Add Nemotron nano v2 vl

.github/workflows/cicd-main.yml

@@ -222,6 +222,7 @@ jobs:
           - script: L2_Launch_models_mistral
           - script: L2_Launch_models_nemotron
           - script: L2_Launch_models_nemotronh
+          - script: L2_Launch_models_nemotron_vl
           - script: L2_Launch_models_olmoe
           - script: L2_Launch_models_qwen
           - script: L2_Launch_models_qwen_vl

examples/conversion/convert_checkpoints.py

@@ -141,6 +141,7 @@ def export_megatron_to_hf(
     megatron_path: str,
     hf_path: str,
     show_progress: bool = True,
+    strict: bool = True,
 ) -> None:
     """
     Export a Megatron checkpoint to HuggingFace format.
@@ -175,14 +176,15 @@ def export_megatron_to_hf(
 
     # For demonstration, we'll create a bridge from a known config
     # This would typically be extracted from the checkpoint metadata
-    bridge = AutoBridge.from_hf_pretrained(hf_model)
+    bridge = AutoBridge.from_hf_pretrained(hf_model, trust_remote_code=True)
 
     # Export using the convenience method
     print("📤 Exporting to HuggingFace format...")
     bridge.export_ckpt(
         megatron_path=megatron_path,
         hf_path=hf_path,
         show_progress=show_progress,
+        strict=strict,
     )
 
     print(f"✅ Successfully exported model to: {hf_path}")
@@ -232,6 +234,9 @@ def main():
         "--hf-path", required=True, help="Directory path where the HuggingFace model will be saved"
     )
     export_parser.add_argument("--no-progress", action="store_true", help="Disable progress bar during export")
+    export_parser.add_argument(
+        "--not-strict", action="store_true", help="Allow source and target checkpoint to have different keys"
+    )
 
     args = parser.parse_args()
 
@@ -254,6 +259,7 @@ def main():
             megatron_path=args.megatron_path,
             hf_path=args.hf_path,
             show_progress=not args.no_progress,
+            strict=not args.not_strict,
         )
     else:
         raise RuntimeError(f"Unknown command: {args.command}")

examples/conversion/hf_megatron_roundtrip_multi_gpu.py

@@ -63,6 +63,7 @@ def main(
     etp: int = 1,
     megatron_save_path: str | None = None,
     megatron_load_path: str | None = None,
+    strict: bool = False,
 ) -> None:
     """Perform round-trip conversion between HuggingFace and Megatron-LM models on multiple GPUs."""
     if os.environ.get("WORLD_SIZE") is None:
@@ -151,7 +152,7 @@ def main(
         console.print(table)
         console.print(f"Saving HF-ckpt in {save_path}...")
 
-    bridge.save_hf_pretrained(megatron_model, save_path)
+    bridge.save_hf_pretrained(megatron_model, save_path, strict=strict)
 
     # Save in Megatron format if path is provided
     if megatron_save_path:
@@ -188,6 +189,7 @@ def main(
         default=None,
         help="Path to load the model in Megatron checkpoint format. If provided, model will not start from HF checkpoint.",
     )
+    parser.add_argument("--not-strict", action="store_true", help="Perform loose validation during weight export")
     args = parser.parse_args()
     main(
         args.hf_model_id,