[Spec Decode] Enable Gemma4 DSpark adaptive verification with FlashInfer (#57263)

Signed-off-by: zixi-qi <zixi@inferact.ai>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
qizixi
2026-09-26 17:01:17 +08:00
committed by GitHub
co-authored by OpenAI Codex Claude Opus 5.5 Nick Hill
parent 3b4566c5cf
commit a4eb3f25d6
24 changed files with 894 additions and 135 deletions
+6 -5
View File
@@ -52,7 +52,11 @@ from vllm.models.minimax_m3.common.sparse_attention import (
_minimax_m3_aiter_sparse_pa_requested,
)
from vllm.platforms import current_platform
from vllm.v1.attention.backend import AttentionBackend, CommonAttentionMetadata
from vllm.v1.attention.backend import (
AttentionBackend,
CommonAttentionMetadata,
max_decode_query_len,
)
from vllm.v1.attention.backends.utils import split_decodes_and_prefills
from vllm.v1.kv_cache_interface import AttentionSpec
@@ -97,10 +101,7 @@ def aiter_indexer_max_decode_query_len(vllm_config: VllmConfig) -> int:
since that is what the builder splits the batch on and therefore what the
decode kernel will actually be handed.
"""
spec = vllm_config.speculative_config
if spec is None or spec.num_speculative_tokens is None:
return 1
return 1 + (2 if spec.parallel_drafting else 1) * spec.num_speculative_tokens
return max_decode_query_len(vllm_config)
@cache