Skip to content

Decode PoC: move to vLLM 0.30.0 — one base for MiniMax, DeepSeek and GLM #1852

Closed @baychak opened 2026-09-25 16:52 UTC 1 comment Updated 2026-10-09 12:04 UTC

One decode-PoC base for all three models: vLLM 0.30.0, which carries GLM-5.3-Flash upstream. Integration and design stay on #1688. The decode-PoC release ships on 0.30.0 (@vbgd0, 2026-09-23). Out of scope: A100 (none on chain) and a new settings search.

Done when

  • residual on v0.30.0: gonka-ai/vllm#114 into release/v0.30-decode-int
  • plugin on v0.30.0: gonka-ai/gonka-vllm-plugins#19 into decode/vlm030
  • all three models on B300 and H100: start, validate against the previous reference artifacts, throughput measured. GLM runs with two prefill-kernel flags that keep the 0.28.1 references valid
  • PRs merged, testnet built

Needs: @vbgd0 — review and testnet build.

Estimate: done on our side; waiting for review.

Handover: treated as delivered when the checklist above is complete. Anything skipped is listed here with the reason before handover.

💬 Comments (1)

@baychak commented 2026-10-09 12:04 UTC

Status: done; closing.

Delivered: the decode-PoC move to vLLM 0.30 landed through @vbgd0's PRs: gonka-ai/vllm#116 and gonka-ai/gonka-vllm-plugins#21 on 2026-09-29, with follow-ups gonka-ai/gonka-vllm-plugins#25 and gonka-ai/gonka-vllm-plugins#26 on 2026-10-06. Our gonka-ai/vllm#114 and gonka-ai/gonka-vllm-plugins#19 are closed as superseded.

Limitations

  • GLM runs with --language-model-only. The GLM reference artifacts were re-taken on the release container on 2026-10-03: kaitakuai/experiments/2026-10.
  • A testnet run is not confirmed on this issue. The decode-PoC release goes out in v0.2.17 (#1743).

🔄 Auto-synced from Issue #1852 every hour.