Add support for accelerate in the pyo3 bindings. (huggingface#1167)

Author: LaurentMazare

candle-pyo3/Cargo.toml

@@ -14,16 +14,18 @@ name = "candle"
 crate-type = ["cdylib"]
 
 [dependencies]
+accelerate-src = { workspace = true, optional = true }
 candle = { path = "../candle-core", version = "0.3.0", package = "candle-core" }
 candle-nn = { path = "../candle-nn", version = "0.3.0" }
 half = { workspace = true }
-pyo3 = { version = "0.19.0", features = ["extension-module"] }
 intel-mkl-src = { workspace = true, optional = true }
+pyo3 = { version = "0.19.0", features = ["extension-module"] }
 
 [build-dependencies]
 pyo3-build-config = "0.19"
 
 [features]
 default = []
+accelerate = ["dep:accelerate-src", "candle/accelerate"]
 cuda = ["candle/cuda"]
 mkl = ["dep:intel-mkl-src","candle/mkl"]

candle-pyo3/src/lib.rs

@@ -11,6 +11,9 @@ use half::{bf16, f16};
 #[cfg(feature = "mkl")]
 extern crate intel_mkl_src;
 
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
 use ::candle::{quantized::QTensor, DType, Device, Tensor, WithDType};
 
 pub fn wrap_err(err: ::candle::Error) -> PyErr {

candle-pyo3/test.py

@@ -1,5 +1,10 @@
 import candle
 
+print(f"mkl:         {candle.utils.has_mkl()}")
+print(f"accelerate:  {candle.utils.has_accelerate()}")
+print(f"num-threads: {candle.utils.get_num_threads()}")
+print(f"cuda:        {candle.utils.cuda_is_available()}")
+
 t = candle.Tensor(42.0)
 print(t)
 print(t.shape, t.rank, t.device)