[erezzarum]

Overview

Use the released NIXL 1.5.2 wheels in the Rubin CUDA 13.4 image and stop building NIXL from source there. NIXL 1.5.2 ships a NIXL EP extension built for torch 2.16, which the Rubin prerelease uses, so the BUILD_NIXL=true workaround from #60090 is no longer needed. Fixes the gap tracked in ai-dynamo/nixl#2297 (now closed).

Related: #64579 bumps requirements/kv_connectors.txt to the same version for the default images.

Claims

• requirements/rubin-prerelease.txt: nixl 1.4.1 -> 1.5.2.
• requirements/kv_connectors.txt: nixl 1.5.0 -> 1.5.2 (same line as [NIXL] Bump NIXL version to 1.5.2 #64579).
• .buildkite/release-pipeline.yaml: the nightly "aarch64 - Rubin CUDA 13.4" build drops --build-arg BUILD_NIXL=true and its TODO, so it installs the PyPI wheels.

Validation

• nixl-cu13==1.5.2 wheels from PyPI (cp312) contain nixl_ep_cpp_torch216:
  • manylinux_2_28_aarch64: torch 2.12, 2.13, 2.14, 2.15, 2.16
  • manylinux_2_28_x86_64: includes torch 2.16
  • Reproduce: pip download --no-deps --only-binary=:all: --platform manylinux_2_28_aarch64 --python-version 3.12 nixl-cu13==1.5.2 && python -m zipfile -l nixl_cu13-*.whl | grep nixl_ep_cpp_torch
• requirements/rubin-prerelease.txt pins torch==2.16.0.dev20261009+cu134, so the 216 variant is the one that gets loaded.
• TODO before submitting: build the aarch64 Rubin image without BUILD_NIXL and check that the NIXL EP extension loads (import nixl_ep_cu13 on a GPU host, or the same nixl_ep_cpp_torch216 glob the Dockerfile check uses).
• Not run: NixlConnector P/D serving on Rubin hardware.

Details

• The source build added in [CI/Build][NVIDIA] Build NIXL with NIXL EP from source in Rubin image #60090 is kept (BUILD_NIXL ARG, tools/build_nixl_with_torch215_for_rubin.py, docker/versions.json) as an opt-in escape hatch for future torch nightlies that get ahead of NIXL releases. No CI job sets it after this change.
• Updated the installation doc, which said NIXL's release wheels lack NIXL EP for the Rubin PyTorch nightly.

---

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.

🚀

[arpera]

In docker/Dockerfile this block that runs nixl build from source needs to be removed

[arpera]

This block delete as well

[arpera]

This as well
https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile#L1313-L1330

[erezzarum]

@arpera Don't we want to keep it, just in case something like this happens during the prerubin release? it's harmful as we removed BUILD_NIXL=true from the build and it's false by default.

[arpera]

I think we should undo all the changes I temporary made to build nixl in PR #60090

Also, in my opinion, script tools/build_nixl_with_torch215_for_rubin.py can also be removed. This is a dead code since no one use it anymore. This script is not general and it needs to be updated each time nixl changes the way they build wheel from source code. When I was working on PR #60090, I have seen that in vllm we have similar script tools/install_nixl_from_source_ubuntu.py that is expected to do the same thing. But this script failed to build nixl for me because it wasn't in active use in vllm for a while and no one updated it properly.
I think if in the future there will be a need to build nixl from source again it's much easier to generate script from scratch again.

[erezzarum]

I think we should undo all the changes I temporary made to build nixl in PR #60090

Also, in my opinion, script tools/build_nixl_with_torch215_for_rubin.py can also be removed. This is a dead code since no one use it anymore. This script is not general and it needs to be updated each time nixl changes the way they build wheel from source code. When I was working on PR #60090, I have seen that in vllm we have similar script tools/install_nixl_from_source_ubuntu.py that is expected to do the same thing. But this script failed to build nixl for me because it wasn't in active use in vllm for a while and no one updated it properly. I think if in the future there will be a need to build nixl from source again it's much easier to generate script from scratch again.

Done

[arpera]

Forgot to clean docs

• docs/contributing/dockerfile/dockerfile.md
• docs/getting_started/installation/gpu.cuda.inc.md

and docker/versions.json

[erezzarum]

Forgot to clean docs

• docs/contributing/dockerfile/dockerfile.md
• docs/getting_started/installation/gpu.cuda.inc.md

and docker/versions.json

Done, AI would have caught it, just didn't use one ;)

[mergify]

Documentation preview: https://vllm--64585.org.readthedocs.build/en/64585/

[arpera]

/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 #98816 for commit 07fd2e991a4c.

[erezzarum]

Built the image

DOCKER_BUILDKIT=1 docker buildx build --load --platform linux/arm64 \
  --build-arg max_jobs=64 \
  --build-arg CUDA_VERSION=13.4 \
  --build-arg torch_cuda_arch_list="8.0 8.7 8.9 9.0 10.0 11.0 12.0" \
  --build-arg INSTALL_KV_CONNECTORS=true \
  --build-arg INSTALL_RUBIN_PRERELEASE=true \
  --build-arg TRITON_INSTALL_FROM_SOURCE_REPO=https://github.com/triton-lang/triton.git \
  --build-arg TRITON_INSTALL_FROM_SOURCE_REVISION=3f6e41132b5edf639bfb872ad73d4688765e08b8 \
  --build-arg RUN_WHEEL_CHECK=false \
  --build-arg BUILD_BASE_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.4 \
  --build-arg FINAL_BASE_IMAGE=nvidia/cuda:13.4.1-base-ubuntu24.04 \
  --build-arg NCCL_VERSION=2.32.3 \
  --target vllm-openai --progress plain \
  -t vllm:openai-cu134-arm64 -f docker/Dockerfile .

Importing nixl_ep results in no errors

Python 3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import nixl_ep

For testing purposes, reverted to nixl 1.5.0 (no torch 2.15 and 2.16 compatible)

Python 3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import nixl_ep
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.12/dist-packages/nixl_ep/__init__.py", line 51, in <module>
    _pkg = sys.modules[_load_ep_module()]
                       ^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/nixl_ep/__init__.py", line 31, in _load_ep_module
    return importlib.import_module(mod_name).__name__
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/importlib/__init__.py", line 90, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/nixl_ep_cu13/__init__.py", line 51, in <module>
    _nixl_ep_cpp = importlib.import_module(
                   ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/importlib/__init__.py", line 90, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ModuleNotFoundError: No module named 'nixl_ep_cu13.nixl_ep_cpp_torch216'

[arpera]

/ci retry

[github-actions]

✅ Queued 1 failed job(s) for retry in Buildkite CI #98816.

[AndreasKaratzas]

Are there any amd files that need bumping too ?

[erezzarum]

Are there any amd files that need bumping too ?

This is related to the CUDA13.4 (Pre rubin release) variant.

[khluu]

LGTM!

[rabi82]

@khluu could this be cherry-picked into releases/v0.33.0 before v0.33.0 final?

v0.33.0rc3 still pins nixl == 1.5.0 in requirements/kv_connectors.txt and kv_connectors_cu12.txt, because the release branch was cut on 10-26 and this PR merged on 10-31. NIXL 1.5.2 contains three LIBFABRIC fixes that matter for anyone running NixlConnector with backends: ["LIBFABRIC"] on AWS EFA:

1.5.2 has no public API changes compared with 1.5.0, and it removes the libfabric 2.7.0 build bump that caused the 1.5.1 revert (#64085), so the LIBFABRIC plugin needs the same libfabric symbol versions as 1.5.0. AWS validated the 1.5.2 wheel with vLLM before release.

cc: @erezzarum