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/qwen3_vl.py

@@ -53,7 +53,11 @@ 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.qwen3 import Qwen3Model
-from sglang.srt.models.utils import RotaryPosMixin, compute_cu_seqlens_from_grid_numpy
+from sglang.srt.models.utils import (
+    RotaryPosMixin,
+    WeightsMapper,
+    compute_cu_seqlens_from_grid_numpy,
+)
 from sglang.srt.multimodal.mm_utils import run_dp_sharded_mrope_vision_model
 from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import add_prefix, get_int_env_var
@@ -594,6 +598,21 @@ class Qwen3LLMModel(Qwen3Model):
 
 
 class Qwen3VLForConditionalGeneration(nn.Module):
+    # 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: Qwen3VLConfig,