[.ai doc] Refine .ai attention-mask and component-mutation guidance (#13982)

Refine .ai attention-mask and component-mutation guidance from a PR review

- models.md: clarify padding-mask bullet — pass None when padding is
  uniform (mask.all()) or when the model is trained to tolerate padding
  without a mask
- pipelines.md: add gotcha #7 — don't mutate registered component state
  per call; prefer an already-prepared component, otherwise snapshot and
  restore (PAG pattern)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: dg845 <58458699+dg845@users.noreply.github.com>

Author: 3 people

.ai/models.md

@@ -62,7 +62,7 @@ class MyModelAttention(nn.Module, AttentionModuleMixin):
 What you pass as `attn_mask=` to `dispatch_attention_fn` determines which backends work:
 
 - **No mask needed → pass `None`, not an all-zero tensor.** A dense 4D additive float mask of all `0.0` does no math but still hard-raises on `flash` / `_flash_3` / `_sage` (see `attention_dispatch.py:2328, 2544, 3266`). Only materialize a mask when it carries information. This is the Flux / Flux2 / Wan pattern: no mask, works on every backend, relies on the model having been trained tolerating consistent padding.
-- **Padding mask → bool `(B, L)` or `(B, 1, 1, L)`.** Only pass when the batch actually contains different-length sequences (i.e. there is real padding). If all sequences are the same length, set the mask to `None` — many backends (flash, sage, aiter) raise `ValueError` on any non-None mask, and even SDPA-based backends pay unnecessary overhead processing a no-op mask. See `pipeline_qwenimage.py` `encode_prompt` for the pattern: `if mask.all(): mask = None`. When a mask is needed, use bool format — it stays compatible with the `*_varlen` kernels via `_normalize_attn_mask` (`attention_dispatch.py:639`), which reduces bool masks to `cu_seqlens`. Dense additive-float masks *cannot* be reduced this way and so lose the varlen path.
+- **Padding mask → bool `(B, L)` or `(B, 1, 1, L)`.** Only pass when the batch actually contains padding. If all sequences are the same length and padded to max length, set the mask to `None` — many backends (flash, sage, aiter) raise `ValueError` on any non-None mask, and even SDPA-based backends pay unnecessary overhead processing a no-op mask. See `pipeline_qwenimage.py` `encode_prompt` for the pattern: `if mask.all(): mask = None`. Some models are also trained without a mask — pass `None` for these even when padding is present (SD, Flux etc). When a mask is needed, use bool format — it stays compatible with the `*_varlen` kernels via `_normalize_attn_mask` (`attention_dispatch.py:639`), which reduces bool masks to `cu_seqlens`. Dense additive-float masks *cannot* be reduced this way and so lose the varlen path.
 - **Other mask types (structural, BlockMask, etc.)** — if the model requires a different mask pattern, figure out how to support as many backends as possible (e.g. use `window_size` kwarg for sliding window on flash, `BlockMask` for Flex) and document which backends are supported for that model.
 - **Don't declare `attention_mask` (or `encoder_hidden_states_mask`) in the forward signature if you ignore it.** "For API stability with other transformers" is not a reason; readers assume a declared param is honored, and downstream pipelines will pass padding masks that silently get dropped. Some existing models in the repo carry unused mask params for historical reasons — e.g. `QwenDoubleStreamAttnProcessor2_0.__call__` declares `encoder_hidden_states_mask` but never reads it (the joint mask is routed through `attention_mask` instead), and the block-level forward in `transformer_qwenimage.py` declares it but always receives `None`. This is a legacy behavior and should not be replicated in new models.
 

.ai/pipelines.md

@@ -76,3 +76,5 @@ src/diffusers/pipelines/<model>/
 6. **Be deliberate about methods on the pipeline.** `__call__` is the user's mental model. The methods on the class are how they navigate it. Diffusers convention (flux, sdxl, wan, qwenimage) is a flat class body of public lifecycle methods (`__init__`, `check_inputs`, `encode_prompt`, `prepare_latents`, `__call__`). Two principles, not strict rules — use judgment:
     - **If a method is called from `__call__`, and it's a step in the pipeline lifecycle, make it public.** Each call from `__call__` should correspond to a step a user can identify: either a standard one (`encode_prompt`, `prepare_latents`, `set_timesteps`, …) or a pipeline-specific one (`prepare_src_latents`, `prepare_reference_audio_latents`, …). Don't gate these behind a `_`; they're part of the pipeline's API surface alongside their standard siblings.
     - **If a method is only used by another method, make it private (`_foo`) or lift it to a module-level function — and keep the count down.** Before adding one, see if the logic can be absorbed into its caller. Unless you expect the helper to be reused by another method (or another task pipeline), absorbing is usually the better call — especially when the body is small. Avoid a pipeline class littered with private helpers that bury the lifecycle..
+
+7. **Don't modify the state of a registered component on the fly.** From inside `__call__` or other helper methods, don't change the state of `self.text_encoder` / `self.transformer` / `self.vae` — no in-place `.to(dtype/device)`, no setting attributes/buffers or swapping submodules. Components are shared and routinely reused across pipelines, so a per-call mutation may silently change another pipeline's outputs. You should pass a component that's already in the right state, and document that expectation explicitly. Only when that's genuinely inconvenient and you must change state for the duration of a call — e.g. swapping in an attention processor — save the original first and restore it before returning, so the component is left exactly as you found it. The PAG pipelines are the reference for this: `pipeline_pag_sd.py` snapshots `original_attn_proc = self.unet.attn_processors`, installs the PAG processors for the denoising loop, then calls `self.unet.set_attn_processor(original_attn_proc)` at the end of `__call__`.