A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 06:15 UTC
[refactor] Add prefix support to embed_tokens in DeepSeek MTP (#30788)
Signed-off-by: zzhx1 <zzh_201018@outlook.com>
zzhxxx committed
Dec 18, 2025 at 12:45 UTC
b166ef20e1e5256913b34456507d89850a8dba38
1 file changed
+1
vllm/model_executor/models/deepseek_mtp.py
+1
@@ -141,6 +141,7 @@ class DeepSeekMultiTokenPredictor(nn.Module):
141
self.embed_tokens = VocabParallelEmbedding(
142
config.vocab_size,
143
config.hidden_size,
144
+ prefix=maybe_prefix(prefix, "embed_tokens"),
145
)
146
self.logits_processor = LogitsProcessor(config.vocab_size)
147