[agents docs] add pipelines.md etc (#13567)
* [agents docs] add pipelines.md and restructure review rules
- Add .ai/pipelines.md: pipeline conventions and gotchas (config-derived
values, no_grad discipline, reinventing scheduler logic, subclassing
variants, # Copied from annotations).
- models.md: add Attention masks subsection inside Attention pattern;
fold reference-implementations skim into conventions; consolidate
__init__.py / _import_structure gotchas; trim gotchas covered by
AGENTS.md (silent fallbacks, config serialization gap) or pipelines.md
(no_grad, guider/scheduler reuse).
- review-rules.md: collapse to a short reviewer checklist that points
into AGENTS / models / pipelines / modular gotchas; only LLM-specific
pattern (ephemeral context) lives here directly.
- AGENTS.md: collapse defensive-code / unused-params / backwards-compat
/ deprecation rules into one umbrella bullet; replace inline pipeline
bullet list with a pointer to pipelines.md.
- SKILL.md (model-integration): trim pre-PR self-review to a one-line
pointer.
Sourced from the ACE-Step PR (#13095) review.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Apply suggestions from code review
Co-authored-by: YiYi Xu <yixu310@gmail.com>
* Apply suggestion from @yiyixuxu
* Apply suggestions from code review
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
* fix capability-flags gotcha: drop fake attrs, tighten to real failure modes
`_supports_quantization` and `_supports_cache_class` don't exist in diffusers
(sayak flagged the first; the second was also fabricated). Replaced with the
two flags where the "advertised but unbacked" pattern is a real mistake:
`_supports_gradient_checkpointing` (needs `if self.gradient_checkpointing:`
branches in forward) and `_no_split_modules` (needs correct block class names
for `device_map`). Dropped `_supports_group_offloading` — its realistic
failure mode is forgetting to opt out, not opt in.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: yiyi@huggingface.co <yiyi@ip-26-0-160-103.ec2.internal>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
