diff --git a/vllm/model_executor/kernels/linear/mxfp8/flashinfer.py b/vllm/model_executor/kernels/linear/mxfp8/flashinfer.py index d26e5579ed..3150d93f99 100644 --- a/vllm/model_executor/kernels/linear/mxfp8/flashinfer.py +++ b/vllm/model_executor/kernels/linear/mxfp8/flashinfer.py @@ -11,7 +11,7 @@ from vllm.model_executor.layers.quantization.utils.mxfp8_utils import ( ) from vllm.platforms import current_platform from vllm.utils import flashinfer as vllm_flashinfer -from vllm.utils.flashinfer import has_flashinfer_cutedsl +from vllm.utils.flashinfer import has_flashinfer, has_flashinfer_cutedsl from .Mxfp8LinearKernel import Mxfp8LinearKernel, Mxfp8LinearLayerConfig @@ -23,9 +23,13 @@ class FlashInferCutlassMxfp8LinearKernel(Mxfp8LinearKernel): def is_supported( cls, compute_capability: int | None = None ) -> tuple[bool, str | None]: - if current_platform.has_device_capability(100): - return True, None - return False, "requires >=sm_100 (Blackwell)" + if not ( + current_platform.is_cuda() and current_platform.has_device_capability(100) + ): + return False, "requires >=sm_100 (Blackwell)" + if not has_flashinfer(): + return False, "requires FlashInfer" + return True, None @classmethod def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]: diff --git a/vllm/model_executor/layers/quantization/utils/mxfp8_utils.py b/vllm/model_executor/layers/quantization/utils/mxfp8_utils.py index 932bf92352..1e45ea9ab0 100644 --- a/vllm/model_executor/layers/quantization/utils/mxfp8_utils.py +++ b/vllm/model_executor/layers/quantization/utils/mxfp8_utils.py @@ -177,8 +177,9 @@ def _mxfp8_e4m3_quantize_impl( alignment: int = 0, ) -> tuple[torch.Tensor, torch.Tensor]: from vllm.platforms import current_platform + from vllm.utils.flashinfer import has_flashinfer - if current_platform.has_device_capability(100): + if current_platform.has_device_capability(100) and has_flashinfer(): from flashinfer import mxfp8_quantize as flashinfer_mxfp8_quantize x_q, x_scales = flashinfer_mxfp8_quantize(