[CPU] enable CI for PRs, add Dockerfile and auto build task

[ZailiWang]

Motivation

The PR is for enabling the docker env setup and CI processes for running SGLang on Xeon CPU servers.

Modifications

Add the dockerfile, and yml files for CPU part of CI process and auto docker image build-up.
The test files are updated to enable device specific tests, as well as the CPU test cases.

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

[mingfeima]

can docker file accepts arguements? if yes, we can change TORCH_VERSION as an arguement and set it default to be 2.6

[ZailiWang]

can docker file accepts arguements? if yes, we can change TORCH_VERSION as an arguement and set it default to be 2.6

Yeah, by setting ARG VER_TORCH=2.6.0 in the Dockerfile (L5 in the added Dockerfile in this PR), it is a settable argument and can be changed by e.g. --build-arg VER_TORCH=2.7.0 in docker build command.

[zhyncs]

ref #6408 (comment)

[mingfeima]

@ZailiWang can we satisfy:

• skip sgl-kernel cpu tests when CPU doesn't have AMX
• skip certain tests when vllm is not installed (since some test cases, e.g. rope, uses reference from sglang)

[mingfeima]

@ZailiWang change the tile and provide some details of what this PR is for?

[ZailiWang]

@ZailiWang can we satisfy:

• skip sgl-kernel cpu tests when CPU doesn't have AMX
• skip certain tests when vllm is not installed (since some test cases, e.g. rope, uses reference from sglang)

Hi @zhyncs ,
We have drafted this PR for CPU CI and docker env build-up. Yet there are still something that we are unsure. Your comment and guidance is highly appreciated!

• For satisfying this

• skip sgl-kernel cpu tests when CPU doesn't have AMX

We added a Check AMX Support step in CPU CI process and CPU test cases are triggered only when the check is passed here. Is this good or would you comment with some better solution?

• For CPU docker build-up, we created release-docker-xeon.yml here by imitating "release-docker" files for other devices. Please correct if anything still required to be updated in this file.

• In 2 CPU test cases (test_rope.py and test_topk.py), vllm implementation is still invoked when calculating the reference results. Since in SGLang we do not install vllm as a dependency now, such errors would be thrown:

File "/sgl-workspace/sglang/python/sglang/srt/layers/quantization/utils.py", line 14, in <module>
    from vllm._custom_ops import scaled_fp8_quant
ModuleNotFoundError: No module named 'vllm'

File "/sgl-workspace/sglang/python/sglang/srt/layers/rotary_embedding.py", line 87, in __init__
    from vllm._custom_ops import rotary_embedding
ModuleNotFoundError: No module named 'vllm'

Hence these 2 test cases are not added into run_suite.py yet.

[ZailiWang]

In 2 CPU test cases (test_rope.py and test_topk.py), vllm implementation is still invoked when calculating the reference results. Since in SGLang we do not install vllm as a dependency now, such errors would be thrown:

We have another PR to mitigate vllm dependency issue now. Will add back the 2 test cases when #6614 merged.

[zhyncs]

Currently this has not been set up in the sglang repo as a self-hosted runner.