[Feature] Support MTP speculative decoding for Bailing hybrid models (#44880)

Signed-off-by: zc02384840 <zc02384840@antgroup.com>
Co-authored-by: zc02384840 <zc02384840@antgroup.com>
This commit is contained in:
alex101-ops
2026-07-06 10:38:50 +08:00
committed by GitHub
co-authored by zc02384840
parent 95a248faed
commit f2aaf59151
11 changed files with 1200 additions and 6 deletions
+12 -2
View File
@@ -132,6 +132,9 @@ from vllm.v1.attention.backend import (
CommonAttentionMetadata,
)
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
from vllm.v1.attention.backends.linear_attn import (
BailingLinearAttentionMetadataBuilder,
)
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder
from vllm.v1.attention.backends.utils import (
NULL_BLOCK_ID,
@@ -2442,9 +2445,16 @@ class GPUModelRunner(
extra_attn_metadata_args = {}
if use_spec_decode and isinstance(
builder, (Mamba2AttentionMetadataBuilder, GDNAttentionMetadataBuilder)
builder,
(
Mamba2AttentionMetadataBuilder,
GDNAttentionMetadataBuilder,
BailingLinearAttentionMetadataBuilder,
),
):
assert ubid is None, "UBatching not supported with GDN yet"
assert ubid is None, (
"UBatching not supported with GDN or linear attn yet"
)
extra_attn_metadata_args = dict(
num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs_padded],
num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[