Update app.py

app.py

@@ -130,6 +130,8 @@ def pil_image_to_base64(pil_image):
     return base64_image
 
 image_cache = {}
+ln0_weight = model.w['blocks.0.ln0.weight'].to(torch.float32).to(device)
+ln0_bias = model.w['blocks.0.ln0.bias'].to(torch.float32).to(device)
 def compute_image_state(image):
     base64_image = pil_image_to_base64(image)
     if base64_image in image_cache:
@@ -139,10 +141,10 @@ def compute_image_state(image):
         image = image.to(device)
         image_features = visual_encoder.encode_images(image.unsqueeze(0)).squeeze(0) # [L, D]
         # apply layer norm to image feature, very important
-        image_features = F.layer_norm(image_features.to(torch.float32), 
+        image_features = F.layer_norm(image_features, 
                                     (image_features.shape[-1],), 
-                                    weight=model.w['blocks.0.ln0.weight'], 
-                                    bias=model.w['blocks.0.ln0.bias'])
+                                    weight=ln0_weight, 
+                                    bias=ln0_bias)
         _, image_state = model.forward(embs=image_features, state=None)
         image_cache[base64_image] = image_state
     return image_state