multimodalart/FLUX.1-dev-quantized

1import gradio as gr2import numpy as np3import random4import spaces5import torch6from diffusers import  DiffusionPipeline, FlowMatchEulerDiscreteScheduler, AutoencoderTiny, AutoencoderKL7from transformers import CLIPTextModel, CLIPTokenizer,T5EncoderModel, T5TokenizerFast8from live_preview_helpers import calculate_shift, retrieve_timesteps, flux_pipe_call_that_returns_an_iterable_of_images9from torchao.quantization.quant_api import Int8WeightOnlyConfig, quantize_10 11dtype = torch.bfloat1612device = "cuda" if torch.cuda.is_available() else "cpu"13 14taef1 = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to(device)15good_vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae", torch_dtype=dtype).to(device)16pipe = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=dtype, vae=taef1).to(device)17quantize_(pipe.transformer, Int8WeightOnlyConfig())18 19torch.cuda.empty_cache()20 21MAX_SEED = np.iinfo(np.int32).max22MAX_IMAGE_SIZE = 204823 24pipe.flux_pipe_call_that_returns_an_iterable_of_images = flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe)25 26@spaces.GPU(duration=75)27def infer(prompt, seed=42, randomize_seed=False, width=1024, height=1024, guidance_scale=3.5, num_inference_steps=28, progress=gr.Progress(track_tqdm=True)):28    if randomize_seed:29        seed = random.randint(0, MAX_SEED)30    generator = torch.Generator().manual_seed(seed)31    32    for img in pipe.flux_pipe_call_that_returns_an_iterable_of_images(33            prompt=prompt,34            guidance_scale=guidance_scale,35            num_inference_steps=num_inference_steps,36            width=width,37            height=height,38            generator=generator,39            output_type="pil",40            good_vae=good_vae,41        ):42            yield img, seed43    44examples = [45    "a tiny astronaut hatching from an egg on the moon",46    "a cat holding a sign that says hello world",47    "an anime illustration of a wiener schnitzel",48]49 50css="""51#col-container {52    margin: 0 auto;53    max-width: 520px;54}55"""56 57with gr.Blocks(css=css) as demo:58    59    with gr.Column(elem_id="col-container"):60        gr.Markdown(f"""# FLUX.1 [dev] 8-bit quantized6112B param rectified flow transformer guidance-distilled from [FLUX.1 [pro]](https://blackforestlabs.ai/)  62[[non-commercial license](https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md)] [[blog](https://blackforestlabs.ai/announcing-black-forest-labs/)] [[model](https://huggingface.co/black-forest-labs/FLUX.1-dev)]63        """)64        65        with gr.Row():66            67            prompt = gr.Text(68                label="Prompt",69                show_label=False,70                max_lines=1,71                placeholder="Enter your prompt",72                container=False,73            )74            75            run_button = gr.Button("Run", scale=0)76        77        result = gr.Image(label="Result", show_label=False)78        79        with gr.Accordion("Advanced Settings", open=False):80            81            seed = gr.Slider(82                label="Seed",83                minimum=0,84                maximum=MAX_SEED,85                step=1,86                value=0,87            )88            89            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)90            91            with gr.Row():92                93                width = gr.Slider(94                    label="Width",95                    minimum=256,96                    maximum=MAX_IMAGE_SIZE,97                    step=32,98                    value=1024,99                )100                101                height = gr.Slider(102                    label="Height",103                    minimum=256,104                    maximum=MAX_IMAGE_SIZE,105                    step=32,106                    value=1024,107                )108            109            with gr.Row():110 111                guidance_scale = gr.Slider(112                    label="Guidance Scale",113                    minimum=1,114                    maximum=15,115                    step=0.1,116                    value=3.5,117                )118  119                num_inference_steps = gr.Slider(120                    label="Number of inference steps",121                    minimum=1,122                    maximum=50,123                    step=1,124                    value=28,125                )126        127        gr.Examples(128            examples = examples,129            fn = infer,130            inputs = [prompt],131            outputs = [result, seed],132            cache_examples="lazy"133        )134 135    gr.on(136        triggers=[run_button.click, prompt.submit],137        fn = infer,138        inputs = [prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],139        outputs = [result, seed]140    )141 142demo.launch()