cat(&[&lhs3, &lhs3], 0)?;
+    let mm4 = rhs.forward(&lhs4)?;
+    assert_eq!(mm4.shape().dims(), [12, 6]);
+    let diff4 = (mm4.i(..6)? - &mm3)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    if dev.is_cuda() {
+        // We use a different kernel for sizes from 1 to 8 on cuda which explains
+        // the difference here.
+        assert!(0. < diff4 && diff4 < 1e-4)
+    } else {
+        assert_eq!(diff4, 0.0)
+    };
+    let diff4 = (mm4.i(6..)? - &mm4.i(..6)?)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff4, 0.0);
+    Ok(())
+}
+
+test_device!(quantized_matmul, qmm_cpu, qmm_cuda, qmm_metal);
+test_device!(quantized_matmul_neg, qmm_n_cpu, qmm_n_cuda, qmm_n_metal);
+test_device!(qmm_batch, qmm_b_cpu, qmm_b_cuda, qmm_b_metal);
 
 fn quantize_q4_0(device: &Device) -> Result<()> {
     let src = (0..32 * 4).map(|v| v as f32).collect::<Vec<_>>();
@@ -183,6 +225,13 @@ fn quantize_q4_0(device: &Device) -> Result<()> {
     let src = Tensor::from_slice(&src, (32 * 4,), device)?;
     let quant = quantized::QTensor::quantize(&src, GgmlDType::Q4_0)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     assert_eq!(
         dst.to_vec1::<f32>()?,
         &[
@@ -209,6 +258,13 @@ fn quantize_q4_1(device: &Device) -> Result<()> {
     let src = Tensor::from_slice(&src, (32 * 4,), device)?;
     let quant = quantized::QTensor::quantize(&src, GgmlDType::Q4_1)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     assert_eq!(
         round_vector(&dst.to_vec1::<f32>()?),
         &[
@@ -235,6 +291,13 @@ fn quantize_q5_0(device: &Device) -> Result<()> {
     let src = Tensor::from_slice(&src, (32 * 4,), device)?;
     let quant = quantized::QTensor::quantize(&src, GgmlDType::Q5_0)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     assert_eq!(
         round_vector(&dst.to_vec1::<f32>()?),
         &[
@@ -261,6 +324,13 @@ fn quantize_q5_1(device: &Device) -> Result<()> {
     let src = Tensor::from_slice(&src, (32 * 4,), device)?;
     let quant = quantized::QTensor::quantize(&src, GgmlDType::Q5_1)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     assert_eq!(
         round_vector(&dst.to_vec1::<f32>()?),
         &[
@@ -345,6 +415,13 @@ fn ggml_quantization_error_test(dtype: GgmlDType, device: &Device, max_error: f3
     let src = Tensor::from_slice(&src, (GGML_TEST_SIZE,), device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
     let error = calculate_rmse(&src.to_vec1::<f32>()?, &dst.to_vec1::<f32>()?);
     if error > max_error {
         bail!(
@@ -362,6 +439,13 @@ fn quantize_q2k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -381,6 +465,13 @@ fn quantize_q2k(device: &Device) -> Result<()> {
     let src_big = get_test_vector2(128.0, 1024, device)?;
     let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
     let dst_big = quant_big.dequantize(device)?;
+    let dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -395,6 +486,13 @@ fn quantize_q3k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -414,6 +512,13 @@ fn quantize_q3k(device: &Device) -> Result<()> {
     let src_big = get_test_vector2(128.0, 1024, device)?;
     let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
     let dst_big = quant_big.dequantize(device)?;
+    let dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -428,6 +533,13 @@ fn quantize_q4k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -447,6 +559,13 @@ fn quantize_q4k(device: &Device) -> Result<()> {
     let src_big = get_test_vector2(128.0, 1024, device)?;
     let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
     let dst_big = quant_big.dequantize(device)?;
+    let dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -461,6 +580,13 @@ fn quantize_q5k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -480,6 +606,13 @@ fn quantize_q5k(device: &Device) -> Result<()> {
     let src_big = get_test_vector2(128.0, 1024, device)?;
     let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
     let dst_big = quant_big.dequantize(device)?;
+    let dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -494,6 +627,13 @@ fn quantize_q6k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -513,6 +653,13 @@ fn quantize_q6k(device: &Device) -> Result<()> {
     let src_big = get_test_vector2(128.0, 1024, device)?;
     let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
     let dst_big = quant_big.dequantize(device)?;
+    let dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;
@@ -527,6 +674,13 @@ fn quantize_q8k(device: &Device) -> Result<()> {
     let src = get_test_vector2(0.5, 1024, device)?;
     let quant = quantized::QTensor::quantize(&src, dtype)?;
     let dst = quant.dequantize(device)?;
+    let dst_f16 = quant.dequantize_f16(device)?;
+    let diff = (dst.to_dtype(DType::F16)? - dst_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src = src.to_vec1::<f32>()?;
     let dst = dst.to_vec1::<f32>()?;
@@ -546,6 +700,13 @@ fn quantize_q8k(device: &Device) -> Result<()> {
     let src_big = get_test_vector2(128.0, 1024, device)?;
     let quant_big = quantized::QTensor::quantize(&src_big, dtype)?;
     let dst_big = quant_big.dequantize(device)?;
+    let dst_big_f16 = quant_big.dequantize_f16(device)?;
+    let diff = (dst_big.to_dtype(DType::F16)? - dst_big_f16)?
+        .to_dtype(DType::F32)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
 
     let src_big = src_big.to_vec1::<f32>()?;
     let dst_big = dst_big.to_vec1::<f32>()?;

candle-core/tests/serialization_tests.rs

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

candle-core/tests/tensor_tests.rs

@@ -29,6 +29,36 @@ 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]],
     );
+    assert_eq!(
+        Tensor::ones((2, 3), DType::F16, device)?.to_vec2::<half::f16>()?,
+        [
+            [
+                half::f16::from_f32(1.0),
+                half::f16::from_f32(1.0),
+                half::f16::from_f32(1.0)
+            ],
+            [
+                half::f16::from_f32(1.0),
+                half::f16::from_f32(1.0),
+                half::f16::from_f32(1.0)
+            ]
+        ],
+    );
+    assert_eq!(
+        Tensor::ones((2, 3), DType::BF16, device)?.to_vec2::<half::bf16>()?,
+        [
+            [
+                half::bf16::from_f32(1.0),
+                half::bf16::from_f32(1.0),
+                half::bf16::from_f32(1.0)
+            ],
+            [
+                half::bf16::from_f32(1.0),
+                half::bf16::from_f32(1.0),
+                half::bf16::from_f32(1.0)
+            ]
+        ],
+    );
     Ok(())
 }
 
@@ -96,6 +126,40 @@ fn clamp(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn asort(device: &Device) -> Result<()> {
+    let data = &[[3f32, 1., 4., 1.1, 5.], [2.1, 1., 7., 8., 2.]];
+    let tensor = Tensor::new(data, device)?;
+    let indexes = tensor.arg_sort_last_dim(true)?;
+    assert_eq!(
+        indexes.to_vec2::<u32>()?,
+        [[1, 3, 0, 2, 4], [1, 4, 0, 2, 3]],
+    );
+    let indexes = tensor.arg_sort_last_dim(false)?;
+    assert_eq!(
+        indexes.to_vec2::<u32>()?,
+        [[4, 2, 0, 3, 1], [3, 2, 0, 4, 1]],
+    );
+    let (sorted, indexes) = tensor.sort_last_dim(true)?;
+    assert_eq!(
+        indexes.to_vec2::<u32>()?,
+        [[1, 3, 0, 2, 4], [1, 4, 0, 2, 3]],
+    );
+    assert_eq!(
+        sorted.to_vec2::<f32>()?,
+        [[1.0, 1.1, 3.0, 4.0, 5.0], [1.0, 2.0, 2.1, 7.0, 8.0]]
+    );
+    let (sorted, indexes) = tensor.sort_last_dim(false)?;
+    assert_eq!(
+        indexes.to_vec2::<u32>()?,
+        [[4, 2, 0, 3, 1], [3, 2, 0, 4, 1]],
+    );
+    assert_eq!(
+        sorted.to_vec2::<f32>()?,
+        [[5.0, 4.0, 3.0, 1.1, 1.0], [8.0, 7.0, 2.1, 2.0, 1.0]]
+    );
+    Ok(())
+}
+
 fn unary_op(device: &Device) -> Result<()> {
     let data = &[[-3f32, 1., 4., -0.1, 0.5], [2.7, -1.8, -0.28, 1.8, 2.8]];
     let tensor = Tensor::new(data, device)?;
@@ -159,6 +223,19 @@ fn unary_op(device: &Device) -> Result<()> {
         tensor.sign()?.to_vec1::<f32>()?,
         [-1., -1., -1., 0., 0., 1., 1., 1., 1.]
     );
+    let tensor = Tensor::new(&[-1.0f32, 0., -2., 3.], device)?;
+    let y = tensor.elu(2.)?;
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [-1.2642, 0.0000, -1.7293, 3.0000]
+    );
+    // This test failed on metal prior to the following PR:
+    // https://github.com/huggingface/candle/pull/2490
+    let y = tensor.reshape((2, 2))?.t()?.elu(2.)?.flatten_all()?;
+    assert_eq!(
+        test_utils::to_vec1_round(&y, 4)?,
+        [-1.2642, -1.7293, 0.0000, 3.0000]
+    );
     Ok(())
 }
 
@@ -631,6 +708,30 @@ fn broadcast(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn slice_set(device: &Device) -> Result<()> {
+    let (b, h, max_t, d) = (2, 4, 7, 3);
+    let cache = Tensor::zeros((b, h, max_t, d), DType::F32, device)?;
+    let tensor = Tensor::randn(0f32, 1f32, (b, h, 4, d), device)?;
+    cache.slice_set(&tensor, 2, 0)?;
+    let cache_t = cache.narrow(2, 0, 4)?;
+    let diff = (cache_t - &tensor)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
+    cache.slice_set(&tensor, 2, 1)?;
+    let cache_t = cache.narrow(2, 1, 4)?;
+    let diff = (cache_t - &tensor)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
+    let ones = Tensor::ones((b, h, 1, d), DType::F32, device)?;
+    cache.slice_set(&ones, 2, 6)?;
+    let diff = cache.narrow(2, 5, 1)?.abs()?.sum_all()?.to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
+    let diff = (cache.narrow(2, 6, 1)? - 1.)?
+        .abs()?
+        .sum_all()?
+        .to_vec0::<f32>()?;
+    assert_eq!(diff, 0.);
+    Ok(())
+}
+
 fn cat(device: &Device) -> Result<()> {
     // 1D
     let t1 = Tensor::new(&[3f32, 1., 4.], device)?;
@@ -946,6 +1047,280 @@ fn gather(device: &Device) -> Result<()> {
     let ids = Tensor::new(&[[0u32, 2u32, 0u32], [0u32, 1u32, 1u32]], device)?;
     let hs = t.gather(&ids, 0)?;
     assert_eq!(hs.to_vec2::<f32>()?, &[[0.0, 7.0, 2.0], [0.0, 4.0, 5.0]]);
+
+    // Random data
+
+    // Dim: 0
+    let t = Tensor::new(
+        &[
+            [
+                [108_f32, -47., 16., -56., -83., -130., 210.],
+                [253., 95., 151., 228., -210., -123., -127.],
+                [-9., -217., 2., -78., 163., 245., -204.],
+                [-246., 79., -238., 88., -226., -184., 171.],
+                [8., -48., -153., 234., -34., 166., -153.],
+                [124., 0., -10., -61., -242., -15., -238.],
+            ],
+            [
+                [12., -64., -199., 244., -240., 156., -128.],
+                [173., -57., 4., -198., 233., -110., 238.],
+                [95., 82., 0., 240., 53., -211., 209.],
+                [-122., 167., -212., 227., -144., 61., 118.],
+                [-63., -146., 200., 244., 168., -167., 116.],
+                [-125., -147., 110., -253., -178., -250., -18.],
+            ],
+            [
+                [57., 86., -50., 56., 92., 205., -78.],
+                [-137., -156., -18., 248., -61., -239., 14.],
+                [-248., -30., -50., -70., -251., 250., -83.],
+                [-221., 67., 72., 59., -24., -154., 232.],
+                [-144., -23., -74., 5., 93., 171., 205.],
+                [46., -77., -38., -226., 246., 161., -17.],
+            ],
+            [
+                [-153., -231., -236., 161., 126., 2., -22.],
+                [-229., -41., 209., 164., 234., 160., 57.],
+                [223., 254., -186., -162., -46., -160., -102.],
+                [65., 30., 213., -253., 59., 224., -154.],
+                [-82., -203., -177., 17., 31., -256., -246.],
+                [176., -135., -65., 54., -56., 210., 76.],
+            ],
+            [
+                [-10., -245., 168., 124., -14., -33., -178.],
+                [25., -43., -39., 132., -89., 169., 179.],
+                [187., -215., 32., -133., 87., -7., -168.],
+                [-224., -215., -5., -230., -58., -162., 128.],
+                [158., -137., -122., -100., -202., -83., 136.],
+                [30., -185., -144., 250., 209., -40., 127.],
+            ],
+            [
+                [-196., 108., -245., 122., 146., -228., 62.],
+                [-1., -66., 160., 137., 13., -172., -21.],
+                [244., 199., -164., 28., 119., -175., 198.],
+                [-62., 253., -162., 195., -95., -230., -211.],
+                [123., -72., -26., -107., -139., 64., 245.],
+                [11., -126., -182., 108., -12., 184., -127.],
+            ],
+            [
+                [-159., 126., 176., 161., 73., -111., -138.],
+                [-187., 214., -217., -33., -223., -201., -212.],
+                [-61., -120., -166., -172., -95., 53., 196.],
+                [-33., 86., 134., -152., 154., -53., 74.],
+                [186., -28., -154., -174., 141., -109., 217.],
+                [82., 35., 252., 145., 181., 74., -87.],
+            ],
+        ],
+        device,
+    )?;
+
+    let ids = Tensor::new(
+        &[
+            [
+                [6_u32, 6, 4, 3, 4, 4, 6],
+                [3, 3, 2, 4, 4, 4, 6],
+                [3, 3, 0, 2, 4, 6, 4],
+                [2, 5, 1, 2, 6, 6, 1],
+                [2, 1, 6, 5, 3, 2, 3],
+                [6, 1, 0, 1, 0, 2, 6],
+            ],
+            [
+                [4, 6, 4, 3, 3, 3, 2],
+                [4, 3, 2, 4, 4, 4, 6],
+                [2, 3, 0, 2, 4, 6, 4],
+                [6, 5, 1, 2, 6, 6, 1],
+                [4, 1, 6, 5, 3, 2, 3],
+                [1, 1, 0, 1, 0, 2, 6],
+            ],
+            [
+                [3, 6, 4, 3, 3, 3, 2],
+                [2, 3, 2, 4, 4, 4, 6],
+                [4, 3, 0, 2, 4, 6, 4],
+                [0, 5, 1, 2, 6, 6, 1],
+                [6, 1, 6, 5, 3, 2, 3],
+                [4, 1, 0, 1, 0, 2, 6],
+            ],
+            [
+                [0, 6, 4, 3, 3, 3, 2],
+                [5, 3, 2, 4, 4, 4, 6],
+                [0, 3, 0, 2, 4, 6, 4],
+                [3, 5, 1, 2, 6, 6, 1],
+                [0, 1, 6, 5, 3, 2, 3],
+                [3, 1, 0, 1, 0, 2, 6],
+            ],
+        ],
+        device,
+    )?;
+
+    let hs = t.gather(&ids, 0)?;
+    assert_eq!(
+        hs.to_vec3::<f32>()?,
+        &[
+            [
+                [-159_f32, 126., 168., 161., -14., -33., -138.],
+                [-229., -41., -18., 132., -89., 169., -212.],
+                [223., 254., 2., -70., 87., 53., -168.],
+                [-221., 253., -212., 59., 154., -53., 118.],
+                [-144., -146., -154., -107., 31., 171., -246.],
+                [82., -147., -10., -253., -242., 161., -87.]
+            ],
+            [
+                [-10., 126., 168., 161., 126., 2., -78.],
+                [25., -41., -18., 132., -89., 169., -212.],
+                [-248., 254., 2., -70., 87., 53., -168.],
+                [-33., 253., -212., 59., 154., -53., 118.],
+                [158., -146., -154., -107., 31., 171., -246.],
+                [-125., -147., -10., -253., -242., 161., -87.]
+            ],
+            [
+                [-153., 126., 168., 161., 126., 2., -78.],
+                [-137., -41., -18., 132., -89., 169., -212.],
+                [187., 254., 2., -70., 87., 53., -168.],
+                [-246., 253., -212., 59., 154., -53., 118.],
+                [186., -146., -154., -107., 31., 171., -246.],
+                [30., -147., -10., -253., -242., 161., -87.]
+            ],
+            [
+                [108., 126., 168., 161., 126., 2., -78.],
+                [-1., -41., -18., 132., -89., 169., -212.],
+                [-9., 254., 2., -70., 87., 53., -168.],
+                [65., 253., -212., 59., 154., -53., 118.],
+                [8., -146., -154., -107., 31., 171., -246.],
+                [176., -147., -10., -253., -242., 161., -87.]
+            ]
+        ]
+    );
+
+    // Dim: 1
+    let t = Tensor::new(
+        &[
+            [
+                [-117_f32, -175., 69., -163.],
+                [200., 242., -21., -67.],
+                [179., 150., -126., -75.],
+                [-118., 38., -138., -13.],
+                [-221., 136., -185., 180.],
+                [58., 182., -204., -149.],
+            ],
+            [
+                [3., -148., -58., -154.],
+                [-43., 45., -108., 4.],
+                [-69., -249., -71., -21.],
+                [80., 110., -152., -235.],
+                [-88., 7., 92., -250.],
+                [-186., 207., -242., 98.],
+            ],
+            [
+                [238., 19., 64., -242.],
+                [-150., -97., 218., 58.],
+                [111., -233., 204., -212.],
+                [-242., -232., 83., 42.],
+                [153., 62., -251., 219.],
+                [-117., 36., -119., 10.],
+            ],
+            [
+                [215., 159., -169., -27.],
+                [-83., 101., -88., 169.],
+                [-205., 93., 225., -64.],
+                [-162., 240., 214., 23.],
+                [-112., 6., 21., 245.],
+                [-38., 113., 93., 215.],
+            ],
+            [
+                [91., -188., -148., 101.],
+                [74., 203., -35., 55.],
+                [-116., -130., -153., -96.],
+                [58., 22., -45., -194.],
+                [-221., -134., 73., 159.],
+                [-203., -254., 31., 235.],
+            ],
+            [
+                [105., -53., 61., 186.],
+                [-195., 234., 75., -1.],
+                [51., 139., 160., -108.],
+                [-173., -167., 161., 19.],
+                [83., -246., 156., -222.],
+                [109., 39., -149., 137.],
+            ],
+        ],
+        device,
+    )?;
+
+    let ids = Tensor::new(
+        &[
+            [[4_u32, 4, 4, 2]],
+            [[0, 4, 4, 3]],
+            [[1, 5, 3, 4]],
+            [[0, 3, 3, 2]],
+            [[1, 1, 5, 2]],
+            [[1, 4, 5, 4]],
+        ],
+        device,
+    )?;
+
+    let hs = t.gather(&ids, 1)?;
+    assert_eq!(
+        hs.to_vec3::<f32>()?,
+        &[
+            [[-221., 136., -185., -75.]],
+            [[3., 7., 92., -235.]],
+            [[-150., 36., 83., 219.]],
+            [[215., 240., 214., -64.]],
+            [[74., 203., 31., -96.]],
+            [[-195., -246., -149., -222.]]
+        ]
+    );
+
+    // Dim: 2
+    let t = Tensor::new(
+        &[
+            [[-162_f32, 202.], [-126., -39.], [35., -65.], [1., 80.]],
+            [[37., 248.], [-191., 89.], [117., -40.], [-217., 220.]],
+        ],
+        device,
+    )?;
+
+    let ids = Tensor::new(&[[[1_u32], [0], [1], [1]], [[0], [1], [0], [1]]], device)?;
+
+    let hs = t.gather(&ids, 2)?;
+    assert_eq!(
+        hs.to_vec3::<f32>()?,
+        &[
+            [[202.], [-126.], [-65.], [80.]],
+            [[37.], [89.], [117.], [220.]]
+        ]
+    );
+
+    let t = Tensor::new(
+        &[
+            [[-21_f32, -197.], [194., 122.]],
+            [[255., -106.], [-191., 250.]],
+            [[33., -117.], [43., 10.]],
+            [[-130., 238.], [-217., -92.]],
+        ],
+        device,
+    )?;
+
+    let ids = Tensor::new(
+        &[
+            [[0_u32, 1], [1, 0]],
+            [[1, 0], [0, 1]],
+            [[0, 1], [0, 1]],
+            [[1, 0], [1, 0]],
+        ],
+        device,
+    )?;
+
+    let hs = t.gather(&ids, 2)?;
+    assert_eq!(
+        hs.to_vec3::<f32>()?,
+        &[
+            [[-21., -197.], [122., 194.]],
+            [[-106., 255.], [-191., 250.]],
+            [[33., -117.], [43., 10.]],
+            [[238., -130.], [-92., -217.]]
+        ]
+    );
+
     Ok(())
 }
 
@@ -1083,6 +1458,27 @@ fn randn(device: &Device) -> Result<()> {
     Ok(())
 }
 
+fn zero_dim(device: &Device) -> Result<()> {
+    let t = Tensor::zeros((4, 0, 1), DType::F32, device)?;
+    assert_eq!(t.dims3()?, (4, 0, 1));
+    let t2 = Tensor::zeros((4, 3, 1), DType::F32, device)?;
+    let t_cat = Tensor::cat(&[&t, &t2], 1)?;
+    assert_eq!(t_cat.dims3()?, (4, 3, 1));
+    let t_cat = Tensor::cat(&[&t, &t], 1)?;
+    assert_eq!(t_cat.dims3()?, (4, 0, 1));
+    let t_unary = t.sqrt()?;
+    assert_eq!(t_unary.dims3()?, (4, 0, 1));
+    let t_plus = (&t + 1.)?;
+    assert_eq!(t_plus.dims3()?, (4, 0, 1));
+    let t_mm = t2.matmul(&t.t()?)?;
+    assert_eq!(t_mm.dims3()?, (4, 3, 0));
+    let t_mm = t.matmul(&t2.t()?)?;
+    assert_eq!(t_mm.dims3()?, (4, 0, 3));
+    let t_mm = t.t()?.matmul(&t)?;
+    assert_eq!(t_mm.dims3()?, (4, 1, 1));
+    Ok(())
+}
+
 test_device!(zeros, zeros_cpu, zeros_gpu, zeros_metal);
 test_device!(ones, ones_cpu, ones_gpu, ones_metal);
 test_device!(full, full_cpu, full_gpu, full_metal);
@@ -1091,6 +1487,7 @@ test_device!(add_mul, add_mul_cpu, add_mul_gpu, add_mul_metal);
 test_device!(tensor_2d, tensor_2d_cpu, tensor_2d_gpu, tensor_2d_metal);
 test_device!(narrow, narrow_cpu, narrow_gpu, narrow_metal);
 test_device!(broadcast, broadcast_cpu, broadcast_gpu, broadcast_metal);
+test_device!(slice_set, ss_cpu, ss_gpu, ss_metal);
 test_device!(cat, cat_cpu, cat_gpu, cat_metal);
 test_device!(sum, sum_cpu, sum_gpu, sum_metal);
 test_device!(min, min_cpu, min_gpu, min_metal);
@@ -1130,7 +1527,9 @@ test_device!(
 );
 test_device!(randn, randn_cpu, randn_gpu, randn_metal);
 test_device!(clamp, clamp_cpu, clamp_gpu, clamp_metal);
+test_device!(asort, asort_cpu, asort_gpu, asort_metal);
 test_device!(var, var_cpu, var_gpu, var_metal);
+test_device!(zero_dim, zero_dim_cpu, zero_dim_gpu, zero_dim_metal);
 
 // There was originally a bug on the CPU implementation for randn
 // https://github.com/huggingface/candle/issues/381
@@ -1244,11 +1643,29 @@ fn assert_close(a: &Tensor, b: &Tensor, epsilon: f64) -> Result<()> {
 
 #[test]
 fn log_sum_exp() -> Result<()> {
-    let input = Tensor::new(&[[1f64, 2., 3.], [4., 5., 6.]], &Device::Cpu)?;
+    let input = Tensor::new(
+        &[
+            [[1f64, 2., 3.], [4., 5., 6.]],
+            [[-1000.0, -999.0, -1001.0], [1000.0, 999.0, 1001.0]],
+        ],
+        &Device::Cpu,
+    )?;
+
     let output = input.log_sum_exp(D::Minus1)?;
     // The expectations obtained from pytorch.
-    let expected = Tensor::new(&[3.4076, 6.4076], &Device::Cpu)?;
-    assert_close(&output, &expected, 0.00001)?;
+    let expected = Tensor::new(&[[3.4076, 6.4076], [-998.5924, 1001.4076]], &Device::Cpu)?;
+    assert_eq!(output.dims(), expected.dims());
+    assert_close(&output.flatten_all()?, &expected.flatten_all()?, 0.00001)?;
+
+    assert_eq!(
+        input.log_sum_exp((0, 1))?.to_vec1::<f64>()?,
+        [1000.0, 999.0, 1001.0]
+    );
+    assert_eq!(
+        input.log_sum_exp(())?.to_vec3::<f64>()?,
+        input.to_vec3::<f64>()?
+    );
+
     Ok(())
 }
 

candle-datasets/src/vision/mnist.rs

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

candle-examples/Cargo.toml

@@ -25,7 +25,9 @@ hf-hub = { workspace = true, features = ["tokio"] }
 image = { workspace = true }
 intel-mkl-src = { workspace = true, optional = true }
 num-traits = { workspace = true }
-pyo3 = { version = "0.21.0", features = ["auto-initialize"], optional = true }
+palette = { version = "0.7.6", optional = true }
+enterpolation = { version = "0.2.1", optional = true}
+pyo3 = { version = "0.22.0", features = ["auto-initialize"], optional = true }
 rayon = { workspace = true }
 rubato = { version = "0.15.0", optional = true }
 safetensors = { workspace = true }
@@ -33,7 +35,8 @@ serde = { workspace = true }
 serde_json = { workspace = true }
 symphonia = { version = "0.5.3", features = ["all"], optional = true }
 tokenizers = { workspace = true, features = ["onig"] }
-cpal= { version = "0.15.2", optional = true }
+cpal = { version = "0.15.2", optional = true }
+pdf2image = { version = "0.1.2" , optional = true}
 
 [dev-dependencies]
 anyhow = { workspace = true }
@@ -63,8 +66,10 @@ mkl = ["dep:intel-mkl-src", "candle/mkl", "candle-nn/mkl", "candle-transformers/
 nccl = ["cuda", "cudarc/nccl", "dep:half"]
 onnx = ["candle-onnx"]
 metal = ["candle/metal", "candle-nn/metal"]
-microphone = ["cpal"]
+microphone = ["cpal", "rubato"]
 encodec = ["cpal", "symphonia", "rubato"]
+mimi = ["cpal", "symphonia", "rubato"]
+depth_anything_v2 = ["palette", "enterpolation"]
 
 [[example]]
 name = "llama_multiprocess"
@@ -98,6 +103,22 @@ required-features = ["candle-datasets"]
 name = "llama2-c"
 required-features = ["candle-datasets"]
 
+[[example]]
+name = "mimi"
+required-features = ["mimi"]
+
 [[example]]
 name = "encodec"
 required-features = ["encodec"]
+
+[[example]]
+name = "depth_anything_v2"
+required-features = ["depth_anything_v2"]
+
+[[example]]
+name = "silero-vad"
+required-features = ["onnx"]
+
+[[example]]
+name = "colpali"
+required-features = ["pdf2image"]

candle-examples/examples/based/README.md

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

candle-examples/examples/based/main.rs

@@ -0,0 +1,275 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::{Error as E, Result};
+use clap::{Parser, ValueEnum};
+
+use candle_transformers::models::based::Model;
+
+use candle::{DType, Device, Tensor};
+use candle_examples::token_output_stream::TokenOutputStream;
+use candle_nn::VarBuilder;
+use candle_transformers::generation::LogitsProcessor;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::Tokenizer;
+
+struct TextGeneration {
+    model: Model,
+    device: Device,
+    tokenizer: TokenOutputStream,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+}
+
+impl TextGeneration {
+    #[allow(clippy::too_many_arguments)]
+    fn new(
+        model: Model,
+        tokenizer: Tokenizer,
+        seed: u64,
+        temp: Option<f64>,
+        top_p: Option<f64>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        device: &Device,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            tokenizer: TokenOutputStream::new(tokenizer),
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            device: device.clone(),
+        }
+    }
+
+    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
+        use std::io::Write;
+        self.tokenizer.clear();
+        let mut tokens = self
+            .tokenizer
+            .tokenizer()
+            .encode(prompt, true)
+            .map_err(E::msg)?
+            .get_ids()
+            .to_vec();
+        for &t in tokens.iter() {
+            if let Some(t) = self.tokenizer.next_token(t)? {
+                print!("{t}")
+            }
+        }
+        std::io::stdout().flush()?;
+
+        let mut generated_tokens = 0usize;
+        let eos_token = match self.tokenizer.get_token("<|endoftext|>") {
+            Some(token) => token,
+            None => anyhow::bail!("cannot find the <|endoftext|> token"),
+        };
+        let start_gen = std::time::Instant::now();
+        for index in 0..sample_len {
+            let context_size = if index > 0 { 1 } else { tokens.len() };
+            let start_pos = tokens.len().saturating_sub(context_size);
+            let ctxt = &tokens[start_pos..];
+            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+            let logits = self.model.forward(&input, start_pos)?;
+            let logits = logits.squeeze(0)?.squeeze(0)?.to_dtype(DType::F32)?;
+            let logits = if self.repeat_penalty == 1. {
+                logits
+            } else {
+                let start_at = tokens.len().saturating_sub(self.repeat_last_n);
+                candle_transformers::utils::apply_repeat_penalty(
+                    &logits,
+                    self.repeat_penalty,
+                    &tokens[start_at..],
+                )?
+            };
+
+            let next_token = self.logits_processor.sample(&logits)?;
+            tokens.push(next_token);
+            generated_tokens += 1;
+            if next_token == eos_token {
+                break;
+            }
+            if let Some(t) = self.tokenizer.next_token(next_token)? {
+                print!("{t}");
+                std::io::stdout().flush()?;
+            }
+        }
+        let dt = start_gen.elapsed();
+        if let Some(rest) = self.tokenizer.decode_rest().map_err(E::msg)? {
+            print!("{rest}");
+        }
+        std::io::stdout().flush()?;
+        println!(
+            "\n{generated_tokens} tokens generated ({:.2} token/s)",
+            generated_tokens as f64 / dt.as_secs_f64(),
+        );
+        Ok(())
+    }
+}
+
+#[derive(Clone, Debug, Copy, PartialEq, Eq, ValueEnum)]
+enum Which {
+    #[value(name = "360m")]
+    W360m,
+    #[value(name = "1b")]
+    W1b,
+    #[value(name = "1b-50b")]
+    W1b50b,
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    #[arg(long)]
+    prompt: String,
+
+    /// The temperature used to generate samples.
+    #[arg(long)]
+    temperature: Option<f64>,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// The seed to use when generating random samples.
+    #[arg(long, default_value_t = 299792458)]
+    seed: u64,
+
+    /// The length of the sample to generate (in tokens).
+    #[arg(long, short = 'n', default_value_t = 10000)]
+    sample_len: usize,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long, default_value = "refs/pr/1")]
+    revision: String,
+
+    #[arg(long)]
+    config_file: Option<String>,
+
+    #[arg(long)]
+    tokenizer_file: Option<String>,
+
+    #[arg(long)]
+    weight_files: Option<String>,
+
+    /// Penalty to be applied for repeating tokens, 1. means no penalty.
+    #[arg(long, default_value_t = 1.1)]
+    repeat_penalty: f32,
+
+    /// The context size to consider for the repeat penalty.
+    #[arg(long, default_value_t = 64)]
+    repeat_last_n: usize,
+
+    #[arg(long, default_value = "360m")]
+    which: Which,
+}
+
+fn main() -> Result<()> {
+    use tracing_chrome::ChromeLayerBuilder;
+    use tracing_subscriber::prelude::*;
+
+    let args = Args::parse();
+    let _guard = if args.tracing {
+        let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
+        tracing_subscriber::registry().with(chrome_layer).init();
+        Some(guard)
+    } else {
+        None
+    };
+    println!(
+        "avx: {}, neon: {}, simd128: {}, f16c: {}",
+        candle::utils::with_avx(),
+        candle::utils::with_neon(),
+        candle::utils::with_simd128(),
+        candle::utils::with_f16c()
+    );
+    println!(
+        "temp: {:.2} repeat-penalty: {:.2} repeat-last-n: {}",
+        args.temperature.unwrap_or(0.),
+        args.repeat_penalty,
+        args.repeat_last_n
+    );
+
+    let start = std::time::Instant::now();
+    let api = Api::new()?;
+    let model_id = match args.model_id {
+        Some(model_id) => model_id,
+        None => match args.which {
+            Which::W360m => "hazyresearch/based-360m".to_string(),
+            Which::W1b => "hazyresearch/based-1b".to_string(),
+            Which::W1b50b => "hazyresearch/based-1b-50b".to_string(),
+        },
+    };
+    let repo = api.repo(Repo::with_revision(
+        model_id,
+        RepoType::Model,
+        args.revision,
+    ));
+    let config_file = match args.config_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("config.json")?,
+    };
+    let filenames = match args.weight_files {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        None => vec![repo.get("model.safetensors")?],
+    };
+
+    let repo = api.model("openai-community/gpt2".to_string());
+    let tokenizer_file = match args.tokenizer_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("tokenizer.json")?,
+    };
+
+    println!("retrieved the files in {:?}", start.elapsed());
+    let tokenizer = Tokenizer::from_file(tokenizer_file).map_err(E::msg)?;
+
+    let start = std::time::Instant::now();
+    let config = serde_json::from_reader(std::fs::File::open(config_file)?)?;
+    let device = candle_examples::device(args.cpu)?;
+    let dtype = if device.is_cuda() {
+        DType::BF16
+    } else {
+        DType::F32
+    };
+
+    let mut vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
+    if args.which == Which::W1b50b {
+        vb = vb.pp("model");
+    };
+
+    let model = Model::new(&config, vb)?;
+
+    println!("loaded the model in {:?}", start.elapsed());
+
+    let mut pipeline = TextGeneration::new(
+        model,
+        tokenizer,
+        args.seed,
+        args.temperature,
+        args.top_p,
+        args.repeat_penalty,
+        args.repeat_last_n,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}

candle-examples/examples/beit/README.md

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

candle-examples/examples/beit/main.rs

@@ -0,0 +1,79 @@
+//! BEiT: BERT Pre-Training of Image Transformers
+//! https://github.com/microsoft/unilm/tree/master/beit
+
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use clap::Parser;
+
+use candle::{DType, Device, IndexOp, Result, Tensor, D};
+use candle_nn::{Module, VarBuilder};
+use candle_transformers::models::beit;
+
+/// Loads an image from disk using the image crate, this returns a tensor with shape
+/// (3, 384, 384). Beit special normalization is applied.
+pub fn load_image384_beit_norm<P: AsRef<std::path::Path>>(p: P) -> Result<Tensor> {
+    let img = image::ImageReader::open(p)?
+        .decode()
+        .map_err(candle::Error::wrap)?
+        .resize_to_fill(384, 384, image::imageops::FilterType::Triangle);
+    let img = img.to_rgb8();
+    let data = img.into_raw();
+    let data = Tensor::from_vec(data, (384, 384, 3), &Device::Cpu)?.permute((2, 0, 1))?;
+    let mean = Tensor::new(&[0.5f32, 0.5, 0.5], &Device::Cpu)?.reshape((3, 1, 1))?;
+    let std = Tensor::new(&[0.5f32, 0.5, 0.5], &Device::Cpu)?.reshape((3, 1, 1))?;
+    (data.to_dtype(candle::DType::F32)? / 255.)?
+        .broadcast_sub(&mean)?
+        .broadcast_div(&std)
+}
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    image: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    let device = candle_examples::device(args.cpu)?;
+
+    let image = load_image384_beit_norm(args.image)?.to_device(&device)?;
+    println!("loaded image {image:?}");
+
+    let model_file = match args.model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model("vincent-espitalier/candle-beit".into());
+            api.get("beit_base_patch16_384.in22k_ft_in22k_in1k.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+    let model = beit::vit_base(vb)?;
+    println!("model built");
+    let logits = model.forward(&image.unsqueeze(0)?)?;
+    let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
+        .i(0)?
+        .to_vec1::<f32>()?;
+    let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
+    prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
+    for &(category_idx, pr) in prs.iter().take(5) {
+        println!(
+            "{:24}: {:.2}%",
+            candle_examples::imagenet::CLASSES[category_idx],
+            100. * pr
+        );
+    }
+    Ok(())
+}

candle-examples/examples/bert/main.rs

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

candle-examples/examples/blip/main.rs

@@ -55,7 +55,7 @@ const SEP_TOKEN_ID: u32 = 102;
 /// Loads an image from disk using the image crate, this returns a tensor with shape
 /// (3, 384, 384). OpenAI normalization is applied.
 pub fn load_image<P: AsRef<std::path::Path>>(p: P) -> Result<Tensor> {
-    let img = image::io::Reader::open(p)?
+    let img = image::ImageReader::open(p)?
         .decode()
         .map_err(candle::Error::wrap)?
         .resize_to_fill(384, 384, image::imageops::FilterType::Triangle);

candle-examples/examples/chinese_clip/main.rs

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

candle-examples/examples/clip/README.md

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

candle-examples/examples/clip/main.rs

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

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

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

candle-examples/examples/codegeex4-9b/main.rs

@@ -0,0 +1,252 @@
+use candle_transformers::models::codegeex4_9b::*;
+use clap::Parser;
+
+use candle::{DType, Device, Tensor};
+use candle_nn::VarBuilder;
+use candle_transformers::generation::LogitsProcessor;
+use hf_hub::{Repo, RepoType};
+use tokenizers::Tokenizer;
+
+struct TextGeneration {
+    model: Model,
+    device: Device,
+    tokenizer: Tokenizer,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+    verbose_prompt: bool,
+    dtype: DType,
+}
+
+impl TextGeneration {
+    #[allow(clippy::too_many_arguments)]
+    fn new(
+        model: Model,
+        tokenizer: Tokenizer,
+        seed: u64,
+        temp: Option<f64>,
+        top_p: Option<f64>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        verbose_prompt: bool,
+        device: &Device,
+        dtype: DType,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            tokenizer,
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            verbose_prompt,
+            device: device.clone(),
+            dtype,
+        }
+    }
+
+    fn run(&mut self, prompt: &str, sample_len: usize) -> anyhow::Result<()> {
+        use std::io::Write;
+        println!("starting the inference loop");
+        let tokens = self.tokenizer.encode(prompt, true).expect("tokens error");
+        if tokens.is_empty() {
+            panic!("Empty prompts are not supported in the chatglm model.")
+        }
+        if self.verbose_prompt {
+            for (token, id) in tokens.get_tokens().iter().zip(tokens.get_ids().iter()) {
+                let token = token.replace('▁', " ").replace("<0x0A>", "\n");
+                println!("{id:7} -> '{token}'");
+            }
+        }
+        let eos_token = match self.tokenizer.get_vocab(true).get("<|endoftext|>") {
+            Some(token) => *token,
+            None => panic!("cannot find the endoftext token"),
+        };
+        let mut tokens = tokens.get_ids().to_vec();
+        let mut generated_tokens = 0usize;
+
+        print!("{prompt}");
+        std::io::stdout().flush().expect("output flush error");
+        let start_gen = std::time::Instant::now();
+
+        println!("\n start_gen");
+        println!("samplelen {}", sample_len);
+        let mut count = 0;
+        let mut result = vec![];
+        for index in 0..sample_len {
+            count += 1;
+            let context_size = if index > 0 { 1 } else { tokens.len() };
+            let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
+            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+            let logits = self.model.forward(&input)?;
+            let logits = logits.squeeze(0)?.to_dtype(self.dtype)?;
+            let logits = if self.repeat_penalty == 1. {
+                logits
+            } else {
+                let start_at = tokens.len().saturating_sub(self.repeat_last_n);
+                candle_transformers::utils::apply_repeat_penalty(
+                    &logits,
+                    self.repeat_penalty,
+                    &tokens[start_at..],
+                )?
+            };
+
+            let next_token = self.logits_processor.sample(&logits)?;
+            tokens.push(next_token);
+            generated_tokens += 1;
+            if next_token == eos_token {
+                break;
+            }
+            let token = self
+                .tokenizer
+                .decode(&[next_token], true)
+                .expect("Token error");
+            if self.verbose_prompt {
+                println!(
+                    "[Count: {}] [Raw Token: {}] [Decode Token: {}]",
+                    count, next_token, token
+                );
+            }
+            result.push(token);
+            std::io::stdout().flush()?;
+        }
+        let dt = start_gen.elapsed();
+        println!(
+            "\n{generated_tokens} tokens generated ({:.2} token/s)",
+            generated_tokens as f64 / dt.as_secs_f64(),
+        );
+        println!("Result:");
+        for tokens in result {
+            print!("{tokens}");
+        }
+        Ok(())
+    }
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(name = "cache", short, long, default_value = ".")]
+    cache_path: String,
+
+    #[arg(long)]
+    cpu: bool,
+
+    /// Display the token for the specified prompt.
+    #[arg(long)]
+    verbose_prompt: bool,
+
+    #[arg(long)]
+    prompt: String,
+
+    /// The temperature used to generate samples.
+    #[arg(long)]
+    temperature: Option<f64>,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// The seed to use when generating random samples.
+    #[arg(long, default_value_t = 299792458)]
+    seed: u64,
+
+    /// The length of the sample to generate (in tokens).
+    #[arg(long, short = 'n', default_value_t = 5000)]
+    sample_len: usize,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long)]
+    revision: Option<String>,
+
+    #[arg(long)]
+    weight_file: Option<String>,
+
+    #[arg(long)]
+    tokenizer: Option<String>,
+
+    /// Penalty to be applied for repeating tokens, 1. means no penalty.
+    #[arg(long, default_value_t = 1.1)]
+    repeat_penalty: f32,
+
+    /// The context size to consider for the repeat penalty.
+    #[arg(long, default_value_t = 64)]
+    repeat_last_n: usize,
+}
+
+fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+    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(0.95),
+        args.repeat_penalty,
+        args.repeat_last_n
+    );
+
+    let start = std::time::Instant::now();
+    println!("cache path {}", args.cache_path);
+    let api = hf_hub::api::sync::ApiBuilder::from_cache(hf_hub::Cache::new(args.cache_path.into()))
+        .build()
+        .map_err(anyhow::Error::msg)?;
+
+    let model_id = match args.model_id {
+        Some(model_id) => model_id.to_string(),
+        None => "THUDM/codegeex4-all-9b".to_string(),
+    };
+    let revision = match args.revision {
+        Some(rev) => rev.to_string(),
+        None => "main".to_string(),
+    };
+    let repo = api.repo(Repo::with_revision(model_id, RepoType::Model, revision));
+    let tokenizer_filename = match args.tokenizer {
+        Some(file) => std::path::PathBuf::from(file),
+        None => api
+            .model("THUDM/codegeex4-all-9b".to_string())
+            .get("tokenizer.json")
+            .map_err(anyhow::Error::msg)?,
+    };
+    let filenames = match args.weight_file {
+        Some(weight_file) => vec![std::path::PathBuf::from(weight_file)],
+        None => candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.json")?,
+    };
+    println!("retrieved the files in {:?}", start.elapsed());
+    let tokenizer = Tokenizer::from_file(tokenizer_filename).expect("Tokenizer Error");
+
+    let start = std::time::Instant::now();
+    let config = Config::codegeex4();
+    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,
+        args.verbose_prompt,
+        &device,
+        dtype,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}

candle-examples/examples/colpali/README.md

@@ -0,0 +1,18 @@
+# Colpali
+
+[HuggingFace Model Card](https://huggingface.co/vidore/colpali-v1.2-merged)
+
+```
+wget https://arxiv.org/pdf/1706.03762.pdf
+cargo run --features cuda,pdf2image --release --example colpali -- --prompt "What is Positional Encoding" --pdf "1706.03762.pdf"
+```
+
+```
+Prompt: what is position encoding?
+top 3 page numbers that contain similarity to the prompt
+-----------------------------------
+Page: 6
+Page: 11
+Page: 15
+-----------------------------------
+```

candle-examples/examples/colpali/main.rs

@@ -0,0 +1,268 @@
+use anyhow::{Error as E, Result};
+use candle::{DType, Device, Tensor};
+use candle_nn::VarBuilder;
+use candle_transformers::models::colpali::Model;
+use candle_transformers::models::{colpali, paligemma};
+use clap::Parser;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use image::DynamicImage;
+use pdf2image::{RenderOptionsBuilder, PDF};
+use tokenizers::Tokenizer;
+
+struct PageRetriever {
+    model: Model,
+    config: paligemma::Config,
+    pdf: PDF,
+    device: Device,
+    tokenizer: Tokenizer,
+    range: pdf2image::Pages,
+    batch_size: usize,
+    top_k: usize,
+}
+
+impl PageRetriever {
+    fn new(
+        model: Model,
+        config: paligemma::Config,
+        pdf: PDF,
+        tokenizer: Tokenizer,
+        device: &Device,
+        range: Option<pdf2image::Pages>,
+        batch_size: usize,
+        top_k: usize,
+    ) -> Self {
+        let page_count = pdf.page_count();
+        Self {
+            model,
+            config,
+            pdf,
+            device: device.clone(),
+            tokenizer,
+            range: range.unwrap_or_else(|| pdf2image::Pages::Range(1..=page_count)),
+            batch_size,
+            top_k,
+        }
+    }
+
+    fn get_images_from_pdf(&self) -> Result<Vec<DynamicImage>> {
+        let pages = self
+            .pdf
+            .render(self.range.clone(), RenderOptionsBuilder::default().build()?)?;
+        Ok(pages)
+    }
+
+    fn tokenize_batch(&self, prompts: Vec<&str>) -> Result<Tensor> {
+        let tokens = self.tokenizer.encode_batch(prompts, true).map_err(E::msg)?;
+        let token_ids = tokens
+            .iter()
+            .map(|tokens| {
+                let tokens = tokens.get_ids().to_vec();
+                Tensor::new(tokens.as_slice(), &self.device)
+            })
+            .collect::<candle::Result<Vec<_>>>()?;
+        let input = Tensor::stack(&token_ids, 0)?;
+        Ok(input)
+    }
+
+    fn images_to_tensor(
+        &self,
+        pages: &[DynamicImage],
+        image_size: usize,
+    ) -> anyhow::Result<Tensor> {
+        let mut images = vec![];
+        for page in pages.iter() {
+            let img = page.resize_to_fill(
+                image_size as u32,
+                image_size as u32,
+                image::imageops::FilterType::Triangle,
+            );
+            let img = img.to_rgb8();
+            let img = img.into_raw();
+            let img = Tensor::from_vec(img, (image_size, image_size, 3), &Device::Cpu)?
+                .permute((2, 0, 1))?
+                .to_dtype(DType::F32)?
+                .affine(2. / 255., -1.)?;
+            images.push(img);
+        }
+        let images = Tensor::stack(&images, 0)?;
+        Ok(images)
+    }
+
+    fn retrieve(&mut self, prompt: &str) -> Result<Vec<usize>> {
+        let dtype = if self.device.is_cuda() {
+            DType::BF16
+        } else {
+            DType::F32
+        };
+
+        let dummy_prompt: &str = "Describe the image";
+
+        let input = self.tokenize_batch(vec![prompt])?;
+        let dummy_input = self.tokenize_batch(vec![dummy_prompt])?;
+
+        let pages = self.get_images_from_pdf()?;
+        let mut all_scores = Vec::new();
+        for batch in pages.chunks(self.batch_size) {
+            let page_images = self
+                .images_to_tensor(batch, self.config.vision_config.image_size)?
+                .to_device(&self.device)?
+                .to_dtype(dtype)?;
+            let dummy_input = dummy_input.repeat((page_images.dims()[0], 0))?;
+
+            let image_embeddings = self.model.forward_images(&page_images, &dummy_input)?;
+            let text_embeddings = self.model.forward_text(&input)?;
+
+            let scores = text_embeddings
+                .unsqueeze(1)?
+                .broadcast_matmul(&image_embeddings.unsqueeze(0)?.transpose(3, 2)?)?
+                .max(3)?
+                .sum(2)?;
+            let batch_scores: Vec<f32> = scores
+                .to_dtype(DType::F32)?
+                .to_vec2()?
+                .into_iter()
+                .flatten()
+                .collect();
+            all_scores.extend(batch_scores);
+        }
+
+        let mut indices: Vec<usize> = (0..all_scores.len()).collect();
+        indices.sort_by(|a, b| all_scores[*b].partial_cmp(&all_scores[*a]).unwrap());
+
+        let top_k_indices = indices[0..self.top_k].to_vec();
+
+        Ok(top_k_indices)
+    }
+}
+
+#[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,
+
+    /// number of top pages to show.
+    #[arg(long, default_value_t = 3)]
+    top_k: usize,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long, default_value = "main")]
+    revision: String,
+
+    #[arg(long)]
+    tokenizer_file: Option<String>,
+
+    #[arg(long)]
+    weight_files: Option<String>,
+
+    #[arg(long)]
+    pdf: String,
+
+    #[arg(long)]
+    start: Option<u32>,
+
+    #[arg(long)]
+    end: Option<u32>,
+}
+
+fn main() -> Result<()> {
+    use tracing_chrome::ChromeLayerBuilder;
+    use tracing_subscriber::prelude::*;
+
+    let args = Args::parse();
+    let _guard = if args.tracing {
+        let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
+        tracing_subscriber::registry().with(chrome_layer).init();
+        Some(guard)
+    } else {
+        None
+    };
+    println!(
+        "avx: {}, neon: {}, simd128: {}, f16c: {}",
+        candle::utils::with_avx(),
+        candle::utils::with_neon(),
+        candle::utils::with_simd128(),
+        candle::utils::with_f16c()
+    );
+
+    let api = Api::new()?;
+    let model_id = match &args.model_id {
+        Some(model_id) => model_id.to_string(),
+        None => "vidore/colpali-v1.2-merged".to_string(),
+    };
+    let repo = api.repo(Repo::with_revision(
+        model_id,
+        RepoType::Model,
+        args.revision,
+    ));
+
+    let tokenizer_filename = match args.tokenizer_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => api
+            .repo(Repo::with_revision(
+                "vidore/colpali".to_string(),
+                RepoType::Model,
+                "main".to_string(),
+            ))
+            .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")?,
+    };
+
+    let start = std::time::Instant::now();
+
+    let config: paligemma::Config = paligemma::Config::paligemma_3b_448();
+
+    println!("retrieved the files in {:?}", start.elapsed());
+
+    let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
+    let device = candle_examples::device(false)?;
+    let dtype = if device.is_cuda() {
+        DType::BF16
+    } else {
+        DType::F32
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
+    let model = colpali::Model::new(&config, vb)?;
+
+    let pdf = PDF::from_file(args.pdf)?;
+
+    // check if start and end given in arg
+    let range = if let (Some(start), Some(end)) = (args.start, args.end) {
+        pdf2image::Pages::Range(start..=end)
+    } else {
+        pdf2image::Pages::Range(1..=pdf.page_count()) // can use pdf2image::Pages::All but there is a bug in the library which causes the first page to rendered twice.
+    };
+
+    let mut retriever =
+        PageRetriever::new(model, config, pdf, tokenizer, &device, Some(range), 4, 3);
+    let top_k_indices = retriever.retrieve(&args.prompt)?;
+
+    println!("Prompt: {}", args.prompt);
+    println!(
+        "top {} page numbers that contain similarity to the prompt",
+        retriever.top_k
+    );
+    println!("-----------------------------------");
+    for index in top_k_indices {
+        println!("Page: {:?}", index + 1);
+    }
+    println!("-----------------------------------");
+    Ok(())
+}

candle-examples/examples/depth_anything_v2/README.md

@@ -0,0 +1,13 @@
+# candle-dinov2
+
+[Depth Anything V2] is a model for Monocular Depth Estimation (MDE, i.e. just using a single image) which
+builds on the [DINOv2](https://github.com/facebookresearch/dinov2) vision transformer.
+
+This example first instantiates the DINOv2 model and then proceeds to create DepthAnythingV2 and run it.
+
+## Running an example with color map and CUDA
+
+```bash
+cargo run --features cuda,depth_anything_v2 --package candle-examples --example depth_anything_v2 -- --color-map --image candle-examples/examples/yolo-v8/assets/bike.jpg 
+```
+

candle-examples/examples/depth_anything_v2/color_map.rs

@@ -0,0 +1,50 @@
+use enterpolation::linear::ConstEquidistantLinear;
+use enterpolation::Generator;
+use palette::LinSrgb;
+
+use candle::Tensor;
+
+pub struct SpectralRColormap {
+    gradient: ConstEquidistantLinear<f32, LinSrgb, 9>,
+}
+
+impl SpectralRColormap {
+    pub(crate) fn new() -> Self {
+        // Define a colormap similar to 'Spectral_r' by specifying key colors.
+        // got the colors from ChatGPT-4o
+        let gradient = ConstEquidistantLinear::<f32, _, 9>::equidistant_unchecked([
+            LinSrgb::new(0.3686, 0.3098, 0.6353), // Dark blue
+            LinSrgb::new(0.1961, 0.5333, 0.7412), // Blue
+            LinSrgb::new(0.4000, 0.7608, 0.6471), // Cyan
+            LinSrgb::new(0.6706, 0.8667, 0.6431), // Green
+            LinSrgb::new(0.9020, 0.9608, 0.5961), // Yellow
+            LinSrgb::new(0.9961, 0.8784, 0.5451), // Orange
+            LinSrgb::new(0.9922, 0.6824, 0.3804), // Red
+            LinSrgb::new(0.9569, 0.4275, 0.2627), // Dark red
+            LinSrgb::new(0.8353, 0.2431, 0.3098), // Dark purple
+        ]);
+        Self { gradient }
+    }
+
+    fn get_color(&self, value: f32) -> LinSrgb {
+        self.gradient.gen(value)
+    }
+
+    pub fn gray2color(&self, gray: &Tensor) -> candle::Result<Tensor> {
+        println!("Gray: {:?}", gray.dims());
+        let gray_values: Vec<f32> = gray.flatten_all()?.to_vec1()?;
+        let rgb_values: Vec<f32> = gray_values
+            .iter()
+            .map(|g| self.get_color(*g))
+            .flat_map(|rgb| [rgb.red, rgb.green, rgb.blue])
+            .collect();
+
+        let [.., height, width] = gray.dims() else {
+            candle::bail!("Not enough dims!")
+        };
+
+        let color = Tensor::from_vec(rgb_values, (*height, *width, 3), gray.device())?;
+
+        color.permute((2, 0, 1))
+    }
+}

candle-examples/examples/depth_anything_v2/main.rs

@@ -0,0 +1,187 @@
+//! Depth Anything V2
+//! https://huggingface.co/spaces/depth-anything/Depth-Anything-V2
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+use std::ffi::OsString;
+use std::path::PathBuf;
+
+use clap::Parser;
+
+use candle::DType::{F32, U8};
+use candle::{DType, Device, Module, Result, Tensor};
+use candle_examples::{load_image, load_image_and_resize, save_image};
+use candle_nn::VarBuilder;
+use candle_transformers::models::depth_anything_v2::{DepthAnythingV2, DepthAnythingV2Config};
+use candle_transformers::models::dinov2;
+
+use crate::color_map::SpectralRColormap;
+
+mod color_map;
+
+// taken these from: https://huggingface.co/spaces/depth-anything/Depth-Anything-V2/blob/main/depth_anything_v2/dpt.py#L207
+const MAGIC_MEAN: [f32; 3] = [0.485, 0.456, 0.406];
+const MAGIC_STD: [f32; 3] = [0.229, 0.224, 0.225];
+
+const DINO_IMG_SIZE: usize = 518;
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    dinov2_model: Option<PathBuf>,
+
+    #[arg(long)]
+    depth_anything_v2_model: Option<PathBuf>,
+
+    #[arg(long)]
+    image: PathBuf,
+
+    #[arg(long)]
+    output_dir: Option<PathBuf>,
+
+    #[arg(long)]
+    cpu: bool,
+
+    #[arg(long)]
+    color_map: bool,
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+    let device = candle_examples::device(args.cpu)?;
+
+    let dinov2_model_file = match args.dinov2_model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model("lmz/candle-dino-v2".into());
+            api.get("dinov2_vits14.safetensors")?
+        }
+        Some(dinov2_model) => dinov2_model,
+    };
+    println!("Using file {:?}", dinov2_model_file);
+
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[dinov2_model_file], F32, &device)? };
+    let dinov2 = dinov2::vit_small(vb)?;
+    println!("DinoV2 model built");
+
+    let depth_anything_model_file = match args.depth_anything_v2_model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model("jeroenvlek/depth-anything-v2-safetensors".into());
+            api.get("depth_anything_v2_vits.safetensors")?
+        }
+        Some(depth_anything_model) => depth_anything_model,
+    };
+    println!("Using file {:?}", depth_anything_model_file);
+
+    let vb = unsafe {
+        VarBuilder::from_mmaped_safetensors(&[depth_anything_model_file], DType::F32, &device)?
+    };
+
+    let config = DepthAnythingV2Config::vit_small();
+    let depth_anything = DepthAnythingV2::new(&dinov2, &config, vb)?;
+
+    let (original_height, original_width, image) = load_and_prep_image(&args.image, &device)?;
+
+    println!("Loaded image {image:?}");
+
+    let depth = depth_anything.forward(&image)?;
+
+    println!("Got predictions {:?}", depth.shape());
+
+    let output_image = post_process_image(&depth, original_height, original_width, args.color_map)?;
+
+    let output_path = full_output_path(&args.image, &args.output_dir);
+    println!("Saving image to {}", output_path.to_string_lossy());
+    save_image(&output_image, output_path)?;
+
+    Ok(())
+}
+
+fn full_output_path(image_path: &PathBuf, output_dir: &Option<PathBuf>) -> PathBuf {
+    let input_file_name = image_path.file_name().unwrap();
+    let mut output_file_name = OsString::from("depth_");
+    output_file_name.push(input_file_name);
+    let mut output_path = match output_dir {
+        None => image_path.parent().unwrap().to_path_buf(),
+        Some(output_path) => output_path.clone(),
+    };
+    output_path.push(output_file_name);
+
+    output_path
+}
+
+fn load_and_prep_image(
+    image_path: &PathBuf,
+    device: &Device,
+) -> anyhow::Result<(usize, usize, Tensor)> {
+    let (_original_image, original_height, original_width) = load_image(&image_path, None)?;
+
+    let image = load_image_and_resize(&image_path, DINO_IMG_SIZE, DINO_IMG_SIZE)?
+        .unsqueeze(0)?
+        .to_dtype(F32)?
+        .to_device(&device)?;
+
+    let max_pixel_val = Tensor::try_from(255.0f32)?
+        .to_device(&device)?
+        .broadcast_as(image.shape())?;
+    let image = (image / max_pixel_val)?;
+    let image = normalize_image(&image, &MAGIC_MEAN, &MAGIC_STD)?;
+
+    Ok((original_height, original_width, image))
+}
+
+fn normalize_image(image: &Tensor, mean: &[f32; 3], std: &[f32; 3]) -> Result<Tensor> {
+    let mean_tensor =
+        Tensor::from_vec(mean.to_vec(), (3, 1, 1), &image.device())?.broadcast_as(image.shape())?;
+    let std_tensor =
+        Tensor::from_vec(std.to_vec(), (3, 1, 1), &image.device())?.broadcast_as(image.shape())?;
+    image.sub(&mean_tensor)?.div(&std_tensor)
+}
+
+fn post_process_image(
+    image: &Tensor,
+    original_height: usize,
+    original_width: usize,
+    color_map: bool,
+) -> Result<Tensor> {
+    let out = image.interpolate2d(original_height, original_width)?;
+    let out = scale_image(&out)?;
+
+    let out = if color_map {
+        let spectral_r = SpectralRColormap::new();
+        spectral_r.gray2color(&out)?
+    } else {
+        let rgb_slice = [&out, &out, &out];
+        Tensor::cat(&rgb_slice, 0)?.squeeze(1)?
+    };
+
+    let max_pixel_val = Tensor::try_from(255.0f32)?
+        .to_device(out.device())?
+        .broadcast_as(out.shape())?;
+    let out = (out * max_pixel_val)?;
+
+    out.to_dtype(U8)
+}
+
+fn scale_image(depth: &Tensor) -> Result<Tensor> {
+    let flat_values: Vec<f32> = depth.flatten_all()?.to_vec1()?;
+
+    let min_val = flat_values.iter().min_by(|a, b| a.total_cmp(b)).unwrap();
+    let max_val = flat_values.iter().max_by(|a, b| a.total_cmp(b)).unwrap();
+
+    let min_val_tensor = Tensor::try_from(*min_val)?
+        .to_device(depth.device())?
+        .broadcast_as(depth.shape())?;
+    let depth = (depth - min_val_tensor)?;
+
+    let range = max_val - min_val;
+    let range_tensor = Tensor::try_from(range)?
+        .to_device(depth.device())?
+        .broadcast_as(depth.shape())?;
+
+    depth / range_tensor
+}

candle-examples/examples/dinov2reg4/README.md

@@ -0,0 +1,25 @@
+# candle-dinov2-reg4
+
+[DINOv2-reg4](https://arxiv.org/abs/2309.16588) is the lastest version of DINOv2 with registers.
+In this example, it is used as an plant species classifier: the model returns the
+probability for the image to belong to each of the 7806 PlantCLEF2024 categories.
+
+## Running some example
+
+```bash
+# Download classes names and a plant picture to identify
+curl https://huggingface.co/vincent-espitalier/dino-v2-reg4-with-plantclef2024-weights/raw/main/species_id_mapping.txt --output candle-examples/examples/dinov2reg4/species_id_mapping.txt
+curl https://bs.plantnet.org/image/o/bd2d3830ac3270218ba82fd24e2290becd01317c --output candle-examples/examples/dinov2reg4/bd2d3830ac3270218ba82fd24e2290becd01317c.jpg
+
+# Perform inference
+cargo run --example dinov2reg4 --release -- --image candle-examples/examples/dinov2reg4/bd2d3830ac3270218ba82fd24e2290becd01317c.jpg
+
+> Orchis simia Lam.       : 45.55%
+> Orchis × bergonii Nanteuil: 9.80%
+> Orchis italica Poir.    : 9.66%
+> Orchis × angusticruris Franch.: 2.76%
+> Orchis × bivonae Tod.   : 2.54%
+
+```
+
+![Orchis Simia](https://bs.plantnet.org/image/o/bd2d3830ac3270218ba82fd24e2290becd01317c)

candle-examples/examples/dinov2reg4/main.rs

@@ -0,0 +1,70 @@
+//! DINOv2 reg4 finetuned on PlantCLEF 2024
+//! https://arxiv.org/abs/2309.16588
+//! https://huggingface.co/spaces/BVRA/PlantCLEF2024
+//! https://zenodo.org/records/10848263
+
+#[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::dinov2reg4;
+
+#[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_image518(args.image)?.to_device(&device)?;
+    println!("loaded image {image:?}");
+
+    let f_species_id_mapping = "candle-examples/examples/dinov2reg4/species_id_mapping.txt";
+    let classes: Vec<String> = std::fs::read_to_string(f_species_id_mapping)
+        .expect("missing classes file")
+        .split('\n')
+        .map(|s| s.to_string())
+        .collect();
+
+    let model_file = match args.model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api =
+                api.model("vincent-espitalier/dino-v2-reg4-with-plantclef2024-weights".into());
+            api.get(
+                "vit_base_patch14_reg4_dinov2_lvd142m_pc24_onlyclassifier_then_all.safetensors",
+            )?
+        }
+        Some(model) => model.into(),
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+    let model = dinov2reg4::vit_base(vb)?;
+    println!("model built");
+    let logits = model.forward(&image.unsqueeze(0)?)?;
+    let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
+        .i(0)?
+        .to_vec1::<f32>()?;
+    let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
+    prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
+    for &(category_idx, pr) in prs.iter().take(5) {
+        println!("{:24}: {:.2}%", classes[category_idx], 100. * pr);
+    }
+    Ok(())
+}

candle-examples/examples/encodec/README.md

@@ -7,7 +7,7 @@ quantization.
 ## Running one example
 
 ```bash
-cargo run --example encodec --features symphonia --release -- code-to-audio \
+cargo run --example encodec --features encodec --release -- code-to-audio \
     candle-examples/examples/encodec/jfk-codes.safetensors \
     jfk.wav
 ```

candle-examples/examples/encodec/audio_io.rs

@@ -1,4 +1,3 @@
-#![allow(unused)]
 use anyhow::{Context, Result};
 use std::sync::{Arc, Mutex};
 

candle-examples/examples/eva2/README.md

@@ -0,0 +1,21 @@
+# candle-eva2
+
+[EVA-02](https://arxiv.org/abs/2303.11331) is a computer vision model.
+In this example, it is used as an ImageNet classifier: the model returns the
+probability for the image to belong to each of the 1000 ImageNet categories.
+
+## Running some example
+
+```bash
+cargo run --example eva2 --release -- --image candle-examples/examples/yolo-v8/assets/bike.jpg
+
+> mountain bike, all-terrain bike, off-roader: 37.09%
+> maillot                 : 8.30%
+> alp                     : 2.13%
+> bicycle-built-for-two, tandem bicycle, tandem: 0.84%
+> crash helmet            : 0.73%
+
+
+```
+
+![Leading group, Giro d'Italia 2021](../yolo-v8/assets/bike.jpg)

candle-examples/examples/eva2/main.rs

@@ -0,0 +1,82 @@
+//! EVA-02: Explore the limits of Visual representation at scAle
+//! https://github.com/baaivision/EVA
+
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use clap::Parser;
+
+use candle::{DType, Device, IndexOp, Result, Tensor, D};
+use candle_nn::{Module, VarBuilder};
+use candle_transformers::models::eva2;
+
+/// Loads an image from disk using the image crate, this returns a tensor with shape
+/// (3, 448, 448). OpenAI normalization is applied.
+pub fn load_image448_openai_norm<P: AsRef<std::path::Path>>(p: P) -> Result<Tensor> {
+    let img = image::ImageReader::open(p)?
+        .decode()
+        .map_err(candle::Error::wrap)?
+        .resize_to_fill(448, 448, image::imageops::FilterType::Triangle);
+    let img = img.to_rgb8();
+    let data = img.into_raw();
+    let data = Tensor::from_vec(data, (448, 448, 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)
+}
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    image: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    let device = candle_examples::device(args.cpu)?;
+
+    let image = load_image448_openai_norm(args.image)?.to_device(&device)?;
+    println!("loaded image {image:?}");
+
+    let model_file = match args.model {
+        None => {
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model("vincent-espitalier/candle-eva2".into());
+            api.get("eva02_base_patch14_448.mim_in22k_ft_in22k_in1k_adapted.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+
+    let model = eva2::vit_base(vb)?;
+    println!("model built");
+    let logits = model.forward(&image.unsqueeze(0)?)?;
+    let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
+        .i(0)?
+        .to_vec1::<f32>()?;
+    let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
+    prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
+    for &(category_idx, pr) in prs.iter().take(5) {
+        println!(
+            "{:24}: {:.2}%",
+            candle_examples::imagenet::CLASSES[category_idx],
+            100. * pr
+        );
+    }
+    Ok(())
+}

candle-examples/examples/fastvit/README.md

@@ -0,0 +1,20 @@
+# candle-fastvit
+
+[FastViT: A Fast Hybrid Vision Transformer using Structural Reparameterization](https://arxiv.org/abs/2303.14189).
+This candle implementation uses a pre-trained FastViT network for inference. The
+classification head has been trained on the ImageNet dataset and returns the
+probabilities for the top-5 classes.
+
+## Running an example
+
+```
+$ cargo run --example fastvit --release -- --image candle-examples/examples/yolo-v8/assets/bike.jpg --which sa12
+
+loaded image Tensor[dims 3, 256, 256; f32]
+model built
+mountain bike, all-terrain bike, off-roader: 52.67%
+bicycle-built-for-two, tandem bicycle, tandem: 7.93%
+unicycle, monocycle     : 3.46%
+maillot                 : 1.32%
+crash helmet            : 1.28%
+```

candle-examples/examples/fastvit/main.rs

@@ -0,0 +1,102 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use clap::{Parser, ValueEnum};
+
+use candle::{DType, IndexOp, D};
+use candle_nn::{Module, VarBuilder};
+use candle_transformers::models::fastvit;
+
+#[derive(Clone, Copy, Debug, ValueEnum)]
+enum Which {
+    T8,
+    T12,
+    S12,
+    SA12,
+    SA24,
+    SA36,
+    MA36,
+}
+
+impl Which {
+    fn model_filename(&self) -> String {
+        let name = match self {
+            Self::T8 => "t8",
+            Self::T12 => "t12",
+            Self::S12 => "s12",
+            Self::SA12 => "sa12",
+            Self::SA24 => "sa24",
+            Self::SA36 => "sa36",
+            Self::MA36 => "ma36",
+        };
+        format!("timm/fastvit_{}.apple_in1k", name)
+    }
+
+    fn config(&self) -> fastvit::Config {
+        match self {
+            Self::T8 => fastvit::Config::t8(),
+            Self::T12 => fastvit::Config::t12(),
+            Self::S12 => fastvit::Config::s12(),
+            Self::SA12 => fastvit::Config::sa12(),
+            Self::SA24 => fastvit::Config::sa24(),
+            Self::SA36 => fastvit::Config::sa36(),
+            Self::MA36 => fastvit::Config::ma36(),
+        }
+    }
+}
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    image: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    #[arg(value_enum, long, default_value_t=Which::S12)]
+    which: Which,
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    let device = candle_examples::device(args.cpu)?;
+
+    let image = candle_examples::imagenet::load_image(args.image, 256)?.to_device(&device)?;
+    println!("loaded image {image:?}");
+
+    let model_file = match args.model {
+        None => {
+            let model_name = args.which.model_filename();
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model(model_name);
+            api.get("model.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+    let model = fastvit::fastvit(&args.which.config(), 1000, vb)?;
+    println!("model built");
+    let logits = model.forward(&image.unsqueeze(0)?)?;
+    let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
+        .i(0)?
+        .to_vec1::<f32>()?;
+    let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
+    prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
+    for &(category_idx, pr) in prs.iter().take(5) {
+        println!(
+            "{:24}: {:.2}%",
+            candle_examples::imagenet::CLASSES[category_idx],
+            100. * pr
+        );
+    }
+    Ok(())
+}

candle-examples/examples/flux/README.md

@@ -0,0 +1,19 @@
+# candle-flux: image generation with latent rectified flow transformers
+
+![rusty robot holding a candle](./assets/flux-robot.jpg)
+
+Flux is a 12B rectified flow transformer capable of generating images from text
+descriptions,
+[huggingface](https://huggingface.co/black-forest-labs/FLUX.1-schnell),
+[github](https://github.com/black-forest-labs/flux),
+[blog post](https://blackforestlabs.ai/announcing-black-forest-labs/).
+
+
+## Running the model
+
+```bash
+cargo run --features cuda --example flux -r -- \
+    --height 1024 --width 1024 \
+    --prompt "a rusty robot walking on a beach holding a small torch, the robot has the word "rust" written on it, high quality, 4k"
+```
+

candle-examples/examples/flux/main.rs

@@ -0,0 +1,262 @@
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+use candle_transformers::models::{clip, flux, t5};
+
+use anyhow::{Error as E, Result};
+use candle::{IndexOp, Module, Tensor};
+use candle_nn::VarBuilder;
+use clap::Parser;
+use tokenizers::Tokenizer;
+
+#[derive(Parser)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// The prompt to be used for image generation.
+    #[arg(long, default_value = "A rusty robot walking on a beach")]
+    prompt: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Use the quantized model.
+    #[arg(long)]
+    quantized: bool,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    /// The height in pixels of the generated image.
+    #[arg(long)]
+    height: Option<usize>,
+
+    /// The width in pixels of the generated image.
+    #[arg(long)]
+    width: Option<usize>,
+
+    #[arg(long)]
+    decode_only: Option<String>,
+
+    #[arg(long, value_enum, default_value = "schnell")]
+    model: Model,
+
+    /// Use the slower kernels.
+    #[arg(long)]
+    use_dmmv: bool,
+
+    /// The seed to use when generating random samples.
+    #[arg(long)]
+    seed: Option<u64>,
+}
+
+#[derive(Debug, Clone, Copy, clap::ValueEnum, PartialEq, Eq)]
+enum Model {
+    Schnell,
+    Dev,
+}
+
+fn run(args: Args) -> Result<()> {
+    use tracing_chrome::ChromeLayerBuilder;
+    use tracing_subscriber::prelude::*;
+
+    let Args {
+        prompt,
+        cpu,
+        height,
+        width,
+        tracing,
+        decode_only,
+        model,
+        quantized,
+        ..
+    } = args;
+    let width = width.unwrap_or(1360);
+    let height = height.unwrap_or(768);
+
+    let _guard = if tracing {
+        let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
+        tracing_subscriber::registry().with(chrome_layer).init();
+        Some(guard)
+    } else {
+        None
+    };
+
+    let api = hf_hub::api::sync::Api::new()?;
+    let bf_repo = {
+        let name = match model {
+            Model::Dev => "black-forest-labs/FLUX.1-dev",
+            Model::Schnell => "black-forest-labs/FLUX.1-schnell",
+        };
+        api.repo(hf_hub::Repo::model(name.to_string()))
+    };
+    let device = candle_examples::device(cpu)?;
+    if let Some(seed) = args.seed {
+        device.set_seed(seed)?;
+    }
+    let dtype = device.bf16_default_to_f32();
+    let img = match decode_only {
+        None => {
+            let t5_emb = {
+                let repo = api.repo(hf_hub::Repo::with_revision(
+                    "google/t5-v1_1-xxl".to_string(),
+                    hf_hub::RepoType::Model,
+                    "refs/pr/2".to_string(),
+                ));
+                let model_file = repo.get("model.safetensors")?;
+                let vb =
+                    unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], dtype, &device)? };
+                let config_filename = repo.get("config.json")?;
+                let config = std::fs::read_to_string(config_filename)?;
+                let config: t5::Config = serde_json::from_str(&config)?;
+                let mut model = t5::T5EncoderModel::load(vb, &config)?;
+                let tokenizer_filename = api
+                    .model("lmz/mt5-tokenizers".to_string())
+                    .get("t5-v1_1-xxl.tokenizer.json")?;
+                let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
+                let mut tokens = tokenizer
+                    .encode(prompt.as_str(), true)
+                    .map_err(E::msg)?
+                    .get_ids()
+                    .to_vec();
+                tokens.resize(256, 0);
+                let input_token_ids = Tensor::new(&tokens[..], &device)?.unsqueeze(0)?;
+                println!("{input_token_ids}");
+                model.forward(&input_token_ids)?
+            };
+            println!("T5\n{t5_emb}");
+            let clip_emb = {
+                let repo = api.repo(hf_hub::Repo::model(
+                    "openai/clip-vit-large-patch14".to_string(),
+                ));
+                let model_file = repo.get("model.safetensors")?;
+                let vb =
+                    unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], dtype, &device)? };
+                // https://huggingface.co/openai/clip-vit-large-patch14/blob/main/config.json
+                let config = clip::text_model::ClipTextConfig {
+                    vocab_size: 49408,
+                    projection_dim: 768,
+                    activation: clip::text_model::Activation::QuickGelu,
+                    intermediate_size: 3072,
+                    embed_dim: 768,
+                    max_position_embeddings: 77,
+                    pad_with: None,
+                    num_hidden_layers: 12,
+                    num_attention_heads: 12,
+                };
+                let model =
+                    clip::text_model::ClipTextTransformer::new(vb.pp("text_model"), &config)?;
+                let tokenizer_filename = repo.get("tokenizer.json")?;
+                let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
+                let tokens = tokenizer
+                    .encode(prompt.as_str(), true)
+                    .map_err(E::msg)?
+                    .get_ids()
+                    .to_vec();
+                let input_token_ids = Tensor::new(&tokens[..], &device)?.unsqueeze(0)?;
+                println!("{input_token_ids}");
+                model.forward(&input_token_ids)?
+            };
+            println!("CLIP\n{clip_emb}");
+            let img = {
+                let cfg = match model {
+                    Model::Dev => flux::model::Config::dev(),
+                    Model::Schnell => flux::model::Config::schnell(),
+                };
+                let img = flux::sampling::get_noise(1, height, width, &device)?.to_dtype(dtype)?;
+                let state = if quantized {
+                    flux::sampling::State::new(
+                        &t5_emb.to_dtype(candle::DType::F32)?,
+                        &clip_emb.to_dtype(candle::DType::F32)?,
+                        &img.to_dtype(candle::DType::F32)?,
+                    )?
+                } else {
+                    flux::sampling::State::new(&t5_emb, &clip_emb, &img)?
+                };
+                let timesteps = match model {
+                    Model::Dev => {
+                        flux::sampling::get_schedule(50, Some((state.img.dim(1)?, 0.5, 1.15)))
+                    }
+                    Model::Schnell => flux::sampling::get_schedule(4, None),
+                };
+                println!("{state:?}");
+                println!("{timesteps:?}");
+                if quantized {
+                    let model_file = match model {
+                        Model::Schnell => api
+                            .repo(hf_hub::Repo::model("lmz/candle-flux".to_string()))
+                            .get("flux1-schnell.gguf")?,
+                        Model::Dev => todo!(),
+                    };
+                    let vb = candle_transformers::quantized_var_builder::VarBuilder::from_gguf(
+                        model_file, &device,
+                    )?;
+
+                    let model = flux::quantized_model::Flux::new(&cfg, vb)?;
+                    flux::sampling::denoise(
+                        &model,
+                        &state.img,
+                        &state.img_ids,
+                        &state.txt,
+                        &state.txt_ids,
+                        &state.vec,
+                        &timesteps,
+                        4.,
+                    )?
+                    .to_dtype(dtype)?
+                } else {
+                    let model_file = match model {
+                        Model::Schnell => bf_repo.get("flux1-schnell.safetensors")?,
+                        Model::Dev => bf_repo.get("flux1-dev.safetensors")?,
+                    };
+                    let vb = unsafe {
+                        VarBuilder::from_mmaped_safetensors(&[model_file], dtype, &device)?
+                    };
+                    let model = flux::model::Flux::new(&cfg, vb)?;
+                    flux::sampling::denoise(
+                        &model,
+                        &state.img,
+                        &state.img_ids,
+                        &state.txt,
+                        &state.txt_ids,
+                        &state.vec,
+                        &timesteps,
+                        4.,
+                    )?
+                }
+            };
+            flux::sampling::unpack(&img, height, width)?
+        }
+        Some(file) => {
+            let mut st = candle::safetensors::load(file, &device)?;
+            st.remove("img").unwrap().to_dtype(dtype)?
+        }
+    };
+    println!("latent img\n{img}");
+
+    let img = {
+        let model_file = bf_repo.get("ae.safetensors")?;
+        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], dtype, &device)? };
+        let cfg = match model {
+            Model::Dev => flux::autoencoder::Config::dev(),
+            Model::Schnell => flux::autoencoder::Config::schnell(),
+        };
+        let model = flux::autoencoder::AutoEncoder::new(&cfg, vb)?;
+        model.decode(&img)?
+    };
+    println!("img\n{img}");
+    let img = ((img.clamp(-1f32, 1f32)? + 1.0)? * 127.5)?.to_dtype(candle::DType::U8)?;
+    candle_examples::save_image(&img.i(0)?, "out.jpg")?;
+    Ok(())
+}
+
+fn main() -> Result<()> {
+    let args = Args::parse();
+    #[cfg(feature = "cuda")]
+    candle::quantized::cuda::set_force_dmmv(args.use_dmmv);
+    run(args)
+}

candle-examples/examples/flux/t5_tokenizer.py

@@ -0,0 +1,6 @@
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+BASE_MODEL = "google/t5-v1_1-xxl"
+tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
+# The tokenizer will be saved in /tmp/tokenizer/tokenizer.json
+tokenizer.save_pretrained("/tmp/tokenizer/")

candle-examples/examples/gemma/README.md

@@ -1,27 +1,27 @@
 # candle-gemma: 2b and 7b LLMs from Google DeepMind
 
 [Gemma](https://ai.google.dev/gemma/docs) is a collection of lightweight open
-models published by Google Deepmind with a 2b and a 7b variant.
-
-In order to use the example below, you have to accept the license on the
-[HuggingFace Hub Gemma repo](https://huggingface.co/google/gemma-7b) and set up
-your access token via the [HuggingFace cli login
-command](https://huggingface.co/docs/huggingface_hub/guides/cli#huggingface-cli-login).
+models published by Google Deepmind with a 2b and a 7b variant for the first
+version, and a 2b and a 9b variant for v2.
 
 ## Running the example
 
 ```bash
-$ cargo run --example gemma --release -- --prompt "fn count_primes(max_n: usize)"
-fn count_primes(max_n: usize) -> usize {
-    let mut primes = vec![true; max_n];
-    for i in 2..=max_n {
-        if primes[i] {
-            for j in i * i..max_n {
-                primes[j] = false;
-             }
-         }
-    }
-    primes.len()
-}
+$ cargo run --example gemma --features cuda -r -- \
+    --prompt "Here is a proof that square root of 2 is not rational: "
+
+Here is a proof that square root of 2 is not rational:
+
+Let us assume it to be rational. Then, we can write √2 = p/q where q ≠ 0 and p and q are integers with no common factors other than 1. Squaring both sides gives us (p/q)^2 = 2 or p^2/q^2 = 2. This implies that p^2 is divisible by 2, which means that p must be even. Let us write p = 2m where m is an integer. Substituting this in the above equation we get:
+
+(p^2)/q^2 = 2 or (4m^2)/q^2 = 2 or q^2/2m^2 = 1 which implies that q^2 must be divisible by 2, and hence q is even. This contradicts our assumption that p and q have no common factors other than 1. Hence we conclude that √2 cannot be rational.
 ```
 
+## Access restrictions
+
+In order to use the v1 examples, you have to accept the license on the
+[HuggingFace Hub Gemma repo](https://huggingface.co/google/gemma-7b) and set up
+your access token via the [HuggingFace cli login
+command](https://huggingface.co/docs/huggingface_hub/guides/cli#huggingface-cli-login).
+
+

candle-examples/examples/gemma/main.rs

@@ -7,7 +7,8 @@ extern crate accelerate_src;
 use anyhow::{Error as E, Result};
 use clap::Parser;
 
-use candle_transformers::models::gemma::{Config, Model};
+use candle_transformers::models::gemma::{Config as Config1, Model as Model1};
+use candle_transformers::models::gemma2::{Config as Config2, Model as Model2};
 
 use candle::{DType, Device, Tensor};
 use candle_examples::token_output_stream::TokenOutputStream;
@@ -30,6 +31,54 @@ enum Which {
     InstructV1_1_2B,
     #[value(name = "1.1-7b-it")]
     InstructV1_1_7B,
+    #[value(name = "code-2b")]
+    CodeBase2B,
+    #[value(name = "code-7b")]
+    CodeBase7B,
+    #[value(name = "code-2b-it")]
+    CodeInstruct2B,
+    #[value(name = "code-7b-it")]
+    CodeInstruct7B,
+    #[value(name = "2-2b")]
+    BaseV2_2B,
+    #[value(name = "2-2b-it")]
+    InstructV2_2B,
+    #[value(name = "2-9b")]
+    BaseV2_9B,
+    #[value(name = "2-9b-it")]
+    InstructV2_9B,
+}
+
+impl Which {
+    fn is_v1(&self) -> bool {
+        match self {
+            Self::Base2B
+            | Self::Base7B
+            | Self::Instruct2B
+            | Self::Instruct7B
+            | Self::InstructV1_1_2B
+            | Self::InstructV1_1_7B
+            | Self::CodeBase2B
+            | Self::CodeBase7B
+            | Self::CodeInstruct2B
+            | Self::CodeInstruct7B => true,
+            Self::BaseV2_2B | Self::InstructV2_2B | Self::BaseV2_9B | Self::InstructV2_9B => false,
+        }
+    }
+}
+
+enum Model {
+    V1(Model1),
+    V2(Model2),
+}
+
+impl Model {
+    fn forward(&mut self, input_ids: &Tensor, pos: usize) -> candle::Result<Tensor> {
+        match self {
+            Self::V1(m) => m.forward(input_ids, pos),
+            Self::V2(m) => m.forward(input_ids, pos),
+        }
+    }
 }
 
 struct TextGeneration {
@@ -183,8 +232,11 @@ struct Args {
     repeat_last_n: usize,
 
     /// The model to use.
-    #[arg(long, default_value = "2b")]
+    #[arg(long, default_value = "2-2b")]
     which: Which,
+
+    #[arg(long)]
+    use_flash_attn: bool,
 }
 
 fn main() -> Result<()> {
@@ -224,6 +276,14 @@ fn main() -> Result<()> {
             Which::Base7B => "google/gemma-7b".to_string(),
             Which::Instruct2B => "google/gemma-2b-it".to_string(),
             Which::Instruct7B => "google/gemma-7b-it".to_string(),
+            Which::CodeBase2B => "google/codegemma-2b".to_string(),
+            Which::CodeBase7B => "google/codegemma-7b".to_string(),
+            Which::CodeInstruct2B => "google/codegemma-2b-it".to_string(),
+            Which::CodeInstruct7B => "google/codegemma-7b-it".to_string(),
+            Which::BaseV2_2B => "google/gemma-2-2b".to_string(),
+            Which::InstructV2_2B => "google/gemma-2-2b-it".to_string(),
+            Which::BaseV2_9B => "google/gemma-2-9b".to_string(),
+            Which::InstructV2_9B => "google/gemma-2-9b-it".to_string(),
         },
     };
     let repo = api.repo(Repo::with_revision(
@@ -248,7 +308,6 @@ fn main() -> Result<()> {
     };
     println!("retrieved the files in {:?}", start.elapsed());
     let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
-    let config: Config = serde_json::from_reader(std::fs::File::open(config_filename)?)?;
 
     let start = std::time::Instant::now();
     let device = candle_examples::device(args.cpu)?;
@@ -258,7 +317,15 @@ fn main() -> Result<()> {
         DType::F32
     };
     let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
-    let model = Model::new(&config, vb)?;
+    let model = if args.which.is_v1() {
+        let config: Config1 = serde_json::from_reader(std::fs::File::open(config_filename)?)?;
+        let model = Model1::new(args.use_flash_attn, &config, vb)?;
+        Model::V1(model)
+    } else {
+        let config: Config2 = serde_json::from_reader(std::fs::File::open(config_filename)?)?;
+        let model = Model2::new(args.use_flash_attn, &config, vb)?;
+        Model::V2(model)
+    };
 
     println!("loaded the model in {:?}", start.elapsed());
 

candle-examples/examples/glm4/README.org

@@ -0,0 +1,77 @@
+* GLM4
+GLM-4-9B is the open-source version of the latest generation of pre-trained models in the GLM-4 series launched by Zhipu AI. 
+
+- [[https://github.com/THUDM/GLM4][Github]]
+- [[https://huggingface.co/THUDM/glm-4-9b][huggingface]]  
+
+** Running with ~cuda~
+
+#+begin_src shell
+  cargo run --example glm4 --release --features cuda 
+#+end_src
+
+** Running with ~cpu~
+#+begin_src shell
+  cargo run --example glm4 --release -- --cpu
+#+end_src
+
+** Output Example
+#+begin_src shell
+cargo run  --example glm4 --release --features cuda -- --sample-len 500 --cache .
+    Finished release [optimized] target(s) in 0.24s
+     Running `/root/candle/target/release/examples/glm4 --sample-len 500 --cache .`
+avx: true, neon: false, simd128: false, f16c: true
+temp: 0.60 repeat-penalty: 1.20 repeat-last-n: 64
+cache path .
+retrieved the files in 6.88963ms
+loaded the model in 6.113752297s
+starting the inference loop
+[欢迎使用GLM-4,请输入prompt]
+请你告诉我什么是FFT
+266 tokens generated (34.50 token/s)
+Result:
+。Fast Fourier Transform (FFT) 是一种快速计算离散傅里叶变换（DFT）的方法，它广泛应用于信号处理、图像处理和数据分析等领域。
+
+具体来说，FFT是一种将时域数据转换为频域数据的算法。在数字信号处理中，我们通常需要知道信号的频率成分，这就需要进行傅立叶变换。传统的傅立叶变换的计算复杂度较高，而 FFT 则大大提高了计算效率，使得大规模的 DFT 换成为可能。
+
+以下是使用 Python 中的 numpy 进行 FFT 的简单示例：
+
+```python
+import numpy as np
+
+# 创建一个时域信号
+t = np.linspace(0, 1, num=100)
+f = np.sin(2*np.pi*5*t) + 3*np.cos(2*np.pi*10*t)
+
+# 对该信号做FFT变换，并计算其幅值谱
+fft_result = np.fft.fftshift(np.abs(np.fft.fft(f)))
+
+```
+
+在这个例子中，我们首先创建了一个时域信号 f。然后我们对这个信号进行了 FFT 换，得到了一个频域结果 fft_result。
+#+end_src
+
+This example will read prompt from stdin
+
+*  Citation
+#+begin_src
+  @misc{glm2024chatglm,
+      title={ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools}, 
+      author={Team GLM and Aohan Zeng and Bin Xu and Bowen Wang and Chenhui Zhang and Da Yin and Diego Rojas and Guanyu Feng and Hanlin Zhao and Hanyu Lai and Hao Yu and Hongning Wang and Jiadai Sun and Jiajie Zhang and Jiale Cheng and Jiayi Gui and Jie Tang and Jing Zhang and Juanzi Li and Lei Zhao and Lindong Wu and Lucen Zhong and Mingdao Liu and Minlie Huang and Peng Zhang and Qinkai Zheng and Rui Lu and Shuaiqi Duan and Shudan Zhang and Shulin Cao and Shuxun Yang and Weng Lam Tam and Wenyi Zhao and Xiao Liu and Xiao Xia and Xiaohan Zhang and Xiaotao Gu and Xin Lv and Xinghan Liu and Xinyi Liu and Xinyue Yang and Xixuan Song and Xunkai Zhang and Yifan An and Yifan Xu and Yilin Niu and Yuantao Yang and Yueyan Li and Yushi Bai and Yuxiao Dong and Zehan Qi and Zhaoyu Wang and Zhen Yang and Zhengxiao Du and Zhenyu Hou and Zihan Wang},
+      year={2024},
+      eprint={2406.12793},
+      archivePrefix={arXiv},
+      primaryClass={id='cs.CL' full_name='Computation and Language' is_active=True alt_name='cmp-lg' in_archive='cs' is_general=False description='Covers natural language processing. Roughly includes material in ACM Subject Class I.2.7. Note that work on artificial languages (programming languages, logics, formal systems) that does not explicitly address natural-language issues broadly construed (natural-language processing, computational linguistics, speech, text retrieval, etc.) is not appropriate for this area.'}
+}
+#+end_src
+
+#+begin_src
+    @misc{wang2023cogvlm,
+      title={CogVLM: Visual Expert for Pretrained Language Models}, 
+      author={Weihan Wang and Qingsong Lv and Wenmeng Yu and Wenyi Hong and Ji Qi and Yan Wang and Junhui Ji and Zhuoyi Yang and Lei Zhao and Xixuan Song and Jiazheng Xu and Bin Xu and Juanzi Li and Yuxiao Dong and Ming Ding and Jie Tang},
+      year={2023},
+      eprint={2311.03079},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
+#+end_src

candle-examples/examples/glm4/main.rs

@@ -0,0 +1,255 @@
+use candle_transformers::models::glm4::*;
+use clap::Parser;
+
+use candle::{DType, Device, Tensor};
+use candle_nn::VarBuilder;
+use candle_transformers::generation::LogitsProcessor;
+use hf_hub::{Repo, RepoType};
+use tokenizers::Tokenizer;
+
+struct TextGeneration {
+    model: Model,
+    device: Device,
+    tokenizer: Tokenizer,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+    verbose_prompt: bool,
+    dtype: DType,
+}
+
+impl TextGeneration {
+    #[allow(clippy::too_many_arguments)]
+    fn new(
+        model: Model,
+        tokenizer: Tokenizer,
+        seed: u64,
+        temp: Option<f64>,
+        top_p: Option<f64>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        verbose_prompt: bool,
+        device: &Device,
+        dtype: DType,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            tokenizer,
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            verbose_prompt,
+            device: device.clone(),
+            dtype,
+        }
+    }
+
+    fn run(&mut self, sample_len: usize) -> anyhow::Result<()> {
+        use std::io::BufRead;
+        use std::io::BufReader;
+        use std::io::Write;
+        println!("starting the inference loop");
+        println!("[欢迎使用GLM-4,请输入prompt]");
+        let stdin = std::io::stdin();
+        let reader = BufReader::new(stdin);
+        for line in reader.lines() {
+            let line = line.expect("Failed to read line");
+
+            let tokens = self.tokenizer.encode(line, true).expect("tokens error");
+            if tokens.is_empty() {
+                panic!("Empty prompts are not supported in the chatglm model.")
+            }
+            if self.verbose_prompt {
+                for (token, id) in tokens.get_tokens().iter().zip(tokens.get_ids().iter()) {
+                    let token = token.replace('▁', " ").replace("<0x0A>", "\n");
+                    println!("{id:7} -> '{token}'");
+                }
+            }
+            let eos_token = match self.tokenizer.get_vocab(true).get("<|endoftext|>") {
+                Some(token) => *token,
+                None => panic!("cannot find the endoftext token"),
+            };
+            let mut tokens = tokens.get_ids().to_vec();
+            let mut generated_tokens = 0usize;
+
+            std::io::stdout().flush().expect("output flush error");
+            let start_gen = std::time::Instant::now();
+
+            let mut count = 0;
+            let mut result = vec![];
+            for index in 0..sample_len {
+                count += 1;
+                let context_size = if index > 0 { 1 } else { tokens.len() };
+                let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
+                let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+                let logits = self.model.forward(&input)?;
+                let logits = logits.squeeze(0)?.to_dtype(self.dtype)?;
+                let logits = if self.repeat_penalty == 1. {
+                    logits
+                } else {
+                    let start_at = tokens.len().saturating_sub(self.repeat_last_n);
+                    candle_transformers::utils::apply_repeat_penalty(
+                        &logits,
+                        self.repeat_penalty,
+                        &tokens[start_at..],
+                    )?
+                };
+
+                let next_token = self.logits_processor.sample(&logits)?;
+                tokens.push(next_token);
+                generated_tokens += 1;
+                if next_token == eos_token {
+                    break;
+                }
+                let token = self
+                    .tokenizer
+                    .decode(&[next_token], true)
+                    .expect("Token error");
+                if self.verbose_prompt {
+                    println!(
+                        "[Count: {}] [Raw Token: {}] [Decode Token: {}]",
+                        count, next_token, token
+                    );
+                }
+                result.push(token);
+                std::io::stdout().flush()?;
+            }
+            let dt = start_gen.elapsed();
+            println!(
+                "\n{generated_tokens} tokens generated ({:.2} token/s)",
+                generated_tokens as f64 / dt.as_secs_f64(),
+            );
+            println!("Result:");
+            for tokens in result {
+                print!("{tokens}");
+            }
+            self.model.reset_kv_cache(); // clean the cache
+        }
+        Ok(())
+    }
+}
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(name = "cache", short, long, default_value = ".")]
+    cache_path: String,
+
+    #[arg(long)]
+    cpu: bool,
+
+    /// Display the token for the specified prompt.
+    #[arg(long)]
+    verbose_prompt: bool,
+
+    /// 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 = 8192)]
+    sample_len: usize,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long)]
+    revision: Option<String>,
+
+    #[arg(long)]
+    weight_file: Option<String>,
+
+    #[arg(long)]
+    tokenizer: Option<String>,
+
+    /// Penalty to be applied for repeating tokens, 1. means no penalty.
+    #[arg(long, default_value_t = 1.2)]
+    repeat_penalty: f32,
+
+    /// The context size to consider for the repeat penalty.
+    #[arg(long, default_value_t = 64)]
+    repeat_last_n: usize,
+}
+
+fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+    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(0.6),
+        args.repeat_penalty,
+        args.repeat_last_n
+    );
+
+    let start = std::time::Instant::now();
+    println!("cache path {}", args.cache_path);
+    let api = hf_hub::api::sync::ApiBuilder::from_cache(hf_hub::Cache::new(args.cache_path.into()))
+        .build()
+        .map_err(anyhow::Error::msg)?;
+
+    let model_id = match args.model_id {
+        Some(model_id) => model_id.to_string(),
+        None => "THUDM/glm-4-9b".to_string(),
+    };
+    let revision = match args.revision {
+        Some(rev) => rev.to_string(),
+        None => "main".to_string(),
+    };
+    let repo = api.repo(Repo::with_revision(model_id, RepoType::Model, revision));
+    let tokenizer_filename = match args.tokenizer {
+        Some(file) => std::path::PathBuf::from(file),
+        None => api
+            .model("THUDM/codegeex4-all-9b".to_string())
+            .get("tokenizer.json")
+            .map_err(anyhow::Error::msg)?,
+    };
+    let filenames = match args.weight_file {
+        Some(weight_file) => vec![std::path::PathBuf::from(weight_file)],
+        None => candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.json")?,
+    };
+    println!("retrieved the files in {:?}", start.elapsed());
+    let tokenizer = Tokenizer::from_file(tokenizer_filename).expect("Tokenizer Error");
+
+    let start = std::time::Instant::now();
+    let config = Config::glm4();
+    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,
+        args.verbose_prompt,
+        &device,
+        dtype,
+    );
+    pipeline.run(args.sample_len)?;
+    Ok(())
+}

candle-examples/examples/granite/README.md

@@ -0,0 +1,20 @@
+# candle-granite LLMs from IBM Research
+
+[Granite](https://www.ibm.com/granite) is a family of Large Language Models built for business, to help drive trust and scalability in AI-driven applications.
+
+## Running the example
+
+```bash
+$ cargo run --example granite --features metal -r -- --model-type "granite7b-instruct" \
+    --prompt "Explain how quantum computing differs from classical computing, focusing on key concepts like qubits, superposition, and entanglement. Describe two potential breakthroughs in the fields of drug discovery and cryptography. Offer a convincing argument for why businesses and governments should invest in quantum computing research now, emphasizing its future benefits and the risks of falling behind"
+
+    Explain how quantum computing differs from classical computing, focusing on key concepts like qubits, superposition, and entanglement. Describe two potential breakthroughs in the fields of drug discovery and cryptography. Offer a convincing argument for why businesses and governments should invest in quantum computing research now, emphasizing its future benefits and the risks of falling behind competitors.
+
+    In recent years, there has been significant interest in quantum computing due to its potential to revolutionize various fields, including drug discovery, cryptography, and optimization problems. Quantum computers, which leverage the principles of quantum mechanics, differ fundamentally from classical computers. Here are some of the key differences:
+```
+
+## Supported Models
+There are two different modalities for the Granite family models: Language and Code.
+
+### Granite for language
+1. [Granite 7b Instruct](https://huggingface.co/ibm-granite/granite-7b-instruct)

candle-examples/examples/granite/main.rs

@@ -0,0 +1,251 @@
+// An implementation of different Granite models https://www.ibm.com/granite
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+use anyhow::{bail, Error as E, Result};
+use clap::{Parser, ValueEnum};
+
+use candle::{DType, Tensor};
+use candle_nn::VarBuilder;
+use candle_transformers::generation::{LogitsProcessor, Sampling};
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use std::io::Write;
+
+use candle_transformers::models::granite as model;
+use model::{Granite, GraniteConfig};
+
+use std::time::Instant;
+
+const EOS_TOKEN: &str = "</s>";
+const DEFAULT_PROMPT: &str = "How Fault Tolerant Quantum Computers will help humanity?";
+
+#[derive(Clone, Debug, Copy, PartialEq, Eq, ValueEnum)]
+enum GraniteModel {
+    Granite7bInstruct,
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// The temperature used to generate samples.
+    #[arg(long, default_value_t = 0.8)]
+    temperature: f64,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// Only sample among the top K samples.
+    #[arg(long)]
+    top_k: Option<usize>,
+
+    /// 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(short = 'n', long, default_value_t = 10000)]
+    sample_len: usize,
+
+    /// Disable the key-value cache.
+    #[arg(long)]
+    no_kv_cache: bool,
+
+    /// The initial prompt.
+    #[arg(long)]
+    prompt: Option<String>,
+
+    /// Use different dtype than f16
+    #[arg(long)]
+    dtype: Option<String>,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long)]
+    revision: Option<String>,
+
+    #[arg(long, default_value = "granite7b-instruct")]
+    model_type: GraniteModel,
+
+    #[arg(long)]
+    use_flash_attn: bool,
+
+    /// 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 = 128)]
+    repeat_last_n: usize,
+}
+
+fn main() -> Result<()> {
+    use tokenizers::Tokenizer;
+    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
+    };
+
+    let device = candle_examples::device(args.cpu)?;
+    let dtype = match args.dtype.as_deref() {
+        Some("f16") => DType::F16,
+        Some("bf16") => DType::BF16,
+        Some("f32") => DType::F32,
+        Some(dtype) => bail!("Unsupported dtype {dtype}"),
+        None => DType::F16,
+    };
+    let (granite, tokenizer_filename, mut cache, config) = {
+        let api = Api::new()?;
+        let model_id = args.model_id.unwrap_or_else(|| match args.model_type {
+            GraniteModel::Granite7bInstruct => "ibm-granite/granite-7b-instruct".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: GraniteConfig = serde_json::from_slice(&std::fs::read(config_filename)?)?;
+        let config = config.into_config(args.use_flash_attn);
+
+        let filenames = match args.model_type {
+            GraniteModel::Granite7bInstruct => {
+                candle_examples::hub_load_safetensors(&api, "model.safetensors.index.json")?
+            }
+        };
+        let cache = model::Cache::new(!args.no_kv_cache, dtype, &config, &device)?;
+
+        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
+        (
+            Granite::load(vb, &config)?,
+            tokenizer_filename,
+            cache,
+            config,
+        )
+    };
+    let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
+    let eos_token_id = config.eos_token_id.or_else(|| {
+        tokenizer
+            .token_to_id(EOS_TOKEN)
+            .map(model::GraniteEosToks::Single)
+    });
+
+    let default_prompt = match args.model_type {
+        GraniteModel::Granite7bInstruct => DEFAULT_PROMPT,
+    };
+
+    let prompt = args.prompt.as_ref().map_or(default_prompt, |p| p.as_str());
+    let mut tokens = tokenizer
+        .encode(prompt, true)
+        .map_err(E::msg)?
+        .get_ids()
+        .to_vec();
+    let mut tokenizer = candle_examples::token_output_stream::TokenOutputStream::new(tokenizer);
+
+    println!("Starting the inference loop:");
+    print!("{prompt}");
+    let mut logits_processor = {
+        let temperature = args.temperature;
+        let sampling = if temperature <= 0. {
+            Sampling::ArgMax
+        } else {
+            match (args.top_k, args.top_p) {
+                (None, None) => Sampling::All { temperature },
+                (Some(k), None) => Sampling::TopK { k, temperature },
+                (None, Some(p)) => Sampling::TopP { p, temperature },
+                (Some(k), Some(p)) => Sampling::TopKThenTopP { k, p, temperature },
+            }
+        };
+        LogitsProcessor::from_sampling(args.seed, sampling)
+    };
+
+    let mut start_gen = std::time::Instant::now();
+    let mut index_pos = 0;
+    let mut token_generated = 0;
+    let use_cache_kv = cache.use_kv_cache;
+
+    (0..args.sample_len)
+        .inspect(|index| {
+            if *index == 1 {
+                start_gen = Instant::now();
+            }
+        })
+        .try_for_each(|index| -> Result<()> {
+            let (context_size, context_index) = if use_cache_kv && index > 0 {
+                (1, index_pos)
+            } else {
+                (tokens.len(), 0)
+            };
+            let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
+            let input = Tensor::new(ctxt, &device)?.unsqueeze(0)?;
+            let logits = granite
+                .forward(&input, context_index, &mut cache)?
+                .squeeze(0)?;
+
+            let logits = if args.repeat_penalty == 1. {
+                logits
+            } else {
+                let start_at = tokens.len().saturating_sub(args.repeat_last_n);
+                candle_transformers::utils::apply_repeat_penalty(
+                    &logits,
+                    args.repeat_penalty,
+                    &tokens[start_at..],
+                )?
+            };
+
+            index_pos += ctxt.len();
+
+            let next_token = logits_processor.sample(&logits)?;
+            token_generated += 1;
+            tokens.push(next_token);
+
+            if let Some(model::GraniteEosToks::Single(eos_tok_id)) = eos_token_id {
+                if next_token == eos_tok_id {
+                    return Err(E::msg("EOS token found"));
+                }
+            } else if let Some(model::GraniteEosToks::Multiple(ref eos_ids)) = eos_token_id {
+                if eos_ids.contains(&next_token) {
+                    return Err(E::msg("EOS token found"));
+                }
+            }
+
+            if let Some(t) = tokenizer.next_token(next_token)? {
+                print!("{t}");
+                std::io::stdout().flush()?;
+            }
+            Ok(())
+        })
+        .unwrap_or(());
+
+    if let Some(rest) = tokenizer.decode_rest().map_err(E::msg)? {
+        print!("{rest}");
+    }
+
+    let dt = start_gen.elapsed();
+    println!(
+        "\n\n{} tokens generated ({} token/s)\n",
+        token_generated,
+        (token_generated - 1) as f64 / dt.as_secs_f64(),
+    );
+    Ok(())
+}

candle-examples/examples/gte-qwen/README.md

@@ -0,0 +1,19 @@
+# gte-Qwen1.5-7B-instruct
+
+gte-Qwen1.5-7B-instruct is a variant of the GTE embedding model family.
+
+- [Model card](https://huggingface.co/Alibaba-NLP/gte-Qwen1.5-7B-instruct) on the HuggingFace Hub.
+- [Technical report](https://arxiv.org/abs/2308.03281) *Towards General Text Embeddings with Multi-stage Contrastive Learning*
+
+
+## Running the example
+
+Automatically download the model from the HuggingFace hub:
+```bash
+$ cargo run --example gte-qwen --release
+```
+
+or, load the model from a local directory:
+```bash
+cargo run --example gte-qwen --release --features cuda -- --local-repo /path/to/gte_Qwen1.5-7B-instruct/
+```

candle-examples/examples/gte-qwen/main.rs

@@ -0,0 +1,178 @@
+#[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::qwen2::{Config, Model};
+
+use candle::{DType, Tensor};
+use candle_nn::VarBuilder;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::{
+    utils::padding::{PaddingDirection, PaddingParams, PaddingStrategy},
+    Tokenizer,
+};
+
+// gte-Qwen1.5-7B-instruct use EOS token as padding token
+const EOS_TOKEN: &str = "<|endoftext|>";
+const EOS_TOKEN_ID: u32 = 151643;
+
+#[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, default_value = "Alibaba-NLP/gte-Qwen1.5-7B-instruct")]
+    model_id: String,
+
+    #[arg(long, default_value = "main")]
+    revision: String,
+
+    #[arg(long)]
+    local_repo: Option<String>,
+}
+
+#[derive(Debug)]
+struct ConfigFiles {
+    pub config: std::path::PathBuf,
+    pub tokenizer: std::path::PathBuf,
+    pub weights: Vec<std::path::PathBuf>,
+}
+
+// Loading the model from the HuggingFace Hub. Network access is required.
+fn load_from_hub(model_id: &str, revision: &str) -> Result<ConfigFiles> {
+    let api = Api::new()?;
+    let repo = api.repo(Repo::with_revision(
+        model_id.to_string(),
+        RepoType::Model,
+        revision.to_string(),
+    ));
+    Ok(ConfigFiles {
+        config: repo.get("config.json")?,
+        tokenizer: repo.get("tokenizer.json")?,
+        weights: candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.json")?,
+    })
+}
+
+// Loading the model from a local directory.
+fn load_from_local(local_path: &str) -> Result<ConfigFiles> {
+    let local_path = std::path::PathBuf::from(local_path);
+    let weight_path = local_path.join("model.safetensors.index.json");
+    let json: serde_json::Value = serde_json::from_str(&std::fs::read_to_string(weight_path)?)?;
+    let weight_map = match json.get("weight_map") {
+        Some(serde_json::Value::Object(map)) => map,
+        Some(_) => panic!("`weight map` is not a map"),
+        None => panic!("`weight map` not found"),
+    };
+    let mut safetensors_files = std::collections::HashSet::new();
+    for value in weight_map.values() {
+        safetensors_files.insert(
+            value
+                .as_str()
+                .expect("Weight files should be parsed as strings"),
+        );
+    }
+    let safetensors_paths = safetensors_files
+        .iter()
+        .map(|v| local_path.join(v))
+        .collect::<Vec<_>>();
+    Ok(ConfigFiles {
+        config: local_path.join("config.json"),
+        tokenizer: local_path.join("tokenizer.json"),
+        weights: safetensors_paths,
+    })
+}
+
+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
+    };
+
+    // Fetch the model. Do this offline if local path provided.
+    println!("Fetching model files...");
+    let start = std::time::Instant::now();
+    let config_files = match args.local_repo {
+        Some(local_path) => load_from_local(&local_path)?,
+        None => load_from_hub(&args.model_id, &args.revision)?,
+    };
+    println!("Model file retrieved in {:?}", start.elapsed());
+
+    // Inputs will be padded to the longest sequence in the batch.
+    let padding = PaddingParams {
+        strategy: PaddingStrategy::BatchLongest,
+        direction: PaddingDirection::Left,
+        pad_to_multiple_of: None,
+        pad_id: EOS_TOKEN_ID,
+        pad_type_id: 0,
+        pad_token: String::from(EOS_TOKEN),
+    };
+
+    // Tokenizer setup
+    let mut tokenizer = Tokenizer::from_file(config_files.tokenizer).map_err(E::msg)?;
+    tokenizer.with_padding(Some(padding));
+
+    // Model initialization
+    let device = candle_examples::device(args.cpu)?;
+    let dtype = if device.is_cuda() {
+        DType::BF16
+    } else {
+        DType::F32
+    };
+    let config: Config = serde_json::from_slice(&std::fs::read(config_files.config)?)?;
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&config_files.weights, dtype, &device)? };
+    let mut model = Model::new(&config, vb)?;
+    println!("Model loaded in {:?}", start.elapsed());
+
+    // Encode the queries and the targets
+    let instruct = "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: ";
+    let documents = vec![
+        format!("{instruct}how much protein should a female eat{EOS_TOKEN}"),
+        format!("{instruct}summit define{EOS_TOKEN}"),
+        format!("As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.{EOS_TOKEN}"),
+        format!("Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments.{EOS_TOKEN}"),
+    ];
+    let encoded = tokenizer.encode_batch(documents, true).map_err(E::msg)?;
+    let tokens: Vec<&[u32]> = encoded.iter().map(|x| x.get_ids()).collect();
+    let tokens = Tensor::new(tokens, &device)?;
+    let mask: Vec<&[u32]> = encoded.iter().map(|x| x.get_attention_mask()).collect();
+    let mask = Tensor::new(mask, &device)?;
+
+    // Inference
+    let start_gen = std::time::Instant::now();
+    let logits = model.forward(&tokens, 0, Some(&mask))?;
+
+    // Extract the last hidden states as embeddings since inputs are padded left.
+    let (_, seq_len, _) = logits.dims3()?;
+    let embd = logits
+        .narrow(1, seq_len - 1, 1)?
+        .squeeze(1)?
+        .to_dtype(DType::F32)?;
+
+    // Calculate the relativity scores. Note the embeddings should be normalized.
+    let norm = embd.broadcast_div(&embd.sqr()?.sum_keepdim(1)?.sqrt()?)?;
+    let scores = norm.narrow(0, 0, 2)?.matmul(&norm.narrow(0, 2, 2)?.t()?)?;
+
+    // Print the results
+    println!("Embedding done in {:?}", start_gen.elapsed());
+    println!("Scores: {:?}", scores.to_vec2::<f32>()?);
+
+    Ok(())
+}

candle-examples/examples/hiera/README.md

@@ -0,0 +1,18 @@
+# hiera
+
+[Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles](https://arxiv.org/abs/2306.00989)
+This candle implementation uses pre-trained Hiera models from timm for inference.
+The classification head has been trained on the ImageNet dataset and returns the probabilities for the top-5 classes.
+
+## Running an example
+
+```
+$ cargo run --example  hiera --release  -- --image candle-examples/examples/yolo-v8/assets/bike.jpg --which tiny
+loaded image Tensor[dims 3, 224, 224; f32]
+model built
+mountain bike, all-terrain bike, off-roader: 71.15%
+unicycle, monocycle     : 7.11%
+knee pad                : 4.26%
+crash helmet            : 1.48%
+moped                   : 1.07%
+```

candle-examples/examples/hiera/main.rs

@@ -0,0 +1,99 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use clap::{Parser, ValueEnum};
+
+use candle::{DType, IndexOp, D};
+use candle_nn::{Module, VarBuilder};
+use candle_transformers::models::hiera;
+
+#[derive(Clone, Copy, Debug, ValueEnum)]
+enum Which {
+    Tiny,
+    Small,
+    Base,
+    BasePlus,
+    Large,
+    Huge,
+}
+
+impl Which {
+    fn model_filename(&self) -> String {
+        let name = match self {
+            Self::Tiny => "tiny",
+            Self::Small => "small",
+            Self::Base => "base",
+            Self::BasePlus => "base_plus",
+            Self::Large => "large",
+            Self::Huge => "huge",
+        };
+        format!("timm/hiera_{}_224.mae_in1k_ft_in1k", name)
+    }
+
+    fn config(&self) -> hiera::Config {
+        match self {
+            Self::Tiny => hiera::Config::tiny(),
+            Self::Small => hiera::Config::small(),
+            Self::Base => hiera::Config::base(),
+            Self::BasePlus => hiera::Config::base_plus(),
+            Self::Large => hiera::Config::large(),
+            Self::Huge => hiera::Config::huge(),
+        }
+    }
+}
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    image: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    #[arg(value_enum, long, default_value_t=Which::Tiny)]
+    which: Which,
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    let device = candle_examples::device(args.cpu)?;
+
+    let image = candle_examples::imagenet::load_image224(args.image)?.to_device(&device)?;
+    println!("loaded image {image:?}");
+
+    let model_file = match args.model {
+        None => {
+            let model_name = args.which.model_filename();
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model(model_name);
+            api.get("model.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+    let model = hiera::hiera(&args.which.config(), 1000, vb)?;
+    println!("model built");
+    let logits = model.forward(&image.unsqueeze(0)?)?;
+    let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
+        .i(0)?
+        .to_vec1::<f32>()?;
+    let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
+    prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
+    for &(category_idx, pr) in prs.iter().take(5) {
+        println!(
+            "{:24}: {:.2}%",
+            candle_examples::imagenet::CLASSES[category_idx],
+            100. * pr
+        );
+    }
+    Ok(())
+}

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

@@ -4,7 +4,7 @@ extern crate intel_mkl_src;
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
 
-use candle_transformers::models::jina_bert::{BertModel, Config};
+use candle_transformers::models::jina_bert::{BertModel, Config, PositionEmbeddingType};
 
 use anyhow::Error as E;
 use candle::{DType, Module, Tensor};
@@ -39,32 +39,47 @@ struct Args {
 
     #[arg(long)]
     model: Option<String>,
+
+    #[arg(long)]
+    model_file: 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 {
+        let model_name = match self.model.as_ref() {
+            Some(model) => model.to_string(),
+            None => "jinaai/jina-embeddings-v2-base-en".to_string(),
+        };
+
+        let model = match &self.model_file {
             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,
-                ))
+                .repo(Repo::new(model_name.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,
-                ))
+                .repo(Repo::new(model_name.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 config = Config::new(
+            tokenizer.get_vocab_size(true),
+            768,
+            12,
+            12,
+            3072,
+            candle_nn::Activation::Gelu,
+            8192,
+            2,
+            0.02,
+            1e-12,
+            0,
+            PositionEmbeddingType::Alibi,
+        );
         let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model], DType::F32, &device)? };
         let model = BertModel::new(vb, &config)?;
         Ok((model, tokenizer))
@@ -101,14 +116,20 @@ fn main() -> anyhow::Result<()> {
             .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());
+        let start = std::time::Instant::now();
+        let embeddings = model.forward(&token_ids)?;
+        let (_n_sentence, n_tokens, _hidden_size) = embeddings.dims3()?;
+        let embeddings = (embeddings.sum(1)? / (n_tokens as f64))?;
+        println!("pooled_embeddigns: {embeddings}");
+        let embeddings = if args.normalize_embeddings {
+            normalize_l2(&embeddings)?
+        } else {
+            embeddings
+        };
+        if args.normalize_embeddings {
+            println!("normalized_embeddings: {embeddings}");
         }
+        println!("Took {:?}", start.elapsed());
     } else {
         let sentences = [
             "The cat sits outside",