ComfyUI-Launcher-Environments

Standalone environment definitions consumed by ComfyUI Desktop 2.0 to provision Python environments for ComfyUI installations.

Torch index stacks manifest

torch-index-stacks.json lists the index-served PyTorch stacks (pip-applied tuples with no bundle artifact) that the desktop app offers in its PyTorch picker, e.g. legacy CUDA builds for GPUs the current bundles dropped. The desktop app fetches it from R2 on every update check, so new stacks ship by editing this file - no app release needed.

• Edit via pull request; CI validates the file with scripts/validate_torch_index_stacks.py (the app silently drops invalid entries, so validation runs here where a mistake fails loudly).
• On merge to main, the Publish Torch Index Stacks workflow uploads it to standalone-environments/torch-index-stacks.json on R2. The desktop app also has a GCS fallback mirror that is synced outside this repository - until that sync is automated (#10), fallback clients can serve a stale manifest after a withdrawal, so update the mirror alongside urgent changes.
• The desktop app validates entries default-deny at runtime; this file cannot point installs at arbitrary indexes.
• An explicit empty "stacks": [] withdraws every index-served stack from all apps; deleting entries selectively withdraws just those.
• Stable entries never carry PEP 440 dev (nightly) versions - validation rejects them. Nightlies use "kind": "pytorch-nightly-index" and are managed exclusively by the scheduled Refresh Nightly Torch Stacks workflow (scripts/refresh_nightly_stacks.py), which re-resolves exact dated tuples from the live nightly indexes daily, validates, commits, and publishes. Do not hand-edit nightly entries: dated nightly wheels are purged from pytorch.org's index after roughly 60 days, so they are pins that decay - validation rejects any nightly entry older than 7 days, and the desktop app stops offering entries ~45 days after their wheel date. The tag allowlist (NIGHTLY_TAGS, currently cu132) lives in the refresh script; old app versions ignore the nightly kind entirely.

Building and testing environments locally (no publishing)

scripts/build_local.py reproduces the Build Standalone Environment workflow for one variant on your machine, and scripts/serve_local.py serves the result with the same URL layout as R2 so a dev desktop can ingest it. Nothing is uploaded anywhere.

Prerequisites: Python 3, git, 7-Zip (7z on PATH, or the default Windows install location), network access, and tens of GB of free disk. Only variants matching the host OS can be built (pip resolves wheels for the running platform) - for other platforms, run the workflow via workflow_dispatch on a fork without R2 secrets, which skips every publish step and leaves the archives on the draft GitHub release.

# 1. Build a variant (output lands in dist/<variant>/<tag>/)
python scripts/build_local.py win-nvidia --comfyui-ref v0.4.61

# 2. Generate R2-style metadata and serve dist/ on 127.0.0.1:8000.
#    Also serves the working-tree torch-index-stacks.json, so local
#    manifest edits are testable the same way.
python scripts/serve_local.py

# 3. Point an unpackaged desktop at it (packaged builds ignore this)
cd ..\ComfyUI-Launcher
$env:COMFY_STANDALONE_BASE_URL = 'http://127.0.0.1:8000'; pnpm dev

While the override is active the desktop's GCS mirror fallback is disabled, so a failure against the local server fails loudly instead of silently falling back to production metadata. Local builds default to a <ref>-local1 tag; keep a -local suffix so a local artifact can never shadow a real release tag. Keep build_local.py's constants in sync with the workflow matrix when it changes; scripts/test_local_tooling.py covers the metadata generation (python -m unittest scripts.test_local_tooling).