Issue fix

app.py

@@ -31,8 +31,8 @@ from torch.cuda.amp import autocast
 # Set default dtypes
 torch.set_default_dtype(torch.float32)  # Main computations in float32
 MODEL_DTYPE = torch.float16 if torch.cuda.is_available() else torch.float32
-
-warnings.filterwarnings("ignore", category=FutureWarning, module="timm")
+import warnings
+warnings.filterwarnings("ignore", category=FutureWarning)
 warnings.filterwarnings("ignore", category=UserWarning)
 logging.set_verbosity_error()
 token = os.getenv("HF_TOKEN")
@@ -354,6 +354,7 @@ class SkinGPT4(nn.Module):
         input_embeddings = self.llama.model.embed_tokens(inputs.input_ids)
         # projected_features = self.llama_proj(aligned_features.mean(dim=1, keepdim=True))
         visual_embeds = aligned_features.mean(dim=1, keepdim=True)  # [1, 1, 5120]
+        visual_embeds = visual_embeds.to(input_embeddings.dtype)
         input_embeddings[image_token_pos] = visual_embeds
 
         outputs = self.llama.generate(