Build process - gh200 with pytorch uvm patch

docker build examples with "Drone + Gitea" for gh200 pytorch (with uvm), torchvision, triton, xformers

Steps

1. build pytorch version required for vllm version

   • apply gh200 uvm patch before build (https://github.com/feuler/gh200-pytorch/tree/v2.5.1)

2. build torchvision version matching the torch version with the custom built torch pre-installed

3. build xformers (optional) with custom torch and torchvision whl packages pre-installed

4. build triton

   • pre-install custom torch whl package
   • checkout triton version defined in pytorch version file ".ci/docker/ci_commit_pins/triton.txt" (version number in ".ci/docker/triton_version.txt")
   • clone & checkout & build llvm version defined in triton version file "triton/cmake/llvm-hash.txt"
   • use llvm build in env as llvm paths (lib/include/root)
   • build triton

5. build vllm package with above whl packages pre-installed and use "use_existing_torch.py" vllm script

6. build vllm-docker container by installing above whl packages for torch, torchvision, xformers, triton and vllm

Current build environment for all involved builds

• CUDA: 12.4.1
• Python: 3.11
• Docker image: nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04