Skip to content

docs: state the torch 2.14 floor where the compile doors still said 2.13 - #104

Merged
yonghakim merged 1 commit into
mainfrom
claude/torch-floor-wording
Oct 27, 2026
Merged

yonghakim merged 1 commit into
mainfrom
claude/torch-floor-wording

Conversation

@yonghakim

Copy link
Copy Markdown
Contributor

The torch floor has been torch>=2.14 since c438709. That is the first release with the Inductor eye-write miscompile fixed (pytorch/pytorch#182012). The refusal that torch.compile(fullgraph=True) prints still said Inductor "(torch 2.13) silently miscompiles", as if the defect were current.

The doors stay. Compiling wavebit buys no speed: the solve is about 98% LAPACK, which Inductor forwards as is. The doors also hold the by-name refusals of fullgraph, torch.export and torch.jit.trace. This PR only corrects the text.

Changes

  • EAGER_REASON: now gives the current reason (no speed gain, plus a past miscompile on 2.13 that 2.14 fixed).
  • Door docstring: describes the defect in the past tense, says why the doors stay, and cites the upstream fix.
  • asarray docstring: re-measured on 2.14.1. Arrays with negative strides still raise, and requires_grad and the graph are still kept.
  • torch-compile.md:
    • Says where the floor stands now (2.14).
    • Quotes the new refusal text.
    • The torch.linalg.solve forward-rule defect is now dated "2.13 and 2.14" (it is wrong on 2.14.1 too, per torch-upstream-defects.md).
  • torch-upstream-defects.md: the floor is now 2.14, and the defect-1 canary is listed under its current name.
  • backends-v1.md: the floor is now 2.14, with the reason.

Historical measurements on 2.13.0 are left as written.

Checks

tests/api/test_torch_compile.py and tests/api/test_docs_contract.py: 186 passed on torch 2.14.1.

🤖 Generated with Claude Code

The floor is torch>=2.14 since c438709, the first release with the
Inductor eye-write miscompile fixed, but the refusal a fullgraph compile
prints still said Inductor "(torch 2.13) silently miscompiles". The doors
stay; their reason now says why: compiling wavebit buys no speed, and
Inductor has miscompiled it before (2.13, fixed in 2.14).

- EAGER_REASON and the door docstring: past tense for the defect, the
  reason the doors stay.
- asarray docstring: re-measured on 2.14.1 (negative strides still raise;
  requires_grad and the graph are still kept).
- torch-compile.md, torch-upstream-defects.md, backends-v1.md: the floor
  is 2.14; torch.linalg.solve's forward-rule defect holds through 2.14;
  the defect-1 canary's current name.

Historical measurements on 2.13.0 are left as written.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@yonghakim
yonghakim merged commit 3b4817e into main Oct 27, 2026
56 checks passed
@yonghakim
yonghakim deleted the claude/torch-floor-wording branch October 27, 2026 06:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant