fix swinv2 for TPU saving model

keras_cv_attention_models/swin_transformer_v2/swin_transformer_v2.py

@@ -89,7 +89,9 @@ def window_multi_head_self_attention(inputs, filters=-1, num_heads=4, meta_hidde
     if mask is not None:
         query_blocks = attn.shape[2]
         attn = tf.reshape(attn, [-1, mask.shape[0], num_heads, query_blocks, query_blocks])
-        attn += tf.expand_dims(tf.expand_dims(mask, 1), 0)  # expand dims on batch and num_heads
+        # attn += tf.expand_dims(tf.expand_dims(mask, 1), 0)  # expand dims on batch and num_heads
+        mask = tf.expand_dims(tf.expand_dims(mask, 1), 0)  # expand dims on batch and num_heads
+        attn = keras.layers.Add()([attn, mask])
         attn = tf.reshape(attn, [-1, num_heads, query_blocks, query_blocks])
     attention_scores = keras.layers.Softmax(axis=-1, name=name and name + "attention_scores")(attn)
 
@@ -107,7 +109,7 @@ def window_multi_head_self_attention(inputs, filters=-1, num_heads=4, meta_hidde
 
 
 def make_window_attention_mask(height, width, window_height, window_width, shift_height, shift_width):
-    float_dtype = tf.keras.mixed_precision.global_policy().compute_dtype
+    # float_dtype = tf.keras.mixed_precision.global_policy().compute_dtype
     hh_split = [0, height - window_height, height - shift_height, height]
     ww_split = [0, width - window_width, width - shift_width, width]
     mask_value, total_ww, mask = 0, len(ww_split) - 1, []
@@ -123,7 +125,7 @@ def make_window_attention_mask(height, width, window_height, window_width, shift
     mask = tf.transpose(mask, [0, 2, 1, 3])
     mask = tf.reshape(mask, [-1, window_height * window_width])
     attn_mask = tf.expand_dims(mask, 1) - tf.expand_dims(mask, 2)
-    return tf.cast(tf.where(attn_mask != 0, -100, 0), float_dtype)
+    return tf.cast(tf.where(attn_mask != 0, -100, 0), "float32")
 
 
 def shifted_window_attention(inputs, window_size, num_heads=4, shift_size=0, name=""):