[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
@@ -22,6 +22,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.models.interfaces import LocalArgmaxMixin
from vllm.model_executor.models.qwen3_5 import Qwen3_5DecoderLayer, Qwen3_5RMSNorm
from vllm.model_executor.models.qwen3_next import QwenNextMixtureOfExperts
from vllm.sequence import IntermediateTensors
@@ -353,7 +354,7 @@ class Qwen3_5MultiTokenPredictor(nn.Module):
"hidden_states": 0,
}
)
class Qwen3_5MTP(nn.Module, SupportsMultiModal):
class Qwen3_5MTP(LocalArgmaxMixin, nn.Module, SupportsMultiModal):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",