diff --git a/vllm/model_executor/layers/attention/mm_encoder_attention.py b/vllm/model_executor/layers/attention/mm_encoder_attention.py index 1731cc26bc..2ca051ad9e 100644 --- a/vllm/model_executor/layers/attention/mm_encoder_attention.py +++ b/vllm/model_executor/layers/attention/mm_encoder_attention.py @@ -396,8 +396,9 @@ class MMEncoderAttention(CustomOp): if not is_flashinfer_cudnn_fp8_prefill_attn_supported(): raise ValueError( "mm_encoder_attn_dtype='fp8' requires the FlashInfer " - "cuDNN backend with cuDNN >= 9.17.1 on a GPU with native " - "FP8 support." + "cuDNN backend with cuDNN >= 9.17.1 on Blackwell (SM 100) " + "or newer. cuDNN's FP8 SDPA path with bf16/fp16 output is " + "not available on Hopper (H100/H200) or earlier." ) self.fp8_enabled = True diff --git a/vllm/utils/flashinfer.py b/vllm/utils/flashinfer.py index 95f8b4b7ec..e051827786 100644 --- a/vllm/utils/flashinfer.py +++ b/vllm/utils/flashinfer.py @@ -934,20 +934,27 @@ def should_use_flashinfer_for_blockscale_fp8_gemm( return should_use_flashinfer -_MIN_CUDNN_FP8 = 91701 # cuDNN >= 9.17.1 required for FP8 attention +_MIN_CUDNN_FP8 = 91701 # cuDNN >= 9.17.1 required for FP8 ViT attention @functools.cache def is_flashinfer_cudnn_fp8_prefill_attn_supported() -> bool: """Check if FP8 ViT attention is supported on this platform. - Requires native FP8 hardware support, the FlashInfer cuDNN backend, + Requires Blackwell (SM 100) or newer, the FlashInfer cuDNN backend, and cuDNN >= 9.17.1. + + cuDNN's FP8 SDPA forward path with bf16/fp16 output (used by + ``MMEncoderAttention._forward_flashinfer``) gates internally on + ``prop.major >= 10``; on Hopper it raises a misleading + ``cudnnGraphNotSupportedError: ... cuDNN version 9.13.0 and newer`` + even when the installed cuDNN is new enough. See PR #38065 for the + original Blackwell-only design intent. """ from vllm.v1.attention.backends.registry import AttentionBackendEnum - # cuDNN SDPA FP8 requires Hopper (SM 90) or newer. - if not current_platform.has_device_capability(90): + # cuDNN SDPA FP8 with bf16/fp16 output requires Blackwell (SM 100) or newer. + if not current_platform.has_device_capability(100): return False try: