[PR #15] [MERGED] Offer one stable stack per PyTorch minor (2.9+) on every index card #19

New issue

Closed

opened 2026-10-26 22:22:59 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:22:59 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/ComfyUI-Standalone-Environments/pull/15
Author: @Kosinkadink
Created: 7/28/2026
Status: ✅ Merged
Merged: 7/28/2026
Merged by: @Kosinkadink

Base: main ← Head: feat/one-stack-per-torch-minor

📝 Commits (2)

0af28c7 Offer one stable stack per PyTorch minor (2.9+) on every index card
56d32fe Drop the xpu 2.9.1 entry: torch import fails on Windows

📊 Changes

1 file changed (+267 additions, -0 deletions)

View changed files

📝 torch-index-stacks.json (+267 -0)

📄 Description

What

Expands the stable catalog from 14 to 34 stacks so every index card offers one stable release per PyTorch minor starting at 2.9, wherever the complete tuple actually exists upstream AND works.

Card	Before	After
cu126	2.9, 2.11, 2.13	2.9, 2.10, 2.11, 2.12, 2.13
cu128	2.9, 2.11	2.9, 2.10, 2.11 (index tops out at 2.11)
cu130	2.11, 2.13	2.9, 2.10, 2.11, 2.12, 2.13
rocm7.1 (linux)	2.11	2.10, 2.11, 2.12, 2.13 (no 2.9 on that index)
rocm7.14.0 multi-arch	2.10, 2.11	2.10, 2.11, 2.12 (index serves 2.10-2.12)
xpu	2.11, 2.13	2.10, 2.11, 2.12, 2.13 (2.9.1 dropped - see below)
cpu	2.11	2.9, 2.10, 2.11, 2.12, 2.13
pypi/mps	2.11	2.9, 2.10, 2.11, 2.12, 2.13

Verification

Every added tuple (torch + matching torchvision + torchaudio) was verified against the live index listings for the card's platforms and cp310-cp314 ABIs (PEP 503 pages for pytorch.org and repo.amd.com; PyPI JSON API for the mps card, including the cpXY-none-macosx wheels that 2.9/2.10 macOS torch ships).
torchaudio publishes nothing past 2.11.0 on any index, so 2.12/2.13 tuples pair with torchaudio 2.11.0 - same convention as the pre-existing 2.13 entries.
Dates are the actual torch release upload dates from PyPI. The new AMD multi-arch 2.12.0 entry uses the ROCm 7.14.0 GA date (2026-07-15) like its siblings; since the desktop sorts by date and ties keep manifest order, the E2E-validated 2.11.0 entry stays the picker default.
python scripts/validate_torch_index_stacks.py torch-index-stacks.json: OK (34 stacks)
python -m unittest scripts.test_torch_index_scripts: 45 tests, OK

Hardware E2E validation of new entries

AMD multi-arch 2.12.0 (Radeon PRO W7800, Windows): picker switch, clean package reconciliation, real Z-Image-Turbo inference (8/8 steps), and switch-back to the universal baseline all passed. Picker default stayed 2.11.0 as intended.
xpu 2.12.1 (Arc B570, Windows): picker switch, real SD1.5 inference on xpu:0, all passed.
xpu 2.9.1: REMOVED from this PR. The wheels exist upstream, but on a current Windows Arc setup (B570, Python 3.13.12, driver 32.0.101.8860) importing torch fails with WinError 127 loading torch_python.dll after a successful pip install, so the switch verification deterministically fails and rolls back. A tuple that exists but cannot pass verification must not be offered.
Linux AMD validation (rocm7.1 new entries + first Linux exercise of the multi-arch path) is in progress on a Linux/AMD box.

No validator or schema changes needed - additions only.

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/Comfy-Org/ComfyUI-Standalone-Environments/pull/15 **Author:** [@Kosinkadink](https://github.com/Kosinkadink) **Created:** 7/28/2026 **Status:** ✅ Merged **Merged:** 7/28/2026 **Merged by:** [@Kosinkadink](https://github.com/Kosinkadink) **Base:** `main` ← **Head:** `feat/one-stack-per-torch-minor` --- ### 📝 Commits (2) - [`0af28c7`](https://github.com/Comfy-Org/ComfyUI-Standalone-Environments/commit/0af28c7e83dd61ca6d6965f166d0d71e0cbfdc17) Offer one stable stack per PyTorch minor (2.9+) on every index card - [`56d32fe`](https://github.com/Comfy-Org/ComfyUI-Standalone-Environments/commit/56d32feebfc4cbbf58c3293d11495d31db095532) Drop the xpu 2.9.1 entry: torch import fails on Windows ### 📊 Changes **1 file changed** (+267 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `torch-index-stacks.json` (+267 -0) </details> ### 📄 Description ## What Expands the stable catalog from 14 to 34 stacks so every index card offers **one stable release per PyTorch minor starting at 2.9**, wherever the complete tuple actually exists upstream AND works. | Card | Before | After | |------|--------|-------| | cu126 | 2.9, 2.11, 2.13 | 2.9, 2.10, 2.11, 2.12, 2.13 | | cu128 | 2.9, 2.11 | 2.9, 2.10, 2.11 (index tops out at 2.11) | | cu130 | 2.11, 2.13 | 2.9, 2.10, 2.11, 2.12, 2.13 | | rocm7.1 (linux) | 2.11 | 2.10, 2.11, 2.12, 2.13 (no 2.9 on that index) | | rocm7.14.0 multi-arch | 2.10, 2.11 | 2.10, 2.11, 2.12 (index serves 2.10-2.12) | | xpu | 2.11, 2.13 | 2.10, 2.11, 2.12, 2.13 (2.9.1 dropped - see below) | | cpu | 2.11 | 2.9, 2.10, 2.11, 2.12, 2.13 | | pypi/mps | 2.11 | 2.9, 2.10, 2.11, 2.12, 2.13 | ## Verification - Every added tuple (torch + matching torchvision + torchaudio) was verified against the live index listings for the card's platforms and cp310-cp314 ABIs (PEP 503 pages for pytorch.org and repo.amd.com; PyPI JSON API for the mps card, including the `cpXY-none-macosx` wheels that 2.9/2.10 macOS torch ships). - torchaudio publishes nothing past 2.11.0 on any index, so 2.12/2.13 tuples pair with torchaudio 2.11.0 - same convention as the pre-existing 2.13 entries. - Dates are the actual torch release upload dates from PyPI. The new AMD multi-arch 2.12.0 entry uses the ROCm 7.14.0 GA date (2026-07-15) like its siblings; since the desktop sorts by date and ties keep manifest order, the E2E-validated 2.11.0 entry stays the picker default. - `python scripts/validate_torch_index_stacks.py torch-index-stacks.json`: OK (34 stacks) - `python -m unittest scripts.test_torch_index_scripts`: 45 tests, OK ## Hardware E2E validation of new entries - **AMD multi-arch 2.12.0** (Radeon PRO W7800, Windows): picker switch, clean package reconciliation, real Z-Image-Turbo inference (8/8 steps), and switch-back to the universal baseline all passed. Picker default stayed 2.11.0 as intended. - **xpu 2.12.1** (Arc B570, Windows): picker switch, real SD1.5 inference on xpu:0, all passed. - **xpu 2.9.1: REMOVED from this PR.** The wheels exist upstream, but on a current Windows Arc setup (B570, Python 3.13.12, driver 32.0.101.8860) importing torch fails with `WinError 127` loading `torch_python.dll` after a successful pip install, so the switch verification deterministically fails and rolls back. A tuple that exists but cannot pass verification must not be offered. - Linux AMD validation (rocm7.1 new entries + first Linux exercise of the multi-arch path) is in progress on a Linux/AMD box. No validator or schema changes needed - additions only. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 22:22:59 +00:00

closed this issue
added the
pull-request
label

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

Comfy-Org/ComfyUI-Standalone-Environments#19

No description provided.

Rows
Columns