worker_runpod.py

import os, json, requests, random, runpod
import torch
from diffusers import AutoencoderKLCogVideoX, CogVideoXImageToVideoPipeline, CogVideoXTransformer3DModel
from cogvideox.utils.lora_utils import merge_lora, unmerge_lora
from diffusers.utils import export_to_video, load_image
from transformers import T5EncoderModel, T5Tokenizer

with torch.inference_mode():
    model_id = "/runpod-volume/model"
    
    # Load models and ensure they are placed on CUDA device
    transformer = CogVideoXTransformer3DModel.from_pretrained(model_id, subfolder="transformer", torch_dtype=torch.float16).to("cuda")
    text_encoder = T5EncoderModel.from_pretrained(model_id, subfolder="text_encoder", torch_dtype=torch.float16).to("cuda")
    vae = AutoencoderKLCogVideoX.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float16).to("cuda")
    tokenizer = T5Tokenizer.from_pretrained(model_id, subfolder="tokenizer")
    
    # Ensure the pipeline is on the same device (CUDA)
    pipe = CogVideoXImageToVideoPipeline.from_pretrained(model_id, tokenizer=tokenizer, text_encoder=text_encoder, transformer=transformer, vae=vae, torch_dtype=torch.float16).to("cuda")
    
    lora_path = "/content/shirtlift.safetensors"
    lora_weight = 1.0
    
    # Merge Lora model and ensure it's on the same device
    pipe = merge_lora(pipe, lora_path, lora_weight).to("cuda")

    # pipe.enable_model_cpu_offload()

def download_file(url, save_dir, file_name):
    os.makedirs(save_dir, exist_ok=True)
    original_file_name = url.split('/')[-1]
    _, original_file_extension = os.path.splitext(original_file_name)
    file_path = os.path.join(save_dir, file_name + original_file_extension)
    response = requests.get(url)
    response.raise_for_status()
    with open(file_path, 'wb') as file:
        file.write(response.content)
    return file_path

@torch.inference_mode()
def generate(input):
    values = input["input"]

    input_image = values['input_image_check']
    input_image = download_file(url=input_image, save_dir='/content/input', file_name='input_image_tost')
    prompt = values['prompt']
    # guidance_scale = values['guidance_scale']
    # use_dynamic_cfg = values['use_dynamic_cfg']
    # num_inference_steps = values['num_inference_steps']
    # fps = values['fps']
    guidance_scale = 6
    use_dynamic_cfg = True
    num_inference_steps = 17
    fps = 9

    image = load_image(input_image)
    video = pipe(image=image, prompt=prompt, guidance_scale=guidance_scale, use_dynamic_cfg=use_dynamic_cfg, num_inference_steps=num_inference_steps).frames[0]
    export_to_video(video, "/content/cogvideox_5b_i2v_tost.mp4", fps=fps)

    result = "/content/cogvideox_5b_i2v_tost.mp4"
    try:
        default_filename = os.path.basename(result)
        print("Video saved to grid, uploading to huggingface")
        hf_api = HfApi()
        repo_id = "meepmoo/h4h4jejdf"  # Set your HF repo
        tokenxf = os.getenv("HF_API_TOKEN")
        hf_api.upload_file(path_or_fileobj=result,path_in_repo=f"{default_filename}.mp4",repo_id=repo_id,token=tokenxf,repo_type="model")
        result_url = f"https://huggingface.co/{repo_id}/blob/main/{default_filename}.mp4"
        return {"jobId": job_id, "result": result_url, "status": "DONE"}
    except Exception as e:
        
        return {"jobId": job_id, "result": f"FAILED: {str(e)}", "status": "FAILED"}
        
    finally:
        if os.path.exists(result):
            os.remove(result)
        if os.path.exists(input_image):
            os.remove(input_image)

runpod.serverless.start({"handler": generate})