ai-toolkit Spark wheel set (Windows on ARM, CUDA 13.4)

Builds the complete native win_arm64 Python wheel set that ai-toolkit uses on NVIDIA RTX Spark (N1X) machines: PyTorch + CUDA 13.4 (sm_121), torchvision, torchaudio, torchcodec, flash-attn, NATTEN, triton (torch.compile), bitsandbytes, llvmlite/numba (librosa), opencv, and the small deps with no upstream arm64 wheels.

Everything is reproducible from pinned sources + patches:

• pins.json — every upstream repo/sdist, ref, version, and build note
• patches/ — our modifications as reviewable diffs
• scripts/ — the per-component build scripts (exact proven flags)
• build.py — orchestrator: python build.py all → dist/*.whl

Prerequisites

One manual install (NVIDIA's developer-preview EULA requires their own click-through; we do not redistribute any NVIDIA components):

1. CUDA 13.4 toolkit (arm64) + RTX Spark developer driver (R616+) — https://developer.nvidia.com/cuda-13-4-0-download-archive?target_os=Windows&target_arch=arm64

Everything else installs unattended (or run python build.py doctor to check):

• VS2022 Build Tools with ARM64 MSVC + Windows 11 SDK (winget)
• cuDNN 9.x arm64 (NVIDIA's official public installer)
• Arm Performance Libraries (winget)
• Rust aarch64-pc-windows-msvc (rustup), pkg-config-lite (winget)
• native arm64 Python 3.12 (uv)

Build

python build.py doctor      # verify toolchain
python build.py --plan      # see the component order
python build.py all         # several hours; torch and LLVM dominate

Wheels land in dist/. Memory guidance: the flash-attn/NATTEN CUTLASS compiles are memory-hungry — the scripts pin low job counts on purpose (learned the hard way on a 54 GB machine).

Publishing

Upload dist/*.whl as a GitHub release, publish a simple find-links index.html (GitHub Pages) linking to the assets, then point ai-toolkit's manager/spec.py SPARK_WHEELS_URL at that page. The ai-toolkit manager handles end-user runtime provisioning (cuDNN/APL/MSVC auto-install, CUDA detection + instructions) by itself.

Updating

Bump the pin in pins.json, re-run the component. A patch that no longer applies fails loudly at exactly the hunk that needs attention. Watch upstream: pytorch#190448 / #189949 (WoA CUDA enablement), triton-windows#49 (official ARM64 wheels — retire our triton when solid), numba cp312 arm64 wheels (retire the LLVM/llvmlite/numba chain).