[BugFix] Fix is_supported of cutlass FP8 linear (selected and fails on A100) (#55884)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
This commit is contained in:
@@ -17,6 +17,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
|||||||
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
|
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
|
||||||
CUTLASS_BLOCK_FP8_SUPPORTED,
|
CUTLASS_BLOCK_FP8_SUPPORTED,
|
||||||
convert_to_channelwise,
|
convert_to_channelwise,
|
||||||
|
cutlass_fp8_supported,
|
||||||
)
|
)
|
||||||
from vllm.model_executor.utils import set_weight_attrs
|
from vllm.model_executor.utils import set_weight_attrs
|
||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
@@ -164,8 +165,8 @@ class CutlassFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
|
|||||||
def is_supported(
|
def is_supported(
|
||||||
cls, compute_capability: int | None = None
|
cls, compute_capability: int | None = None
|
||||||
) -> tuple[bool, str | None]:
|
) -> tuple[bool, str | None]:
|
||||||
if not current_platform.is_cuda():
|
if not cutlass_fp8_supported():
|
||||||
return False, "requires CUDA."
|
return False, "CUTLASS FP8 kernels not available"
|
||||||
return True, None
|
return True, None
|
||||||
|
|
||||||
@classmethod
|
@classmethod
|
||||||
|
|||||||
Reference in New Issue
Block a user