Update app.py

app.py

@@ -1,21 +1,38 @@
 import gradio as gr
 from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
 from qwen_vl_utils import process_vision_info
-import torch
 from PIL import Image
 import cv2
 import numpy as np
 import os
 
+import torch
+
+# Optimize for A100
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+
+# Set the default tensor type to cuda
+if torch.cuda.is_available():
+    torch.set_default_tensor_type('torch.cuda.FloatTensor')
+
+
+
 def load_model():
-    model = Qwen2VLForConditionalGeneration.from_pretrained(
-        "Qwen/Qwen2-VL-2B-Instruct",
-        torch_dtype=torch.float16
-    ).to("cuda")
-    processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
-    return model, processor
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    try:
+        model = Qwen2VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2-VL-2B-Instruct",
+            torch_dtype=torch.float16,  # Use float16 for faster inference on GPU
+            device_map="auto"  # This will automatically handle multi-GPU setups
+        )
+        processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
+        return model, processor, device
+    except Exception as e:
+        print(f"Error loading model: {e}")
+        return None, None, None
 
-model, processor = load_model()
+model, processor, device = load_model()
 
 SYSTEM_PROMPT = """You are an expert technical analyst specializing in identifying bugs, fixing errors, and explaining code functions from visual inputs. When presented with an image or video:
 1. If you see code, analyze it for potential bugs or errors, and suggest fixes.
@@ -54,71 +71,93 @@ def analyze_image(image, prompt):
     return generate_response(messages)
 
 def analyze_video(video_path, prompt, max_frames=16, frame_interval=30, max_resolution=224):
-    cap = cv2.VideoCapture(video_path)
-    frames = []
-    frame_count = 0
-
-    while len(frames) < max_frames:
-        ret, frame = cap.read()
-        if not ret:
-            break
+    try:
+        cap = cv2.VideoCapture(video_path)
+        if not cap.isOpened():
+            return "Error: Could not open video file."
         
-        if frame_count % frame_interval == 0:
-            h, w = frame.shape[:2]
-            if h > w:
-                new_h, new_w = max_resolution, int(w * max_resolution / h)
-            else:
-                new_h, new_w = int(h * max_resolution / w), max_resolution
-            frame = cv2.resize(frame, (new_w, new_h))
+        frames = []
+        frame_count = 0
+
+        while len(frames) < max_frames:
+            ret, frame = cap.read()
+            if not ret:
+                break
             
-            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
-            frame = Image.fromarray(frame)
+            if frame_count % frame_interval == 0:
+                h, w = frame.shape[:2]
+                if h > w:
+                    new_h, new_w = max_resolution, int(w * max_resolution / h)
+                else:
+                    new_h, new_w = int(h * max_resolution / w), max_resolution
+                frame = cv2.resize(frame, (new_w, new_h))
+                
+                frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+                frame = Image.fromarray(frame)
+                
+                frames.append(frame)
             
-            frames.append(frame)
-        
-        frame_count += 1
+            frame_count += 1
 
-    cap.release()
-
-    messages = [
-        {"role": "system", "content": SYSTEM_PROMPT},
-        {
-            "role": "user",
-            "content": [
-                {"type": "video", "video": frames},
-                {"type": "text", "text": f"Based on the system instructions, {prompt}"},
-            ],
-        }
-    ]
-    
-    return generate_response(messages)
+        return generate_response([
+            {"role": "system", "content": SYSTEM_PROMPT},
+            {
+                "role": "user",
+                "content": [
+                    {"type": "video", "video": frames},
+                    {"type": "text", "text": f"Based on the system instructions, {prompt}"},
+                ],
+            }
+        ])
+    except Exception as e:
+        return f"Error processing video: {e}"
+    finally:
+        if 'cap' in locals():
+            cap.release()
+                
 
 def generate_response(messages):
-    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-    image_inputs, video_inputs = process_vision_info(messages)
-    
-    inputs = processor(
-        text=[text],
-        images=image_inputs,
-        videos=video_inputs,
-        padding=True,
-        return_tensors="pt",
-    ).to("cuda")
-    
-    del image_inputs, video_inputs
-    torch.cuda.empty_cache()
-    
-    with torch.no_grad():
-        generated_ids = model.generate(**inputs, max_new_tokens=512)  # Increased token limit for more detailed responses
-    generated_ids_trimmed = [
-        out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
-    ]
-    output_text = processor.batch_decode(
-        generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
-    )
-    
-    return output_text[0]
-
+    try:
+        text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        image_inputs, video_inputs = process_vision_info(messages)
+        
+        inputs = processor(
+            text=[text],
+            images=image_inputs,
+            videos=video_inputs,
+            padding=True,
+            return_tensors="pt"
+        )
+        
+        # Move inputs to GPU
+        inputs = {k: v.to(device) for k, v in inputs.items()}
+        
+        with torch.no_grad():
+            generated_ids = model.generate(
+                **inputs,
+                max_new_tokens=512,
+                do_sample=True,
+                top_k=20,
+                top_p=0.9,
+                temperature=0.7
+            )
+        
+        generated_ids_trimmed = [
+            out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
+        ]
+        output_text = processor.batch_decode(
+            generated_ids_trimmed,
+            skip_special_tokens=True,
+            clean_up_tokenization_spaces=False
+        )
+        
+        # Clear CUDA cache
+        torch.cuda.empty_cache()
+        
+        return output_text[0]
+    except Exception as e:
+        return f"Error generating response: {e}"
+        
 # Gradio interface
 iface = gr.Interface(
     fn=process_content,