huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-21 04:10:46 +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

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(())