[Bugfix] Fix speculative decoding for short_conv (LFM2) models (#50272)

Signed-off-by: jens.luecke <jens.luecke@liquid.ai>
Co-authored-by: Tonoken3 <sakamakismile@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Jens
2026-08-22 13:34:22 +00:00
committed by GitHub
co-authored by Tonoken3 mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 704f12aa1e
commit da329cc303
4 changed files with 43 additions and 11 deletions
@@ -12,6 +12,7 @@ from vllm.config.compilation import CUDAGraphMode
from vllm.triton_utils import tl, triton
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder
from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionMetadataBuilder
from vllm.v1.core.sched.output import NewRequestData
from vllm.v1.kv_cache_interface import KVCacheConfig, MambaSpec
from vllm.v1.utils import CpuGpuBuffer
@@ -48,7 +49,11 @@ class MambaHybridAttnMetadata(ModelSpecificAttnMetadata):
) -> dict[str, Any]:
if not isinstance(
attn_metadata_builder,
(Mamba2AttentionMetadataBuilder, GDNAttentionMetadataBuilder),
(
Mamba2AttentionMetadataBuilder,
GDNAttentionMetadataBuilder,
ShortConvAttentionMetadataBuilder,
),
):
return {}
return {