Update modelopt quantization config parsing (#13919)

Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>

python/sglang/srt/configs/model_config.py

@@ -626,6 +626,17 @@ class ModelConfig:
     # adapted from https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/config.py
     def _parse_quant_hf_config(self):
         quant_cfg = getattr(self.hf_config, "quantization_config", None)
+        if quant_cfg is not None and not isinstance(quant_cfg, dict):
+            quant_cfg = quant_cfg.to_dict()
+        if quant_cfg is not None:
+            # Identify modelopt quantization
+            if "quant_method" not in quant_cfg:
+                parsed_cfg = self._parse_modelopt_quant_config(
+                    {"quantization": quant_cfg}
+                )
+                if parsed_cfg:
+                    quant_cfg.update(parsed_cfg)
+
         if quant_cfg is None:
             # compressed-tensors uses a "compression_config" key
             quant_cfg = getattr(self.hf_config, "compression_config", None)
@@ -736,6 +747,12 @@ class ModelConfig:
 
     def _is_already_quantized(self) -> bool:
         """Check if the model is already quantized based on config files."""
+        # Check for quantization in hf_config (config.json)
+        if getattr(self.hf_config, "quantization_config", None) or getattr(
+            self.hf_config, "compression_config", None
+        ):
+            return True
+
         # Check for HuggingFace quantization config
         from sglang.srt.utils import has_hf_quant_config
 

python/sglang/srt/model_loader/weight_utils.py

@@ -197,6 +197,8 @@ def get_quant_config(
         # compressed-tensors uses a compressions_config
         hf_quant_config = getattr(model_config.hf_config, "compression_config", None)
     if hf_quant_config is not None:
+        if not isinstance(hf_quant_config, dict):
+            hf_quant_config = hf_quant_config.to_dict()
         hf_quant_config["packed_modules_mapping"] = packed_modules_mapping
         return quant_cls.from_config(hf_quant_config)