[Bugfix] Restrict FlashInfer cuDNN FP8 ViT attention gate to Blackwell (SM 100) (#45251)

Signed-off-by: Wentian Byte <3400259131@qq.com>
This commit is contained in:
wentian-byte
2026-06-11 16:39:24 +00:00
committed by GitHub
parent 2ec6594db9
commit b8142294b7
2 changed files with 14 additions and 6 deletions
@@ -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