[NVBUG-6241842][fix] DSA DSL atom-split: guard against MTP draft next… by limin2021 · Pull Request #14891 · NVIDIA/TensorRT-LLM

limin2021 · 2026-06-03T05:46:20Z

fix https://nvbugspro.nvidia.com/bug/6241845

…_n=1

_pick_dsl_expand caches (factor, atom) on the metadata once per prepare() with the invariant factor * atom == 1 + max_draft_tokens (the target-pass next_n). MTPEagle reuses the same attn_metadata for its multi-step draft loop: after iteration i=0 it mutates seq_lens to 1 so subsequent iterations run 1-token incremental decode with next_n=1.

The DSL FP4/FP8 atom-split reshape (num_gen, next_n, H, D) -> (num_gen*factor, atom, H, D) is only valid when the caller actually supplies next_n == factor * atom tokens per request. For MTP=3 + FP4 + atom-split (factor=2, atom=2), the i>=1 draft iteration's q only has num_gen * 1 tokens, so the reshape crashed with shape '[48, 2, 64, 64]' is invalid for input of size 98304 (4x mismatch).

Add an explicit guard next_n == factor * atom to both FP4 and FP8 atom-split entries. Target / draft i=0 satisfy the invariant and continue to benefit from atom-split; MTPEagle i>=1 falls back to the kernel-native next_n=1 path (atom-split has no benefit for 1 token per request anyway).

Summary by CodeRabbit

Bug Fixes
- Fixed sparse attention reshape behavior during multi-step inference scenarios to prevent incompatible operations.
- Corrected behavior in both FP4 and FP8 quantized model paths for improved reliability.

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)
If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.
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.

…_n=1 `_pick_dsl_expand` caches `(factor, atom)` on the metadata once per `prepare()` with the invariant `factor * atom == 1 + max_draft_tokens` (the target-pass next_n). MTPEagle reuses the same `attn_metadata` for its multi-step draft loop: after iteration `i=0` it mutates `seq_lens` to 1 so subsequent iterations run 1-token incremental decode with `next_n=1`. The DSL FP4/FP8 atom-split reshape `(num_gen, next_n, H, D) -> (num_gen*factor, atom, H, D)` is only valid when the caller actually supplies `next_n == factor * atom` tokens per request. For MTP=3 + FP4 + atom-split (`factor=2, atom=2`), the `i>=1` draft iteration's `q` only has `num_gen * 1` tokens, so the reshape crashed with `shape '[48, 2, 64, 64]' is invalid for input of size 98304` (4x mismatch). Add an explicit guard `next_n == factor * atom` to both FP4 and FP8 atom-split entries. Target / draft `i=0` satisfy the invariant and continue to benefit from atom-split; MTPEagle `i>=1` falls back to the kernel-native `next_n=1` path (atom-split has no benefit for 1 token per request anyway). Signed-off-by: Mindy Li <11663212+limin2021@users.noreply.github.com>

limin2021 · 2026-06-03T05:47:36Z

/bot run

coderabbitai · 2026-06-03T05:50:07Z

📝 Walkthrough

Walkthrough

Modified the atom-split reshape guard in both FP4 and FP8 paged-MQA logits decode paths to require next_n == dsl_expand_factor * dsl_atom in addition to checking dsl_expand_factor > 1, preventing reshape execution with incompatible token counts during multi-step draft iterations.

Changes

DSL Paged-MQA Atom-Split Guard

Layer / File(s)	Summary
DSL paged-MQA atom-split reshape eligibility `tensorrt_llm/_torch/attention_backend/sparse/dsa.py`	FP4 and FP8 CuTe DSL paged-MQA logits paths updated with narrower atom-split reshape guards: now require `next_n == dsl_expand_factor * dsl_atom` to ensure token-per-request invariant before reshape execution, with expanded explanatory comments.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The PR description includes a detailed technical explanation of the bug fix, root cause analysis, and solution, but the template sections (Description, Test Coverage, and PR Checklist items) are not properly filled out.	Fill out the Description and Test Coverage sections with concise explanations. Ensure all PR Checklist items are properly reviewed and marked; currently only the final checkbox is checked without addressing the other items.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
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.
Title check	✅ Passed	The title references the bug ID and describes the fix for the DSA DSL atom-split guard, which matches the primary change in the code modification.

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests

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

tensorrt-cicd · 2026-06-03T05:53:11Z

PR_Github #51765 [ run ] triggered by Bot. Commit: 6053105 Link to invocation

tensorrt-cicd · 2026-06-03T08:59:14Z

PR_Github #51765 [ run ] completed with state SUCCESS. Commit: 6053105
/LLM/main/L0_MergeRequest_PR pipeline #41133 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

CI Agent Failure Analysis

Link to invocation

limin2021 · 2026-06-03T11:19:16Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-03T11:25:01Z

PR_Github #51830 [ run ] triggered by Bot. Commit: 6053105 Link to invocation

tensorrt-cicd · 2026-06-03T15:29:31Z

PR_Github #51830 [ run ] completed with state SUCCESS. Commit: 6053105
/LLM/main/L0_MergeRequest_PR pipeline #41192 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

tburt-nv · 2026-06-03T18:17:06Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-03T18:23:30Z

PR_Github #51892 [ run ] triggered by Bot. Commit: 6053105 Link to invocation

tensorrt-cicd · 2026-06-03T20:26:23Z

PR_Github #51892 [ run ] completed with state SUCCESS. Commit: 6053105
/LLM/main/L0_MergeRequest_PR pipeline #41249 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

limin2021 · 2026-06-04T01:45:50Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-04T01:52:24Z

PR_Github #51933 [ run ] triggered by Bot. Commit: 6053105 Link to invocation

tensorrt-cicd · 2026-06-04T03:07:48Z

PR_Github #51933 [ run ] completed with state FAILURE. Commit: 6053105
/LLM/main/L0_MergeRequest_PR pipeline #41285 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

limin2021 · 2026-06-04T03:16:32Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-04T03:22:47Z

PR_Github #51953 [ run ] triggered by Bot. Commit: 6053105 Link to invocation

tensorrt-cicd · 2026-06-04T04:29:00Z

PR_Github #51953 [ run ] completed with state FAILURE. Commit: 6053105
/LLM/main/L0_MergeRequest_PR pipeline #41304 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

CI Agent Failure Analysis

Link to invocation

limin2021 · 2026-06-04T08:16:25Z

/bot run --disable-fail-fast

limin2021 · 2026-06-04T08:16:38Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-04T08:22:17Z

PR_Github #52017 [ run ] triggered by Bot. Commit: 6053105 Link to invocation

tensorrt-cicd · 2026-06-04T08:22:47Z

PR_Github #52018 [ run ] triggered by Bot. Commit: 6053105 Link to invocation

tensorrt-cicd · 2026-06-04T08:22:51Z

PR_Github #52017 [ run ] completed with state ABORTED. Commit: 6053105

Link to invocation

tensorrt-cicd · 2026-06-04T09:31:40Z

PR_Github #52018 [ run ] completed with state SUCCESS. Commit: 6053105
/LLM/main/L0_MergeRequest_PR pipeline #41359 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

CI Agent Failure Analysis

Link to invocation

limin2021 · 2026-06-05T00:37:14Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-05T00:43:29Z

PR_Github #52190 [ run ] triggered by Bot. Commit: 6053105 Link to invocation

tensorrt-cicd · 2026-06-08T01:34:26Z

PR_Github #52611 [ run ] triggered by Bot. Commit: 1434aef Link to invocation

tensorrt-cicd · 2026-06-08T05:04:15Z

PR_Github #52611 [ run ] completed with state FAILURE. Commit: 1434aef
/LLM/main/L0_MergeRequest_PR pipeline #41894 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

CI Agent Failure Analysis

Link to invocation

limin2021 · 2026-06-08T08:16:02Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-06-08T08:21:52Z

PR_Github #52714 [ run ] triggered by Bot. Commit: 1434aef Link to invocation

limin2021 · 2026-06-08T09:08:13Z

/bot run --disable-fail-fast

…atom-split-mtp-thin Signed-off-by: Mindy Li <11663212+limin2021@users.noreply.github.com>

limin2021 · 2026-06-08T10:35:37Z

/bot run

tensorrt-cicd · 2026-06-08T10:42:15Z

PR_Github #52746 [ run ] triggered by Bot. Commit: 5ff1413 Link to invocation

tensorrt-cicd · 2026-06-08T10:47:02Z

PR_Github #52714 [ run ] completed with state ABORTED. Commit: 1434aef

Link to invocation

tensorrt-cicd · 2026-06-08T14:02:38Z

PR_Github #52746 [ run ] completed with state SUCCESS. Commit: 5ff1413
/LLM/main/L0_MergeRequest_PR pipeline #42009 completed with status: 'SUCCESS'
Pipeline passed with automatic retried tests. Check the rerun report for details.

CI Report

Link to invocation

xxi-nv · 2026-06-09T00:40:40Z

/bot --help

github-actions · 2026-06-09T00:40:49Z

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. Supports wildcard * for pattern matching (e.g., "*PerfSanity*" matches all stages containing PerfSanity). Examples: "A10-PyTorch-1, xxx", "PerfSanity". 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. Supports wildcard * for pattern matching. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx", --extra-stage "Post-Merge".

--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.

limin2021 · 2026-06-09T00:48:59Z

/bot --stage-list "DGX_B200-8_GPUs-PyTorch-1,DGX_B200-8_GPUs-PyTorch-2,DGX_B200-8_GPUs-PyTorch-3"

github-actions · 2026-06-09T00:49:07Z