huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-16 18:48:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Splitting the features to enable different mkl linking.

Browse Source
This commit is contained in:
Nicolas Patry
2025-03-28 10:13:13 +01:00
parent cb02b389d5
commit 9862cd3ba2
23 changed files with 42 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
Cargo.toml
View File

@ -51,7 +51,7 @@ half = { version = "2.5.0", features = ["num-traits", "use-intrinsics", "rand_di
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"] }
intel-mkl-src = { version = "0.8.1" }
libc = { version = "0.2.147" }
log = "0.4"
memmap2 = { version = "0.9.3", features = ["stable_deref_trait"] }

3
candle-core/Cargo.toml
View File

@ -45,7 +45,8 @@ criterion = { workspace = true }
default = []
cuda = ["cudarc", "dep:candle-kernels", "dep:ug-cuda"]
cudnn = ["cuda", "cudarc/cudnn"]
mkl = ["dep:libc", "dep:intel-mkl-src"]
_mkl = ["dep:libc", "dep:intel-mkl-src"]
mkl = ["_mkl", "intel-mkl-src?/mkl-static-lp64-iomp"]
accelerate = ["dep:libc", "dep:accelerate-src"]
metal = ["dep:metal", "dep:candle-metal-kernels", "dep:ug-metal"]

2
candle-core/benches/benchmarks/mod.rs
View File

@ -39,7 +39,7 @@ impl BenchDevice for Device {
Device::Cpu => {
let cpu_type = if cfg!(feature = "accelerate") {
"accelerate"
} else if cfg!(feature = "mkl") {
} else if cfg!(feature = "_mkl") {
"mkl"
} else {
"cpu"

2
candle-core/examples/basics.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-core/examples/cuda_basics.rs
View File

@ -1,7 +1,7 @@
#[cfg(feature = "accelerate")]
extern crate accelerate_src;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
use anyhow::Result;

2
candle-core/examples/cuda_sum_benchmark.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-core/examples/metal_basics.rs
View File

@ -1,7 +1,7 @@
#[cfg(feature = "accelerate")]
extern crate accelerate_src;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
use anyhow::Result;

4
candle-core/src/cpu_backend/mod.rs
View File

@ -1246,7 +1246,7 @@ impl MatMul {
impl Map2 for MatMul {
const OP: &'static str = "mat_mul";
#[cfg(all(not(feature = "mkl"), not(feature = "accelerate")))]
#[cfg(all(not(feature = "_mkl"), not(feature = "accelerate")))]
fn f<T: 'static + WithDType + num_traits::Num + Copy>(
&self,
lhs: &[T],
@ -1411,7 +1411,7 @@ impl Map2 for MatMul {
Ok(dst)
}
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
fn f<T: 'static + WithDType + num_traits::Num + Copy>(
&self,
lhs: &[T],

4
candle-core/src/lib.rs
View File

@ -68,7 +68,7 @@ mod indexer;
pub mod layout;
#[cfg(feature = "metal")]
pub mod metal_backend;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
mod mkl;
pub mod npy;
pub mod op;
@ -118,7 +118,7 @@ pub use metal_backend::{MetalDevice, MetalError, MetalStorage};
#[cfg(not(feature = "metal"))]
pub use dummy_metal_backend::{MetalDevice, MetalError, MetalStorage};
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

32
candle-core/src/op.rs
View File

@ -294,16 +294,16 @@ macro_rules! bin_op {
$e(v1, v2)
}
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
const F32_VEC: bool = true;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
const F64_VEC: bool = true;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
#[inline(always)]
fn f32_vec(xs1: &[f32], xs2: &[f32], ys: &mut [f32]) {
crate::mkl::$f32_vec(xs1, xs2, ys)
}
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
#[inline(always)]
fn f64_vec(xs1: &[f64], xs2: &[f64], ys: &mut [f64]) {
crate::mkl::$f64_vec(xs1, xs2, ys)
@ -418,16 +418,16 @@ macro_rules! unary_op {
todo!("no unary function for i64")
}
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
const F32_VEC: bool = true;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
const F64_VEC: bool = true;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
#[inline(always)]
fn f32_vec(xs: &[f32], ys: &mut [f32]) {
crate::mkl::$f32_vec(xs, ys)
}
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
#[inline(always)]
fn f64_vec(xs: &[f64], ys: &mut [f64]) {
crate::mkl::$f64_vec(xs, ys)
@ -518,19 +518,19 @@ impl UnaryOpT for Gelu {
}
const KERNEL: &'static str = "ugelu";
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
const F32_VEC: bool = true;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
#[inline(always)]
fn f32_vec(xs: &[f32], ys: &mut [f32]) {
crate::mkl::vs_gelu(xs, ys)
}
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
const F64_VEC: bool = true;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
#[inline(always)]
fn f64_vec(xs: &[f64], ys: &mut [f64]) {
crate::mkl::vd_gelu(xs, ys)
@ -625,19 +625,19 @@ impl UnaryOpT for Silu {
}
const KERNEL: &'static str = "usilu";
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
const F32_VEC: bool = true;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
#[inline(always)]
fn f32_vec(xs: &[f32], ys: &mut [f32]) {
crate::mkl::vs_silu(xs, ys)
}
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
const F64_VEC: bool = true;
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
#[inline(always)]
fn f64_vec(xs: &[f64], ys: &mut [f64]) {
crate::mkl::vd_silu(xs, ys)

2
candle-core/src/utils.rs
View File

@ -17,7 +17,7 @@ pub fn has_accelerate() -> bool {
}
pub fn has_mkl() -> bool {
cfg!(feature = "mkl")
cfg!(feature = "_mkl")
}
pub fn cuda_is_available() -> bool {

3
candle-nn/Cargo.toml
View File

@ -33,7 +33,8 @@ criterion = { workspace = true }
default = []
accelerate = ["dep:accelerate-src", "candle/accelerate"]
cuda = ["candle/cuda"]
mkl = ["dep:intel-mkl-src", "candle/mkl"]
mkl = ["candle/mkl"]
_mkl = ["dep:intel-mkl-src", "candle/_mkl"]
metal = ["candle/metal", "dep:candle-metal-kernels", "dep:metal"]
[[bench]]

2
candle-nn/benches/benchmarks/mod.rs
View File

@ -34,7 +34,7 @@ impl BenchDevice for Device {
Device::Cpu => {
let cpu_type = if cfg!(feature = "accelerate") {
"accelerate"
} else if cfg!(feature = "mkl") {
} else if cfg!(feature = "_mkl") {
"mkl"
} else {
"cpu"

2
candle-nn/examples/basic_optimizer.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-nn/examples/cpu_benchmarks.rs
View File

@ -1,5 +1,5 @@
/// This example contains some simple benchmarks so that it's easy to run them in perf etc.
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-nn/tests/batch_norm.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-nn/tests/group_norm.rs
View File

@ -18,7 +18,7 @@ t = torch.tensor(
print(group_norm(t, num_groups=2))
print(group_norm(t, num_groups=3))
*/
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-nn/tests/kv_cache.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-nn/tests/layer_norm.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-nn/tests/loss.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-nn/tests/ops.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-nn/tests/optim.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]

2
candle-nn/tests/rnn.rs
View File

@ -1,4 +1,4 @@
#[cfg(feature = "mkl")]
#[cfg(feature = "_mkl")]
extern crate intel_mkl_src;
#[cfg(feature = "accelerate")]