[SpecDecode] Reduce TP communication for large-vocab draft models speculative decoding (#39419)

Signed-off-by: EanWang211123 <wangyiheng@sangfor.com.cn>
This commit is contained in:
yiheng
2026-06-10 07:59:24 +00:00
committed by GitHub
parent bb78168b21
commit bd2d83ff31
7 changed files with 53 additions and 39 deletions
+2 -1
View File
@@ -62,6 +62,7 @@ from vllm.v1.attention.backend import AttentionType
from .adapters import as_embedding_model, as_seq_cls_model
from .interfaces import (
EagleModelMixin,
LocalArgmaxMixin,
SupportsEagle,
SupportsEagle3,
SupportsLoRA,
@@ -487,7 +488,7 @@ class LlamaModel(nn.Module, EagleModelMixin):
class LlamaForCausalLM(
nn.Module, SupportsLoRA, SupportsPP, SupportsEagle, SupportsEagle3
LocalArgmaxMixin, nn.Module, SupportsLoRA, SupportsPP, SupportsEagle, SupportsEagle3
):
packed_modules_mapping = {
"qkv_proj": ["q_proj", "k_proj", "v_proj"],