fix: guard MoE branch when quantization_format is None

[andrewwhitecdw]

This PR addresses the following issue in modelopt/torch/export/unified_export_hf.py: guard MoE branch when quantization_format is None.

Changes

• modelopt/torch/export/unified_export_hf.py: guard MoE branch when quantization_format is None.

Details

--- a/modelopt/torch/export/unified_export_hf.py
+++ b/modelopt/torch/export/unified_export_hf.py
@@ -1,5 +1,6 @@
-        # For MoE models update pre_quant_scale to average pre_quant_scale amongst experts
-        if is_moe(module) and (
-            quantization_format is not QUANTIZATION_NONE
-            and ("awq" in quantization_format or quantization_format == QUANTIZATION_NVFP4_SVDQUANT)
-        ):
+        # For MoE models update pre_quant_scale to average pre_quant_scale amongst experts
+        if is_moe(module) and (
+            quantization_format is not None
+            and quantization_format != QUANTIZATION_NONE
+            and ("awq" in quantization_format or quantization_format == QUANTIZATION_NVFP4_SVDQUANT)
+        ):

Tests

• tests/torch/export/test_unified_export_hf.py

--- /dev/null
+++ b/tests/torch/export/test_unified_export_hf.py
@@ -0,0 +1,41 @@
+import unittest
+from unittest.mock import patch
+
+from modelopt.torch.export.unified_export_hf import requantize_resmooth_fused_llm_layers
+
+
+class TestRequantizeResmoothFusedLLMLayers(unittest.TestCase):
+    def test_moe_branch_handles_none_quantization_format(self):
+        """Regression: MoE branch must not raise TypeError when quantization_format is None."""
+        class FakeModel:
+            pass
+
+        moe_module = object()
+        model = FakeModel()
+        model.named_modules = lambda: [("moe", moe_module)]
+
+        with patch(
+            "modelopt.torch.export.unified_export_hf.get_quantization_format",
+            return_value=None,
+        ) as mock_get_format, \
+            patch(
+            "modelopt.torch.export.unified_export_hf.is_moe",
+            return_value=True,
+        ) as mock_is_moe, \
+            patch(
+            "modelopt.torch.export.unified_export_hf.collect_shared_input_modules",
+            return_value=({}, {}),
+        ) as mock_collect, \
+            patch(
+            "modelopt.torch.export.unified_export_hf.fuse_prequant_to_linear",
+        ) as mock_fuse:
+            requantize_resmooth_fused_llm_layers(model)
+
+        mock_get_format.assert_called_once_with(model)
+        mock_is_moe.assert_called_once_with(moe_module)
+        mock_fuse.assert_not_called()
+        mock_collect.assert_called_once()

Contributor guidelines

Per this repo's CONTRIBUTING.md:

• All commits are signed off (Signed-off-by trailer, DCO).

Summary by CodeRabbit

• Bug Fixes
  • Improved MoE pre-quantization handling when no quantization format is specified.
  • Preserved existing AWQ and SVDQuant processing behavior.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 8d115bd5-207c-4daa-84b3-1ecaeae9406c

📥 Commits

Reviewing files that changed from the base of the PR and between b96841d and 2c976ec.

📒 Files selected for processing (1)

• modelopt/torch/export/unified_export_hf.py

---

📝 Walkthrough

Walkthrough

The MoE pre-quantization resmoothing condition now requires a non-null quantization format before excluding QUANTIZATION_NONE. AWQ and SVDQuant handling remains unchanged.

Changes

MoE resmoothing

Layer / File(s)	Summary
Quantization format guard modelopt/torch/export/unified_export_hf.py	The resmoothing branch now checks that quantization_format is not None before comparing it with QUANTIZATION_NONE.

Estimated code review effort: 1 (Trivial) | ~2 minutes

Mergeability Score: ⚪ Minimal · up to 2c976

This localized change prevents a TypeError when MoE export runs with no quantization format and adds regression coverage; no actionable merge-blocking risk remains after normal checks and review.

Suggested reviewers: meenchen

---

Important

Pre-merge checks failed

Please resolve all errors before merging. Addressing warnings is optional.

❌ Failed checks (1 inconclusive)

Check name	Status	Explanation	Resolution
Security Anti-Patterns	❓ Inconclusive	The working tree has no diff, so PR-introduced security changes cannot yet be verified.	Provide the PR base revision or an applicable git diff for changed modelopt/examples files.

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: guarding the MoE branch when quantization_format is None.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}