Repository navigation
Bump CI Megatron-Bridge to e42042fa8 and test Qwen3.5-VL bridge export with full NVFP4 - #3125
kevalmorabia97 wants to merge 1 commit into
Conversation
…ith full NVFP4 Megatron-Bridge e42042fa8 (NVIDIA-NeMo/Megatron-Bridge#6714) exports quantized GatedDeltaNet in_proj, so the bridge-vs-Megatron-exporter parity test now covers Qwen3.5-VL-MoE with every linear layer in NVFP4 instead of MLP only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
|
No actionable comments were generated in the recent review.
 |
| Layer / File(s) | Summary |
|---|---|
Megatron-Bridge commit selection .github/actions/megatron-from-source/action.yml, examples/megatron_bridge/README.md |
The action default and README checkout command now use commit e42042fa83aed837ce6bc13480c06350ea091f1a. |
Quantized export test configuration tests/gpu_megatron/torch/export/test_quantized_mbridge_export.py |
The Qwen3.5 MoE-VL case now uses nvfp4 and test ID qwen3_5_moe_vl-nvfp4. The Nemotron-H comment now refers only to Mamba projections. |
Priority:
Estimated code review effort: 2 (Simple) | ~10 minutes
Change: Other
Suggested reviewers: chenhanyu, shengliangxu, danielkorzekwa
Merge Risk: 
000fb
-
No actionable merge-blocking risk is established by the supplied change context.
Pre-merge checks: 
-
Check name Status Explanation Description Check Check skipped - CodeRabbit’s high-level summary is enabled. Title check The title clearly and concisely describes the main changes: updating the CI Megatron-Bridge commit and testing Qwen3.5-VL bridge export with full NVFP4 quantization. Docstring Coverage No eligible functions found in the changed files. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff with cyclomatic complexity of at least 2. Documentat… Linked Issues check Check skipped because no linked issues were found for this pull request. Out of Scope Changes check Check skipped because no linked issues were found for this pull request. Security Anti-Patterns PASS. The pull request changes only a Megatron-Bridge commit reference, README checkout instructions, and a test quantization configuration/comment. The added Python lines contain no torch.load(..., w… Full details: Docstring Coverage
-
Explanation
No eligible functions found in the changed files. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff with cyclomatic complexity of at least 2. Documentation-only edits and deletions within surviving functions are included; unchanged diff context and fully deleted functions are excluded. Analyzed 0 functions across 1 files. (2 skipped: 2 unsupported.)
Full details: Security Anti-Patterns
-
Explanation
PASS. The pull request changes only a Megatron-Bridge commit reference, README checkout instructions, and a test quantization configuration/comment. The added Python lines contain no torch.load(..., weights_only=False), numpy.load(..., allow_pickle=True), hardcoded trust_remote_code=True, eval/exec, or # nosec. No dependency manifests changed.
-
Finishing Touches
-
📝 Generate docstrings
-
- Commit to this branch
- Create a new PR
🧪 Generate unit tests (beta)
-
- Commit to this branch
- Create a new PR
-
- Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
Comment @coderabbitai help to get the list of available commands.
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #3125 +/- ##
==========================================
+ Coverage 78.16% 78.72% +0.56%
==========================================
Files 664 664
Lines 76377 76377
==========================================
+ Hits 59698 60130 +432
+ Misses 16679 16247 -432
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
What does this PR do?
Type of change: new tests
.github/actions/megatron-from-source, and the matching checkout inexamples/megatron_bridge/README.md) from 0652ea7 to e42042fa8. The new commit includes feat(quant): export quantized GatedDeltaNet input projections to HF NVIDIA-NeMo/Megatron-Bridge#6714: quantized GatedDeltaNetin_projexport insave_hf_pretrained_modelopt. It also moves Megatron-LM from ed9c2c8 to a296523510.test_quantized_mbridge_export.py: the Qwen3.5-VL-MoE case oftest_matches_megatron_exporter(Test Megatron-Bridge quantized HF export against the Megatron exporter #3066) now quantizes every linear layer with NVFP4 instead of MLP only. The bridge export of the GatedDeltaNet projections is now compared withexport_mcore_gpt_to_hf.Usage
N/A
Testing
Run in
nvcr.io/nvidian/nemo:26.10.rc9with Megatron-Bridge e42042fa8 (and its Megatron-LM) overlaid as the CI action does, on 2× RTX 6000 Ada.tests/gpu_megatron/torch/export+tests/gpu_megatron/torch/quantization: 308 passed, 40 skipped (GPU count, apex, known TODOs), includingtest_matches_megatron_exporter[qwen3_5_moe_vl-nvfp4].[llama-nvfp4]: TE's fused RoPE has no kernel for this GPU in that container, and it fails the same way onmain.tests/examples/megatron_bridge: 96 passed.multi_tensor_apply/fused_attn"no kernel image is available for execution on the device" (the container's TE has no Ada build). They could not be checked locally; CI covers them.export_mcore_gpt_to_hfon a tiny Qwen3.5-VL-MoE with every linear layer in NVFP4 and FP8 matched bit for bit at bridge TP2.Before your PR is "Ready for review"
CONTRIBUTING.md: N/AAdditional Information
Follow-up of #3066 and NVIDIA-NeMo/Megatron-Bridge#6714.
🤖 Generated with Claude Code
Summary by CodeRabbit
nvfp4quantization configuration.