fix qwenvl compressed tensors quantization weight loader (#11914)

Signed-off-by: LHXuuu <xulianhao.xlh@antgroup.com>
Co-authored-by: Peng Zhang <aniz1905@gmail.com>

python/sglang/srt/models/qwen2_5_vl.py

@@ -70,7 +70,7 @@ from sglang.srt.managers.schedule_batch import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2 import Qwen2Model
-from sglang.srt.models.utils import RotaryPosMixin, permute_inv
+from sglang.srt.models.utils import RotaryPosMixin, WeightsMapper, permute_inv
 from sglang.srt.multimodal.mm_utils import run_dp_sharded_mrope_vision_model
 from sglang.srt.multimodal.vit_cuda_graph_runner import ViTCudaGraphRunner
 from sglang.srt.server_args import get_global_server_args
@@ -547,6 +547,24 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
         "up_proj": ("gate_up_proj", 1),
     }
 
+    packed_modules_mapping = {
+        "gate_up_proj": ["gate_proj", "up_proj"],
+    }
+    # To ensure correct weight loading and mapping.
+    hf_to_sglang_mapper = WeightsMapper(
+        orig_to_new_substr={
+            "attn.qkv": "attn.qkv_proj",
+        },
+        orig_to_new_prefix={
+            # mapping for new names in checkpoint saved after transformers v4.52
+            "model.language_model.": "language_model.model.",
+            "model.visual.": "visual.",
+            # mapping for original checkpoint
+            "lm_head.": "language_model.lm_head.",
+            "model.": "language_model.model.",
+        },
+    )
+
     def __init__(
         self,
         config: Qwen2_5_VLConfig,