[chenmiaoming]

What changed

• add a dedicated tvm Conda environment to aistation-13.0.Dockerfile
• install LLVM 21 from conda-forge instead of compiling LLVM from source
• build Apache TVM v0.25.0 from source with CUDA, cuBLAS, cuDNN, RPC, and hidden private symbols enabled
• install and pin PyTorch 2.13.0 from the official cu132 wheel index inside the tvm environment
• install TVM Python dependencies and tvm-ffi in the same environment
• scope TVM_HOME, TVM_LIBRARY_PATH, LLVM_CONFIG, and the TVM-specific LD_LIBRARY_PATH to the tvm Conda environment
• add build-time checks for LLVM, TVM, CUDA build flags, and PyTorch CUDA 13.2 metadata

Why

The previous TVM setup compiled LLVM from source, which made image builds substantially slower. Keeping LLVM, TVM dependencies, and PyTorch in one isolated Conda environment reduces build time and prevents TVM-specific environment variables from affecting the existing base, vLLM, or SGLang environments.

Validation

• branch is one commit ahead of main
• only aistation-13.0.Dockerfile is modified
• PR Docker build CI should build the updated image automatically

The final GPU runtime check still requires running the built container with an NVIDIA GPU; the Docker build itself validates package metadata and TVM build configuration without initializing a GPU.