Optimize for NVIDIA T4 GPU

app.py

@@ -1,11 +1,10 @@
 """
 BubbleScribe - AI Manga & Comic Translator
 Translate manga/comics using Qwen2-VL for OCR + Translation and LaMa for inpainting.
-ZeroGPU optimized
+Optimized for NVIDIA T4 GPU
 """
 
 import gradio as gr
-import spaces
 import torch
 import os
 import json
@@ -19,33 +18,36 @@ from concurrent.futures import ThreadPoolExecutor
 import threading
 
 # ============================================================
-# HARDWARE OPTIMIZATION: ZeroGPU
+# HARDWARE OPTIMIZATION: NVIDIA T4 (16GB VRAM)
 # ============================================================
 
 import cv2
 
+# Enable CUDA optimizations
+torch.backends.cudnn.benchmark = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+
 # Thread pool for parallel operations
 executor = ThreadPoolExecutor(max_workers=4)
 
 # ============================================================
-# MODEL PRELOADING
+# MODEL PRELOADING (Load at startup for faster inference)
 # ============================================================
 
 print("🚀 BubbleScribe starting up...")
-print(f"   Hardware: ZeroGPU")
-print(f"   OCR: Qwen2-VL")
-print(f"   Inpainting: LaMa")
+print(f"   Hardware: NVIDIA T4 (16GB VRAM)")
+print(f"   OCR: Qwen2-VL (API)")
+print(f"   Inpainting: LaMa (GPU)")
 
-# Load LaMa model for inpainting
-lama_model = None
+# Load LaMa model at startup
+print("📦 Loading LaMa model...")
+from simple_lama_inpainting import SimpleLama
+lama_model = SimpleLama()
+print("✅ LaMa model loaded and ready!")
 
 def load_lama():
-    """Load LaMa inpainting model."""
-    global lama_model
-    if lama_model is None:
-        from simple_lama_inpainting import SimpleLama
-        lama_model = SimpleLama()
-        print("✅ LaMa model loaded")
+    """Get LaMa model (already loaded at startup)."""
     return lama_model
 
 # ============================================================
@@ -516,9 +518,8 @@ def draw_detections(image: Image.Image, detections: list) -> Image.Image:
 # MAIN PIPELINE
 # ============================================================
 
-@spaces.GPU
 def translate_manga(image, source_lang, target_lang, show_boxes, apply_inpaint, progress=gr.Progress()):
-    """Main translation pipeline."""
+    """Main translation pipeline (GPU-accelerated on T4)."""
     if image is None:
         return None, None, "Please upload an image"
     
@@ -657,7 +658,7 @@ with gr.Blocks(title="BubbleScribe", css=css, theme=gr.themes.Soft()) as demo:
     
     gr.HTML("""
     <div class="stats">
-        ⚡ <strong>Hardware:</strong> ZeroGPU (H200) • <strong>Models:</strong> Qwen2-VL (OCR & Translation) + LaMa (Inpainting)
+        ⚡ <strong>Hardware:</strong> NVIDIA T4 (16GB) • <strong>Models:</strong> Qwen2-VL (OCR & Translation) + LaMa (Inpainting)
     </div>
     """)