address comments

Signed-off-by: Li Ding <liding@nvidia.com>

Author: liding-nv

examples/conversion/compare_hf_and_megatron/compare.py

@@ -715,9 +715,6 @@ def compare_models_one_step(args) -> None:
         tokenizer, processor, args.image_path, args.prompt, is_vl_model, args.tp
     )
 
-    print(f"{input_ids.shape=}")
-    print(f"{input_ids=}")
-
     # Move to GPU
     input_ids = input_ids.cuda()
     if pixel_values is not None:

examples/models/nemotron_3/super/finetune_nemotron_3_super.py

@@ -64,9 +64,7 @@ def main() -> None:
     """
     args, cli_overrides = parse_cli_args()
 
-    cfg: ConfigContainer = finetune_config(
-        seq_length=args.seq_length, peft=args.peft, packed_sequence=args.packed_sequence
-    )
+    cfg: ConfigContainer = finetune_config(peft=args.peft)
     cfg.model.seq_length = args.seq_length
 
     # Convert the initial Python dataclass to an OmegaConf DictConfig for merging

examples/models/nemotron_3/super/pretrain_nemotron_3_super.py

@@ -58,7 +58,7 @@ def parse_cli_args() -> Tuple[argparse.Namespace, list[str]]:
 
 def main() -> None:
     """
-    Entry point for the Mamba 8B pretraining script.
+    Entry point for the Nemotron 3 Super pretraining script.
     """
     args, cli_overrides = parse_cli_args()
 

examples/models/nemotron_3/super/qat_nemotron_3_super.py

@@ -77,13 +77,9 @@ def main() -> None:
         print(f"[yellow]torchrun examples/models/quantize.py --megatron-save-path {args.megatron_load_path}[/yellow]")
         sys.exit(1)
 
-    cfg: ConfigContainer = finetune_config(
-        pretrained_checkpoint=args.megatron_load_path,
-        seq_length=args.seq_length,
-        peft=None,
-        packed_sequence=args.packed_sequence,
-    )
+    cfg: ConfigContainer = finetune_config(peft=None)
     cfg.model.seq_length = args.seq_length
+    cfg.checkpoint.pretrained_checkpoint = args.megatron_load_path
 
     # Check for ModelOpt state
     checkpoint_to_check = cfg.checkpoint.pretrained_checkpoint or cfg.checkpoint.load