Upload 2 files

app.py

@@ -54,7 +54,7 @@ def encode_video(video_path, max_num_frames=32):
         return []
 
 def load_model():
-    """Load MiniCPM-o 2.6 model and tokenizer with fallback attention implementation"""
+    """Load MiniCPM-o 2.6 model and tokenizer with optimized attention implementation"""
     global model, tokenizer
     
     if model is not None and tokenizer is not None:
@@ -69,14 +69,9 @@ def load_model():
         
         print(f"Loading on device: {device} with dtype: {torch_dtype}")
         
-        # Try to determine best attention implementation
-        attn_implementation = 'sdpa'  # Default fallback
-        try:
-            import flash_attn
-            attn_implementation = 'flash_attention_2'
-            print("✅ Using Flash Attention 2 for optimal performance")
-        except ImportError:
-            print("⚠️ Flash Attention not available, using SDPA (still good performance)")
+        # Use SDPA attention which provides excellent performance without compilation issues
+        attn_implementation = 'sdpa'
+        print("🚀 Using SDPA (Scaled Dot Product Attention) for optimal compatibility and performance")
         
         # Load model with memory optimization for Spaces
         model = AutoModel.from_pretrained(
@@ -94,38 +89,12 @@ def load_model():
             trust_remote_code=True
         )
         
-        print("✅ Model loaded successfully!")
+        print("✅ Model loaded successfully with SDPA!")
         return model, tokenizer
         
     except Exception as e:
-        # If flash_attn fails, try with SDPA as fallback
-        if "flash_attn" in str(e):
-            print("⚠️ Flash Attention failed, retrying with SDPA...")
-            try:
-                model = AutoModel.from_pretrained(
-                    model_path,
-                    trust_remote_code=True,
-                    attn_implementation='sdpa',
-                    torch_dtype=torch_dtype,
-                    device_map="auto",
-                    offload_buffers=True,
-                    low_cpu_mem_usage=True
-                )
-                
-                tokenizer = AutoTokenizer.from_pretrained(
-                    model_path, 
-                    trust_remote_code=True
-                )
-                
-                print("✅ Model loaded successfully with SDPA fallback!")
-                return model, tokenizer
-                
-            except Exception as fallback_e:
-                print(f"❌ Error loading model with fallback: {fallback_e}")
-                raise fallback_e
-        else:
-            print(f"❌ Error loading model: {e}")
-            raise e
+        print(f"❌ Error loading model: {e}")
+        raise e
 
 def analyze_video(video_file, prompt, max_frames):
     """Analyze video using MiniCPM-o 2.6"""
@@ -178,6 +147,7 @@ def analyze_video(video_file, prompt, max_frames):
 **Processing Time:** {processing_time:.2f} seconds  
 **Frames Analyzed:** {len(frames)}  
 **Model:** MiniCPM-o 2.6
+**Attention:** SDPA (Optimized)
 
 ### Analysis:
 {answer}
@@ -243,6 +213,7 @@ def create_interface():
         - 🖼️ Frame-by-frame understanding  
         - 📝 Detailed descriptions
         - 🎨 Creative and marketing insights
+        - 🚀 SDPA optimized for reliable performance
         
         **Supported formats:** MP4, AVI, MOV, WebM
         """)
@@ -333,7 +304,7 @@ def create_interface():
         
         - **Model:** [openbmb/MiniCPM-o-2_6](https://huggingface.co/openbmb/MiniCPM-o-2_6)
         - **Code:** Based on exact official sample implementation
-        - **GPU:** Optimized for Hugging Face Spaces GPU
+        - **GPU:** Optimized for Hugging Face Spaces GPU with SDPA
         
         **Note:** Processing time depends on video length and complexity.
         """)

requirements.txt

@@ -3,9 +3,6 @@ torch==2.3.1
 torchaudio==2.3.1
 torchvision==0.18.1
 
-# Flash Attention - needs torch to be installed first
-flash-attn==2.8.0.post2
-
 # Core ML libraries
 transformers==4.44.2
 accelerate==1.2.1