fix(model): keep GLM-5 export max_position_embeddings from the reference config (#5996)

Signed-off-by: linmuchuiyang <6169320+linmuchuiyang@users.noreply.github.com>
Signed-off-by: Chen Cui <chcui@nvidia.com>
Co-authored-by: linmuchuiyang <6169320+linmuchuiyang@users.noreply.github.com>
Co-authored-by: Chen Cui <chcui@nvidia.com>
Co-authored-by: Sajad Norouzi <snorouzi@nvidia.com>
Co-authored-by: meatybobby <meatybobby@gmail.com>

Author: 4 people

src/megatron/bridge/models/glm_moe_dsa/glm5_bridge.py

@@ -186,6 +186,21 @@ def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> MLAModelProvider
 
         return provider
 
+    @classmethod
+    def megatron_to_hf_config(cls, provider: MLAModelProvider) -> dict:
+        """Convert a GLM-5 provider to an HF config dict without leaking training-only context length.
+
+        The generic ``CONFIG_MAPPING`` maps HF ``max_position_embeddings`` from Megatron ``seq_length``. For a
+        fine-tuned checkpoint that is the *training* sequence length (e.g. 8192), not the model's context
+        capability (1048576 for ``zai-org/GLM-5.2``), and ``conform_config_to_reference`` keeps the
+        Megatron-derived value because the key exists in the reference config. Drop it here so the reference
+        HF config supplies it; when no reference config is available the key is simply absent and transformers
+        falls back to the ``GlmMoeDsaConfig`` default.
+        """
+        hf_config = super().megatron_to_hf_config(provider)
+        hf_config.pop("max_position_embeddings", None)
+        return hf_config
+
     def generate_pipeline_layout(self, num_layers: int, pp: int, mtp_layers: int = 0) -> list[list[str]] | None:
         """Generate a conversion pipeline layout that keeps DSA top-k sharing groups on one stage.
 

tests/unit_tests/models/glm_moe_dsa/test_glm5_bridge.py

@@ -156,6 +156,25 @@ def test_megatron_config_export_keeps_generic_moe_aliases() -> None:
     assert mapped_config["n_routed_experts"] == 8
 
 
+def test_megatron_config_export_does_not_emit_training_seq_length_as_max_position_embeddings() -> None:
+    """``seq_length`` is the fine-tuning context, not the model's context capability; leave it to the reference."""
+    mapped_config = GLM5Bridge.megatron_to_hf_config(SimpleNamespace(num_moe_experts=8, seq_length=8192))
+
+    assert "max_position_embeddings" not in mapped_config
+    assert mapped_config["num_experts"] == 8
+
+
+def test_megatron_config_export_reference_supplies_max_position_embeddings() -> None:
+    """With a reference config, the exported value must be the reference one (1048576 for GLM-5.2), not seq_length."""
+    from megatron.bridge.models.conversion.utils import conform_config_to_reference
+
+    mapped_config = GLM5Bridge.megatron_to_hf_config(SimpleNamespace(num_moe_experts=256, seq_length=8192))
+    reference = {"max_position_embeddings": 1048576, "n_routed_experts": 256}
+    conformed = conform_config_to_reference(mapped_config, reference)
+
+    assert conformed["max_position_embeddings"] == 1048576
+
+
 @pytest.mark.parametrize(
     ("config_overrides", "expected_topk_freq", "expected_skip_topk_offset"),
     [