[Perf] FP8 FlashInfer Attn for ViT (#38065)

Signed-off-by: Zhanda Zhu <zhandazhu@gmail.com>
Co-authored-by: Yubo Gao <ybgao-nvidia@users.noreply.github.com>
This commit is contained in:
Zhanda Zhu
2026-04-27 13:44:15 +08:00
committed by GitHub
co-authored by Yubo Gao
parent 592ae6805c
commit 5d5c776444
18 changed files with 1829 additions and 49 deletions
+35
View File
@@ -772,6 +772,40 @@ def should_use_flashinfer_for_blockscale_fp8_gemm(
return should_use_flashinfer
_MIN_CUDNN_FP8 = 91701 # cuDNN >= 9.17.1 required for FP8 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,
and cuDNN >= 9.17.1.
"""
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):
return False
try:
supported = current_platform.get_supported_vit_attn_backends()
if AttentionBackendEnum.FLASHINFER not in supported:
return False
except (ImportError, AttributeError):
return False
try:
import torch.backends.cudnn as cudnn
if cudnn.is_available() and cudnn.version() < _MIN_CUDNN_FP8:
return False
except (ImportError, AttributeError):
pass
return True
__all__ = [
"has_flashinfer",
"flashinfer_trtllm_fp8_block_scale_moe",
@@ -803,4 +837,5 @@ __all__ = [
"flashinfer_fp8_blockscale_gemm",
"should_use_flashinfer_for_blockscale_fp8_gemm",
"is_flashinfer_fp8_blockscale_gemm_supported",
"is_flashinfer_cudnn_fp8_prefill_attn_supported",
]