[https://nvbugs/6242591][fix] Fix bugs in Beam Search kernels

[wili-65535]

Description

Fix bugs in the beam search V2 path with beam_width >= 9 and beam_search_diversity_rate > 0.

Root cause:

1. addCumLogProbs compares a candidate-slot index i (range 0..nBMIn*nBMOut*2-1) against a vocabulary token ID (endIds[slot]) for the EOS forcing condition on finished beams. The comparison is always false, so those candidates receive inflated scores, rank into the top-nBM every step, and continuously increment numBeamsCBA.
2. A separate overflow guard uses == nBM instead of >= nBM, so once numBeamsCBA exceeds nBM the guard never triggers, leading to unbounded out-of-bounds writes into the CBA output buffer and eventual SIGSEGV.
3. EOS candidate length-penalty scoring uses the post-sort candidate rank i as the parent beam index instead of (topId / nV) % nBM, producing wrong scores in the CBA.
4. When all top candidates for a step are EOS tokens (nBeamForNextStep < nBMOut), the state-update block reads uninitialized outputIdsPtr slots, corrupting parentIdsPtr/outputIdsPtr/sequenceLengths and the traceback chain.

Changed Files

File	Change
cpp/tensorrt_llm/kernels/beamSearchKernels.h	Add pStage1Ids parameter to both addCumLogProbs overloads
cpp/tensorrt_llm/kernels/beamSearchKernels.cu	Use pStage1Ids for EOS token lookup on finished beams
cpp/tensorrt_llm/kernels/beamSearchKernels/beamSearchKernelsTemplate.h	Change three == nBM into >= nBM; Use parentBeam in EOS scoring

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.

[wili-65535]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #55813 [ run ] triggered by Bot. Commit: 6ae77a8 Link to invocation

[tensorrt-cicd]

PR_Github #55813 [ run ] completed with state SUCCESS. Commit: 6ae77a8
/LLM/main/L0_MergeRequest_PR pipeline #44705 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

[wili-65535]

/bot run --disable-fail-fast