chore(candle): Allow enabling accelerate

backend-comparison/Cargo.toml

@@ -14,6 +14,7 @@ default = ["std"]
 std = []
 candle-cpu = ["burn/candle"]
 candle-cuda = ["burn/candle-cuda"]
+candle-accelerate = ["burn/candle-accelerate"]
 ndarray = ["burn/ndarray"]
 ndarray-blas-accelerate = ["burn/ndarray-blas-accelerate"]
 ndarray-blas-netlib = ["burn/ndarray-blas-netlib"]

burn-candle/Cargo.toml

@@ -12,6 +12,7 @@ version = "0.11.0"
 
 [features]
 cuda = ["candle-core/cuda"]
+accelerate = ["candle-core/accelerate"]
 
 [dependencies]
 derive-new = { workspace = true }

burn-candle/README.md

@@ -4,4 +4,11 @@ This crate provides a backend for [Burn](https://github.com/burn-rs/burn) based
 
 It is still in alpha stage, not all operations are supported. It is usable for some use cases, like for inference. 
 
-It can be used with CPU or CUDA. 
+It can be used with CPU or CUDA. On macOS computations can be accelerated by using the Accelerate framework.
+
+## Feature Flags
+
+The following features are supported:
+
+- `cuda` - Cuda GPU device (NVIDIA only)
+- `accelerate` - Accelerate framework (macOS only)

burn-candle/src/ops/tensor.rs

@@ -137,7 +137,8 @@ impl<F: FloatCandleElement, I: IntCandleElement> TensorOps<Self> for Candle<F, I
         lhs: FloatTensor<Self, D>,
         rhs: FloatTensor<Self, D>,
     ) -> FloatTensor<Self, D> {
-        CandleTensor::new(lhs.tensor.broadcast_matmul(&rhs.tensor).unwrap())
+        let rhs_contiguous = rhs.tensor.contiguous().unwrap();
+        CandleTensor::new(lhs.tensor.broadcast_matmul(&rhs_contiguous).unwrap())
     }
 
     fn swap_dims<const D: usize>(

burn-core/Cargo.toml

@@ -55,6 +55,7 @@ tch = ["burn-tch"]
 
 candle = ["burn-candle"]
 candle-cuda = ["candle", "burn-candle/cuda"]
+candle-accelerate = ["candle", "burn-candle/accelerate"]
 
 # Serialization formats
 experimental-named-tensor = ["burn-tensor/experimental-named-tensor"]

burn/Cargo.toml

@@ -51,6 +51,7 @@ wgpu = ["burn-core/wgpu"]
 tch = ["burn-core/tch"]
 candle = ["burn-core/candle"]
 candle-cuda = ["burn-core/candle-cuda"]
+candle-accelerate = ["burn-core/candle-accelerate"]
 
 # Experimental
 experimental-named-tensor = ["burn-core/experimental-named-tensor"]

xtask/src/runchecks.rs

@@ -249,6 +249,13 @@ fn burn_dataset_features_std() {
     cargo_doc(["-p", "burn-dataset", "--all-features"].into());
 }
 
+// Test burn-candle with accelerate (macOS only)
+// Leverages the macOS Accelerate framework: https://developer.apple.com/documentation/accelerate
+#[cfg(target_os = "macos")]
+fn burn_candle_accelerate() {
+    cargo_test(["-p", "burn-candle", "--features", "accelerate"].into());
+}
+
 fn std_checks() {
     // Set RUSTDOCFLAGS environment variable to treat warnings as errors
     // for the documentation build
@@ -284,6 +291,10 @@ fn std_checks() {
     // Test each workspace
     cargo_test(["--workspace"].into());
 
+    // Test burn-candle with accelerate (macOS only)
+    #[cfg(target_os = "macos")]
+    burn_candle_accelerate();
+
     // Test burn-dataset features
     burn_dataset_features_std();