# ------------------------------------------------------------------------- # Copyright (c) Microsoft Corporation. All rights reserved. # Licensed under the MIT License. # -------------------------------------------------------------------------- import logging from typing import Optional, Union import numpy as np from fusion_attention import AttentionMask, FusionAttention from fusion_base import Fusion from fusion_simplified_layernorm import FusionSimplifiedLayerNormalization, FusionSkipSimplifiedLayerNormalization from fusion_utils import NumpyHelper from onnx import NodeProto, TensorProto, helper from onnx_model import OnnxModel from onnx_model_bert import BertOnnxModel logger = logging.getLogger(__name__) class FusionT5Attention(FusionAttention): """ Fuse T5 Attention subgraph into one Attention node. """ def __init__( self, model: OnnxModel, hidden_size: int, num_heads: int, attention_mask: AttentionMask, ): super().__init__( model, hidden_size, num_heads, attention_mask, use_multi_head_attention=False, search_op_types=["SkipSimplifiedLayerNormalization", "Add"], ) self.static_kv = 1 def create_attention_node( self, mask_index: str, q_matmul: NodeProto, k_matmul: NodeProto, v_matmul: NodeProto, num_heads: int, hidden_size: int, input: str, output: str, add_qk_str: str, scale: Optional[float] = None, ) -> Union[NodeProto, None]: """Create an Attention node. Args: mask_index (str): mask input q_matmul (NodeProto): MatMul node in fully connection for Q k_matmul (NodeProto): MatMul node in fully connection for K v_matmul (NodeProto): MatMul node in fully connection for V num_heads (int): number of attention heads. If a model is pruned, it is the number of heads after pruning. hidden_size (int): hidden dimension. If a model is pruned, it is the hidden dimension after pruning. input (str): input name output (str): output name Returns: Union[NodeProto, None]: the node created or None if failed. """ assert num_heads > 0 if hidden_size > 0 and (hidden_size % num_heads) != 0: logger.debug(f"input hidden size {hidden_size} is not a multiple of num of heads {num_heads}") return None q_weight = self.model.get_initializer(q_matmul.input[1]) k_weight = self.model.get_initializer(k_matmul.input[1]) v_weight = self.model.get_initializer(v_matmul.input[1]) if q_weight is None: print( f"{q_matmul.input[1]} is not an initializer. " "Please set do_constant_folding=True in torch.onnx.export to unblock attention fusion" ) return None qw = NumpyHelper.to_array(q_weight) kw = NumpyHelper.to_array(k_weight) vw = NumpyHelper.to_array(v_weight) # assert q and k have same shape as expected assert qw.shape == kw.shape qw_in_size = qw.shape[0] kw_in_size = kw.shape[0] vw_in_size = vw.shape[0] assert qw_in_size == kw_in_size == vw_in_size if hidden_size > 0 and hidden_size != qw_in_size: logger.warning( f"Input hidden size ({hidden_size}) is not same as weight matrix dimension of q,k,v ({qw_in_size}). " "Please provide a correct input hidden size or pass in 0" ) qw_out_size = np.prod(qw.shape[1:]) qkv_weight = np.stack((qw, kw, vw), axis=1) qkv_weight_dim = 3 * qw_out_size attention_node_name = self.model.create_node_name("Attention") weight = helper.make_tensor( name=attention_node_name + "_qkv_weight", data_type=TensorProto.FLOAT, dims=[qw_in_size, qkv_weight_dim], vals=qkv_weight.tobytes(), raw=True, ) self.model.add_initializer(weight, self.this_graph_name) attention_inputs = [ input, attention_node_name + "_qkv_weight", "", ] if mask_index is not None: attention_inputs.append(mask_index) else: attention_inputs.append("") if add_qk_str is not None: attention_inputs.append("") # no past attention_inputs.append(add_qk_str) attention_node = helper.make_node( "Attention", inputs=attention_inputs, outputs=[output], name=attention_node_name, ) attention_node.domain = "com.microsoft" attention_node.attribute.extend([helper.make_attribute("num_heads", num_heads)]) if scale is not None: attention_node.attribute.extend([helper.make_attribute("scale", scale)]) if self.mask_filter_value is not None: attention_node.attribute.extend([helper.make_attribute("mask_filter_value", float(self.mask_filter_value))]) return attention_node def create_mha_node( self, query: str, key: str, value: str, mask_index: str, res_pos_bias: str, past_key: str, past_value: str, output: str, present_key: str, present_value: str, num_heads: int, hidden_size: int, ) -> Union[NodeProto, None]: assert num_heads > 0 if hidden_size > 0 and (hidden_size % num_heads) != 0: logger.debug(f"input hidden size {hidden_size} is not a multiple of num of heads {num_heads}") return None attention_node_name = self.model.create_node_name("MultiHeadAttention") attention_inputs = [ query, "" if key is None else key, # key "" if value is None else value, # value "", # bias ] if mask_index is not None: attention_inputs.append(mask_index) else: attention_inputs.append("") if res_pos_bias is not None: attention_inputs.append(res_pos_bias) else: attention_inputs.append("") if past_key is not None: assert past_value is not None attention_inputs.append(past_key) attention_inputs.append(past_value) attention_outputs = [output] if present_key is not None: assert present_value is not None attention_outputs.append(present_key) attention_outputs.append(present_value) attention_node = helper.make_node( "MultiHeadAttention", inputs=attention_inputs, outputs=attention_outputs, name=attention_node_name, ) attention_node.domain = "com.microsoft" attention_node.attribute.extend([helper.make_attribute("num_heads", num_heads)]) attention_node.attribute.extend([helper.make_attribute("scale", 1.0)]) if self.mask_filter_value is not None: attention_node.attribute.extend([helper.make_attribute("mask_filter_value", float(self.mask_filter_value))]) self.increase_counter("MultiHeadAttention") return attention_node def fuse(self, normalize_node, input_name_to_nodes, output_name_to_node): self.fuse_t5_encoder(normalize_node, input_name_to_nodes, output_name_to_node) self.fuse_t5_decoder(normalize_node, input_name_to_nodes, output_name_to_node) def fuse_t5_encoder(self, normalize_node, input_name_to_nodes, output_name_to_node): if normalize_node.op_type != "SkipSimplifiedLayerNormalization" and normalize_node.op_type != "Add": return qkv_nodes = self.model.match_parent_path( normalize_node, ["MatMul", "Reshape", "Transpose", "MatMul"], [1, 0, 0, 0], ) if qkv_nodes is None: return _, reshape_qkv, transpose_qkv, matmul_qkv = qkv_nodes qkv_shape_nodes = self.model.match_parent_path( reshape_qkv, ["Concat", "Unsqueeze", "Gather", "Shape"], [1, 0, 0, 0], ) if qkv_shape_nodes is None: return input_shape_node = qkv_shape_nodes[-1] v_nodes = self.model.match_parent_path( matmul_qkv, ["Transpose", "Reshape", "MatMul"], [1, 0, 0], ) if v_nodes is None: return _, reshape_v, matmul_v = v_nodes # todo: check reshape_v parent nodes qk_nodes = self.model.match_parent_path( matmul_qkv, ["Softmax", "Add", "MatMul"], [0, 0, 0], ) if qk_nodes is None: return _, add_qk, matmul_qk = qk_nodes mask_index = None mask_nodes = self.model.match_parent_path( add_qk, ["Add", "Mul", "Sub", "Cast", "Unsqueeze", "Unsqueeze"], [1, 1, 0, 1, 0, 0], ) if mask_nodes is None: return mul_node = mask_nodes[1] if mask_nodes[1].op_type != "Mul": return _, mul_val = self.model.get_constant_input(mul_node) if mul_val != -10000: self.mask_filter_value = mul_val mask_index = self.attention_mask.process_mask(mask_nodes[-1].input[0]) res_pos_bias = None rpb_nodes = self.model.match_parent_path( add_qk, ["Add", "RelativePositionBias"], [1, 0], ) if rpb_nodes is None: return rpb_add_node = rpb_nodes[0] res_pos_bias = rpb_add_node.input[0] k_nodes = self.model.match_parent_path( matmul_qk, ["Transpose", "Reshape", "MatMul"], [1, 0, 0], ) if k_nodes is None: return _, reshape_k, matmul_k = k_nodes # todo: check reshape_k parent nodes q_nodes = self.model.match_parent_path( matmul_qk, ["Transpose", "Reshape", "MatMul"], [0, 0, 0], ) if q_nodes is None: return transpose_q, reshape_q, matmul_q = q_nodes # todo: check reshape_q parent nodes if matmul_q.input[0] != input_shape_node.input[0]: return q_num_heads, q_hidden_size = self.get_num_heads_and_hidden_size(reshape_q) new_node = self.create_attention_node( mask_index, matmul_q, matmul_k, matmul_v, q_num_heads, q_hidden_size, input_shape_node.input[0], reshape_qkv.output[0], res_pos_bias, 1.0, ) if new_node is None: return self.nodes_to_add.append(new_node) self.node_name_to_graph_name[new_node.name] = self.this_graph_name self.nodes_to_remove.extend(qkv_nodes[1:]) self.nodes_to_remove.extend(qk_nodes) self.nodes_to_remove.extend(k_nodes[:-1]) if v_nodes is not None: self.nodes_to_remove.extend(v_nodes[:-1]) self.nodes_to_remove.extend(q_nodes[:-1]) self.prune_graph = True def fuse_t5_decoder(self, normalize_node, input_name_to_nodes, output_name_to_node): if normalize_node.op_type != "SkipSimplifiedLayerNormalization" and normalize_node.op_type != "Add": return qkv_nodes = self.model.match_parent_path( normalize_node, ["MatMul", "Reshape", "Transpose", "MatMul"], [1, 0, 0, 0], ) if qkv_nodes is None: return _, reshape_qkv, transpose_qkv, matmul_qkv = qkv_nodes qkv_shape_nodes = self.model.match_parent_path( reshape_qkv, ["Concat", "Unsqueeze", "Gather", "Shape"], [1, 0, 0, 0], ) if qkv_shape_nodes is None: return input_shape_node = qkv_shape_nodes[-1] value = None past_value = None present_value = None v_nodes = self.model.match_parent_path( matmul_qkv, ["Concat", "Transpose", "Reshape", "MatMul"], [1, 1, 0, 0], ) if v_nodes is None: v_nodes = self.model.match_parent_path( matmul_qkv, ["Transpose", "Reshape", "MatMul"], [1, 0, 0], ) if v_nodes is not None: transpose_v, reshape_v, matmul_v = v_nodes value = reshape_v.input[0] present_value = transpose_v.output[0] if "present_value" not in present_value: return if matmul_v.input[0] != input_shape_node.input[0]: self.static_kv = 1 else: self.static_kv = 0 else: past_value = matmul_qkv.input[1] if past_value in output_name_to_node: return if "past_value_cross" not in past_value: return self.static_kv = 1 else: concat_v, _, reshape_v, _ = v_nodes past_value = concat_v.input[0] if past_value in output_name_to_node: return if "past_value_self" not in past_value: return present_value = concat_v.output[0] if "present_value_self" not in present_value: return value = reshape_v.input[0] self.static_kv = 0 qk_nodes = self.model.match_parent_path( matmul_qkv, ["Softmax", "Add", "MatMul"], [0, 0, 0], ) if qk_nodes is None: return _, add_qk, matmul_qk = qk_nodes mask_index = None res_pos_bias = None if self.static_kv == 1: mask_nodes = self.model.match_parent_path( add_qk, ["Add", "Mul", "Sub", "Cast", "Unsqueeze", "Unsqueeze"], [1, 1, 0, 1, 0, 0], ) if mask_nodes is None: return mul_node = mask_nodes[1] if mask_nodes[1].op_type != "Mul": return _, mul_val = self.model.get_constant_input(mul_node) if mul_val != -10000: self.mask_filter_value = mul_val mask_index = self.attention_mask.process_mask(mask_nodes[-1].input[0]) else: rpb_nodes = self.model.match_parent_path( add_qk, ["Add", "Slice"], [1, 0], ) if rpb_nodes is not None: res_pos_bias = add_qk.input[1] else: rpb_nodes = self.model.match_parent_path( add_qk, ["Add", "RelativePositionBias"], [1, 0], ) if rpb_nodes is None: return res_pos_bias = add_qk.input[1] key = None past_key = None present_key = None if self.static_kv == 1: k_nodes = self.model.match_parent_path( matmul_qk, ["Transpose", "Reshape", "MatMul"], [1, 0, 0], ) if k_nodes is not None: transpose_k, reshape_k, _ = k_nodes key = reshape_k.input[0] present_key_transpose_nodes = input_name_to_nodes[reshape_k.output[0]] for present_key_transpose_node in present_key_transpose_nodes: present_key_candidate = self.model.find_graph_output(present_key_transpose_node.output[0]) if present_key_candidate is not None: present_key = present_key_candidate.name break if present_key is None: return if "present_key_cross" not in present_key: return else: k_nodes = self.model.match_parent_path( matmul_qk, ["Transpose"], [1], ) if k_nodes is None: return transpose_k = k_nodes[0] past_key = transpose_k.input[0] if past_key in output_name_to_node: return if "past_key_cross" not in past_key: return else: idx, k_nodes, _ = self.model.match_parent_paths( matmul_qk, [ (["Transpose", "Concat", "Reshape", "MatMul"], [1, 0, 1, 0]), (["Transpose", "Concat", "Transpose", "Reshape", "MatMul"], [1, 0, 1, 0, 0]), ], output_name_to_node, ) past_key_transpose_node = None present_key_transpose_nodes = None if k_nodes is not None: concat_k, reshape_k = k_nodes[1], k_nodes[-2] key = reshape_k.input[0] if idx == 0: past_key_transpose_node = output_name_to_node[concat_k.input[0]] past_key = past_key_transpose_node.input[0] else: past_key = concat_k.input[0] if past_key in output_name_to_node: return if "past_key_self" not in past_key: return if idx == 0: present_key_transpose_nodes = input_name_to_nodes[concat_k.output[0]] for present_key_transpose_node in present_key_transpose_nodes: present_key_candidate = self.model.find_graph_output(present_key_transpose_node.output[0]) if present_key_candidate is not None: present_key = present_key_candidate.name break else: present_key = concat_k.output[0] if present_key is None: return if "present_key_self" not in present_key: return else: k_nodes = self.model.match_parent_path( matmul_qk, ["Transpose", "Reshape", "MatMul"], [1, 0, 0], ) if k_nodes is None: return _, reshape_k, _ = k_nodes key = reshape_k.input[0] present_key_transpose_nodes = input_name_to_nodes[reshape_k.output[0]] for present_key_transpose_node in present_key_transpose_nodes: present_key_candidate = self.model.find_graph_output(present_key_transpose_node.output[0]) if present_key_candidate is not None: present_key = present_key_candidate.name break if present_key is None: return if "present_key_self" not in present_key: return q_nodes = self.model.match_parent_path( matmul_qk, ["Transpose", "Reshape", "MatMul"], [0, 0, 0], ) if q_nodes is None: return transpose_q, reshape_q, matmul_q = q_nodes if matmul_q.input[0] != input_shape_node.input[0]: return q_num_heads, q_hidden_size = self.get_num_heads_and_hidden_size(reshape_q) if self.static_kv == 1 and past_key is not None: key = past_key value = past_value past_key = None past_value = None new_node = self.create_mha_node( matmul_q.output[0], key, value, mask_index, res_pos_bias, past_key, past_value, reshape_qkv.output[0], present_key, present_value, q_num_heads, q_hidden_size, ) if new_node is None: return self.nodes_to_add.append(new_node) self.node_name_to_graph_name[new_node.name] = self.this_graph_name self.nodes_to_remove.extend(qkv_nodes[1:]) self.nodes_to_remove.extend(qk_nodes) self.nodes_to_remove.extend(k_nodes[:-1]) if v_nodes is not None: self.nodes_to_remove.extend(v_nodes[:-1]) self.nodes_to_remove.extend(q_nodes[:-1]) self.prune_graph = True class FusionRelativePositionBiasBlock(Fusion): def __init__(self, model: OnnxModel, max_distance: int): super().__init__(model, "RelativePositionBias", ["Add", "Slice"]) self.max_distance = max_distance # bidirectional=(not self.is_decoder) self.is_bidirectional = False def fuse(self, node, input_name_to_nodes, output_name_to_node): # TODO: Optimization opportunity: only last dimension of relative_position_bias is used in decoder. # Cuda kernel can be optimized to only compute last dimension. if node.op_type != "Add" and node.op_type != "Slice": return compute_bias_nodes = self.model.match_parent_path( node, ["Unsqueeze", "Transpose", "Gather", "Where"], [0, 0, 0, 1] ) if compute_bias_nodes is None: compute_bias_nodes = self.model.match_parent_path( node, ["Unsqueeze", "Transpose", "Gather", "Add", "Where"], [0, 0, 0, 1, 1] ) if compute_bias_nodes is None: return gather = compute_bias_nodes[2] where = compute_bias_nodes[-1] unsqueeze = compute_bias_nodes[0] compute_buckets_nodes = self.model.match_parent_path( where, ["Min", "ConstantOfShape", "Shape", "Add", "Cast", "Mul", "Div", "Log", "Div"], [2, 1, 0, 0, 0, 0, 0, 0, 0], ) if compute_buckets_nodes is None: return div = compute_buckets_nodes[-1] range_nodes = self.model.match_parent_path( div, ["Cast", "Neg", "Min", "ConstantOfShape", "Shape", "Sub", "Unsqueeze", "Range"], [0, 0, 0, 1, 0, 0, 0, 0], ) if range_nodes is None: range_nodes = self.model.match_parent_path( div, ["Cast", "Abs", "Sub", "Unsqueeze", "Range"], [0, 0, 0, 0, 0] ) self.is_bidirectional = True if range_nodes is None: return range_node = range_nodes[-1] self.nodes_to_remove.extend(compute_bias_nodes) self.nodes_to_remove.extend(compute_buckets_nodes) self.nodes_to_remove.extend(range_nodes) node_name_prefix = "encoder" if self.is_bidirectional else "decoder" table_weight_i = self.model.get_initializer(gather.input[0]) table_weight = NumpyHelper.to_array(table_weight_i) table_weight_t = np.transpose(table_weight) bias_table = helper.make_tensor( name=self.model.create_node_name("bias_table_weight", name_prefix=node_name_prefix), data_type=TensorProto.FLOAT, dims=[np.shape(table_weight)[0], np.shape(table_weight)[1]], vals=table_weight_t.tobytes(), raw=True, ) self.model.add_initializer(bias_table, self.this_graph_name) inputs = [bias_table.name, range_node.input[1], range_node.input[1]] outputs = [unsqueeze.output[0]] rpb_node = helper.make_node( "RelativePositionBias", inputs=inputs, outputs=outputs, name=self.model.create_node_name("RelativePositionBias", name_prefix=node_name_prefix), ) rpb_node.domain = "com.microsoft" rpb_node.attribute.extend([helper.make_attribute("max_distance", self.max_distance)]) rpb_node.attribute.extend([helper.make_attribute("is_bidirectional", self.is_bidirectional)]) self.nodes_to_add.append(rpb_node) self.node_name_to_graph_name[rpb_node.name] = self.this_graph_name class T5OnnxModel(BertOnnxModel): def __init__(self, model, num_heads, hidden_size): super().__init__(model, num_heads, hidden_size) self.attention_mask = AttentionMask(self) self.attention_fusion = FusionT5Attention(self, self.hidden_size, self.num_heads, self.attention_mask) self.layer_norm_fusion = FusionSimplifiedLayerNormalization(self) self.skip_layer_norm_fusion = FusionSkipSimplifiedLayerNormalization(self) # TODO: consider retrieve max_distance from model. # math.log(max_distance / (num_buckets // 2)) self.rpb_fusion = FusionRelativePositionBiasBlock(self, 128) def fuse_attention(self): self.attention_fusion.apply() def fuse_layer_norm(self): self.layer_norm_fusion.apply() def fuse_skip_layer_norm(self): self.skip_layer_norm_fusion.apply() # Remove get_extended_attention_mask() since it generates all zeros. def remove_extended_mask_decoder_init(self): nodes_to_remove = [] for node in self.nodes(): if node.op_type == "Add": extended_mask_nodes = self.match_parent_path( node, [ "Mul", "Sub", "Mul", "Unsqueeze", "Cast", "LessOrEqual", "Tile", "Concat", "Unsqueeze", "Gather", "Shape", ], [1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0], ) if extended_mask_nodes is None: continue rpb_nodes = self.match_parent_path(node, ["RelativePositionBias"], [0]) if rpb_nodes is None: continue rpb_node = rpb_nodes[0] rpb_node.output[0] = node.output[0] nodes_to_remove.extend(extended_mask_nodes) nodes_to_remove.append(node) self.remove_nodes(nodes_to_remove) def remove_extended_mask_decoder(self): nodes_to_remove = [] for node in self.nodes(): if node.op_type == "Add": extended_mask_nodes = self.match_parent_path( node, [ "Mul", "Sub", "Mul", "Unsqueeze", "Concat", "Cast", "LessOrEqual", "Tile", "Concat", "Unsqueeze", "Gather", "Shape", ], [1, 0, 1, 0, 0, 1, 0, 0, 1, 1, 0, 0], ) if extended_mask_nodes is None: continue rpb_nodes = self.match_parent_path(node, ["Slice", "RelativePositionBias"], [0, 0]) if rpb_nodes is None: continue rpb_node = rpb_nodes[0] rpb_node.output[0] = node.output[0] nodes_to_remove.extend(extended_mask_nodes) nodes_to_remove.append(node) self.remove_nodes(nodes_to_remove) def preprocess(self): self.adjust_reshape_and_expand() self.rpb_fusion.apply() def postprocess(self): # remove get_extended_attention_mask() since it generates all zeros. self.remove_extended_mask_decoder_init() self.remove_extended_mask_decoder() self.prune_graph()