`scripts/requirements-minimax-h3-cuda.txt` installs diffusers from a `git+https://github.com/huggingface/diffusers.git@2da7040b\` pin because the MiniMax-H3 modular integration (PR #14355) merged to `main` on 2026-08-05, after the v0.39.0 release (2026-07-03) — no tagged release carries it. A git pin means every `INSTALL_MINIMAX_H3_CUDA=1` install clones the diffusers repo to build a wheel, which is slower and needs git on the box, and it will silently keep doing so long after a release makes it unnecessary.
Fix shape: when a diffusers release containing `src/diffusers/modular_pipelines/minimax_h3/` ships, replace the git line with `diffusers==` and drop the explanatory comment above it; nothing else changes, because the runtime's `importProbe` in `server/services/videoGen/runtimes.js` already asserts `MiniMaxH3Transformer3DModel` imports, so a wrong pin reports as not-ready rather than failing inside a render. Check with `curl -s https://api.github.com/repos/huggingface/diffusers/releases | grep tag_name` against the merge date.
Surfaced while adding the CUDA runtime.
Status update (checked 2026-08-14): Still pending — confirmed via both PyPI (`https://pypi.org/pypi/diffusers/json\`) and the GitHub releases API that `v0.39.0` (published 2026-07-03) remains the latest tagged diffusers release, which predates the PR #14355 merge (2026-08-05). No release yet ships `modular_pipelines/minimax_h3/`.
Migrated from PLAN.md by /do:replan --issues.
`scripts/requirements-minimax-h3-cuda.txt` installs diffusers from a `git+https://github.com/huggingface/diffusers.git@2da7040b\` pin because the MiniMax-H3 modular integration (PR #14355) merged to `main` on 2026-08-05, after the v0.39.0 release (2026-07-03) — no tagged release carries it. A git pin means every `INSTALL_MINIMAX_H3_CUDA=1` install clones the diffusers repo to build a wheel, which is slower and needs git on the box, and it will silently keep doing so long after a release makes it unnecessary.
Fix shape: when a diffusers release containing `src/diffusers/modular_pipelines/minimax_h3/` ships, replace the git line with `diffusers==` and drop the explanatory comment above it; nothing else changes, because the runtime's `importProbe` in `server/services/videoGen/runtimes.js` already asserts `MiniMaxH3Transformer3DModel` imports, so a wrong pin reports as not-ready rather than failing inside a render. Check with `curl -s https://api.github.com/repos/huggingface/diffusers/releases | grep tag_name` against the merge date.
Surfaced while adding the CUDA runtime.
Status update (checked 2026-08-14): Still pending — confirmed via both PyPI (`https://pypi.org/pypi/diffusers/json\`) and the GitHub releases API that `v0.39.0` (published 2026-07-03) remains the latest tagged diffusers release, which predates the PR #14355 merge (2026-08-05). No release yet ships `modular_pipelines/minimax_h3/`.
Migrated from PLAN.md by /do:replan --issues.