File size: 4,410 Bytes
1d01e07
 
 
 
31bf3ec
1d01e07
 
 
 
31bf3ec
7a0fd29
 
 
31bf3ec
 
 
7a0fd29
1d01e07
 
 
31bf3ec
1d01e07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31bf3ec
 
3c3ffca
 
 
7a0fd29
1d01e07
7a0fd29
1d01e07
3c3ffca
7a0fd29
 
1d01e07
 
7a0fd29
3c3ffca
1d01e07
31bf3ec
1d01e07
 
 
7a0fd29
1d01e07
3c3ffca
1d01e07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16f2afa
1d01e07
d247f19
1d01e07
 
 
 
13b7826
 
1d01e07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16f2afa
 
1d01e07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
13b7826
 
 
 
 
 
1d01e07
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
import gradio as gr
import PIL
import spaces
import torch
from diffusers import TorchAoConfig as DiffusersTorchAoConfig
from hi_diffusers import HiDreamImagePipeline, HiDreamImageTransformer2DModel
from hi_diffusers.schedulers.flash_flow_match import (
    FlashFlowMatchEulerDiscreteScheduler,
)
from torchao.quantization import Int4WeightOnlyConfig
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
)
from transformers import (
    TorchAoConfig as TransformersTorchAoConfig,
)

# Constants
MODEL_PREFIX: str = "HiDream-ai"
LLAMA_MODEL_NAME: str = "meta-llama/Meta-Llama-3.1-8B-Instruct"
MODEL_PATH = "HiDream-ai/HiDream-I1-Fast"
MODEL_CONFIGS = {
    "guidance_scale": 0.0,
    "num_inference_steps": 16,
    "shift": 3.0,
    "scheduler": FlashFlowMatchEulerDiscreteScheduler,
}


# Supported image sizes
RESOLUTION_OPTIONS: list[str] = [
    "1024 x 1024 (Square)",
    "768 x 1360 (Portrait)",
    "1360 x 768 (Landscape)",
    "880 x 1168 (Portrait)",
    "1168 x 880 (Landscape)",
    "1248 x 832 (Landscape)",
    "832 x 1248 (Portrait)",
]


# Using AOBaseConfig instance (torchao >= 0.10.0)
quant_config = Int4WeightOnlyConfig(group_size=128)
quantization_config = TransformersTorchAoConfig(
    quant_type=quant_config, dtype=torch.bfloat16
)

tokenizer = AutoTokenizer.from_pretrained(LLAMA_MODEL_NAME, use_fast=False)
text_encoder = AutoModelForCausalLM.from_pretrained(
    LLAMA_MODEL_NAME,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    device_map="auto",
    output_hidden_states=True,
    output_attentions=True,
    quantization_config=quantization_config,
).to("cuda", torch.bfloat16)

quantization_config = DiffusersTorchAoConfig("int8wo")
transformer = HiDreamImageTransformer2DModel.from_pretrained(
    MODEL_PATH,
    subfolder="transformer",
    quantization_config=quantization_config,
    torch_dtype=torch.bfloat16,
).to("cuda", dtype=torch.float16)

scheduler = MODEL_CONFIGS["scheduler"](
    num_train_timesteps=1000,
    shift=MODEL_CONFIGS["shift"],
    use_dynamic_shifting=False,
)

pipe = HiDreamImagePipeline.from_pretrained(
    MODEL_PATH,
    scheduler=scheduler,
    tokenizer_4=tokenizer,
    text_encoder_4=text_encoder,
    torch_dtype=torch.bfloat16,
).to("cuda", torch.bfloat16)

pipe.transformer = transformer


@spaces.GPU(duration=120)
def generate_image(
    prompt: str, resolution: str, seed: int, progress=gr.Progress(track_tqdm=True)
) -> tuple[PIL.Image.Image, int]:
    if seed == -1:
        seed = torch.randint(0, 1_000_000, (1,)).item()

    # msg = "ℹ️ This spaces currently crash because of the memory usage. Please help me fix 😅"
    # raise gr.Error(msg, duration=10)
    height, width = tuple(map(int, resolution.replace(" ", "").split("x")))
    generator = torch.Generator("cuda").manual_seed(seed)

    image = pipe(
        prompt=prompt,
        height=height,
        width=width,
        guidance_scale=MODEL_CONFIGS["guidance_scale"],
        num_inference_steps=MODEL_CONFIGS["num_inference_steps"],
        generator=generator,
    ).images[0]

    torch.cuda.empty_cache()
    return image, seed


# Gradio UI
with gr.Blocks(title="HiDream Image Generator Fast") as demo:
    gr.Markdown("## 🌈 HiDream Image Generator Fast")

    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(
                label="Prompt",
                placeholder="e.g. A futuristic city with floating cars at sunset",
                lines=3,
            )

            resolution = gr.Radio(
                choices=RESOLUTION_OPTIONS,
                value=RESOLUTION_OPTIONS[0],
                label="Resolution",
            )

            seed = gr.Number(label="Seed (-1 for random)", value=-1, precision=0)
            generate_btn = gr.Button("Generate Image", variant="primary")
            # generate_btn = gr.Button(
            #     "This space currently crash because of the memory usage. Please help me fix 😅",
            #     variant="primary",
            #     interactive=False,
            # )
            seed_used = gr.Number(label="Seed Used", interactive=False)

        with gr.Column():
            output_image = gr.Image(label="Generated Image", type="pil")

    generate_btn.click(
        fn=generate_image,
        inputs=[prompt, resolution, seed],
        outputs=[output_image, seed_used],
    )

if __name__ == "__main__":
    demo.launch()