fix(docker): /abort_requests abort-all fix + cu13 image variant

[aoshen02]

What

Fix the bundled vLLM /abort_requests endpoint in docker/patch/latest/vllm.patch so that abort-all actually aborts, and add a cu13 Docker image variant (partial port of #307).

The bug

The endpoint (added in #296) handles an empty/missing request_ids by enumerating engine.output_processor.request_states.keys() and calling engine.abort(request_ids).

But request_states is keyed by internal request ids, while EngineClient.abort() defaults to internal=False (treats ids as external, resolving them through the external→internal map). Under the default request-id randomization (internal = f"{external}-{8 random}"), the internal ids are not external keys, so the lookup matches nothing — POST /abort_requests {} returns 200 while aborting nothing.

This was flagged by Codex on the upstream endpoint PR (vllm-project/vllm#47173, review r3516626381); the same code shipped here in the docker patch.

The fix

Abort the all-in-flight list as internal (internal=True); keep external semantics for explicitly-passed request_ids. Also include parent_requests keys (parallel-sampling parents) and return 400 on malformed JSON. Mirrors vllm-project/vllm#47173.

cu13 image variant

Partial port from #307 — excludes the NCCL_CUMEM default flip and glm5.2 scripts:

• ENABLE_CUDA_13 build arg: cu13 dev headers, TE from source, skip cu12 cudnn pin
• just build-cu13 + cu13-* tags via just manifest cu13 ...
• TMS_CUDA_MAJOR auto-detect from torch's CUDA version
• git apply --allow-empty so builds survive once the vllm patch is emptied upstream
• actor_group: preload torch_memory_saver_hook_mode_preload_cu13.abi3.so

Testing

• Verified the patched api_router.py hunk applies cleanly to a pristine vLLM (git apply --check).
• The other patch sections (core.py, all2all_utils.py, disagg/protocol.py) are untouched.
• Full image build / runtime abort behaviour: via CI (Buildkite #430).

AI assistance

Drafted with AI assistance (Claude); reviewed by a human before submission.

[read-the-docs-community]

Documentation build overview

📚 vime | 🛠️ Build #33422724 | 📁 Comparing 60b9ccf against latest (5532763)

  🔍 Preview build  

41 files changed · ± 41 modified

± Modified

• genindex.html
• index.html
• search.html
• advanced/arch-support-beyond-megatron.html
• advanced/delta-weight-sync.html
• advanced/external-rollout-engines.html
• advanced/fault-tolerance.html
• advanced/low-precision.html
• advanced/megatron-config.html
• advanced/observability.html
• and 31 more...

[gemini-code-assist]

Code Review

This pull request updates the vllm.patch file to refine the request abort logic in the dev RLHF API router, separating the handling of external (user-supplied) request IDs from internal request IDs. The review feedback highlights a potential issue where passing an empty list of request IDs could accidentally trigger a mass-cancellation of all active requests, and suggests explicitly checking for None to safely handle empty lists as a no-op.

[gemini-code-assist]

If a client explicitly passes an empty list of request IDs (e.g., {"request_ids": []}), the current logic falls into the else block and aborts all active requests. This can lead to accidental mass-cancellations.

Checking request_ids is None specifically for the "abort all" behavior, and elif request_ids for specific aborts, ensures that an empty list is safely treated as a no-op.

+        if request_ids is None:
+            # The dev RL server runs AsyncLLM; abort everything it is tracking.
+            # request_states is keyed by internal ids, so abort as internal.
+            from vllm.v1.engine.async_llm import AsyncLLM
+
+            assert isinstance(engine, AsyncLLM)
+            request_ids = list(engine.output_processor.request_states.keys())
+            await engine.abort(request_ids, internal=True)
+        elif request_ids:
+            # Body ids are external (user-supplied) request ids.
+            await engine.abort(request_ids)

[CalvinXKY]

Scope note: this PR title only mentions abort fix, but commit 60b9ccf adds the full cu13 variant. Please expand the PR description (or split) so reviewers don't miss half the diff.