Support running FP4 Deepseek on SM120. (#11708)
This commit is contained in:
@@ -86,8 +86,8 @@ def baseline_scaled_mm(
|
||||
).to(out_dtype)
|
||||
|
||||
|
||||
def is_sm100_supported(device=None) -> bool:
|
||||
return (torch.cuda.get_device_capability(device)[0] == 10) and (
|
||||
def is_blackwell_supported(device=None) -> bool:
|
||||
return (torch.cuda.get_device_capability(device)[0] in [10, 12]) and (
|
||||
torch.version.cuda >= "12.8"
|
||||
)
|
||||
|
||||
@@ -99,7 +99,7 @@ def is_sm90_supported(device=None) -> bool:
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
not (is_sm100_supported() or is_sm90_supported()),
|
||||
not (is_blackwell_supported() or is_sm90_supported()),
|
||||
reason="fp8_blockwise_scaled_grouped_mm at sgl-kernel is only supported on sm100 or sm90",
|
||||
)
|
||||
@pytest.mark.parametrize("num_experts", [8, 16, 32, 64, 128])
|
||||
|
||||
Reference in New Issue
Block a user