EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix incorrect LoRA weight loading for fused gate_up_proj (#6734)

Browse Source
This commit is contained in:
Lifu Huang
2025-05-31 13:41:44 -07:00
committed by GitHub
parent 888cb175a6
commit 094fbdacd5
4 changed files with 29 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/models/phi4mm.py
View File

@@ -451,8 +451,8 @@ class Phi4MMForCausalLM(nn.Module):
pattern = MultiModalityDataPaddingPatternMultimodalTokens([im_token_id])
return pattern.pad_input_tokens(input_ids, mm_inputs)
def should_apply_lora(self, module_name: str) -> Optional[str]:
return self.lora_pattern.match(module_name)
def should_apply_lora(self, module_name: str) -> bool:
return bool(self.lora_pattern.match(module_name))
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
stacked_params_mapping = [