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

Fix Gemma3 ModelOpt kv-scale loading (#25794)

Browse Source
This commit is contained in:
Jan Bernlöhr
2026-06-09 17:53:21 -07:00
committed by GitHub
parent f332e52611
commit f3ecc3688f
1 changed files with 10 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/gemma3_causal.py
+10
View File
@@ -857,6 +857,16 @@ class Gemma3ForCausalLM(PreTrainedModel):
params_dict = dict(self.named_parameters()) params_dict = dict(self.named_parameters())
loaded_params: Set[str] = set() loaded_params: Set[str] = set()
for name, loaded_weight in weights: for name, loaded_weight in weights:
remapped_name = maybe_remap_kv_scale_name(name, params_dict)
if remapped_name is None:
continue
if remapped_name != name:
param = params_dict[remapped_name]
weight_loader = getattr(param, "weight_loader", default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(remapped_name)
continue
for param_name, shard_name, shard_id in stacked_params_mapping: for param_name, shard_name, shard_id in stacked_params_mapping:
# if param_name in name: # if param_name in name:
# print(f"{param_name} is already in {name}") # print(f"{param_name} is already in {name}")