[Bugfix] Fix FlashInfer non-causal draft attention (DFlash/DSpark) on Blackwell (#48167)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Michael Goin
2026-07-15 12:44:01 -07:00
committed by GitHub
co-authored by Claude Fable 5
parent 49e777cf08
commit ecf4aa5ce2
12 changed files with 149 additions and 47 deletions
+4 -3
View File
@@ -472,9 +472,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
decode_query_len=self.decode_query_len,
lora_capture_cases=self.lora_capture_cases,
)
if self.speculator is not None:
self.speculator.init_cudagraph_manager(cudagraph_mode)
check_attention_cp_compatibility(self.vllm_config)
if isinstance(self.speculator, DraftModelSpeculator):
# HACK(woosuk)
@@ -485,6 +482,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
self.input_buffers,
self.attn_groups,
)
if self.speculator is not None:
# After set_attn, so the speculator can size its cudagraph mode
# to its own attention support.
self.speculator.init_cudagraph_manager(cudagraph_mode)
self.kv_caches: list[torch.Tensor] = []
kv_caches_dict = init_kv_cache(