[None][fix] Fix DeepSeekV32 test_fp8_blockscale[baseline_mtp1] OOM on Blackwell

[sunnyqgg]

Summary

• Reduce KV cache free_gpu_memory_fraction from 0.6 to 0.55 for MTP-enabled configs on Blackwell (SM 100/103) in TestDeepSeekV32::test_fp8_blockscale
• MTP layers consume extra GPU memory for additional prediction parameters and intermediate tensors during forward pass, causing OOM with ~140 MiB shortage
• Remove corresponding test waive entry

Test plan

• CI DGX_B200-8_GPUs-PyTorch-1 stage passes test_fp8_blockscale[baseline_mtp1]
• Other TestDeepSeekV32::test_fp8_blockscale variants still pass (baseline, baseline_fp8kv, latency, etc.)

Fixes: https://nvbugs/5955792

[sunnyqgg]

/bot run --extra-stage "DGX_B200-8_GPUs-PyTorch-1"

[coderabbitai]

📝 Walkthrough

Walkthrough

Modified the KV cache configuration in a test to conditionally adjust the free_gpu_memory_fraction parameter based on a multi-token prediction condition. Removed a corresponding test waiver entry from the waive-list.

Changes

Cohort / File(s)	Summary
Test Configuration Update tests/integration/defs/accuracy/test_llm_api_pytorch.py	Modified test_fp8_blockscale to set free_gpu_memory_fraction to 0.55 when mtp_nextn > 0, otherwise 0.6.
Waiver Cleanup tests/integration/test_lists/waives.txt	Removed waiver entry for accuracy/test_llm_api_pytorch.py::TestDeepSeekV32::test_fp8_blockscale[baseline_mtp1].

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

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.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: fixing an OOM issue in the DeepSeekV32 test on Blackwell hardware.
Description check	✅ Passed	PR description clearly explains the issue, solution, test coverage, and includes proper NVBugs reference with meaningful context about memory optimization.

_{✏️ 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.}

[sunnyqgg]

/bot run --stage-list "DGX_B200-8_GPUs-PyTorch-1"

[tensorrt-cicd]

PR_Github #42249 [ run ] triggered by Bot. Commit: 7a18667 Link to invocation

[tensorrt-cicd]

PR_Github #42252 [ run ] triggered by Bot. Commit: 7a18667 Link to invocation

[tensorrt-cicd]

PR_Github #42249 [ run ] completed with state ABORTED. Commit: 7a18667

Link to invocation

[tensorrt-cicd]

PR_Github #42252 [ run ] completed with state SUCCESS. Commit: 7a18667
/LLM/main/L0_MergeRequest_PR pipeline #33061 (Partly Tested) completed with status: 'SUCCESS'

CI Report

Link to invocation

[sunnyqgg]

/bot run

[tensorrt-cicd]

PR_Github #42346 [ run ] triggered by Bot. Commit: 7a18667 Link to invocation

[tensorrt-cicd]

PR_Github #42346 [ run ] completed with state SUCCESS. Commit: 7a18667
/LLM/main/L0_MergeRequest_PR pipeline #33132 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