[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
@@ -59,8 +59,6 @@ class DSparkSpeculator(DFlashSpeculator):
self.max_num_tokens, draft_hidden, dtype=self.dtype, device=device
)
self.dflash_causal = False
self._step_cols = torch.arange(
self.num_speculative_steps, dtype=torch.int32, device=device
)
@@ -18,14 +18,13 @@ def load_dspark_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Mo
draft_model_config = speculative_config.draft_model_config
from vllm.compilation.backends import set_model_tag
from vllm.model_executor.models.qwen3_dflash import dflash_has_any_non_causal
# DSpark uses non-causal attention.
causal = False
draft_vllm_config = replace(
vllm_config,
attention_config=replace(
vllm_config.attention_config,
use_non_causal=not causal,
use_non_causal=dflash_has_any_non_causal(draft_model_config.hf_config),
backend=speculative_config.attention_backend,
),
)