[https://nvbugs/6018046][fix] Lower throughput_pp4_mtp max_batch_size to 8 by tianyuxbear · Pull Request #14585 · NVIDIA/TensorRT-LLM

tianyuxbear · 2026-05-26T16:12:54Z

Summary

Fixes TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_pp4_mtp] (NVBug 6018046) by lowering max_batch_size from 32 to 8 to match the already-stable throughput_bs8_mtp configuration. The matching waiver entry is removed.

Root cause

On 4-GPU PP=4 + MTP topology, max_batch_size=32 causes the per-bs CUDA-graph snapshot pool (32 captured graphs) plus per-step activation and NCCL collective buffers to exhaust the physical headroom needed by the lazy cudaMalloc from cuBLAS Lt workspace (DeepseekV3 router GEMM, attention o_proj) on B300. With the default 0.70 KV cache fraction allocating ~100 GiB / rank, only ~3 GiB of physical free pool remains for cuBLAS Lt / NCCL workspaces, which is insufficient under sustained load. The failure manifests as mid-run CUBLAS_STATUS_EXECUTION_FAILED followed by an asynchronous illegal memory access.

Relation to #14201

#14201 fixes the same bug by combining max_batch_size 32→8 with free_gpu_memory_fraction 0.70→0.50. Through isolated controlled experiments I verified the KV-fraction change is not strictly required once bs is lowered:

Experiment	bs	KV fraction	Result
`upstream/main` baseline	32	0.70	❌ crash @ req 1254/4104 (router GEMM)
bs=16 only	16	0.70	❌ crash @ req 1502/4104 (o_proj GEMM)
This PR (3 runs)	8	0.70	✅ pass (531s / 543s / 735s)
#14201 (1 run)	8	0.50	✅ pass (618s)

At bs=8 the per-bs CUDA-graph pool and per-step activation footprint shrink by 4x, restoring enough physical free pool for the lazy cuBLAS Lt / NCCL workspaces even at the default 0.70 KV fraction. This PR is therefore a minimal single-knob alternative to #14201.

Test plan

pytest accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_pp4_mtp] passes 3/3 on B300 4-GPU
Waiver entry removed; no runtime / kernel / LLM-API code touched

Summary by CodeRabbit

Tests
- Modified GPU testing configuration parameters, including batch size and KV cache memory allocation adjustments, to improve stability and compatibility when executing tests across multiple GPUs
- Re-enabled a previously skipped multi-GPU test scenario, ensuring comprehensive test coverage for GPU testing configurations following improvements to GPU resource management and workspace handling compatibility

… to 8 The throughput_pp4_mtp variant uses pp_size=4 + ep_size=1 + mtp_nextn=1 on a 4-GPU topology. With max_batch_size=32 (every other 4-GPU variant in this test uses bs<=16), the per-bs CUDA-graph snapshot pool plus per-step activation and NCCL collective buffers leave no headroom for the lazy cuBLAS Lt workspace allocation. Under sustained load this surfaces as mid-run CUBLAS_STATUS_EXECUTION_FAILED in the DeepseekV3 router / o_proj GEMMs, followed by an asynchronous illegal memory access. Lower max_batch_size from 32 to 8 to match the already-stable throughput_bs8_mtp configuration. With bs=8 the default 0.70 KV cache fraction provides sufficient cuBLAS workspace headroom; the test passes in repeated runs without any KV fraction reduction. Removes the matching waiver entry. Signed-off-by: Tianyu Xiong <117647511+tianyuxbear@users.noreply.github.com>

coderabbitai · 2026-05-26T16:17:37Z

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: 8f57dfec-3958-4c9d-874e-f45322715924

📥 Commits

Reviewing files that changed from the base of the PR and between 1f8312d and 69da278.

📒 Files selected for processing (2)

tests/integration/defs/accuracy/test_llm_api_pytorch.py
tests/integration/test_lists/waives.txt

💤 Files with no reviewable changes (1)

tests/integration/test_lists/waives.txt

📝 Walkthrough

Walkthrough

This PR adjusts a single test parameter in the DeepSeekR1 GPU test suite. The max_batch_size for the throughput_pp4_mtp configuration is reduced from 32 to 8, documented with comments about a cuBLAS/cuBLASLt workspace issue, and the corresponding test waive entry is removed.

Changes

Batch size adjustment for throughput_pp4_mtp test

Layer / File(s)	Summary
Batch size parameter, documentation, and waive removal `tests/integration/defs/accuracy/test_llm_api_pytorch.py`, `tests/integration/test_lists/waives.txt`	Inline comments document the cuBLAS/cuBLASLt workspace headroom issue and KV cache fraction adjustment for the `throughput_pp4_mtp` parameter in `test_nvfp4_multi_gpus`. The `max_batch_size` for that parameter is reduced from 32 to 8. The corresponding skip waive entry is removed from the waives list.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

Possibly related PRs

NVIDIA/TensorRT-LLM#14503: Modifies test waive entries in tests/integration/test_lists/waives.txt for other parametrized test cases in the same accuracy test file.

Suggested reviewers

crazydemo
jieli-matrix
ZhanruiSunCh

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and specifically describes the main change: lowering max_batch_size for the throughput_pp4_mtp configuration to fix the NVBugs issue 6018046.
Description check	✅ Passed	The PR description is comprehensive with Summary, Root Cause, Relation to `#14201`, and Test Plan sections. It clearly explains the issue, the solution, the underlying cause, and test results. All critical sections are well-documented.
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.

_{✏️ 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.}

tianyuxbear · 2026-05-26T16:20:26Z

/bot run --extra-stage "DGX_B300-4_GPUs-PyTorch-Post-Merge-1, DGX_B300-4_GPUs-PyTorch-Post-Merge-2" --disable-fail-fast

tensorrt-cicd · 2026-05-26T16:26:54Z

PR_Github #50365 [ run ] triggered by Bot. Commit: 61e4add Link to invocation

tensorrt-cicd · 2026-05-27T00:42:30Z

PR_Github #50365 [ run ] completed with state FAILURE. Commit: 61e4add
/LLM/main/L0_MergeRequest_PR pipeline #39892 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

tianyuxbear · 2026-05-27T01:53:28Z

/bot run --extra-stage "DGX_B300-4_GPUs-PyTorch-Post-Merge-1, DGX_B300-4_GPUs-PyTorch-Post-Merge-2" --disable-fail-fast

tensorrt-cicd · 2026-05-27T01:58:59Z

PR_Github #50420 [ run ] triggered by Bot. Commit: 61e4add Link to invocation

tensorrt-cicd · 2026-05-27T06:44:42Z

PR_Github #50420 [ run ] completed with state SUCCESS. Commit: 61e4add
/LLM/main/L0_MergeRequest_PR pipeline #39944 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

tianyuxbear · 2026-05-27T08:09:05Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-27T08:15:35Z

PR_Github #50504 [ run ] triggered by Bot. Commit: 61e4add Link to invocation

tensorrt-cicd · 2026-05-27T09:20:24Z

PR_Github #50504 [ run ] completed with state SUCCESS. Commit: 61e4add
/LLM/main/L0_MergeRequest_PR pipeline #40012 completed with status: 'SUCCESS'

CI Report

Link to invocation

tianyuxbear · 2026-05-27T11:02:56Z

/bot run --stage-list "DGX_B300-4_GPUs-PyTorch-Post-Merge-2"

tensorrt-cicd · 2026-05-27T11:09:42Z

PR_Github #50532 [ run ] triggered by Bot. Commit: 61e4add Link to invocation

tensorrt-cicd · 2026-05-27T11:35:46Z

PR_Github #50532 [ run ] completed with state SUCCESS. Commit: 61e4add
/LLM/main/L0_MergeRequest_PR pipeline #40039 (Partly Tested) completed with status: 'SUCCESS'

CI Report

Link to invocation

tianyuxbear · 2026-05-27T11:51:13Z

/bot run --disable-fail-fast

tianyuxbear · 2026-05-28T02:07:54Z

/bot run --disable-fail-fast

tensorrt-cicd · 2026-05-28T02:13:50Z

PR_Github #50660 [ run ] triggered by Bot. Commit: 61e4add Link to invocation

tensorrt-cicd · 2026-05-28T02:40:59Z

PR_Github #50660 [ run ] completed with state SUCCESS. Commit: 61e4add
/LLM/main/L0_MergeRequest_PR pipeline #40150 completed with status: 'SUCCESS'

CI Report

Link to invocation

tianyuxbear requested a review from a team as a code owner May 26, 2026 16:12

github-actions Bot assigned tianyuxbear May 26, 2026

tianyuxbear force-pushed the fix/6018046 branch from 69da278 to 61e4add Compare May 26, 2026 16:15

tianyuxbear changed the title ~~[nvbugs/6018046][fix] Lower throughput_pp4_mtp max_batch_size to 8~~ [https://nvbugs/6018046][fix] Lower throughput_pp4_mtp max_batch_size to 8 May 26, 2026

jieli-matrix approved these changes May 27, 2026

View reviewed changes

tianyuxbear mentioned this pull request May 28, 2026

[https://nvbugs/5970614][fix] Sync CTA before PDL trigger in quantize_with_block_size #14668

Merged

tianyuxbear closed this Jun 10, 2026

Uh oh!

Conversation

tianyuxbear commented May 26, 2026 • edited by coderabbitai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Root cause

Relation to #14201

Test plan

Summary by CodeRabbit

Uh oh!

coderabbitai Bot commented May 26, 2026

Walkthrough

Changes

Estimated code review effort

Possibly related PRs

Suggested reviewers

Uh oh!

tianyuxbear commented May 26, 2026

Uh oh!

tensorrt-cicd commented May 26, 2026

Uh oh!

tensorrt-cicd commented May 27, 2026

Uh oh!

tianyuxbear commented May 27, 2026

Uh oh!

tensorrt-cicd commented May 27, 2026

Uh oh!

tensorrt-cicd commented May 27, 2026

Uh oh!

tianyuxbear commented May 27, 2026

Uh oh!

tensorrt-cicd commented May 27, 2026

Uh oh!

tensorrt-cicd commented May 27, 2026

Uh oh!

tianyuxbear commented May 27, 2026

Uh oh!

tensorrt-cicd commented May 27, 2026

Uh oh!

tensorrt-cicd commented May 27, 2026

Uh oh!

tianyuxbear commented May 27, 2026

Uh oh!

tianyuxbear commented May 28, 2026

Uh oh!

tensorrt-cicd commented May 28, 2026

Uh oh!

tensorrt-cicd commented May 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

tianyuxbear commented May 26, 2026 •

edited by coderabbitai Bot

Loading