biondizzle/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Make various updates and fixes: (#164)

Browse Source 
- Add BF16 support for SM90 and SM100
- Refactor Python APIs
- Other fixes and code refactoring
This commit is contained in:
Ray Wang
2025-08-15 18:32:35 +08:00
committed by GitHub
parent 3254b758e2
commit f85ec649d7
34 changed files with 2293 additions and 495 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/test_fp8.py
View File

@@ -105,7 +105,7 @@ def test_m_grouped_gemm_masked() -> None:
# Test correctness
for i in range(10):
a, b, masked_m, d, ref_d = generate_m_grouped_masked(num_groups, max_m, expected_m_per_group, n, k, use_ue8m0=use_ue8m0)
deep_gemm.fp8_m_grouped_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group, disable_ue8m0_cast=disable_ue8m0_cast)
deep_gemm.m_grouped_fp8_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group, disable_ue8m0_cast=disable_ue8m0_cast)
for j in range(num_groups):
diff = calc_diff(d[j, :masked_m[j].item()], ref_d[j, :masked_m[j].item()])
assert diff < 0.001, f'{max_m=}, {n=}, {k=}, {j=}, masked_m={masked_m[j]}, {kernel_opt}, {num_groups=}, {diff:.5f}'
@@ -115,7 +115,7 @@ def test_m_grouped_gemm_masked() -> None:
# noinspection PyShadowingNames
def test_func():
deep_gemm.fp8_m_grouped_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group, disable_ue8m0_cast=disable_ue8m0_cast)
deep_gemm.m_grouped_fp8_gemm_nt_masked(a, b, d, masked_m, expected_m_per_group, disable_ue8m0_cast=disable_ue8m0_cast)
# Test performance with fixed shapes
valid_m = masked_m.sum().item()