目录

vLLM-metax Repetition Detection Test Contribution

This repository is the competition evidence package for the CCF GPU AI track contribution to MetaX-MACA/vLLM-metax.

Entry Points

Contribution Summary

The contribution ports upstream vLLM v1 core repetition detection tests into the MetaX-adapted vLLM-metax repository.

The new test file is:

  • tests/v1/core/test_repetition_detection.py

The test coverage focuses on the scheduler-side repeated-output detection behavior used to protect generation from pathological repeated token output. It keeps the implementation small and low risk: no runtime code path was changed, and the PR only adds test coverage.

Current Status

  • PR status: open, ready for review
  • Merge status: not merged yet
  • Local commit: 9cca519f3fb075df0e1ace01386a140d334263a8
  • Remote PR head commit: c8601752bceee75a0f52a5728165edc085258fce

Merge evidence will be added to docs/merge_record.md after upstream merge.

Evidence Included

  • docs/pr_record.md: PR metadata, changed files, and contribution scope
  • docs/verification.md: local verification, blockers, and hardware limits
  • docs/public_submission_gap_review.md: comparison against public GPUApps issue records and the supplement checklist
  • logs/local_verification.txt: commands and captured verification results
  • docs/merge_record.md: merge record placeholder for post-merge update
关于

Evidence package for the CCF GPUAI vLLM-metax repetition detection test PR.

42.0 KB
邀请码