[erezzarum]

Overview

Bump the NIXL pin from 1.5.0 to 1.5.2 for both CUDA 13 (requirements/kv_connectors.txt) and CUDA 12 (requirements/kv_connectors_cu12.txt) images. This replaces the reverted 1.5.1 bump (#63817, reverted in #64085).

Claims

• CUDA 12 and CUDA 13 images both install NIXL 1.5.2. The two pins move together.
• No EFA regression: 1.5.2 drops the accidental libfabric 2.7.0 bump that caused the 1.5.1 revert ([release/1.5.2] Revert "build: bump libfabric to v2.7.0 in manylinux Dockerfile" (#2471) ai-dynamo/nixl#2528). The LIBFABRIC plugin needs the same libfabric symbol versions as 1.5.0.
• Picks up these LIBFABRIC fixes:
  • Fallback for context-independent CUDA allocations, such as mapped VMM memory ([release/1.5.1] fix(libfabric): fall back for context-independent CUDA allocations ai-dynamo/nixl#2470).
  • EFA unsolicited write recv is disabled regardless of build headers ([release/1.5.2] libfabric: disable EFA unsolicited write recv regardless of build headers ai-dynamo/nixl#2529).
  • All endpoint/CQ/AV calls are serialized with the rail EP lock ([release/1.5.2] libfabric: serialize all endpoint/CQ/AV calls with the rail EP lock ai-dynamo/nixl#2542).
• The UCX path does not change.

Validation

• Dependency resolution, x86_64 and aarch64, both files: nixl, nixl-cu12 and nixl-cu13 all resolve to 1.5.2.

  uv pip compile --python-version 3.12 --python-platform x86_64-manylinux_2_28 --torch-backend cu129 requirements/kv_connectors.txt
  # repeated for aarch64-manylinux_2_28 and requirements/kv_connectors_cu12.txt

• pre-commit run --files requirements/kv_connectors.txt requirements/kv_connectors_cu12.txt: passed.

• ELF check of libplugin_LIBFABRIC.so in nixl_cu13 cp312 x86_64. None of the wheels bundle libfabric.so.

  Version	Required libfabric symbol versions
  1.5.0	FABRIC_1.0, 1.1, 1.7
  1.5.1 (reverted)	FABRIC_1.0, 1.1, 1.10
  1.5.2	FABRIC_1.0, 1.1, 1.7

• Not yet run: GPU P/D serving with NixlConnector on CUDA 12 and CUDA 13, LIBFABRIC over EFA, and the aarch64 Rubin NIXL source build. requirements/kv_connectors.txt is in the run-all patterns (.buildkite/ci_config.yaml), so CI runs the NIXL integration and P/D accuracy tests.

• Tested on 2 x AWS p6-b300.48xlarge, Kimi K3 disagg

Details

release/1.5.0...v1.5.2 touches only:

• version strings and attributions
• src/plugins/libfabric/libfabric_backend.cpp and src/utils/libfabric/libfabric_rail.{cpp,h}
• a comment in src/utils/libfabric/meson.build
• WHL_TORCH_VERSIONS in contrib/Dockerfile.manylinux, which adds NIXL EP extensions for torch 2.15 and 2.16 ([release/1.5.2] build: include torch 2.15 in NIXL EP wheels ai-dynamo/nixl#2545, [release/1.5.2] build: include torch 2.16 in NIXL EP wheels ai-dynamo/nixl#2546)

The manylinux build stays on libfabric v1.21.0 and the same pinned UCX commit.

• Rubin nightly: with BUILD_NIXL=true, tools/build_nixl_with_torch215_for_rubin.py reads the nixl pin from requirements/kv_connectors.txt, so the aarch64 Rubin nightly now builds NIXL v1.5.2 from source.
• torch extensions: vLLM pins torch==2.14.0, which 1.5.0 already covered. The torch 2.15/2.16 EP extensions only matter for torch-nightly builds.
• Left unchanged on purpose: requirements/rubin-prerelease.txt (nixl==1.4.1) and requirements/tpu.txt.

Not a duplicate: no open PR bumps NIXL to 1.5.2.

AI assistance (Kiro) was used to prepare this PR.

---

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.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment /ci run for upstream CI or /amd-ci run for AMD CI only whenever CI signals are needed.

Once the PR is approved or has the ready label, the PR author can also use the corresponding /ci run, /ci retry, and /ci cancel commands, or their /amd-ci variants. New commits do not start upstream CI automatically.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[ivanium]

/ci run

[github-actions]

✅ @erezzarum, CI is now available for this PR.

• /ci run starts upstream CI; /amd-ci run starts AMD CI only.
• Run commands test your branch with every commit currently on its upstream target branch. If it is behind, the bot merges the target branch into it first (pull before pushing again). Append --allow-stale to a run command to test an outdated branch as is, at your own risk.
• /ci retry retries failed jobs in the CI build for the current PR head. If the current head has no CI build, it starts a new CI build for the current head containing only jobs that failed in the latest earlier CI build for this PR.
• /amd-ci retry retries failed jobs in AMD CI for the current PR head. Use /amd-ci run when the current head has no AMD CI build.
• /ci cancel cancels scheduled or running CI builds for this PR branch; /amd-ci cancel does the same for AMD CI only.

[github-actions]

✅ Triggered Buildkite CI #98846 for commit 391696a8cfda.

ℹ️ This PR is 105 commits behind upstream main. Merged upstream main into this branch as 391696a8cfda before starting CI. Pull the branch before pushing more changes.