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

[Bugfix] Fix PP tied embeddings weight loading for qwen3.5 4B dense model (#21347)

Browse Source
This commit is contained in:
Wenyao Gao
2026-04-01 14:51:03 +08:00
committed by GitHub
parent a188208e9a
commit 2861596fc6
1 changed files with 22 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen3_5.py
+22
View File
@@ -1384,6 +1384,17 @@ class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration):
name = name.replace(r"model.language_model.", r"model.")
if ".self_attn." in name:
name = name.replace(".self_attn", "")
if (
self.config.tie_word_embeddings
and self.pp_group.is_last_rank
and "model.embed_tokens.weight" in name
):
if "lm_head.weight" in params_dict:
lm_head_param = params_dict["lm_head.weight"]
weight_loader = getattr(
lm_head_param, "weight_loader", default_weight_loader
)
weight_loader(lm_head_param, loaded_weight)
layer_id = get_layer_id(name)
if (
layer_id is not None
@@ -1549,6 +1560,17 @@ class Qwen3_5MoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
name = name.replace(r"model.language_model.", r"model.")
if ".self_attn." in name:
name = name.replace(".self_attn", "")
if (
self.config.tie_word_embeddings
and self.pp_group.is_last_rank
and "model.embed_tokens.weight" in name
):
if "lm_head.weight" in params_dict:
lm_head_param = params_dict["lm_head.weight"]
weight_loader = getattr(
lm_head_param, "weight_loader", default_weight_loader
)
weight_loader(lm_head_param, loaded_weight)
layer_id = get_layer_id(name)
if (