fix: use attention dropout with torch SDPA implementation

modeling_bert.py

@@ -356,7 +356,8 @@ class JinaBertSelfAttention(nn.Module):
         if self.attn_implementation == 'torch' and scaled_dot_product_attention is not None:
             b, _, s, _ = query_layer.shape
             new_bias = attention_mask + bias
-            attn = scaled_dot_product_attention(query_layer, key_layer, value_layer, new_bias)
+            dropout_p = self.dropout.p if self.training else 0.0
+            attn = scaled_dot_product_attention(query_layer, key_layer, value_layer, new_bias, dropout_p=dropout_p)
             attn = attn.permute(0, 2, 1, 3).contiguous()
             return (attn.view(b, s, self.all_head_size),)