[Ascend] optimize Qwen-vl on Ascend (#10556)

Co-authored-by: wangqihui01 <wangqh10@163.com>
2025-09-23 03:18:16 +03:00
parent d27a6f7092
commit 095093ee5a
5 changed files with 99 additions and 14 deletions
--- a/python/sglang/srt/model_loader/loader.py
+++ b/python/sglang/srt/model_loader/loader.py
@@ -206,7 +206,10 @@ def _initialize_model(
    if _is_npu:
        packed_modules_mapping.update(
            {
-                "visual": {"qkv_proj": ["qkv"]},
+                "visual": {
+                    "qkv_proj": ["qkv"],
+                    "gate_up_proj": ["gate_proj", "up_proj"],
+                },
                "vision_model": {
                    "qkv_proj": ["q_proj", "k_proj", "v_proj"],
                    "proj": ["out_proj"],