[PR #105] [MERGED] Document the actual PyTorch minimum (>=2.7.0) #146

New issue

Closed

opened 2026-10-26 22:20:51 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:20:51 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/105
Author: @christian-byrne
Created: 8/10/2026
Status: ✅ Merged
Merged: 8/16/2026
Merged by: @christian-byrne

Base: main ← Head: docs/torch-min-version-2.7

📝 Commits (1)

05c5f01 Document the actual PyTorch minimum (>=2.7.0)

📊 Changes

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

View changed files

📝 README.md (+1 -1)

📄 Description

The README lists **PyTorch**: ≥2.5.0, but as of 0.2.28 two things run at import time that need 2.7.0:

backends/eager/na.py:163 registers comfy_kitchen::na3d with PEP 585 annotations (kernel_size: list[int], is_causal: list[bool]). torch.library.infer_schema only accepts builtin generics from 2.7.0 — derived_types gained the GenericAlias(list, ...) entries in v2.7.0 and they are absent in v2.6.0. On 2.5/2.6 the decorator raises ValueError at import.
backends/eager/__init__.py:542 references torch.float8_e8m0fnu inside _build_constraints(), which is called at module scope (line 586). That dtype landed in 2.7.0 (c10/core/ScalarType.h: 0 occurrences at v2.6.0, 6 at v2.7.0).

Both blow up on import comfy_kitchen itself rather than at first use, so the package is unusable below 2.7.0 regardless of backend or device — including CPU-only setups that never touch a kernel.

Verified on clean venvs:

torch	`import comfy_kitchen` (0.2.28)
2.6.0	`ValueError: infer_schema(func): Parameter kernel_size has unsupported type list[int]`
2.7.0+cpu	OK — eager backend available

Docs-only change; no code or behavior is affected. Assumes the higher floor is intended, so this just makes the README match.

Possible follow-ups, out of scope here: pyproject.toml has dependencies = [], so nothing enforces the floor at install time, and the repo has no CI job that imports the package at the documented minimum.

_{🔄 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/comfy-kitchen/pull/105 **Author:** [@christian-byrne](https://github.com/christian-byrne) **Created:** 8/10/2026 **Status:** ✅ Merged **Merged:** 8/16/2026 **Merged by:** [@christian-byrne](https://github.com/christian-byrne) **Base:** `main` ← **Head:** `docs/torch-min-version-2.7` --- ### 📝 Commits (1) - [`05c5f01`](https://github.com/Comfy-Org/comfy-kitchen/commit/05c5f0197859fca3e8c706ca63ae5dda6e4e446e) Document the actual PyTorch minimum (>=2.7.0) ### 📊 Changes **1 file changed** (+1 additions, -1 deletions) <details> <summary>View changed files</summary> 📝 `README.md` (+1 -1) </details> ### 📄 Description The README lists `**PyTorch**: ≥2.5.0`, but as of 0.2.28 two things run at import time that need 2.7.0: - **`backends/eager/na.py:163`** registers `comfy_kitchen::na3d` with PEP 585 annotations (`kernel_size: list[int]`, `is_causal: list[bool]`). `torch.library.infer_schema` only accepts builtin generics from 2.7.0 — `derived_types` gained the `GenericAlias(list, ...)` entries in v2.7.0 and they are absent in v2.6.0. On 2.5/2.6 the decorator raises `ValueError` at import. - **`backends/eager/__init__.py:542`** references `torch.float8_e8m0fnu` inside `_build_constraints()`, which is called at module scope (line 586). That dtype landed in 2.7.0 (`c10/core/ScalarType.h`: 0 occurrences at v2.6.0, 6 at v2.7.0). Both blow up on `import comfy_kitchen` itself rather than at first use, so the package is unusable below 2.7.0 regardless of backend or device — including CPU-only setups that never touch a kernel. Verified on clean venvs: | torch | `import comfy_kitchen` (0.2.28) | |---|---| | 2.6.0 | `ValueError: infer_schema(func): Parameter kernel_size has unsupported type list[int]` | | 2.7.0+cpu | OK — eager backend available | Docs-only change; no code or behavior is affected. Assumes the higher floor is intended, so this just makes the README match. Possible follow-ups, out of scope here: `pyproject.toml` has `dependencies = []`, so nothing enforces the floor at install time, and the repo has no CI job that imports the package at the documented minimum. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 22:20:51 +00:00