Fix Florence-2 model compatibility issue

- Pin transformers to version 4.37.2 for Florence-2 compatibility
- Add attn_implementation="eager" to avoid SDPA issues
- Disable use_cache for better stability
- Remove flash-attn dependency to avoid CPU compatibility issues

This resolves the "'Florence2ForConditionalGeneration' object has no attribute '_supports_sdpa'" error.

app.py

@@ -39,12 +39,18 @@ def load_florence_model():
 
         print(f"Loading Florence-2 model on {device}...")
 
+        # Load model with compatibility fixes
         model = AutoModelForCausalLM.from_pretrained(
             FLORENCE_MODEL_ID,
             torch_dtype=torch.float16 if (torch.cuda.is_available() and not FORCE_CPU) else torch.float32,
-            trust_remote_code=True
+            trust_remote_code=True,
+            attn_implementation="eager"  # Use eager attention for compatibility
         ).to(device)
 
+        # Fix for transformers compatibility issue
+        if hasattr(model, 'config'):
+            model.config.use_cache = False
+
         processor = AutoProcessor.from_pretrained(FLORENCE_MODEL_ID, trust_remote_code=True)
 
         model_cache['model'] = model

requirements.txt

@@ -1,8 +1,8 @@
-# Core dependencies - minimal versions that work
+# Core dependencies - specific versions for Florence-2 compatibility
 gradio>=4.44.0,<5.0.0
 torch>=2.0.0
 torchvision>=0.15.0
-transformers>=4.35.0
+transformers==4.37.2
 Pillow>=9.0.0
 numpy>=1.21.0
 
@@ -11,6 +11,7 @@ timm>=0.9.0
 einops>=0.7.0
 safetensors>=0.4.0
 accelerate>=0.21.0
+# flash-attn>=2.5.0  # Commented out - can cause issues on CPU
 
 # PDF processing
 pdf2image>=1.16.0