Update app.py

app.py

@@ -1,35 +1,43 @@
 import gradio as gr
 import torch
-from transformers import AutoConfig, AutoProcessor, AutoModelForCausalLM
+from transformers import AutoConfig, AutoProcessor
+# Import the custom model class directly.
+from transformers.models.llava.modeling_llava import LlavaQwenForCausalLM
 
-# === Diagnostic Code Start ===
-# Load the configuration with remote code enabled
-config = AutoConfig.from_pretrained("lmms-lab/LLaVA-Video-7B-Qwen2", trust_remote_code=True)
+# --- Diagnostic Print (Optional) ---
+config = AutoConfig.from_pretrained(
+    "lmms-lab/LLaVA-Video-7B-Qwen2",
+    trust_remote_code=True
+)
 print("Configuration type:", type(config))
 print("Configuration architectures:", config.architectures)
-# === Diagnostic Code End ===
+# --- End Diagnostic ---
 
-# Load processor and model with remote code enabled.
+# Load the processor and the model using the custom model class.
 processor = AutoProcessor.from_pretrained(
     "lmms-lab/LLaVA-Video-7B-Qwen2",
     trust_remote_code=True
 )
-model = AutoModelForCausalLM.from_pretrained(
+model = LlavaQwenForCausalLM.from_pretrained(
     "lmms-lab/LLaVA-Video-7B-Qwen2",
     trust_remote_code=True
 )
 
+# Move model to the appropriate device.
 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.
     inputs = processor(text=prompt, video=video_path, return_tensors="pt")
     inputs = {k: v.to(device) for k, v in inputs.items()}
+    # Generate output (assumes the custom model has a generate method).
     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"),