add --no_attention_mask option (#4750)

output producer name and version in optimized model.
avoid removing initializer that existed in graph output

onnxruntime/python/tools/transformers/fusion_attention.py

@@ -17,6 +17,7 @@ class AttentionMaskFormat:
     MaskIndexEnd = 0
     MaskIndexEndAndStart = 1
     AttentionMask = 2
+    NoMask = 3
 
 
 class AttentionMask():
@@ -45,6 +46,9 @@ class AttentionMask():
         return next(iter(self.mask_indice))
 
     def process_mask(self, input):
+        if self.mask_format == AttentionMaskFormat.NoMask:
+            return None
+
         if input in self.mask_indice:
             return self.mask_indice[input]
 
@@ -136,7 +140,10 @@ class FusionAttention(Fusion):
                                   vals=qkv_bias.flatten().tolist())
         self.model.add_initializer(bias)
 
-        attnetion_inputs = [input, attention_node_name + '_qkv_weight', attention_node_name + '_qkv_bias', mask_index]
+        attnetion_inputs = [input, attention_node_name + '_qkv_weight', attention_node_name + '_qkv_bias']
+        if mask_index is not None:
+            attnetion_inputs.append(mask_index)
+
         attention_node = helper.make_node('Attention',
                                           inputs=attnetion_inputs,
                                           outputs=[output],

onnxruntime/python/tools/transformers/onnx_model.py

@@ -628,7 +628,7 @@ class OnnxModel:
         weights_to_remove = []
         weights_to_keep = []
         for initializer in graph.initializer:
-            if initializer.name not in remaining_input_names:
+            if initializer.name not in remaining_input_names and not self.find_graph_output(initializer.name):
                 weights_to_remove.append(initializer)
             else:
                 weights_to_keep.append(initializer.name)

onnxruntime/python/tools/transformers/onnx_model_bert.py

@@ -39,6 +39,9 @@ class BertOptimizationOptions:
     def use_raw_attention_mask(self):
         self.attention_mask_format = AttentionMaskFormat.AttentionMask
 
+    def disable_attention_mask(self):
+        self.attention_mask_format = AttentionMaskFormat.NoMask
+
 
 class BertOnnxModel(OnnxModel):
     def __init__(self, model, num_heads, hidden_size):

onnxruntime/python/tools/transformers/optimizer.py

@@ -192,6 +192,12 @@ def _parse_arguments():
                         help="use raw attention mask instead of mask index in attention operator")
     parser.set_defaults(use_raw_attention_mask=False)
 
+    parser.add_argument('--no_attention_mask',
+                        required=False,
+                        action='store_true',
+                        help="no attention mask. Only works for model_type=bert")
+    parser.set_defaults(no_attention_mask=False)
+
     parser.add_argument('--verbose', required=False, action='store_true')
     parser.set_defaults(verbose=False)
 
@@ -233,6 +239,8 @@ def _get_optimization_options(args):
         optimization_options.enable_gelu_approximation = True
     if args.use_raw_attention_mask:
         optimization_options.use_raw_attention_mask()
+    if args.no_attention_mask:
+        optimization_options.disable_attention_mask()
 
     return optimization_options
 
@@ -295,6 +303,9 @@ def optimize_model(input,
         os.remove(temp_model_path)
         logger.debug("Remove tempoary model: {}".format(temp_model_path))
 
+    optimizer.model.producer_name = "onnxruntime_tools"
+    optimizer.model.producer_version = "1.4"
+
     return optimizer