[discrete diffusion] Add DiffusionGemma pipeline and schedulers

[kashif]

Adds a DiffusionGemma block-diffusion pipeline, alongside the schedulers already on this branch (discrete DDIM, entropy bound, and a uniform mode for block refinement).

DiffusionGemma is an encoder-decoder block-diffusion model: the encoder reads the prompt into a KV cache and the decoder denoises a fixed-size canvas by cross-attending to it. The pipeline runs the outer canvas loop and the inner denoising loop, sampling candidates each step, committing the most confident ones via BlockRefinementScheduler in uniform corruption mode, and renoising the rest. Structure mirrors the LLaDA2 and dflash (#13699) pipelines.

The model itself lives in transformers as DiffusionGemmaForBlockDiffusion (released in 5.12.0).

Tested:

• pipeline unit tests pass (plumbing, callbacks, output types)
• the pipeline drives the real tiny checkpoint end to end without error

Quality on the full google/diffusiongemma-26B-A4B-it checkpoint still needs a GPU run.

[HuggingFaceDocBuilderDev]

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.

[zucchini-nlp]

Looking great! A couple questions from quick skimming

[yiyixuxu]

thanks for the PR! i left a few comments

I reviewed this through the lens of diffuser convention/style. If some of these choices are intentional to keep things familiar for Transformers users, let me know, and we can figure out the right balance together

[yiyixuxu]

I think between prompt and messages, we only need accept prompt since it's a really cheap into messages

it's just this, no?

messages = [{"role": "user", "content": prompt}]

[kashif]

Makes sense. The one wrinkle is image prompts, which we pass through messages today, so I'll fold the prompt/messages simplification into the image input rework so single-image and text both stay clean. Coming in a follow-up.

[kashif]

Made prompt the primary input and dropped the tokenized intermediates. Kept messages for raw multi-turn/multimodal conversations (per the thread below with zucchini), and added a raw image arg for the simple prompt+image case, so it is all raw inputs now.

[zucchini-nlp]

Thanks for working on it, i think the overall latency and quality matches model released in transformers!

[dg845]

When I tried out the example script:

import torch
from transformers import AutoProcessor, DiffusionGemmaForBlockDiffusion
from diffusers import BlockRefinementScheduler, DiffusionGemmaPipeline

model_id = "google/diffusiongemma-26B-A4B-it"
dtype = torch.bfloat16

model = DiffusionGemmaForBlockDiffusion.from_pretrained(model_id, dtype=dtype, device_map="auto")
# model = DiffusionGemmaForBlockDiffusion.from_pretrained(model_id, dtype=dtype)
processor = AutoProcessor.from_pretrained(model_id)
scheduler = BlockRefinementScheduler()

pipe = DiffusionGemmaPipeline(model=model, scheduler=scheduler, processor=processor)
pipe.model.model.decoder = torch.compile(pipe.model.model.decoder, mode="reduce-overhead", fullgraph=True)
# pipe.to("cuda")

output = pipe(
    prompt="Why is the sky blue?",
    gen_length=256,
    num_inference_steps=48,
    cache_implementation="static",
    generator=torch.Generator("cuda").manual_seed(42),
)

print(output.texts[0])

I get the following torch.compile error from the script:

torch._dynamo.exc.Unsupported: Skip calling `torch.compiler.disable()`d function
  Explanation: Skip calling function `<function AlignDevicesHook.pre_forward at 0x73a280162950>` since it was wrapped with `torch.compiler.disable` (reason: None)
  Hint: Remove the `torch.compiler.disable` call

The script works if I either don't torch.compile the model decoder, or use torch.compile without device_map="auto" (using the commented out code above). So I think the torch.compile + device_map="auto" case is what's causing the error, can you look into it?

EDIT: the error can also be mitigated by using fullgraph=False.

[sayakpaul]

@yiyixuxu @dg845 like the rest of our diffusers checkpoints repositories, where we have pipeline components coming from a different repo, could we make this pipeline a diffusers-style checkpoint with something like:

model = DiffusionGemmaForBlockDiffusion.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)
scheduler = BlockRefinementScheduler()

pipe = DiffusionGemmaPipeline(model=model, scheduler=scheduler, processor=processor)
pipe.save_pretrained(...)

[sayakpaul]

Why do we need this? We usually don't use CUDAgraph markers like this explicitly in our pipelines.

[sayakpaul]

@claude could you review this PR and also comment on the usage of the torch ops usage pattern from the lens of efficiency?

[github-actions]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.

[dg845]

Thanks for your work on this PR!