Add PixArt-Alpha modular pipeline

[cgloriacc]

What does this PR do?

This ports the PixArt-Alpha text-to-image pipeline into Modular Diffusers, following the same structure as the existing qwenimage and stable_diffusion_3 modular pipelines.

New block files under src/diffusers/modular_pipelines/pixart_alpha/:

• encoders.py — the T5 text-encoder step. It emits the prompt embeddings and attention mask, plus the negative pair when the guider needs classifier-free guidance, and cleans captions with bs4/ftfy.
• before_denoise.py — per-prompt input expansion, timestep setup, latent preparation, and PixArt micro-conditions. Resolution and aspect-ratio conditions are emitted only when the model's sample size is 128.
• denoise.py — the denoise loop built on the guider abstraction. It also handles the PixArt learned-sigma split, taking the first chunk when out_channels is twice in_channels.
• decoders.py — VAE decode and image post-processing.
• modular_blocks_pixart_alpha.py and modular_pipeline.py — the blocks assembled into PixArtAlphaAutoBlocks and PixArtAlphaModularPipeline, registered in the modular pipeline mapping.

A pipeline-level test is added under tests/modular_pipelines/pixart_alpha/.

Coordination

Fixes #13301. This is tracked under the Modular Diffusers umbrella #13295, which @sayakpaul approved.

Tests run

Run on CPU with CUDA_VISIBLE_DEVICES="", which matches the CPU container the modular fast-test CI uses in pr_modular_tests.yml.

CUDA_VISIBLE_DEVICES="" python -m pytest tests/modular_pipelines/pixart_alpha/test_modular_pipeline_pixart_alpha.py -v

Full pytest output — 11 passed, 3 skipped

========================= test session starts =========================
platform linux -- Python 3.10.20, pytest-9.1.1, pluggy-1.6.0 -- /path/to/python
cachedir: .pytest_cache
rootdir: /path/to/diffusers
configfile: pyproject.toml
plugins: requests-mock-1.10.0, anyio-4.14.1, xdist-3.8.0, timeout-2.4.0
collected 14 items

...::test_inference_batch_consistent PASSED                                   [  7%]
...::test_inference_batch_single_identical PASSED                             [ 14%]
...::test_to_device SKIPPED (test requires a hardware accelerator)            [ 21%]
...::test_inference_is_not_nan_cpu PASSED                                     [ 28%]
...::test_inference_is_not_nan SKIPPED (test requires a hardware accelerator) [ 35%]
...::test_num_images_per_prompt PASSED                                        [ 42%]
...::test_components_auto_cpu_offload_inference_consistent SKIPPED (...)      [ 50%]
...::test_save_from_pretrained PASSED                                         [ 57%]
...::test_load_expected_components_from_pretrained PASSED                     [ 64%]
...::test_load_expected_components_from_save_pretrained PASSED                [ 71%]
...::test_modular_index_consistency PASSED                                    [ 78%]
...::test_workflow_map PASSED                                                 [ 85%]
...::test_pipeline_call_signature PASSED                                      [ 92%]
...::test_float16_inference PASSED                                            [100%]

========================== warnings summary ===========================
<frozen importlib._bootstrap>:241
  DeprecationWarning: builtin type SwigPyPacked has no __module__ attribute
<frozen importlib._bootstrap>:241
  DeprecationWarning: builtin type SwigPyObject has no __module__ attribute

huggingface_hub/utils/_validators.py:205: UserWarning (9 tests)
  The local_dir_use_symlinks argument is deprecated and ignored in hf_hub_download.

scheduling_dpmsolver_multistep.py:438: DeprecationWarning (12 warnings)
  __array__ implementation doesn't accept a copy keyword (numpy 2.0 migration).

============== 11 passed, 3 skipped, 23 warnings in 72.82s ==============

Why the 3 skips. test_to_device, test_inference_is_not_nan, and test_components_auto_cpu_offload_inference_consistent all carry the @require_accelerator decorator, so they skip on a CPU-only run. This is expected because the modular fast-test CI runs on CPU. The
not-NaN check still runs through its CPU companion test_inference_is_not_nan_cpu, which passed.

Why the warnings, none of which come from this PR's code. The SwigPyPacked/SwigPyObject warnings come from a SWIG-based dependency at import time on Python 3.10. The local_dir_use_symlinks warning comes from huggingface_hub while the test downloads the tiny
checkpoint. The numpy 2.0 array warning is raised inside the existing scheduling_dpmsolver_multistep.py scheduler; it is pre-existing and surfaces only because the pipeline uses DPMSolverMultistepScheduler.

The repo-consistency and quality checks that pr_modular_tests.yml runs also pass locally:

make quality
python utils/check_copies.py
python utils/check_dummies.py
python utils/check_support_list.py
python utils/check_forward_call_docstrings.py
make deps_table_check_updated
make modular-autodoctrings

Notes

• Only the text2image workflow is exposed, since PixArt-Alpha is a text-to-image-only model with no image-to-image or inpainting variants.
• Resolution binning is not ported in this first version. The classic pipeline exposes it through use_resolution_binning, and I'm happy to add it as a follow-up.
• The tiny test checkpoint currently lives at idealclx/tiny-pixart-alpha-modular. Per modular.md gotcha add unet ldm in init #6, tiny test models must move under hf-internal-testing before merge. Could a maintainer move it there? I'll update the test path afterwards.
• No per-model docs page is added, which matches the other modular pipelines. The modular docs only cover the framework classes.

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you self-review the diff against .ai/review-rules.md?
• Did you read the contributor guideline?
• Did you read our philosophy doc? (important for complex PRs)
• Was this discussed/approved via a GitHub issue or the forum? See Modular Pipeline: support for PixArtAlphaPipeline #13301 and Modular Diffusers 🧨 #13295.
• Did you make sure to update the documentation with your changes? N/A, there is no per-model modular docs page.
• Did you write any new necessary tests?
• Are you the author (or part of the team) of the model/pipeline? N/A, this ports the existing PixArt-Alpha pipeline.

Who can review?

@yiyixuxu @sayakpaul @asomoza