[train][fullyAsync] Fix abort/pause broken after vllm 0.16.0 bump

[CharlieFRuan]

Summary

• Fix abort_generation() and sleep() abort logic that broke silently after the vllm 0.16.0 bump (Bump vLLM to 0.16.0 with required dep updates #1240)
• Add backward-compatible _get_unfinished_request_ids() helper to resolve internal vs external request ID mismatch
• Fixes [train] Fix pause and resume logic and generation tests after 0.16.0 upgrade #1243

Root Cause

In vllm 0.16.0, InputProcessor.assign_request_id() now creates internal request IDs (with a random suffix) that are distinct from the user-provided external request IDs:

request.external_req_id = request.request_id                       # save original as external
request.request_id = f"{request.external_req_id}-{random_uuid():.8}"  # new internal ID

Our code was reading request IDs from output_processor.request_states.keys() (which are now internal IDs) and passing them to engine.abort() with internal=False (the default). The abort looked them up in the external_req_ids mapping, found nothing, and silently did nothing. Requests completed normally with finish_reason="length" instead of "abort".

This broke fully async RL's pause/resume flow, which relies on abort returning partial outputs with finish_reason="abort" so the retry loop can re-submit with accumulated tokens.

Related vllm changes:

• [RFC]: Extending vLLM Pause/Resume API to support in-flight requests and DPEP scenarios vllm-project/vllm#32103
• [Feat][RL] Pause and Resume with keep requests for single engine vllm-project/vllm#32351
• [Core] Move pause and resume functions into engine vllm-project/vllm#34125
• [Core] Cleanup engine pause/sleep logic  vllm-project/vllm#34528

Fix

Add a _get_unfinished_request_ids() static method on BaseVLLMInferenceEngine that:

• Uses output_processor.external_req_ids.keys() when available (vllm 0.16.0+)
• Falls back to output_processor.request_states.keys() for older vllm versions

Applied to all three abort call sites:

1. AsyncVLLMInferenceEngine.abort_generation() — used by fully async pause/resume
2. AsyncVLLMInferenceEngine.sleep() — cleanup before sleep
3. VLLMInferenceEngine.sleep() — sync engine cleanup before sleep

Test plan

• test_abort_generation_vllm_engine — passes (was failing with assert 'length' == 'abort')
• test_continue_generation_vllm_engine_chat_completion — passes
• test_continue_generation_generate_vllm_engine_generation — passes
• E2E fully async gsm8k (gsm8k_fully_async_ci project) — ran ~12 training steps successfully with pause/resume working correctly

Light blue is the run after this fix (our nightly gsm8k fully async CI) https://wandb.ai/sky-posttraining-uc-berkeley/gsm8k_fully_async_ci

---

[gemini-code-assist]

Code Review

This pull request addresses an issue with abort_generation() and sleep() in the vLLM engine, which was caused by a change in request ID handling in vllm version 0.16.0. The fix introduces a _get_unfinished_request_ids helper method to correctly retrieve request IDs for aborting, ensuring backward compatibility. The changes are logical and well-contained. My feedback includes a minor suggestion to improve type hinting for the new helper method.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 3 additional findings.