[NVIDIA] Auto detect modelopt quant and fix DSR1-FP4 weight loading (#22073)

vllm/transformers_utils/config.py

@@ -449,6 +449,20 @@ def get_config(
         model_type = MODEL_FOR_CAUSAL_LM_MAPPING_NAMES[config.model_type]
         config.update({"architectures": [model_type]})
 
+    # ModelOpt 0.31.0 and after saves the quantization config in the model
+    # config file.
+    quantization_config = config_dict.get("quantization_config", None)
+
+    # ModelOpt 0.29.0 and before saves the quantization config in a separate
+    # "hf_quant_config.json" in the same directory as the model config file.
+    if quantization_config is None \
+        and file_or_path_exists(model, "hf_quant_config.json", revision):
+        quantization_config = get_hf_file_to_dict("hf_quant_config.json",
+                                                  model, revision)
+
+    if quantization_config is not None:
+        config.quantization_config = quantization_config
+
     if hf_overrides_kw:
         logger.debug("Overriding HF config with %s", hf_overrides_kw)
         config.update(hf_overrides_kw)