Allow GLM recurrent cache blocks to divide target blocks

[logprobz]

Permit recurrent blocks to divide target blocks, matching the fine-grained prefix-hit geometry supported by vLLM PR #646. The current launcher rejects target2048 / recurrent256 before the engine can start, although that combination is used by the qualified LP27 cache integration.

This changes only the divisibility check and its error message. Positive-integer and multiple-of-64 validation remain active, and the independent LMCache chunk check is unchanged.

Validation: the original launcher rejects 2048/256. With this change, dry runs accept recurrent256, 2048, and 4096 with target2048, and reject recurrent384, 0, -64, and 65. Bash syntax and whitespace checks pass. The same launcher change is in the four-GPU LP27 image tested with DFlash2 K3, Marlin, TP4/DCP1, and 524288 context.

Related library changes: local-inference-lab/vllm#646, local-inference-lab/vllm#655, local-inference-lab/vllm#663. This launcher update requires an engine that supports the selected geometry.

Prepared with AI assistance and independent source review.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}