app.py

import os
import torch
import gradio as gr
from transformers import AutoModelForCausalLM, AutoProcessor

# Clone the model if not already downloaded
if not os.path.exists("VideoLLaMA3-7B"):
    os.system("apt-get update && apt-get install -y git git-lfs && git lfs install")
    os.system("git clone https://huggingface.co/DAMO-NLP-SG/VideoLLaMA3-7B")

model_path = "./VideoLLaMA3-7B"

# Load model (no flash_attn)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

def describe_video(video, question):
    conversation = [
        {"role": "system", "content": "You are a helpful assistant."},
        {
            "role": "user",
            "content": [
                {"type": "video", "video": {"video_path": video, "fps": 1, "max_frames": 128}},
                {"type": "text", "text": question},
            ]
        },
    ]
    inputs = processor(conversation=conversation, return_tensors="pt")
    inputs = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}
    if "pixel_values" in inputs:
        inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)
    output_ids = model.generate(**inputs, max_new_tokens=128)
    return processor.batch_decode(output_ids, skip_special_tokens=True)[0].strip()

# Gradio interface
demo = gr.Interface(
    fn=describe_video,
    inputs=[
        gr.Video(label="Upload a video"),
        gr.Textbox(label="Question", value="Describe this video in detail."),
    ],
    outputs=gr.Textbox(label="Response"),
)

demo.launch()