Repository navigation
Compress the agent conventions file and document two modular pitfalls - #48586
Merged
Merged
Conversation
`.ai/AGENTS.md` is what code agents are handed as the repo's conventions (serge injects it into every patch prompt), and its modular section stops at "edit the modular file, not the generated one". Two things it does not say cost two wrong patches on the same PR (#48536): - `_no_split_modules = AttributeError()` was read as an invalid placeholder and replaced with `[]`, which changes how the model may be sharded. The idiom is documented in `modular_transformers.md`, but nothing pointed there from here. - the generated `modeling_deepseek_vl.py` was hand-edited to match, and `modeling_deepseek_vl_hybrid.py` — generated from the same modular file — was left stale, so `Check repository consistency` went red on a model the patch never mentioned. Both bullets link to the existing sections rather than restating them. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
This file loads for every agent (the root `AGENTS.md` / `CLAUDE.md` symlinks point here, and serge injects it into every patch prompt), so its length is a fixed cost on every session's context. Same rules, same commands, same links, tighter prose: 4,738 -> 3,160 chars, i.e. 439 chars BELOW what it was before the two modular bullets were added. The three "before opening a PR" sections are merged into one, since coordination, duplicate checks and busywork were all answering the same question. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
tarekziade
changed the title
molbap
approved these changes
Sep 7, 2026
molbap
left a comment
molbap
left a comment
Collaborator
There was a problem hiding this comment.
SGTM, better to reduce the prefix indeed!
| Model files in `src/transformers/models/` avoid inheriting from each other, so duplication is managed two ways: | ||
|
|
||
| ## Copies and Modular Models | ||
| 1. `# Copied from ...` marks a copied class or function. `make fix-repo` re-syncs it, so editing inside such a block is reverted — edit the source it copies from, or deliberately break the link. |
Collaborator
There was a problem hiding this comment.
specify that more Copied from statements should not be added though
Collaborator
Author
There was a problem hiding this comment.
Done in 16c5305 — item 1 now ends with Do not add new # Copied from statements; write a modular file instead, and the intro says which of the two mechanisms is current:
Model files in
src/transformers/models/avoid inheriting from each other, so duplication is managed two ways — modular is the current one, copies are legacy
Review feedback: the section explained how to live with an existing copy but not that the mechanism is closed to new code. Modular is the current one; copies are legacy. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
tarekziade
enabled auto-merge
September 7, 2026 08:30
Contributor
CI recapDashboard: View test results in Grafana |
github-merge-queue
Bot
removed this pull request from the merge queue due to failed status checks
Sep 7, 2026
sbucaille
pushed a commit
to sbucaille/transformers
that referenced
this pull request
Sep 16, 2026
…huggingface#48586) * Document two modular pitfalls in the agent conventions `.ai/AGENTS.md` is what code agents are handed as the repo's conventions (serge injects it into every patch prompt), and its modular section stops at "edit the modular file, not the generated one". Two things it does not say cost two wrong patches on the same PR (huggingface#48536): - `_no_split_modules = AttributeError()` was read as an invalid placeholder and replaced with `[]`, which changes how the model may be sharded. The idiom is documented in `modular_transformers.md`, but nothing pointed there from here. - the generated `modeling_deepseek_vl.py` was hand-edited to match, and `modeling_deepseek_vl_hybrid.py` — generated from the same modular file — was left stale, so `Check repository consistency` went red on a model the patch never mentioned. Both bullets link to the existing sections rather than restating them. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * Compress the file: every rule kept, 12% smaller than before This file loads for every agent (the root `AGENTS.md` / `CLAUDE.md` symlinks point here, and serge injects it into every patch prompt), so its length is a fixed cost on every session's context. Same rules, same commands, same links, tighter prose: 4,738 -> 3,160 chars, i.e. 439 chars BELOW what it was before the two modular bullets were added. The three "before opening a PR" sections are merged into one, since coordination, duplicate checks and busywork were all answering the same question. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * Say that new `# Copied from` statements should not be added Review feedback: the section explained how to live with an existing copy but not that the mechanism is closed to new code. Modular is the current one; copies are legacy. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
.ai/AGENTS.mdis the target of the rootAGENTS.md/CLAUDE.mdsymlinks, so it loads at the start of every agent session (serge also injects it verbatim into every patch prompt — uncapped). Its length is a fixed cost on every one of those sessions, which is the main thing this PR fixes.Compression
4,738 → 3,160 chars, which is 439 chars below the 3,599 it was before the additions below. Every rule, command,
ghinvocation and link is kept — verified one by one. The three separate "coordination / duplicate-work / no busywork" sections are merged into one "Before opening a PR", since they were all answering the same question.Two modular pitfalls added
The modular section stopped at "edit the modular file, not the generated one". Two things it did not say cost two wrong changes in a single patch (#48536):
_no_split_modules = AttributeError()was read as an invalid placeholder and replaced with[], which changes how the model may be sharded across devices. The idiom is documented inmodular_transformers.md, but nothing pointed there from here.modeling_deepseek_vl.pywas then hand-edited to match, leavingmodeling_deepseek_vl_hybrid.py— generated from the same modular file — stale, soCheck repository consistencywent red on a model the patch never mentioned:Both bullets link to the existing sections rather than restating them. Docs only — no code change.