fix(recipes): defer Nano vocabulary to tokenizer

Signed-off-by: Chen Cui <chcui@nvidia.com>

Author: cuichenx

src/megatron/bridge/recipes/nemotronh/gb200/nemotron_3_nano.py

@@ -41,6 +41,10 @@ def nemotron_3_nano_pretrain_8gpu_gb200_bf16_config() -> ConfigContainer:
     cfg = _pretrain_common()
 
     cfg.model = AutoBridge.from_hf_pretrained(_NEMOTRON_3_NANO_MODEL_ID).to_megatron_provider(load_weights=False)
+    # Pretraining may use a tokenizer other than the HF checkpoint tokenizer.
+    # Defer the model vocabulary size to the runtime tokenizer, matching the
+    # pre-migration MambaModelProvider recipe behavior.
+    cfg.model.vocab_size = None
     cfg.tokenizer.tokenizer_model = _NEMOTRON_3_NANO_MODEL_ID
 
     cfg.model.seq_length = 4096

tests/unit_tests/recipes/test_nemotronh_recipes.py

@@ -139,6 +139,13 @@ def test_nemotronh_recipe_rejects_unknown_cli_override():
     assert not hasattr(cfg.model, "not_a_real_field")
 
 
+def test_nemotron_3_nano_gb200_defers_vocab_size_to_training_tokenizer():
+    """The GB200 pretraining model vocabulary must follow its runtime tokenizer."""
+    cfg = _nemotronh_module.nemotron_3_nano_pretrain_8gpu_gb200_bf16_config()
+
+    assert cfg.model.vocab_size is None
+
+
 def test_nemotron_nano_9b_v2_lora_defaults():
     """Test that Nemotron Nano 9B v2 LoRA has correct default parallelism."""
     from megatron.bridge.recipes.nemotronh import nemotron_nano_9b_v2_peft_config