[BugFix] Fix is_supported of cutlass FP8 linear (selected and fails on A100) (#55884)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
danisereb committed
Sep 16, 2026 at 22:31 UTC
2bdbbc80804b2199cbf39b75e78cf7269d0383a4
1 file changed
+3
-2
vllm/model_executor/kernels/linear/scaled_mm/cutlass.py
+3
-2
@@ -17,6 +17,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
17
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
18
CUTLASS_BLOCK_FP8_SUPPORTED,
19
convert_to_channelwise,
20
+ cutlass_fp8_supported,
21
)
22
from vllm.model_executor.utils import set_weight_attrs
23
from vllm.platforms import current_platform
@@ -164,8 +165,8 @@ class CutlassFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
165
def is_supported(
166
cls, compute_capability: int | None = None
167
) -> tuple[bool, str | None]:
167
- if not current_platform.is_cuda():
168
- return False, "requires CUDA."
168
+ if not cutlass_fp8_supported():
169
+ return False, "CUTLASS FP8 kernels not available"
170
return True, None
171
172
@classmethod