fix(rollout): abort vLLM rollout via delete-type /abort_requests

[aoshen02]

What

Replace the pause/resume-based rollout abort with a delete-type abort, so that under --partial-rollout the long tail is truncated to partial (and resumable next step) instead of either deadlocking or running to completion.

Why — and why this is not a duplicate of #294

abort() previously did pause?mode=abort -> drain -> resume. Under --partial-rollout, a /generate that races in after the pause parks in vLLM's PAUSED_NEW waiting queue and never returns until /resume, which ran after the drain → deadlock.

#294 reordered to pause -> resume -> drain, which fixes the hang. But resume reopens the whole waiting queue, so the raced-in long-tail requests run to completion and return full samples — defeating partial rollout's "truncate the tail, continue it next step" purpose.

This PR supersedes #294 with a delete-type abort that truncates the tail correctly and has no pause/resume (so no deadlock by construction). It keeps #294's partial-sample-collection test and replaces its pause/resume-ordering test. Credit to @Josephasafg (PR #294 author), retained as the primary commit author.

How

• vLLM (docker/patch/latest/vllm.patch): add POST /abort_requests to the RLHF control router → EngineClient.abort(), which removes queued requests from the waiting queue and finish-aborts running ones (their partial output returns on the original /generate stream). It does not pause the scheduler, so there is no /resume and no deadlock window. Reuses the engine primitive already used on client-disconnect and mirrors the existing /abort_requests in the disagg router.
• vime: vllm_utils/server_control.abort_inflight_requests() (the vLLM counterpart of slime's sglang_utils/server_control); abort() re-issues the sweep across drain waves and converges on state.pendings (vLLM has no /v1/loads to poll), with a small timeout bounding how long a late multi-turn straggler can run before being truncated.

Tests

• tests/test_vllm_rollout.py:
  • test_abort_deletes_inflight_without_pause_resume — asserts only /abort_requests is posted (never /pause or /resume) and state.pendings is drained.
  • test_abort_collects_partial_samples_when_partial_rollout — kept from fix(rollout): Resume vLLM workers before draining in abort #294.
• Verified locally: python -m py_compile on all changed files; git apply --check of the appended vllm.patch hunk against a pristine vLLM tree (applies cleanly).
• Full unit suite + image build run in CI (no GPU/build env available locally).

AI assistance

This change was developed with AI assistance (Claude) and reviewed by a human before submission.

[read-the-docs-community]

Documentation build overview

📚 vime | 🛠️ Build #33344145 | 📁 Comparing 817a336 against latest (429b3d2)

  🔍 Preview build  

34 files changed · ± 34 modified

± Modified

• genindex.html
• index.html
• search.html
• advanced/arch-support-beyond-megatron.html
• advanced/fault-tolerance.html
• advanced/low-precision.html
• advanced/megatron-config.html
• advanced/pd-disaggregation.html
• advanced/reproducibility.html
• advanced/speculative-decoding.html
• and 24 more...

[gemini-code-assist]

Code Review

This pull request refactors the abort mechanism for the vLLM rollout backend to drop in-flight requests directly via a new /abort_requests endpoint on the workers, eliminating the need to pause and resume the scheduler. The changes include adding the endpoint to the vLLM patch, implementing the control-plane helper, updating the rollout logic to periodically re-sweep during draining, and adding corresponding unit tests. Feedback on the changes highlights critical robustness issues in the vLLM patch's endpoint, specifically regarding JSON type safety when parsing request bodies, potential string iteration bugs if a single string is passed as request_ids, and import safety when catching JSON decoding errors.

[gemini-code-assist]

There are a few robustness and correctness issues in this section of the patch:

1. JSON Type Safety: raw_request.json() can return any valid JSON type (such as a list, string, number, or boolean). If a client sends a non-dictionary JSON payload, calling body.get(...) will raise an AttributeError: 'list' object has no attribute 'get'. We should check isinstance(body, dict) to prevent this.
2. String Iteration Bug: If request_ids is passed as a single string (e.g., {"request_ids": "req_1"}), iterating over it or passing it directly to engine.abort() might treat it as an iterable of characters (['r', 'e', 'q', '_', '1']), attempting to abort non-existent single-character request IDs. Wrapping a single string in a list prevents this.
3. Import Safety: Catching ValueError instead of json.JSONDecodeError is safer because json.JSONDecodeError inherits from ValueError, and this avoids any potential NameError if json is not imported in api_router.py.

+    try:
+        body = await raw_request.json()
+        if not isinstance(body, dict):
+            body = {}
+    except ValueError:
+        body = {}
+
+    request_ids = body.get("request_ids")
+    if isinstance(request_ids, str):
+        request_ids = [request_ids]
+    elif not request_ids:
+        request_ids = list(engine.output_processor.request_states.keys())