app.py

import gradio as gr
import torch
import importlib
from transformers import AutoConfig, AutoProcessor, AutoModelForCausalLM
from transformers.models.llava.configuration_llava import LlavaConfig

# --- Diagnostic: Load the configuration ---
config = AutoConfig.from_pretrained("lmms-lab/LLaVA-Video-7B-Qwen2", trust_remote_code=True)
print("Configuration type:", type(config))
print("Configuration architectures:", config.architectures)

# Expecting the architecture name to be "LlavaQwenForCausalLM"
arch = config.architectures[0]  # This should be "LlavaQwenForCausalLM"

# --- Dynamic Import: Retrieve the model class by name ---
# Import the module that (should) contain the custom model class.
module = importlib.import_module("transformers.models.llava.modeling_llava")
try:
    model_cls = getattr(module, arch)
    print("Successfully imported model class:", model_cls)
except AttributeError:
    raise ImportError(f"Cannot find class {arch} in module transformers.models.llava.modeling_llava")

# --- Register the Custom Model Class ---
# This tells the auto loader that for LlavaConfig, use our dynamically imported model class.
AutoModelForCausalLM.register(LlavaConfig, model_cls)

# --- Load Processor and Model ---
processor = AutoProcessor.from_pretrained(
    "lmms-lab/LLaVA-Video-7B-Qwen2",
    trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
    "lmms-lab/LLaVA-Video-7B-Qwen2",
    trust_remote_code=True
)

# Move model to GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

def analyze_video(video_path):
    prompt = "Analyze this video of a concert and determine the moment when the crowd is most engaged."
    # Process the text and video input
    inputs = processor(text=prompt, video=video_path, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}
    # Generate output (assuming the custom model implements generate)
    outputs = model.generate(**inputs, max_new_tokens=100)
    answer = processor.decode(outputs[0], skip_special_tokens=True)
    return answer

# Create the Gradio Interface
iface = gr.Interface(
    fn=analyze_video,
    inputs=gr.Video(label="Upload Concert/Event Video", type="filepath"),
    outputs=gr.Textbox(label="Engagement Analysis"),
    title="Crowd Engagement Analyzer",
    description="Upload a video of a concert or event and the model will analyze the moment when the crowd is most engaged."
)

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