qqtang/sskj
1
3
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
sskj/experiments/b300/dsv4_flash_glm53_sglang_text_matrix/README.md

3.0 KiB
Raw Blame History

B300 DeepSeek-V4-Flash / GLM-5.3 SGLang text matrix

This directory archives the test code and all files written by the B300 text-matrix run through its manual cutoff and cleanup.

Snapshot

  • Snapshot time: 2026-09-10 02:43:55 UTC (2026-09-10 10:43:55 Asia/Shanghai)
  • Run ID: b300-dsv4-glm53-dev4-20260909-084427
  • B300 code path: /data/b300-deepseek-v4-glm53
  • B300 result path: /data/b300-dsv4-glm53-text-matrix/b300-dsv4-glm53-dev4-20260909-084427
  • Main log: /data/b300-dsv4-glm53-dev4-20260909-084427.log
  • Follow-up log: /data/b300-dsv4-glm53-dev4-20260909-084427-followup.log
  • Source archive SHA-256: 8676f28c6b9df7676ad93d5a4a71777bb504539b0ea6e3b129ccec69f5a563fd

Incremental synchronization:

  • Delta time: 2026-09-10 07:23:10 UTC (2026-09-10 15:23:10 Asia/Shanghai)
  • Delta archive SHA-256: c1bc269866152024afbdefbb15265832a773bfbf07ea2375e318dc252b25bdbb
  • At this point DeepSeek-V4-Flash had produced 53 Low-Latency and 18 Balanced point JSON files.

Final synchronization:

  • Stop time: 2026-09-10 07:46:07 UTC (2026-09-10 15:46:07 Asia/Shanghai)
  • Final delta archive SHA-256: f6ba8dbce5db850d43e7583a40eb8fc7cc95aab91c08fc5352fa69ec1d7d1117
  • Final status: 185/236 points reached a terminal state (78.39%): 164 pass, 3 timeout-pass, 18 fail markers, and 51 pending.
  • Mode progress: DSV4 Low-Latency 54/59, DSV4 Balanced 31/59, GLM-5.3 Low-Latency 54/59, and GLM-5.3 High-Throughput 46/59.
  • Both benchmark tmux sessions were stopped, the test container was removed, and nvidia-smi reported no remaining compute process.

The 17 GLM-5.3 fail markers on OSL=1 are validation false negatives: the benchmark command returned zero and produced complete JSON, while TPOT is mathematically zero because a one-token response has no inter-token interval. The DSV4 Low-Latency 512k_1_c1 failure is a real GPU OOM. Raw status markers are preserved exactly as emitted; use final-summary/ together with the point JSON and logs when interpreting them.

Layout

  • GLM-5.3_B300全量文本性能阶段报告.md: interim GLM-5.3 analysis.
  • snapshot-20260910-024355/b300-deepseek-v4-glm53/: runner, summarizer, model patch, and plan copied from the B300 host.
  • snapshot-20260910-024355/b300-dsv4-glm53-text-matrix/: raw run outputs, point logs, service logs, commands, status markers, and summaries.
  • snapshot-20260910-024355/*.log: top-level orchestration logs.
  • snapshot-20260910-024355/b300-dsv4-glm53-text-matrix/b300-dsv4-glm53-dev4-20260909-084427/final-summary/: final summary.md, summary.csv, comparison.md, and progress.json.

The repository normally excludes raw benchmark logs. They are intentionally included here because this snapshot is an explicit full-run archive.

Compressed JSON

Raw benchmark JSON files larger than 5 MiB are stored losslessly as *.json.gz to keep every individual Git object below common GitLab limits. Restore one file with:

gzip -dk path/to/point.json.gz

Small JSON files remain uncompressed and can be inspected directly.