mirror of
https://github.com/deepseek-ai/DeepGEMM
synced 2025-06-26 23:15:49 +00:00
Minor fix
This commit is contained in:
parent
3bd234e79c
commit
b3726b423c
@ -215,16 +215,15 @@ def test_m_grouped_gemm_contiguous() -> None:
|
|||||||
def test_m_grouped_gemm_masked() -> None:
|
def test_m_grouped_gemm_masked() -> None:
|
||||||
print('Testing grouped masked GEMM:')
|
print('Testing grouped masked GEMM:')
|
||||||
|
|
||||||
max_m = 4096
|
|
||||||
for num_groups, expected_m_per_group in ((1, 1024), (2, 512), (4, 256)):
|
for num_groups, expected_m_per_group in ((1, 1024), (2, 512), (4, 256)):
|
||||||
for k, n in ((7168, 4096), (2048, 7168), ):
|
for k, n in ((7168, 4096), (2048, 7168), ):
|
||||||
# Test correctness
|
# Test correctness
|
||||||
for i in range(10):
|
for i in range(10):
|
||||||
x_fp8, y_fp8, masked_m, out, ref_out = construct_masked_grouped(num_groups, max_m, expected_m_per_group, k, n)
|
x_fp8, y_fp8, masked_m, out, ref_out = construct_masked_grouped(num_groups, 4096, expected_m_per_group, k, n)
|
||||||
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_masked(x_fp8, y_fp8, out, masked_m, expected_m_per_group)
|
deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_masked(x_fp8, y_fp8, out, masked_m, expected_m_per_group)
|
||||||
for j in range(num_groups):
|
for j in range(num_groups):
|
||||||
diff = calc_diff(out[j, :masked_m[j].item()], ref_out[j, :masked_m[j].item()])
|
diff = calc_diff(out[j, :masked_m[j].item()], ref_out[j, :masked_m[j].item()])
|
||||||
assert diff < 0.001, f'{max_m=}, {k=}, {n=}, {j=}, masked_m={masked_m[j]}, {num_groups=}, {diff:.5f}'
|
assert diff < 0.001, f'{expected_m_per_group=}, {k=}, {n=}, {j=}, masked_m={masked_m[j]}, {num_groups=}, {diff:.5f}'
|
||||||
|
|
||||||
# noinspection PyShadowingNames
|
# noinspection PyShadowingNames
|
||||||
def test_func():
|
def test_func():
|
||||||
|
|||||||
Loading…
Reference in New Issue
Block a user