example_generate.py

import torch
from transformers import (
    DiffusionGemmaForBlockDiffusion,
    DiffusionGemmaGenerationConfig,
    EntropyBoundSamplerConfig,
    PreTrainedTokenizerFast,
)


MODEL_PATH = "."
MODEL_SUBFOLDER = "hf"
PROMPT = "Once upon"


def main() -> None:
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    tokenizer = PreTrainedTokenizerFast.from_pretrained(
        MODEL_PATH,
        subfolder=MODEL_SUBFOLDER,
    )
    model = DiffusionGemmaForBlockDiffusion.from_pretrained(
        MODEL_PATH,
        subfolder=MODEL_SUBFOLDER,
        dtype=torch.float32,
    ).to(device)
    model.eval()

    generation_config = DiffusionGemmaGenerationConfig(
        max_new_tokens=64,
        max_denoising_steps=64,
        sampler_config=EntropyBoundSamplerConfig(entropy_bound=1.0),
        t_min=0.4,
        t_max=0.8,
        stability_threshold=3,
        confidence_threshold=0.05,
        bos_token_id=tokenizer.bos_token_id,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
        cache_implementation="dynamic",
        return_dict_in_generate=True,
    )
    input_ids = torch.tensor(
        [
            [tokenizer.bos_token_id]
            + tokenizer.encode(PROMPT, add_special_tokens=False)
        ],
        dtype=torch.long,
        device=device,
    )

    with torch.no_grad():
        output = model.generate(
            input_ids=input_ids,
            generation_config=generation_config,
        )

    sequences = output.sequences if hasattr(output, "sequences") else output
    print(tokenizer.decode(sequences[0].tolist(), skip_special_tokens=True))


if __name__ == "__main__":
    main()