huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-21 12:20:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add a minimal test for the metal bf16 matmul. (#2381)

Browse Source
This commit is contained in:
Laurent Mazare
2024-08-01 10:22:46 +01:00
committed by GitHub
parent 957d604a78
commit d4b6f6eef6
1 changed files with 20 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
candle-core/tests/matmul_tests.rs
View File

@ -49,6 +49,20 @@ fn matmul(device: &Device) -> Result<()> {
Ok(()) Ok(())
} }
fn matmul_bf16(device: &Device) -> Result<()> {
if !device.supports_bf16() {
return Ok(());
}
let data = vec![1.0f32, 2.0, 3.0, 4.0];
let a = Tensor::from_slice(&data, (2, 2), device)?.to_dtype(DType::BF16)?;
let data = vec![1.0f32, 2.0, 3.0, 4.0];
let b = Tensor::from_slice(&data, (2, 2), device)?.to_dtype(DType::BF16)?;
let c = a.matmul(&b)?.to_dtype(DType::F32)?;
assert_eq!(c.to_vec2::<f32>()?, &[[7.0f32, 10.0], [15.0, 22.0]]);
Ok(())
}
fn broadcast_matmul(device: &Device) -> Result<()> { fn broadcast_matmul(device: &Device) -> Result<()> {
let lhs = Tensor::randn(0f32, 1f32, (3, 1, 4, 5), device)?; let lhs = Tensor::randn(0f32, 1f32, (3, 1, 4, 5), device)?;
let rhs = Tensor::randn(0f32, 1f32, (6, 5, 2), device)?; let rhs = Tensor::randn(0f32, 1f32, (6, 5, 2), device)?;
@ -96,6 +110,12 @@ fn mm_layout(device: &Device) -> Result<()> {
} }
test_device!(matmul, matmul_cpu, matmul_gpu, matmul_metal); test_device!(matmul, matmul_cpu, matmul_gpu, matmul_metal);
test_device!(
matmul_bf16,
matmul_bf16_cpu,
matmul_bf16_gpu,
matmul_bf16_metal
);
test_device!( test_device!(
broadcast_matmul, broadcast_matmul,
broadcast_matmul_cpu, broadcast_matmul_cpu,