[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
+5 -1
View File
@@ -70,7 +70,11 @@ class DFlashProposer(SpecDecodeBaseProposer):
# For DFlash we use the input embeddings to embed the mask token
self.parallel_drafting_hidden_state_tensor = None
self.dflash_causal = self.dflash_config.get("causal", False)
from vllm.model_executor.models.qwen3_dflash import dflash_has_any_non_causal
self.dflash_causal = not dflash_has_any_non_causal(
self.draft_model_config.hf_config
)
@override
def _create_draft_vllm_config(self) -> VllmConfig: