Fix Quark/AWQ weight loading for Qwen3-VL-4B text model

src/python/py/models/builders/qwen.py

@@ -900,7 +900,11 @@ def process_cache(input_name, name_suffix):
         return flat_cos, flat_sin
 
     def load_weights(self, input_path):
-        # Load the Hugging Face model
+        # For quantized models (e.g., Quark, AWQ, GPTQ) or GGUF, use base class logic
+        # which loads weights directly via QuantModel
+        if self.quant_type is not None or input_path.endswith(".gguf"):
+            return super().load_weights(input_path)
+
         print("Loading Qwen3VLForConditionalGeneration model...")
         return Qwen3VLForConditionalGeneration.from_pretrained(
             self.model_name_or_path,

src/python/py/models/quantized_model.py

@@ -218,6 +218,17 @@ def __init__(self, quant_type, input_path, quant_attrs, q_size, kv_size, interme
 
                 # Map weights to modules
                 for name, tensor in weights.items():
+                    # Skip vision tower weights in VLM checkpoints
+                    if name.startswith(("model.visual.", "model.vision.", "visual.")):
+                        continue
+                    # Normalize common VLM prefix so existing LLM regex + parsing keeps working
+                    if name.startswith("model.language_model."):
+                        name = "model." + name[len("model.language_model."):]
+
+                    # Normalize Quark weight_quantizer.* naming to flat weight_* naming
+                    name = name.replace(".weight_quantizer.scale", ".weight_scale")
+                    name = name.replace(".weight_quantizer.zero_point", ".weight_zero_point")
+
                     # Per-layer quantization support
                     local_bits = self.get_layer_bits(name)  # codeql[py/init-calls-subclass]
                     local_group_size = self.get_layer_group_size(name)  # codeql[py/init-calls-subclass]