[RFC]: Is huggingface-cli[hf_xet] needed for vllm build?

[Shafi-Hussain]

Motivation.

Recently huggingface-cli[hf_xet] was added in requirements/common.txt and requirements/test.in.

Is it needed in requirements/common.txt ?
If not, can it be moved to rocm build/test requirements?

Ref: 15969

Proposed Change.

Moving huggingface-cli[hf_xet] from requirements/common.txt to requirements/rocm.txt or requirements/rocm-test.txt

Feedback Period.

No response

CC List.

@hmellor

Any Other Things.

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[hmellor]

huggingface-cli[hf_xet] is not device specific and has no particular affinity to ROCm.

Xet allows vLLM to download models which use Xet storage from the Hugging Face Hub faster.

[Shafi-Hussain]

@hmellor Yes, but it's not a build-time requirement vllm. The only reason I suggested rocm, was because it was mentioned to resolve installation issues of ROCm.

Adding it as a build time dependency in common.txt breaks builds for ppc64le.
It should ideally be a test dependency, needed only at runtime.

[hmellor]

Ok, thank you for clarifying.

vLLM does not currently differentiate between build time and runtime dependencies. The same requirements are used for both.

We could add a condition in common.txt to ensure that the Xet version is not installed if the platform is ppc64le?

[Nash-123]

@hmellor We're experiencing the same issue on IBM Z (s390x) as well. Adding s390x to that platform condition alongside ppc64le would solve our build problem as well.

18.26 💥 maturin failed 18.26 Caused by: Failed to normalize python source path python18.26 Caused by: No such file or directory (os error 2) 18.26 Error: command ['maturin', 'pep517', 'build-wheel', '-i', 18.26 '/root/.cache/uv/builds-v0/.tmp0aHjDE/bin/python', '--compatibility', 18.26 'off'] returned non-zero exit status 1 18.26 18.26 hint: This usually indicates a problem with the package or the build 18.26 environment. 18.26 help:hf-xet(v1.0.0) was included becausehuggingface-hub[hf-xet]18.26 (v0.30.1) depends onhf-xet18.26 DEBUG Released lock at/opt/vllm/.lockDockerfile.s390x:118

[Shafi-Hussain]

@hmellor thanks for the confirmation. We wanted to understand if it was indeed needed to build vllm. We are having discussions on building hf-xet from source for ppc64le. We can keep the commons.txt as it is for now.

@Nash-123 we are having discussions for building on power

[hmellor]

I can confirm that it's not needed for building vLLM, it is just to improve model checkpoint download speeds at runtime.

[Akashcodes732]

Hi @Shafi-Hussain, I am facing the same issue and the build is breaking for ppc64le. Were you able to figure out any workaround, otherwise I was planning on adding some conditional check in case of ppc64le.

[Shafi-Hussain]

Raised a PR where hf-xet is built from source: #16048

[Nash-123]

@hmellor thanks for the confirmation. We wanted to understand if it was indeed needed to build vllm. We are having discussions on building hf-xet from source for ppc64le. We can keep the commons.txt as it is for now.

@Nash-123 we are having discussions for building on power

Thanks for the update @Shafi-Hussain !
For s390x (IBM Z), we’ve also followed the same approach to built hf-xet from source.