[#12183][fix] Fix TRTLLM-Gen NVFP4 MoE scales for mixed-precision che… by tcherckez-nvidia · Pull Request #12240 · NVIDIA/TensorRT-LLM

tcherckez-nvidia · 2026-03-16T08:00:11Z

…ckpoints

Pass per-expert fc2_alpha (no min normalization) so w2 input scale variation across experts no longer distorts logits when sampling.
For fc1: require w1/w3 input scales to match across experts; use first expert's scale when same, use min when allow_different_input_scales.
Note: the trtllm kernel was renamed to cutlass.
More test flavors were added to accuracy testing.
Moved to use the official Super v3 NVFP4 model.

Made-with: Cursor

Summary by CodeRabbit

Configuration
- Changed default NVFP4 MoE fusion backend from trtllm to cutlass across all configurations.
Bug Fixes
- Improved FP8 NVFP4 input scale handling with flexible configuration and fallback warnings instead of strict failures.
Tests
- Expanded test coverage with parameterized MOE backend selection for comprehensive validation.
- Updated model variant for NVFP4 accuracy testing.

Description

Test Coverage

PR Checklist

Please review the following before submitting your PR:

PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
Update tava architecture diagram if there is a significant design change in PR.
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

tcherckez-nvidia · 2026-03-16T08:04:03Z

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_B200-AutoDeploy-1"

coderabbitai · 2026-03-16T08:06:12Z

📝 Walkthrough

Walkthrough

The changes update the NVFP4 MoE fusion backend configuration from trtllm to cutlass as the default, rework input scale handling logic with conditional fallback behavior for inconsistent scales, update test configurations with a new model variant, and introduce parametrized moe_backend testing for the Nemotron SuperV3 accuracy tests.

Changes

Cohort / File(s)	Summary
Configuration Updates `examples/auto_deploy/super_v3.yaml`, `tensorrt_llm/_torch/auto_deploy/config/default.yaml`	Backend changed from trtllm to trtllm_gen and cutlass respectively for the fuse_nvfp4_moe transform.
Core MoE Fusion Implementation `tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.py`	Reworked FP8 NVFP4 input scale handling with conditional fallback for different scales; updated FuseNVFP4MoeConfig.backend from Literal["trtllm", "trtllm_gen"] to Literal["cutlass", "trtllm_gen"] with default "cutlass"; changed backend dispatch logic to use cutlass path; modified per-expert vs global scale decision flow with per-expert fc2_alpha handling and adjusted gated MLP scale checks.
Test Implementation `tests/integration/defs/accuracy/test_llm_api_autodeploy.py`	Updated NVFP4 model path variant; added moe_backend parametrized parameter to test_accuracy method signature; wired moe_backend into test configuration for transforms.fuse_nvfp4_moe.backend.
Test Coverage `tests/integration/test_lists/test-db/l0_b200.yml`, `tests/integration/test_lists/test-db/l0_dgx_b200.yml`	Expanded test coverage by replacing single test entries with multiple variants covering different attn_dp configurations and backend combinations (trtllm and trtllm_gen).

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~35 minutes

Possibly related PRs

[None][fix] Refactor nanoV3+superV3 accuracy tests to load example config #11458: Both PRs modify the Nemotron SuperV3 accuracy test configuration and parametrization in tests/integration/defs/accuracy/test_llm_api_autodeploy.py, including model path updates and test signature changes.

Suggested reviewers

suyoggupta
marinayanov
jieli-matrix

🚥 Pre-merge checks | ❌ 3

❌ Failed checks (2 warnings, 1 inconclusive)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.
Description check	⚠️ Warning	The PR description is missing a proper title following the required format and lacks detailed explanation in the Description section.	Add a properly formatted title like [`#12240`][fix] Fix NVFP4 MoE scales for mixed-precision checkpoints. Expand the Description section with detailed rationale and implementation details.
Title check	❓ Inconclusive	The PR title references issue `#12183` and identifies the change type as [fix], but is truncated with ellipsis and doesn't clearly convey the main objective of fixing NVFP4 MoE scales for mixed-precision checkpoints.	Complete the title to clearly state the full objective, e.g., '[`#12183`][fix] Fix NVFP4 MoE scales handling for mixed-precision checkpoints and rename trtllm backend to cutlass.'

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Post copyable unit tests in a comment

📝 Coding Plan

Generate coding plan for human review comments

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

coderabbitai

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@tests/integration/defs/accuracy/test_llm_api_autodeploy.py`:
- Line 425: The test's parametrize uses invalid backend value "trtllm" causing
Pydantic validation failures when instantiating FuseNVFP4MoeConfig (backend is
Literal["cutlass","trtllm_gen"]); update the pytest.mark.parametrize call for
"moe_backend" to use the allowed values (e.g., ["cutlass","trtllm_gen"]) so
tests provide valid inputs for FuseNVFP4MoeConfig.backend and avoid
instantiation errors.

In `@tests/integration/test_lists/test-db/l0_b200.yml`:
- Around line 250-252: Test entries use an invalid backend value "trtllm" which
violates FuseNVFP4MoeConfig.backend (Literal["cutlass","trtllm_gen"]); update
the test parametrization that lists ["trtllm","trtllm_gen"] and any YAML entries
like nvfp4-1-attn_dp_off-trtllm-trtllm to use a valid literal (replace "trtllm"
with either "trtllm_gen" or "cutlass") so the pydantic validation for
FuseNVFP4MoeConfig.backend succeeds.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 6ee8ad2a-104a-4a0b-8bbb-51c4c54dfe43

📥 Commits

Reviewing files that changed from the base of the PR and between 503d678 and a30596b.

📒 Files selected for processing (6)

examples/auto_deploy/super_v3.yaml
tensorrt_llm/_torch/auto_deploy/config/default.yaml
tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.py
tests/integration/defs/accuracy/test_llm_api_autodeploy.py
tests/integration/test_lists/test-db/l0_b200.yml
tests/integration/test_lists/test-db/l0_dgx_b200.yml

tests/integration/defs/accuracy/test_llm_api_autodeploy.py

tests/integration/test_lists/test-db/l0_b200.yml

tensorrt-cicd · 2026-03-16T08:10:59Z

PR_Github #39056 [ run ] triggered by Bot. Commit: a30596b Link to invocation

tensorrt-cicd · 2026-03-16T08:46:58Z

PR_Github #39056 [ run ] completed with state SUCCESS. Commit: a30596b
/LLM/main/L0_MergeRequest_PR pipeline #30325 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation

tcherckez-nvidia · 2026-03-16T09:13:18Z

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_B200-AutoDeploy-1"

tensorrt-cicd · 2026-03-16T09:19:15Z

PR_Github #39069 [ run ] triggered by Bot. Commit: d2d92ce Link to invocation

tensorrt-cicd · 2026-03-16T09:57:59Z

PR_Github #39069 [ run ] completed with state FAILURE. Commit: d2d92ce
/LLM/main/L0_MergeRequest_PR pipeline #30334 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation

tcherckez-nvidia · 2026-03-16T13:10:16Z

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_B200-AutoDeploy-1"

tensorrt-cicd · 2026-03-16T13:17:14Z

PR_Github #39089 [ run ] triggered by Bot. Commit: 473a984 Link to invocation

tensorrt-cicd · 2026-03-16T15:21:12Z

PR_Github #39089 [ run ] completed with state SUCCESS. Commit: 473a984
/LLM/main/L0_MergeRequest_PR pipeline #30351 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation

tests/integration/defs/accuracy/test_llm_api_autodeploy.py

tcherckez-nvidia · 2026-03-18T06:51:45Z

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_B200-AutoDeploy-1"

tensorrt-cicd · 2026-03-18T06:57:27Z

PR_Github #39411 [ run ] triggered by Bot. Commit: 473a984 Link to invocation

tensorrt-cicd · 2026-03-18T07:52:45Z

PR_Github #39411 [ run ] completed with state SUCCESS. Commit: 473a984
/LLM/main/L0_MergeRequest_PR pipeline #30640 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

Please check the failed tests and fix your PR
If you cannot view the failures, ask the CI triggerer to share details
Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation

tcherckez-nvidia · 2026-03-19T06:22:28Z

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_B200-AutoDeploy-1"

tensorrt-cicd · 2026-03-19T06:28:06Z

PR_Github #39552 [ run ] triggered by Bot. Commit: 2a2d2e8 Link to invocation

tensorrt-cicd · 2026-03-19T11:45:43Z

PR_Github #39552 [ run ] completed with state SUCCESS. Commit: 2a2d2e8
/LLM/main/L0_MergeRequest_PR pipeline #30769 completed with status: 'SUCCESS'

CI Report

Link to invocation

tcherckez-nvidia · 2026-03-19T13:01:15Z

/bot reuse-pipeline

tcherckez-nvidia · 2026-03-19T13:01:42Z

/bot help

github-actions · 2026-03-19T13:01:50Z

GitHub Bot Help

/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...

Provide a user friendly way for developers to interact with a Jenkins server.

Run /bot [-h|--help] to print this help message.

See details below for each supported subcommand.

Details

run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental) --high-priority]

Launch build/test pipelines. All previously running jobs will be killed.

--reuse-test (optional)pipeline-id (OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.

--disable-reuse-test (OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.

--disable-fail-fast (OPTIONAL) : Disable fail fast on build/tests/infra failures.

--skip-test (OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.

--stage-list "A10-PyTorch-1, xxx" (OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.

--gpu-type "A30, H100_PCIe" (OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.

--test-backend "pytorch, cpp" (OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.

--only-multi-gpu-test (OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.

--disable-multi-gpu-test (OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.

--add-multi-gpu-test (OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.

--post-merge (OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.

--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" (OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".

--detailed-log (OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.

--debug (OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in the stage-list parameter to access the appropriate container environment. Note: Does NOT update GitHub check status.

--high-priority (OPTIONAL) : Run the pipeline with high priority. This option is restricted to authorized users only and will route the job to a high-priority queue.

kill

kill

Kill all running builds associated with pull request.

skip

skip --comment COMMENT

Skip testing for latest commit on pull request. --comment "Reason for skipping build/test" is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.

reuse-pipeline

reuse-pipeline

Reuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.

tensorrt-cicd · 2026-03-19T13:09:07Z

PR_Github #39599 [ reuse-pipeline ] triggered by Bot. Commit: 2a2d2e8 Link to invocation

tcherckez-nvidia · 2026-03-19T13:10:07Z

/bot reuse-pipeline

tensorrt-cicd · 2026-03-19T13:16:21Z

PR_Github #39599 [ reuse-pipeline ] completed with state SUCCESS. Commit: 2a2d2e8
Reusing PR_Github #39552 for commit 2a2d2e8

Link to invocation

tensorrt-cicd · 2026-03-19T13:17:27Z

PR_Github #39600 [ reuse-pipeline ] triggered by Bot. Commit: abf8107 Link to invocation

tensorrt-cicd · 2026-03-19T13:24:15Z

PR_Github #39600 [ reuse-pipeline ] completed with state SUCCESS. Commit: abf8107
Reusing PR_Github #39552 for commit abf8107

Link to invocation

…on checkpoints - Pass per-expert fc2_alpha (no min normalization) so w2 input scale variation across experts no longer distorts logits when sampling. - For fc1: require w1/w3 input scales to match across experts; use first expert's scale when same, use min when allow_different_input_scales. - Note: the trtllm kernel was renamed to cutlass. - Moved to use the official Super v3 NVFP4 model. Signed-off-by: Tal Cherckez <127761168+tcherckez-nvidia@users.noreply.github.com>

tcherckez-nvidia · 2026-03-19T14:36:22Z

/bot reuse-pipeline

tensorrt-cicd · 2026-03-19T14:42:26Z

PR_Github #39608 [ reuse-pipeline ] triggered by Bot. Commit: 5e40d9c Link to invocation

tensorrt-cicd · 2026-03-19T14:49:38Z

PR_Github #39608 [ reuse-pipeline ] completed with state SUCCESS. Commit: 5e40d9c
Reusing PR_Github #39552 for commit 5e40d9c

Link to invocation

tcherckez-nvidia requested review from a team as code owners March 16, 2026 08:00

tcherckez-nvidia requested a review from suyoggupta March 16, 2026 08:00

github-actions bot assigned tcherckez-nvidia Mar 16, 2026

coderabbitai bot reviewed Mar 16, 2026

View reviewed changes

tests/integration/defs/accuracy/test_llm_api_autodeploy.py Outdated Show resolved Hide resolved

tests/integration/test_lists/test-db/l0_b200.yml Outdated Show resolved Hide resolved

tcherckez-nvidia force-pushed the fix-trtllmgen branch from a30596b to d2d92ce Compare March 16, 2026 09:13

tcherckez-nvidia force-pushed the fix-trtllmgen branch from d2d92ce to 473a984 Compare March 16, 2026 13:09

xinhe-nv approved these changes Mar 17, 2026

View reviewed changes

suyoggupta approved these changes Mar 17, 2026

View reviewed changes

galagam reviewed Mar 17, 2026

View reviewed changes

tests/integration/defs/accuracy/test_llm_api_autodeploy.py Show resolved Hide resolved

galagam approved these changes Mar 17, 2026

View reviewed changes

tcherckez-nvidia force-pushed the fix-trtllmgen branch from 473a984 to 2a2d2e8 Compare March 19, 2026 06:21

tcherckez-nvidia force-pushed the fix-trtllmgen branch from 2a2d2e8 to abf8107 Compare March 19, 2026 13:09

tcherckez-nvidia force-pushed the fix-trtllmgen branch from abf8107 to 5e40d9c Compare March 19, 2026 14:34

tcherckez-nvidia enabled auto-merge (squash) March 19, 2026 14:39

tcherckez-nvidia disabled auto-merge March 19, 2026 14:39

tcherckez-nvidia enabled auto-merge (squash) March 19, 2026 14:39

tcherckez-nvidia merged commit 9f0c204 into NVIDIA:main Mar 19, 2026
5 checks passed

Conversation

tcherckez-nvidia commented Mar 16, 2026 • edited by coderabbitai bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Description

Test Coverage

PR Checklist

GitHub Bot Help

Uh oh!

tcherckez-nvidia commented Mar 16, 2026

Uh oh!

coderabbitai bot commented Mar 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Estimated code review effort

Possibly related PRs

Suggested reviewers

❌ Failed checks (2 warnings, 1 inconclusive)

Uh oh!

coderabbitai bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

tensorrt-cicd commented Mar 16, 2026

Uh oh!

tensorrt-cicd commented Mar 16, 2026

Uh oh!

tcherckez-nvidia commented Mar 16, 2026

Uh oh!

tensorrt-cicd commented Mar 16, 2026

Uh oh!

tensorrt-cicd commented Mar 16, 2026

Uh oh!

tcherckez-nvidia commented Mar 16, 2026

Uh oh!

tensorrt-cicd commented Mar 16, 2026

Uh oh!

tensorrt-cicd commented Mar 16, 2026

Uh oh!

Uh oh!

tcherckez-nvidia commented Mar 18, 2026

Uh oh!

tensorrt-cicd commented Mar 18, 2026

Uh oh!

tensorrt-cicd commented Mar 18, 2026

Uh oh!

tcherckez-nvidia commented Mar 19, 2026

Uh oh!

tensorrt-cicd commented Mar 19, 2026

Uh oh!

tensorrt-cicd commented Mar 19, 2026

Uh oh!

tcherckez-nvidia commented Mar 19, 2026

Uh oh!

tcherckez-nvidia commented Mar 19, 2026

Uh oh!

github-actions bot commented Mar 19, 2026

GitHub Bot Help

kill

skip

reuse-pipeline

Uh oh!

tensorrt-cicd commented Mar 19, 2026

Uh oh!

tcherckez-nvidia commented Mar 19, 2026

Uh oh!

tensorrt-cicd commented Mar 19, 2026

Uh oh!

tensorrt-cicd commented Mar 19, 2026

Uh oh!

tensorrt-cicd commented Mar 19, 2026

Uh oh!

tcherckez-nvidia commented Mar 19, 2026

Uh oh!

tensorrt-cicd commented Mar 19, 2026

Uh oh!

tensorrt-cicd commented Mar 19, 2026

Uh oh!

tcherckez-nvidia commented Mar 16, 2026 •

edited by coderabbitai bot

Loading

coderabbitai bot commented Mar 16, 2026 •

edited

Loading