app.py

import spaces
import gradio as gr
import sys
import time
import os
import random

os.environ["CUDA_VISIBLE_DEVICES"] = ""

# Create the gr.State component *outside* the gr.Blocks context
predictor_state = gr.State(None)

def get_transformer_model_id(task_type: str) -> str:
    if task_type == "i2v":
        return "Skywork/skyreels-v1-Hunyuan-i2v"
    else:
        return "Skywork/skyreels-v1-Hunyuan-t2v"
        
@spaces.GPU(duration=120)
def init_predictor(task_type: str):
    # ALL IMPORTS NOW INSIDE THIS FUNCTION
    import torch
    from skyreelsinfer import TaskType
    from skyreelsinfer.offload import OffloadConfig
    from skyreelsinfer.skyreels_video_infer import SkyReelsVideoInfer
    from huggingface_hub.utils import RepositoryNotFoundError, RevisionNotFoundError, EntryNotFoundError

    try:
        predictor = SkyReelsVideoInfer(
            task_type=TaskType.I2V if task_type == "i2v" else TaskType.T2V,
            model_id=get_transformer_model_id(task_type),
            quant_model=True,
            is_offload=True,
            offload_config=OffloadConfig(
                high_cpu_memory=True,
                parameters_level=True,
            ),
            use_multiprocessing=False,
        )
        return "Model loaded successfully!", predictor  # Return predictor

    except (RepositoryNotFoundError, RevisionNotFoundError, EntryNotFoundError) as e:
        return f"Error: Model not found. Details: {e}", None
    except Exception as e:
        return f"Error loading model: {e}", None

@spaces.GPU(duration=80)
def generate_video(prompt, seed, image, task_type, predictor): # predictor as argument
    # IMPORTS INSIDE THIS FUNCTION TOO
    from diffusers.utils import export_to_video
    from diffusers.utils import load_image
    import os

    if task_type == "i2v" and not isinstance(image, str):
        return "Error: For i2v, provide image path.", "{}"
    if not isinstance(prompt, str) or not isinstance(seed, (int, float)):
        return "Error: Invalid inputs.", "{}"

    if seed == -1:
        random.seed(time.time())
        seed = int(random.randrange(4294967294))

    kwargs = {
        "prompt": prompt,
        "height": 256,
        "width": 256,
        "num_frames": 24,
        "num_inference_steps": 30,
        "seed": int(seed),
        "guidance_scale": 7.0,
        "embedded_guidance_scale": 1.0,
        "negative_prompt": "bad quality, blur",
        "cfg_for": False,
    }

    if task_type == "i2v":
        if image is None or not os.path.exists(image):
            return "Error: Image not found.", "{}"
        try:
            kwargs["image"] = load_image(image=image)
        except Exception as e:
            return f"Error loading image: {e}", "{}"

    try:
        if predictor is None:
            return "Error: Model not init.", "{}"

        output = predictor.inference(kwargs)
        frames = output

        save_dir = f"./result/{task_type}"
        os.makedirs(save_dir, exist_ok=True)
        video_out_file = f"{save_dir}/{prompt[:100]}_{int(seed)}.mp4"
        print(f"Generating video: {video_out_file}")
        export_to_video(frames, video_out_file, fps=24)
        return video_out_file

    except Exception as e:
        return f"Error: {e}", "{}"

# --- Minimal Gradio Interface ---
with gr.Blocks() as demo:
    task_type_dropdown = gr.Dropdown(
        choices=["i2v", "t2v"], label="Task", value="t2v", elem_id="task_type"
    )
    load_model_button = gr.Button("Load Model", elem_id="load_button")
    prompt_textbox = gr.Textbox(label="Prompt", elem_id="prompt")
    generate_button = gr.Button("Generate", elem_id="generate_button")
    output_textbox = gr.Textbox(label="Output", elem_id="output") # Just a textbox
    output_video = gr.Video(label="Output Video", elem_id="output_video") # Just a textbox

    load_model_button.click(
        fn=init_predictor,
        inputs=[task_type_dropdown],
        outputs=[output_textbox, predictor_state], # Correct order of outputs
    )

    generate_button.click(
        fn=generate_video,
        inputs=[prompt_textbox, task_type_dropdown, predictor_state],
        outputs=[output_video],
    )

demo.launch()