[chenmiaoming]

Summary

Update aistation-13.0.Dockerfile without changing the published image name.

Changes

• upgrade vLLM from 0.24.0 to 0.25.1
• upgrade SGLang from 0.5.14 to 0.5.15.post1
• remove the explicit PyTorch CUDA backend selector and let each framework resolve its declared PyTorch dependency
• download Miniforge from the official GitHub latest/download release URL
• keep vLLM and SGLang in separate Conda environments while installing the frameworks from PyPI with uv
• remove the global PYTHONUNBUFFERED=1 setting
• scope FLASHINFER_USE_CUDA_NORM=1 to the sglang Conda environment
• fix the exec-form SSH command

Packaging decision

conda-forge currently publishes vLLM, but its available version trails the PyPI release used here. No suitable conda-forge SGLang package was identified. Using Conda for environment isolation and uv/PyPI for the rapidly changing inference stacks keeps both environments independent and current.

Release behavior

This PR targets main. PR CI builds the modified image without publishing it. The GHCR release workflow runs only after the PR is merged into main.

Validation

The existing PR Docker build workflow validates dependency resolution and the complete image build.