EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Misc] Use pytest.mark.skipif in sgl-kernel test (#5137)

Browse Source
This commit is contained in:
yinfan98
2025-04-08 12:35:14 +08:00
committed by GitHub
parent ade714a67f
commit 9798e72baa
3 changed files with 29 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
sgl-kernel/tests/test_fp4_gemm.py
View File

@@ -2,11 +2,7 @@ import pytest
import torch
from sgl_kernel import cutlass_scaled_fp4_mm, scaled_fp4_quant
if torch.cuda.get_device_capability() < (10, 0):
pytest.skip(
reason="Nvfp4 Requires compute capability of 10 or above.",
allow_module_level=True,
)
skip_condition = torch.cuda.get_device_capability() < (10, 0)
DTYPES = [torch.float16, torch.bfloat16]
# m, n, k
@@ -108,6 +104,9 @@ def get_ref_results(
return torch.matmul(a_in_dtype, b_in_dtype.t())
@pytest.mark.skipif(
skip_condition, reason="Nvfp4 Requires compute capability of 10 or above."
)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("shape", SHAPES)
@torch.inference_mode()
@@ -149,3 +148,7 @@ def test_nvfp4_gemm(
)
torch.testing.assert_close(out, expected_out.to(dtype=dtype), atol=1e-1, rtol=1e-1)
if __name__ == "__main__":
pytest.main([__file__])

16
sgl-kernel/tests/test_fp4_quantize.py
View File

@@ -2,11 +2,7 @@ import pytest
import torch
from sgl_kernel import scaled_fp4_quant
if torch.cuda.get_device_capability() < (10, 0):
pytest.skip(
reason="Nvfp4 Requires compute capability of 10 or above.",
allow_module_level=True,
)
skip_condition = torch.cuda.get_device_capability() < (10, 0)
DTYPES = [torch.float16, torch.bfloat16]
SHAPES = [(128, 64), (128, 128), (256, 64), (256, 128)]
@@ -115,6 +111,9 @@ def recover_swizzled_scales(scale, m, n):
return result[:m, :scale_n]
@pytest.mark.skipif(
skip_condition, reason="Nvfp4 Requires compute capability of 10 or above."
)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("shape", SHAPES)
@torch.inference_mode()
@@ -140,6 +139,9 @@ def test_quantize_to_fp4(
torch.testing.assert_close(scale_ans, scale_ref)
@pytest.mark.skipif(
skip_condition, reason="Nvfp4 Requires compute capability of 10 or above."
)
@pytest.mark.parametrize("pad_shape", PAD_SHAPES)
@torch.inference_mode()
def test_quantize_to_fp4_padded(pad_shape: tuple[int, int]) -> None:
@@ -162,3 +164,7 @@ def test_quantize_to_fp4_padded(pad_shape: tuple[int, int]) -> None:
torch.testing.assert_close(out_ans, out_ref)
torch.testing.assert_close(scale_ans, scale_ref)
if __name__ == "__main__":
pytest.main([__file__])