minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[bug fix][pp] fix qwen3 model load (#15223)

Browse Source 
Signed-off-by: Xuchun Shang <xuchun.shang@gmail.com>
This commit is contained in:
Xuchun Shang
2025-12-17 17:26:22 +08:00
committed by GitHub
parent 888594333e
commit e9abb52576
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen3.py
+3 -3
View File
@@ -392,13 +392,13 @@ class Qwen3ForCausalLM(nn.Module):
if self.pp_group.world_size > 1 and config.tie_word_embeddings: if self.pp_group.world_size > 1 and config.tie_word_embeddings:
if self.pp_group.is_first_rank: if self.pp_group.is_first_rank:
self.pp_group.send( self.pp_group.send(
self.model.embed_tokens.weight, dst=self.pp_group.last_rank self.model.embed_tokens.weight, dst=self.pp_group.world_size - 1
) )
elif self.pp_group.is_last_rank: elif self.pp_group.is_last_rank:
emb_token_weight = self.pp_group.recv( emb_token_weight = self.pp_group.recv(
size=(config.vocab_size, config.hidden_size), size=self.lm_head.weight.shape,
dtype=next(self.model.parameters()).dtype, dtype=next(self.model.parameters()).dtype,
src=self.pp_group.first_rank, src=0,
) )
self.lm_head.weight.copy_(emb_token_weight) self.lm_head.weight.copy_(emb_token_weight)