huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-16 02:38:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix the matmul layout for accelerate & mkl. (#2011)

Browse Source 
* Fix the matmul layout for accelerate & mkl.

* Reduce the required precision for pow (because of accelerate).

* And a fix the gelu f16 test.
This commit is contained in:
Laurent Mazare
2024-04-04 19:18:03 +02:00
committed by GitHub
parent 5aebe53dd2
commit e6a5b82ba6
3 changed files with 8 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
candle-core/src/cpu_backend/mod.rs
View File

@ -1330,7 +1330,7 @@ impl Map2 for MatMul {
let lhs_m1 = lhs_stride[lhs_stride.len() - 1];
let lhs_m2 = lhs_stride[lhs_stride.len() - 2];
let (lda, transa) = if rhs_m1 == 1 && rhs_m2 == n {
let (lda, transa) = if (rhs_m1 == 1 || n == 1) && (rhs_m2 == n || k == 1) {
(n as i32, b'N')
} else if rhs_m1 == k && rhs_m2 == 1 {
(k as i32, b'T')
@ -1338,7 +1338,7 @@ impl Map2 for MatMul {
Err(self.striding_error(lhs_l, rhs_l, "non-contiguous rhs"))?
};
// The b tensor has dims batching, m, k (lhs)
let (ldb, transb) = if lhs_m1 == 1 && lhs_m2 == k {
let (ldb, transb) = if (lhs_m1 == 1 || k == 1) && (lhs_m2 == k || m == 1) {
(k as i32, b'N')
} else if lhs_m1 == m && lhs_m2 == 1 {
(m as i32, b'T')
@ -1421,7 +1421,7 @@ impl Map2 for MatMul {
let lhs_m1 = lhs_stride[lhs_stride.len() - 1];
let lhs_m2 = lhs_stride[lhs_stride.len() - 2];
let (lda, transa) = if rhs_m1 == 1 && rhs_m2 == n {
let (lda, transa) = if (rhs_m1 == 1 || n == 1) && (rhs_m2 == n || k == 1) {
(n as i32, b'N')
} else if rhs_m1 == k && rhs_m2 == 1 {
(k as i32, b'T')
@ -1429,7 +1429,7 @@ impl Map2 for MatMul {
Err(self.striding_error(lhs_l, rhs_l, "non-contiguous rhs"))?
};
// The b tensor has dims batching, m, k (lhs)
let (ldb, transb) = if lhs_m1 == 1 && lhs_m2 == k {
let (ldb, transb) = if (lhs_m1 == 1 || k == 1) && (lhs_m2 == k || m == 1) {
(k as i32, b'N')
} else if lhs_m1 == m && lhs_m2 == 1 {
(m as i32, b'T')

13
candle-core/tests/matmul_tests.rs
View File

@ -73,20 +73,7 @@ fn squeeze_mm(device: &Device) -> Result<()> {
let seq_len = 8_usize;
let a = Tensor::zeros((1, seq_len, 16), DType::F32, device)?;
let x = a.i((.., seq_len - 1, ..))?;
println!(
"x shape:{:?}, stride:{:?}, is_contiguous:{}",
x.shape(),
x.stride(),
x.is_contiguous()
);
let w = Tensor::zeros((32, 16), DType::F32, device)?.t()?;
println!(
"w shape:{:?}, stride:{:?}, is_contiguous:{}",
w.shape(),
w.stride(),
w.is_contiguous()
);
let x = x.matmul(&w)?;
assert_eq!(x.dims(), &[1, 32]);
Ok(())

13
candle-core/tests/tensor_tests.rs
View File

@ -107,13 +107,8 @@ fn unary_op(device: &Device) -> Result<()> {
]
);
let t_f16 = tensor.to_dtype(DType::F16)?.gelu()?.to_dtype(DType::F32)?;
assert_eq!(
test_utils::to_vec2_round(&t_f16, 2)?,
[
[-0.0, 0.84, 4.0, -0.05, 0.35],
[2.69, -0.07, -0.11, 1.73, 2.79]
],
);
let max_diff = (tensor.gelu()? - t_f16)?.flatten_all()?.max(0)?;
assert!(max_diff.to_vec0::<f32>()? < 5e-3);
assert_eq!(
test_utils::to_vec2_round(&tensor.gelu_erf()?, 4)?,
[
@ -1255,8 +1250,8 @@ fn pow() -> Result<()> {
let rhs = (&lhs - 2.)?;
let res = lhs.pow(&rhs)?;
assert_eq!(
test_utils::to_vec2_round(&res, 4)?,
[[1.0, 1.0, 3.0], [16.0, 125.0, 1296.0001]]
test_utils::to_vec2_round(&res, 3)?,
[[1.0, 1.0, 3.0], [16.0, 125.0, 1296.0]]
);
Ok(())
}