Fix model loading: use modelopt get_model() instead of raw AutoModelForCausalLM

Raw from_pretrained OOMs during weight conversion — torch.cat on expert
gate_up_proj tries to allocate 31.5GB on a GPU with only 25.9GB free.
modelopt's get_model() handles max_memory/device_map properly for models
that need sequential device mapping.

scripts/quantize_nvfp4.py

@@ -259,7 +259,7 @@ def run_calibration(model_path, export_dir, calib_save_path, amax_snapshot_path,
     os.environ["HUGGING_FACE_HUB_TOKEN"] = HF_TOKEN
 
     from hf_ptq import (
-        get_model, get_tokenizer, make_calib_dataloader,
+        get_model as modelopt_get_model, get_tokenizer, make_calib_dataloader,
         build_quant_cfg, load_mtp_weights, copy_custom_model_files,
         QUANT_CFG_CHOICES,
     )
@@ -267,22 +267,24 @@ def run_calibration(model_path, export_dir, calib_save_path, amax_snapshot_path,
     from modelopt.torch.quantization.config import need_calibration
     from modelopt.torch.utils.dataset_utils import get_max_batch_size
     from modelopt.torch.export import export_hf_checkpoint
-    from transformers import AutoModelForCausalLM, AutoTokenizer
+    from transformers import AutoTokenizer
 
     apply_patches()
 
     # ── Load model ──
+    # Use modelopt's get_model() instead of raw AutoModelForCausalLM.from_pretrained.
+    # The raw call OOMs during weight conversion (torch.cat on experts needs 31.5GB,
+    # only 25.9GB free). modelopt's loader handles max_memory/device_map properly.
     print(f"\nLoading model from {model_path}...")
     t0 = time.time()
 
-    model = AutoModelForCausalLM.from_pretrained(
+    model = modelopt_get_model(
         model_path,
+        gpu_mem_percentage=GPU_MEM_PCT,
         trust_remote_code=True,
-        torch_dtype=torch.bfloat16,
-        device_map="sequential",
-        offload_folder="offload",
+        use_seq_device_map=True,
     )
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    tokenizer = get_tokenizer(model_path, trust_remote_code=True)
     print(f"✓ Model loaded in {time.time()-t0:.0f}s")
 
     # ── Setup quantization config ──
@@ -411,16 +413,15 @@ def run_export_only(calib_save_path, amax_snapshot_path, model_path, export_dir)
 
     apply_patches()
 
-    from transformers import AutoModelForCausalLM, AutoTokenizer
+    from hf_ptq import get_model as modelopt_get_model, get_tokenizer
 
     print(f"Loading model skeleton from {model_path}...")
-    model = AutoModelForCausalLM.from_pretrained(
+    model = modelopt_get_model(
         model_path,
+        device="cpu",
         trust_remote_code=True,
-        torch_dtype=torch.bfloat16,
-        device_map="cpu",
     )
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    tokenizer = get_tokenizer(model_path, trust_remote_code=True)
 
     print(f"Loading calibrated state from {calib_save_path}...")
     state = torch.load(calib_save_path, map_location='cpu')