Add CUDA 12.8 Docker runtime with pinned dependencies

[LingLambda]

Summary


When I first tried to use the Docker runtime in this repository, the image could not start due to several dependency issues. After resolving those conflicts, I also ran into compatibility problems on an RTX 50-series GPU because the existing Dockerfile is based on an older CUDA/PyTorch runtime.

This PR keeps the existing Dockerfile unchanged and adds a separate CUDA 12.8 runtime Dockerfile, so users on older GPUs or older drivers are not forced onto the newer runtime. In local testing, the default runtime would not start successfully until these dependency issues were resolved.


Changes



• Add runtime/python/Dockerfile.cuda128 based on pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime
• Keep torch and torchaudio pinned to 2.8.0 in the CUDA 12.8 image
• Use onnxruntime-gpu==1.22.0 to avoid CUDA 11 library issues such as missing libcublasLt.so.11
• Avoid dependency conflicts from packages that are incompatible with the CUDA 12 runtime
• Keep setuptools<80 to avoid pkg_resources-related build/import issues
• Install openai-whisper without pulling conflicting dependencies
• Update the README with instructions for building the CUDA 12.8 image
• Keep the original Dockerfile unchanged for existing users
  

Validation


Built and tested the CUDA 12.8 image locally on an RTX 5060 Ti. Verified that PyTorch can use CUDA and that the FastAPI service starts with CosyVoice2-0.5B.

Thank you to the maintainers for all your hard work and continued dedication to this project.

Related issues: #1844, #1123, #1153