[CPU][Fix CI] Solidate torch version for sgl-kernel-cpu and fix device orientation error

[ZailiWang]

Motivation

In sgl-kernel/pyproject_cpu.toml the requirement of PyTorch version is >=2.7.1. However this triggers the installation of the latest torch==2.10.0, leading to symbol errors for build output, as sglang is built with torch==2.9.0.

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/opt/.venv/lib/python3.12/site-packages/sgl_kernel/__init__.py", line 5, in <module>
    common_ops = _load_architecture_specific_ops()
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/.venv/lib/python3.12/site-packages/sgl_kernel/load_utils.py", line 188, in _load_architecture_specific_ops
    raise ImportError(error_msg)
ImportError: 
[sgl_kernel] CRITICAL: Could not load any common_ops library!

Attempted locations:
1. Architecture-specific pattern: /opt/.venv/lib/python3.12/site-packages/sgl_kernel/sm100/common_ops.* - found files: []
2. Fallback pattern: /opt/.venv/lib/python3.12/site-packages/sgl_kernel/common_ops.* - found files: ['/opt/.venv/lib/python3.12/site-packages/sgl_kernel/common_ops.cpython-312-x86_64-linux-gnu.so']
3. Standard Python import: common_ops - failed

GPU Info:
- Compute capability: None
- Expected variant: CPU/No GPU detected (using precise math)

Please ensure sgl_kernel is properly installed with:
pip install --upgrade sgl_kernel

Error details from previous import attempts:
- ImportError: /opt/.venv/lib/python3.12/site-packages/sgl_kernel/common_ops.cpython-312-x86_64-linux-gnu.so: undefined symbol: _ZN3c1013MessageLogger6streamB5cxx11Ev
- ModuleNotFoundError: No module named 'common_ops'

Modifications

Solidate torch version in sgl-kernel/pyproject_cpu.toml as the same with sglang dependency.

Accuracy Tests

N/A

Benchmarking and Profiling

N/A

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[ZailiWang]

/tag-run-ci-label

[ZailiWang]

Fix another CI error introduced by #11657 in which a 'cuda' device was hard-coded.
The fix is to put forward the device orientation judgement code block and change 'cuda' to the deducted device name.

[ZailiWang]

Hi @ch-wan @ympcMark @ShangmingCai would you help review the device orientation fix? Current PR test shows that the CI failure on CPU/XPU can be fixed, but please confirm if the fix is still suitable for the feature in #11657 for CUDA as well as other devices. Thanks

[ZailiWang]

Added some necessary changes in .toml files, for CPU specific wheel file publishment on PyPI.

• The -cpu suffix is added into the output wheel package names for sglang and sgl-kernel.
• The scikit-build.wheel.packages is explicitly set as the wheel name and package folder name are different.

[ZailiWang]

Add torchao installation as it is a mandatory requirement. Also added torchaudio installation originally in #15859 to avoid conflict.