[https://nvbugs/6442073][fix] Lower free_gpu_memory_fraction from 0.6 → 0.4 (aligning with sibling…

[trtllm-agent]

Summary

• Root cause: TestQwen3_5_35B_A3B_VL.kv_cache_config uses free_gpu_memory_fraction=0.6 — with ~65 GB weights + ~8 GB non-torch overhead + 32-batch CUDA-graph replay buffers, real inference peak exceeds the profiled peak on 80–93 GB devices, and cudaMalloc for the next attention scratch buffer fails.
• Fix: Lower free_gpu_memory_fraction from 0.6 → 0.4 (aligning with sibling large-MoE VLM tests) and remove both H100/H20 waivers now that the test passes. KV cache drops 13.51 → 9.01 GiB, freeing ~4.5 GiB headroom.
• Automated fix generated by repair-bot

Test plan

• Verify fix on the same GPU type as the original failure
• Check for regressions in related tests

Links

• Bug: https://nvbugs/6442073

Summary by CodeRabbit

• Bug Fixes
  • Improved multimodal model test stability by reducing GPU memory usage during validation.
  • Restored automated accuracy checks for the Qwen multimodal model on H100 and H20 environments.

[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: 90b1da23-d720-4f72-8dd2-834dd71c63a6

📥 Commits

Reviewing files that changed from the base of the PR and between 53d2bde and 47ace75.

📒 Files selected for processing (2)

• tests/integration/defs/accuracy/test_llm_api_pytorch_multimodal.py
• tests/integration/test_lists/waives.txt

💤 Files with no reviewable changes (1)

• tests/integration/test_lists/waives.txt

---

📝 Walkthrough

Walkthrough

The Qwen3 multimodal accuracy test now uses a lower KV-cache GPU memory fraction, and its H100 and H20 test_auto_dtype waiver entries are removed.

Changes

Multimodal accuracy test updates

Layer / File(s)	Summary
Cache configuration and waiver removal tests/integration/defs/accuracy/test_llm_api_pytorch_multimodal.py, tests/integration/test_lists/waives.txt	Reduces free_gpu_memory_fraction from 0.6 to 0.4 and removes the H100 and H20 test_auto_dtype SKIP entries.

Estimated code review effort: 1 (Trivial) | ~5 minutes

Possibly related PRs

• NVIDIA/TensorRT-LLM#16271: Modifies waivers for the same multimodal accuracy cases.
• NVIDIA/TensorRT-LLM#16283: Updates waivers for the same H100 and H20 test cases.
• NVIDIA/TensorRT-LLM#16292: Changes waiver entries for the same Qwen3 multimodal tests.

Suggested reviewers: jieli-matrix, dongxuy04, pcastonguay, yuxianq

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title matches the main fix and follows the repo’s ticket-plus-type format.
Description check	✅ Passed	The description includes the issue, fix, test plan, and bug link, but it omits explicit Description and PR Checklist sections.
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.}

[Wanli-Jiang]

not the real fix, waiting for further bug comments.

[Wanli-Jiang]

the bug might not be caused by wrong memory configuration, please deep-diving into the architectures to see whether we can reproduce that, and which components consumed unusual memory footprint. After that, please fix the real issue rather than simplely tune the memory footprint config.