[Perf] spec: Let the FlashAttention backend capture the draft-extend CUDA graph (#40762)
This commit is contained in:
@@ -554,6 +554,13 @@ class EagleDraftWorker(EagleDraftWorkerBase):
|
||||
|
||||
graph_supported_backend_types.append(DeepseekV4AttnBackend)
|
||||
if _is_cuda:
|
||||
# FlashAttention is CUDA-only here; import lazily so MUSA and NPU
|
||||
# builds don't pull sgl_kernel in at import time.
|
||||
from sglang.srt.layers.attention.flashattention_backend import (
|
||||
FlashAttentionBackend,
|
||||
)
|
||||
|
||||
graph_supported_backend_types.append(FlashAttentionBackend)
|
||||
# FlashMLA is CUDA-only; import lazily so CPU builds don't pull
|
||||
# sgl_kernel.flash_mla at import time.
|
||||
from sglang.srt.layers.attention.flashmla_backend import FlashMLABackend
|
||||
|
||||
Reference in New Issue
Block a user