[RoadMap] SGL-Kernel-NPU binary release

[iforgetmyname]

Background

We are happy to announce that SGL-Kernel-NPU has achieved 100k+ lines of code 🎉

However, this arises the problem that compilation takes quite a long time, which is not affordable for prepare stage in pr-test-npu of the SGLang repository. To fix this problem, SGL-Kernel-NPU will gradually switch to binary release for now on, and here's the roadmap for this project.

Naming Conventions

SGL-Kernel-NPU will be released in zipped packages containing three different python package releases:

• DeepEP on NPU
• Torch-Memory-Saver on NPU
• SGL-Kernel-NPU

Depending on our current package dependencies, the zipped package will be named after:

sgl-kernel-npu-torch{TORCH_VERSION}-py{PYTHON_VERSION}-cann{CANN_VERSION}-{NPU_PLATFORM}-{OS_ARCHITECTURE}.zip

e.g.
sgl-kernel-npu-torch2.8.0-py311-cann8.5.0-910b-aarch64.zip

Key Milestones

• Uniforms open-source licenses
• Uniforms releasing version with YYYY.MM.DD
  • sgl_kernel_npu using new release version sgl-kernel-npu add release version #253
• Switches to pyproject.toml package description and deprecates outdated setup.py