Summary

• Installs Triton automatically where PyTorch doesn’t provide it by default:
  • Windows + CUDA (cu* / nightly/cu*): installs triton-windows
  • Linux + ROCm 6.x+: installs pytorch-triton-rocm (from https://download.pytorch.org/whl)
  • Linux + XPU: installs pytorch-triton-xpu (from https://download.pytorch.org/whl)
• Adds missing runtime dependency packaging (required by torchruntime.platform_detection).
• Updates docs and adds installer unit tests.

Refs: #5

Why

torch.compile (and many third-party kernels) require Triton. On some platforms Torch bundles it (e.g. Linux CUDA), but on others users end up without Triton even after installing a GPU build of Torch.

Implementation

• torchruntime/installer.py appends an extra pip install command for the platform-specific Triton package.
• Kept logic minimal and platform-gated (no changes to platform detection).

Testing

• Unit tests: python -m pytest -q.
• Real GPU smoke test (Windows, RTX 3060 Ti, Python 3.10):
  1. Create clean venv
  2. Install torch/torchvision/torchaudio from https://download.pytorch.org/whl/cu128
  3. Verify import triton fails
  4. Run python -m torchruntime install -> installs triton-windows
  5. Run a torch.compile CUDA smoke test -> OK

Request For Testing (hardware help wanted)

If you have one of these setups, please try:

• Linux ROCm (6.x): python -m torchruntime install then verify import triton and run a small torch.compile test.
• Linux Intel XPU: same as above.
  Also welcome: Windows CUDA users on different GPUs/Python versions.