fix(checkpoint): wait for restore admission under capacity pressure (#100 + review fixes)

[voipmonitor]

This is #100 by @ktsaou (commits unchanged) plus fixes for our review findings. It pairs with vLLM LMCache#929.

What #100 does. When a restore cannot reserve its GPU pages, the bridge keeps the answered manifest and waits for capacity. Before, it admitted the request to recompute its whole prompt ("admitted without the restore").

Review fixes on top:

1. Compatibility gate. At creation the bridge checks for vLLM's restore-admission API: can_admit_external_boundary_request, external_boundary_admission_ready, release_external_boundary_admission, and the reserve_admission parameter. Without them it logs one WARNING and keeps the previous behaviour. Before this, an older vLLM hit an AttributeError on every finished request, which killed the engine.
2. Revalidation. It happens only if the GPU cache no longer holds a checkpoint at least as long as the selected one. A replaced or longer local checkpoint is used without another directory round trip.
3. Visibility. A waiting restore is logged when the wait starts and every 30 s after that, with free blocks against needed blocks, or copies in flight. report_status() counts waits, and the waiter entry is cleared on every exit path.
4. Fresh pages during long waits. A wait longer than half of lookup_timeout sends the lookup again, so the checkpoint's pages stay recent.
5. Documentation and bypass. The per-attempt lookup timeout is documented. A complete local hit again skips a pending lookup, unless its own reserved copy is in flight.
6. Smaller fixes. The manifest is parsed once per answer, log wording is corrected, and the fragment's models are GLM-5.3-Flash and Qwen3.8-Flash-Next.

Tests (CPU).

• With vLLM LMCache does not support enabling TLS configuration when connecting with remote offloading LMCache/LMCache#929: 46 passed, 2 CUDA-only skips.
• On an older vLLM: the 16 tests that need LMCache does not support enabling TLS configuration when connecting with remote offloading LMCache/LMCache#929 now skip with a reason. The one remaining failure (test_checkpoint_multimodal_roots) fails on the base too.
• The 16 new tests fail against the PR-head bridge.
• pre-commit passes.

E2E (with vLLM LMCache#929; setup and table in vllm#929): 8 agents whose contexts outgrow the KV pool.

	Beta	With the fixes
Aggregate decode	17.2 tok/s	138.8 tok/s
Prompt tokens from cache	14.3%	98.1%
Later turns with a restore	13/82	650/650
Mean turn time	112 s	14.6 s

The bridge logged 648 "waiting for GPU capacity" lines and no "admitted without the restore".

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (1)

• dev/*

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: 78fa6e9f-165c-47fc-b684-adf1bad068c0

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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.}