Fix meta device crash: use pure Python for drop_path_rate

image_encoder.py

@@ -695,7 +695,7 @@ class VisionTransformer(nn.Module):
       dpr = [drop_path_rate] * depth
     else:
       dpr = [
-          x.item() for x in torch.linspace(0, drop_path_rate, depth)
+          drop_path_rate * i / max(depth - 1, 1) for i in range(depth)
       ]  # stochastic depth decay rule
 
     if ffn_layer == "mlp":