Preserve original rope scaling type in export due to transformers library AutoConfig issue

examples/llm_ptq/hf_ptq.py

@@ -249,6 +249,8 @@ def main(args):
             use_seq_device_map=args.use_seq_device_map,
             attn_implementation=args.attn_implementation,
         )
+        # Store original model path for config restoration
+        model._original_model_path = args.pyt_ckpt_path
     else:
         assert args.qformat in QUANT_CFG_CHOICES, (
             f"Quantization format is not supported for low memory mode. Supported formats: {QUANT_CFG_CHOICES.keys()}"
@@ -270,6 +272,8 @@ def main(args):
                 args.pyt_ckpt_path,
                 **model_kwargs,
             )
+            # Store original model path for config restoration
+            model._original_model_path = args.pyt_ckpt_path
         calibration_only = True
     model_is_already_quantized = is_quantized(model)
 

modelopt/torch/export/model_config_utils.py

@@ -227,6 +227,54 @@ def model_config_from_dict(d: dict) -> ModelConfig:
     return _from_dict(config_type, d)
 
 
+def restore_original_rope_scaling(config_data: dict, original_model_path: str) -> dict:
+    """Restore original rope_scaling configuration if it was modified by transformers.
+
+    Some VLM models like Qwen2.5-VL have their rope_scaling configuration modified
+    by the transformers library during loading (e.g., from "mrope" to "default" with
+    additional fields). This function restores the original configuration.
+
+    Args:
+        config_data: The model configuration dictionary to restore
+        original_model_path: Path to the original model directory
+
+    Returns:
+        The config_data dictionary with restored rope_scaling (modified in-place)
+    """
+    import json
+    import warnings
+    from pathlib import Path
+
+    try:
+        original_config_file = Path(original_model_path) / "config.json"
+        if original_config_file.exists():
+            with open(original_config_file) as f:
+                raw_original_config = json.load(f)
+
+            # Check if rope_scaling was modified from mrope to default
+            if (
+                "rope_scaling" in raw_original_config
+                and "rope_scaling" in config_data
+                and raw_original_config["rope_scaling"].get("type") == "mrope"
+                and config_data["rope_scaling"].get("type") == "default"
+                and "rope_type" in config_data["rope_scaling"]
+            ):
+                print(f"Restoring original rope_scaling configuration from {original_model_path}")
+                config_data["rope_scaling"] = raw_original_config["rope_scaling"]
+
+                # Also restore rope_scaling in text_config if it exists
+                if (
+                    "text_config" in config_data
+                    and "rope_scaling" in config_data["text_config"]
+                    and config_data["text_config"]["rope_scaling"].get("type") == "default"
+                ):
+                    config_data["text_config"]["rope_scaling"] = raw_original_config["rope_scaling"]
+    except Exception as e:
+        warnings.warn(f"Could not restore original rope_scaling configuration: {e}")
+
+    return config_data
+
+
 def pad_weights(weights, tp_size):
     """Returns the padded weights to tp_size."""
     assert len(weights.shape) > 1

modelopt/torch/export/unified_export_hf.py

@@ -54,6 +54,7 @@
     QUANTIZATION_W4A8_AWQ,
     QUANTIZATION_W4A8_NVFP4_FP8,
 )
+from .model_config_utils import restore_original_rope_scaling
 from .plugins import export_spec_ckpt_config, export_spec_ckpt_state_dict, spec_opt_only
 from .quant_utils import (
     fuse_prequant_layernorm,
@@ -541,6 +542,11 @@ def export_hf_checkpoint(
         with open(original_config) as file:
             config_data = json.load(file)
 
+        # Preserve original rope_scaling configuration if it was modified by transformers
+        original_model_path = getattr(model, "_original_model_path", None)
+        if original_model_path is not None:
+            config_data = restore_original_rope_scaling(config_data, original_model_path)
+
         config_data["quantization_config"] = hf_quant_config
 
         with open(original_config, "w") as file: