[Bugfix][Core] Return False for retryable prefix-cache reset failures

[Vegetog]

Overview

Related to #60552 (partial fix). Return False when a forced prefix-cache reset still has referenced blocks, so /reset_prefix_cache?reset_running_requests=true returns the documented retryable 200 {"success": false} instead of HTTP 500.

Merging this PR alone fixes the reset response, but does not make the full #60552 offloading reproducer pass. success: false means the cache reset did not complete; referenced blocks remain in use and the caller can retry later. When a preempted request resumes, the existing OffloadingConnector assertion can still crash the engine and interrupt generation.

The full reproduced scenario also needs the companion connector fix in #60574, or an equivalent fix for the same bug covered by #57810. This is a requirement for end-to-end recovery, not a required merge order: the two PRs address separate root causes.

Applied fixes	Observed effect
This PR alone	The reset returns the retryable HTTP response, but repeated runs still encounter the connector preempt/resume assertion.
This PR + #60574	The reset returns the retryable HTTP response, streams complete, and subsequent generation remains healthy in the combined tests below.

A parked streaming-input session (WAITING_FOR_STREAMING_REQ, #60744) is a different holder: it waits on the client rather than on a transfer, so with this PR alone the reset keeps returning False however often it is retried; #60764 reclaims those sessions in the same function, so the two changes are complementary.

Claims

• Preserve blocks held by in-flight KV transfers and allow a later reset to succeed after the transfer completes.
• Preserve the early exit before connector reset, and the existing failure checks that prevent cache-clearing pause/sleep from proceeding after an unsuccessful reset.

Validation

Standalone patch on 0f112d180033f3757cf283723f5593a40b02ed82:

.venv/bin/python -m pytest \
  tests/v1/core/test_scheduler.py \
  tests/v1/core/test_async_scheduler.py \
  tests/v1/engine/test_engine_core.py \
  -k 'reset_prefix_cache or inflight_remote_kv or aux_output_reset or reset_connector_cache or kv_cache_release or pause_synchronizes' -v

19 passed. The four new regression cases failed before the fix. Pre-commit passed for both changed files.

The original #60552 reproducer returned HTTP 500 in 5/5 flagged trials on the initial baseline fb2ac824. With this patch alone, the retryable HTTP response was observed, but repeated runs also exposed the pre-existing offloading preempt/resume crash described in #57810. That crash is addressed separately by #60574 and is not fixed by this PR.

Combined serving validation on the same base, with both this patch and #60574 applied: one RTX 3070, Qwen3-0.6B, eager execution with TRITON_ATTN, 320 GPU blocks and 4 GB CPU offload. The Python checkout reused the official v0.31.0+cu129 CUDA extensions.

• Strict replay, using ignore_eos=True and streamed usage checks: 3 flagged and 3 control resets returned 200 {"success": false}; all 12 streams completed 64 tokens, all 6 follow-up health/generation checks passed, and every trial recorded positive offload-load bytes.
• The original [Bug]: POST /reset_prefix_cache?reset_running_requests=true returns HTTP 500 while an OffloadingConnector load is in flight #60552 client then completed 5 flagged and 5 control trials on the same server, all HTTP 200 and healthy.

These are combined-patch serving results, not a claim that this PR alone fixes the separate connector crash.

No model-quality or performance claim is made. The serving checks validate HTTP responses, stream completion and continued generation; the results above are local validation, not a claim of passing upstream test CI.

Details

Preempting self.running does not release blocks owned by WAITING_FOR_REMOTE_KVS requests. The cache manager already reports this condition as False; the scheduler currently converts it into an exception only when reset_running_requests=True.

Return False at the same point without changing preemption or transfer ownership. Keep the early return so a failed forced reset does not start a connector reset. EngineCore._reset_caches() already checks the result, so no pause/sleep caller changes are needed.

The duplicate-work check before opening this PR found no existing fix for this HTTP return-value contract. Companion #60574 and the overlapping fix in #57810 address the separate same-step offloading resume crash. This PR changes only the scheduler and its existing tests.

Developed and validated with OpenAI Codex assistance.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.