Propagate rope_scaling_factor from args to GPT model construction

gpt_builders.py

@@ -97,6 +97,7 @@ def gpt_builder(args, pre_process, post_process, vp_stage=None, config=None, pg_
         rotary_percent=args.rotary_percent,
         rotary_base=args.rotary_base,
         rope_scaling=args.use_rope_scaling,
+        rope_scaling_factor=args.rope_scaling_factor,
         mtp_block_spec=mtp_block_spec,
         vp_stage=vp_stage,
         pg_collection=pg_collection,

megatron/post_training/model_builder.py

@@ -389,6 +389,7 @@ def modelopt_gpt_hybrid_builder(
             "rotary_percent": args.rotary_percent,
             "rotary_base": args.rotary_base,
             "rope_scaling": args.use_rope_scaling,
+            "rope_scaling_factor": args.rope_scaling_factor,
             "mtp_block_spec": mtp_block_spec,
             "pg_collection": pg_collection,
         }

megatron/training/argument_utils.py

@@ -462,6 +462,7 @@ def gpt_config_from_args(
     kwargs["rotary_base"] = args.rotary_base
     kwargs["make_vocab_size_divisible_by"] = args.make_vocab_size_divisible_by
     kwargs["rope_scaling"] = args.use_rope_scaling
+    kwargs["rope_scaling_factor"] = args.rope_scaling_factor
 
     kwargs["seq_len_interpolation_factor"] = args.rotary_seq_len_interpolation_factor
     kwargs["seq_length"] = args.max_position_embeddings

tests/unit_tests/models/test_gpt_model.py

@@ -615,3 +615,38 @@ def test_get_transformer_layer_spec_forwards_use_te_activation_func():
         assert (
             call_kwargs.get('use_te_activation_func') is True
         ), "use_te_activation_func must be forwarded from config"
+
+
+def test_gpt_builder_forwards_rope_scaling_factor():
+    """Test that gpt_builder forwards rope_scaling_factor to GPTModel.
+
+    Regression test for https://github.com/NVIDIA/Megatron-LM/issues/6305
+    The --rope-scaling-factor flag was silently ignored because gpt_builder
+    passed rope_scaling but not rope_scaling_factor, so GPTModel always fell
+    back to its default factor of 8.0.
+    """
+    mock_config = MagicMock()
+
+    mock_args = MagicMock()
+    mock_args.spec = None
+    mock_args.transformer_impl = "transformer_engine"
+    mock_args.experimental_attention_variant = None
+    mock_args.num_experts = None
+    mock_args.heterogeneous_layers_config_path = None
+    mock_args.mtp_num_layers = None
+    mock_args.use_rope_scaling = True
+    mock_args.rope_scaling_factor = 32.0
+
+    with (
+        patch('gpt_builders.GPTModel') as mock_gpt_model,
+        patch('gpt_builders._get_transformer_layer_spec'),
+    ):
+        from gpt_builders import gpt_builder
+
+        gpt_builder(mock_args, pre_process=True, post_process=True, config=mock_config)
+
+        mock_gpt_model.assert_called_once()
+        _, call_kwargs = mock_gpt_model.call_args
+        assert (
+            call_kwargs.get('rope_scaling_factor') == 32.0
+        ), "rope_scaling_factor must be forwarded from args"

tests/unit_tests/training/models/test_gpt_builder.py

@@ -1,6 +1,7 @@
 # Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
 
 import inspect
+import sys
 from unittest.mock import Mock, call, patch
 
 import pytest
@@ -12,6 +13,8 @@
     HeterogeneousTransformerConfig,
 )
 from megatron.core.transformer.transformer_config import TransformerConfig
+from megatron.training.argument_utils import gpt_config_from_args
+from megatron.training.arguments import parse_args, validate_args
 from megatron.training.models.gpt import (
     GPTModelBuilder,
     GPTModelConfig,
@@ -898,3 +901,38 @@ def test_use_transformer_engine_false_when_impl_not_te(self, mock_get_mtp):
             mtp_block_spec(config, spec)
 
         assert mock_get_mtp.call_args.kwargs["use_transformer_engine"] is False
+
+
+# =============================================================================
+# Section 5 — gpt_config_from_args
+# =============================================================================
+
+
+class TestGPTConfigFromArgs:
+    """Tests for argument propagation through ``gpt_config_from_args``."""
+
+    def _make_args(self, **overrides):
+        sys.argv = ['test_gpt_builder.py']
+        args = parse_args()
+        args.num_layers = 2
+        args.hidden_size = 128
+        args.num_attention_heads = 8
+        args.micro_batch_size = 1
+        args.seq_length = 128
+        args.max_position_embeddings = 131072
+        args.padded_vocab_size = 32000
+        args.position_embedding_type = 'rope'
+        args.apply_rope_fusion = False
+        for name, value in overrides.items():
+            setattr(args, name, value)
+        validate_args(args)
+        return args
+
+    def test_forwards_rope_scaling_factor_from_args(self):
+        """--rope-scaling-factor must reach the model config instead of the default."""
+        args = self._make_args(use_rope_scaling=True, rope_scaling_factor=32.0)
+
+        config = gpt_config_from_args(args)
+
+        assert config.rope_scaling is True
+        assert config.rope_scaling_factor == 32.0