ComfyUI Install Script (Linux / Windows)

<ul><li>Added<ul><li>Easy, Advanced, and Skip configuration modes with context-aware <code>-</code> preservation and selective embedded-default saving</li><li>NVIDIA, Linux ROCm, and CPU backend selection with validated PyTorch companion-package resolution</li><li>Step 12 compatibility audit with <code>uv pip check</code>, conservative directional repair for known shared intersections, core runtime import probes, and explicit unresolved-conflict reporting</li><li>Toolchain-gated SageAttention 2.2.0 builds with a verified 1.0.6 fallback</li><li>Native Windows AMD ROCm 7.2.1 installation with official AMD SDK and PyTorch wheels, Windows 11 and Python 3.12 validation, preserved-environment detection, and a HIP GPU runtime probe</li></ul>Changed<ul><li>ComfyUI-Manager nodes under <code>custom_nodes/.disabled/</code> and legacy <code>*.disabled</code> directories are preserved, excluded from cloning, and skipped during dependency installation</li><li>Flash Attention now installs only an exact official upstream Linux wheel, is force-replaced and import-tested, and is removed when unavailable or ABI-incompatible; Windows no longer uses community wheels or an automatic source build</li><li>Package upgrade work uses package-scoped uv upgrades so ultralytics and GGUF updates preserve compatible Filelock, OpenCV, Packaging, AV, protobuf, inference, tokenizers, and related runtime stacks</li><li>Compatibility repair prioritizes the managed runtime when legacy inference packages conflict with current inference-cli, Hugging Face, Pillow, Aiohttp, Click, or typing requirements</li><li>Installation progress and selection now consistently cover 12 steps while preserving launcher behavior as step 11</li><li>Windows Advanced configuration now offers AMD/ROCm alongside NVIDIA and CPU, while disabling NVIDIA-only Nunchaku and attention packages for AMD installations</li></ul>Fixed<ul><li>Prevented a Flash Attention wheel compiled for an older Torch ABI from remaining installed after a Torch upgrade and breaking Kornia and ComfyUI core imports</li><li>Existing non-Git custom-node directories are reported as unmanaged instead of receiving a failing clone attempt</li><li>Audit-only and other partial-run summaries report versions from the configured venv, mark untouched ComfyUI checkouts, and omit nonexistent launcher paths</li><li>Step 9 keeps <code>nvidia-ml-py</code> below 13 because the installed inference packages require the 12.x API</li><li>Windows step 9 now writes its managed constraint file before using it</li></ul></li></ul>

comfyuiInstallScript_013.zip

Files with SHA256: 691681ad80bc045cf522deb810c3861fd79882a71d7c94223436166564a665f9