[fix]: Dockerfile.ppc64le fixes for opencv-python and hf-xet

[Shafi-Hussain]

Changes:

• bumped up opencv-python version to 4.11.0.86 (in sync with x86)
• (temporary) added opencv patch for version 86 (already fixed on main branch, will get resolved in next release)
• building and installing hf-xet from source

[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.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[Shafi-Hussain]

@DarkLight1337 could you please enable ppc64le specific tests for this PR to have them verified?

[DarkLight1337]

Hmm. @khluu I think you need to enable this over at https://github.com/vllm-project/buildkite-ci

[Shafi-Hussain]

@khluu @DarkLight1337 is the build/test for ppc64le enabled for this PR?
For reference, I wanted these jobs to run: #15402

[DarkLight1337]

Unblocking ppc64le test. Sorry for the delay!

[DarkLight1337]

cc @khluu why are the scripts missing from the CI?

[khluu]

I think this branch is stale (it was created when the script was not moved yet). Can you rebase this branch with main?

[hmellor]

I've unblocked the ppc64le test

[hmellor]

LGTM if the test passes

[hmellor]

It failed, but because of Triton this time, rather than because of hf_xet.

@Shafi-Hussain do you want to fix that in this PR or in a follow up?

[Shafi-Hussain]

Hi @hmellor, I'd prefer to have it fixed in this PR itself. Could we please hold off on merging for a while? Also, if you could share which exact step failed, it would be helpful to debug. Triton should not have been built for ppc64le builds

[hmellor]

Sure, I'll change the review so that the merge button cannot be pressed.

[hmellor]

You can see the logs in https://buildkite.com/vllm/ci/builds/17352/steps?jid=01961f98-719f-47a7-b20a-04d69d7ad280

 => ERROR [vllmcache-builder 5/5] RUN --mount=type=cache,target=/root/.cache/uv     --mount=type=bind,from=torch-builder,source=/torchwheels/,target=/t  111.1s
------
 > [vllmcache-builder 5/5] RUN --mount=type=cache,target=/root/.cache/uv     --mount=type=bind,from=torch-builder,source=/torchwheels/,target=/torchwheels/,ro     --mount=type=bind,from=arrow-builder,source=/arrowwheels/,target=/arrowwheels/,ro     --mount=type=bind,from=cv-builder,source=/opencvwheels/,target=/opencvwheels/,ro     --mount=type=bind,src=.,dst=/src/,rw     source /opt/rh/gcc-toolset-13/enable &&     uv pip install /opencvwheels/*.whl /arrowwheels/*.whl /torchwheels/*.whl &&     sed -i -e 's/.*torch.*//g' /src/pyproject.toml /src/requirements/*.txt &&     uv pip install pandas pythran pybind11 /hf_wheels/*.whl &&     export PKG_CONFIG_PATH=$(find / -type d -name "pkgconfig" 2>/dev/null | tr '\n' ':') &&     uv pip install -r /src/requirements/common.txt -r /src/requirements/cpu.txt -r /src/requirements/build.txt --no-build-isolation &&     cd /src/ &&     uv build --wheel --out-dir /vllmwheel/ --no-build-isolation &&     uv pip install /vllmwheel/*.whl:
0.278 Using Python 3.12.9 environment at: /opt/vllm
0.965 Resolved 16 packages in 683ms
2.510    Building pillow==11.1.0
23.53       Built pillow==11.1.0
23.55 Prepared 6 packages in 22.58s
24.39 Installed 15 packages in 837ms
24.39  + filelock==3.18.0
24.39  + fsspec==2025.3.2
24.39  + jinja2==3.1.6
24.39  + markupsafe==3.0.2
24.39  + mpmath==1.3.0
24.39  + networkx==3.4.2
24.39  + numpy==2.2.4
24.39  + opencv-python-headless==4.11.0.86 (from file:///opencvwheels/opencv_python_headless-4.11.0.86-cp312-cp312-linux_ppc64le.whl)
24.39  + pillow==11.1.0
24.39  + pyarrow==19.0.1 (from file:///arrowwheels/pyarrow-19.0.1-cp312-cp312-linux_ppc64le.whl)
24.39  + sympy==1.13.1
24.39  + torch==2.6.0 (from file:///torchwheels/torch-2.6.0-cp312-cp312-linux_ppc64le.whl)
24.39  + torchaudio==2.6.0 (from file:///torchwheels/torchaudio-2.6.0-cp312-cp312-linux_ppc64le.whl)
24.39  + torchvision==0.21.0 (from file:///torchwheels/torchvision-0.21.0-cp312-cp312-linux_ppc64le.whl)
24.39  + typing-extensions==4.13.1
24.41 Using Python 3.12.9 environment at: /opt/vllm
25.16 Resolved 13 packages in 748ms
25.18 Downloading pythran (4.1MiB)
26.07    Building pandas==2.2.3
26.95  Downloaded pythran
108.6       Built pandas==2.2.3
108.6 Prepared 10 packages in 1m 23s
108.9 Installed 11 packages in 294ms
108.9  + beniget==0.4.2.post1
108.9  + gast==0.6.0
108.9  + hf-xet==1.0.3 (from file:///hf_wheels/hf_xet-1.0.3-cp37-abi3-linux_ppc64le.whl)
108.9  + pandas==2.2.3
108.9  + ply==3.11
108.9  + pybind11==2.13.6
108.9  + python-dateutil==2.9.0.post0
108.9  + pythran==0.17.0
108.9  + pytz==2025.2
108.9  + six==1.17.0
108.9  + tzdata==2025.2
109.5 Using Python 3.12.9 environment at: /opt/vllm
110.2   × No solution found when resolving dependencies:
110.2   ╰─▶ Because triton==3.2.0 has no wheels with a matching platform tag (e.g.,
110.2       `manylinux_2_34_ppc64le`) and you require triton==3.2.0, we can conclude
110.2       that your requirements are unsatisfiable.
110.2
110.2       hint: Wheels are available for `triton` (v3.2.0) on the following
110.2       platforms: `manylinux_2_17_x86_64`, `manylinux2014_x86_64`

[Shafi-Hussain]

@hmellor created an issue for trtiton
#16413

[Shafi-Hussain]

@hmellor can we have this PR merged? Since the triton conflicts have been resolved

[hmellor]

Can you rebase so that we can run the test again to verify that this fix works?

[hmellor]

Test unblocked and running here

[hmellor]

Test failed, this time seemingly because of opentelemetry

[Shafi-Hussain]

@hmellor fixed the errors. It doesn't seem like the ppc64le build got triggered. Could you please check and confirm?

[hmellor]

https://buildkite.com/vllm/ci/builds/17464/steps?sid=0196250d-6a97-4d7a-addc-0dd88742c26e