[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
+8 -2
View File
@@ -48,7 +48,13 @@ from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.config import set_default_rope_theta
from vllm.v1.attention.backend import AttentionType
from .interfaces import SupportsEagle, SupportsEagle3, SupportsLoRA, SupportsPP
from .interfaces import (
LocalArgmaxMixin,
SupportsEagle,
SupportsEagle3,
SupportsLoRA,
SupportsPP,
)
from .qwen2 import Qwen2MLP as Qwen3MLP
from .qwen2 import Qwen2Model
from .utils import AutoWeightsLoader, PPMissingLayer, extract_layer_index, maybe_prefix
@@ -259,7 +265,7 @@ class Qwen3Model(Qwen2Model):
class Qwen3ForCausalLM(
nn.Module, SupportsLoRA, SupportsPP, SupportsEagle, SupportsEagle3
LocalArgmaxMixin, nn.Module, SupportsLoRA, SupportsPP, SupportsEagle, SupportsEagle3
):
packed_modules_mapping = {
"qkv_proj": [