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

Support LoRA in TestOpenAIVisionServer and fix fused kv_proj loading bug. (#6861)

Browse Source
This commit is contained in:
Lifu Huang
2025-06-04 22:08:30 -07:00
committed by GitHub
parent 499f5e620c
commit 4474eaf552
4 changed files with 66 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
python/sglang/srt/lora/lora.py
View File

@@ -165,14 +165,19 @@ class LoRAAdapter(nn.Module):
self.base_hf_config.hidden_size
// self.base_hf_config.num_attention_heads
)
weights[q_name], weights[kv_name] = torch.split(
weights[q_name], k_proj_weight, v_proj_weight = torch.split(
weights[qkv_name],
[
head_size * self.base_hf_config.num_attention_heads,
head_size * self.base_hf_config.num_key_value_heads * 2,
head_size * self.base_hf_config.num_key_value_heads,
head_size * self.base_hf_config.num_key_value_heads,
],
dim=0,
)
weights[kv_name] = torch.stack(
[k_proj_weight, v_proj_weight],
dim=0,
)
def normalize_gate_up_proj(
self, weight_names: List[str], weights: Dict[str, torch.Tensor]