# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. # Copyright 2018 Mesh TensorFlow authors, T5 Authors and HuggingFace Inc. team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from __future__ import annotations import copy import math from typing import Optional, Tuple import numpy as np import paddle import paddle.nn as nn import paddle.nn.functional as F from paddle import Tensor try: from paddle.amp.auto_cast import amp_state except ImportError: from paddle.fluid.dygraph.amp.auto_cast import amp_state from paddle.distributed.fleet.utils import recompute from ...utils.converter import StateDictNameMapping from ...utils.log import logger from ..activations import ACT2FN from ..model_outputs import ( BaseModelOutput, BaseModelOutputWithPastAndCrossAttentions, Seq2SeqLMOutput, Seq2SeqModelOutput, convert_encoder_output, ) from ..model_utils import PretrainedModel, register_base_model from .configuration import ( T5_PRETRAINED_INIT_CONFIGURATION, T5_PRETRAINED_RESOURCE_FILES_MAP, T5Config, ) __all__ = ["T5Model", "T5PretrainedModel", "T5ForConditionalGeneration", "T5EncoderModel"] T5_PRETRAINED_MODEL_ARCHIVE_LIST = [ "t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b", ] DATA_TYPE_MAP = { paddle.int64: "int64", paddle.int32: "int32", paddle.float32: "float32", paddle.float64: "float64", paddle.float16: "float16", } def data_type_converter(tensor): return DATA_TYPE_MAP[tensor.dtype] def finfo(dtype): if dtype == paddle.float32: return np.finfo(np.float32) if dtype == paddle.float16: return np.finfo(np.float16) if dtype == paddle.float64: return np.finfo(np.float64) class T5LayerNorm(nn.Layer): """ Construct a layernorm module in the T5 style No bias and no subtraction of mean. """ def __init__(self, hidden_size, eps=1e-6): super().__init__() self.weight = self.create_parameter(shape=[hidden_size], default_initializer=nn.initializer.Constant(1.0)) self.variance_epsilon = eps def forward(self, hidden_states): # layer norm should always be calculated in float32 variance = paddle.pow(hidden_states.astype(paddle.float32), 2).mean(axis=-1, keepdim=True) hidden_states = hidden_states * paddle.rsqrt(variance + self.variance_epsilon) # convert into float16 if necessary if amp_state() or self.weight.dtype == paddle.float16: hidden_states = hidden_states.astype(paddle.float16) return self.weight * hidden_states class T5DenseReluDense(nn.Layer): """ Construct a dense-relu-dense module. """ def __init__(self, config: T5Config): super().__init__() self.wi = nn.Linear(config.d_model, config.d_ff, bias_attr=False) self.wo = nn.Linear(config.d_ff, config.d_model, bias_attr=False) self.dropout = nn.Dropout(config.dropout_rate) def forward(self, hidden_states): hidden_states = self.wi(hidden_states) hidden_states = F.relu(hidden_states) hidden_states = self.dropout(hidden_states) hidden_states = self.wo(hidden_states) return hidden_states class T5DenseGatedGeluDense(nn.Layer): """ Construct a dense-gated_gelu-dense module. """ def __init__(self, config: T5Config): super().__init__() self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias_attr=False) self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias_attr=False) self.wo = nn.Linear(config.d_ff, config.d_model, bias_attr=False) self.dropout = nn.Dropout(config.dropout_rate) self.gelu_act = ACT2FN["gelu_new"] def forward(self, hidden_states): hidden_gelu = self.gelu_act(self.wi_0(hidden_states)) hidden_linear = self.wi_1(hidden_states) hidden_states = hidden_gelu * hidden_linear hidden_states = self.dropout(hidden_states) hidden_states = self.wo(hidden_states) return hidden_states class T5DenseGatedSiluDense(nn.Layer): """ Construct a dense-gated_gelu-dense module. """ def __init__(self, config: T5Config): super().__init__() self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias_attr=False) self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias_attr=False) self.wo = nn.Linear(config.d_ff, config.d_model, bias_attr=False) self.dropout = nn.Dropout(config.dropout_rate) def forward(self, hidden_states): hidden_silu = F.silu(self.wi_0(hidden_states)) hidden_linear = self.wi_1(hidden_states) hidden_states = hidden_silu * hidden_linear hidden_states = self.dropout(hidden_states) hidden_states = self.wo(hidden_states) return hidden_states class T5LayerFF(nn.Layer): def __init__(self, config: T5Config): super().__init__() if config.feed_forward_proj == "relu": self.DenseReluDense = T5DenseReluDense(config) elif config.feed_forward_proj == "gated-gelu": self.DenseReluDense = T5DenseGatedGeluDense(config) elif config.feed_forward_proj == "gated-silu": self.DenseReluDense = T5DenseGatedSiluDense(config) else: raise ValueError(f"{config.feed_forward_proj} is not supported. Choose between `relu` and `gated-gelu`") self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) def forward(self, hidden_states): forwarded_states = self.layer_norm(hidden_states) forwarded_states = self.DenseReluDense(forwarded_states) hidden_states = hidden_states + self.dropout(forwarded_states) return hidden_states class T5Attention(nn.Layer): def __init__(self, config: T5Config, has_relative_attention_bias: bool = False): super().__init__() self.is_decoder = config.is_decoder self.has_relative_attention_bias = has_relative_attention_bias self.relative_attention_num_buckets = config.relative_attention_num_buckets self.d_model = config.d_model self.key_value_proj_dim = config.d_kv self.n_heads = config.num_heads self.dropout = config.dropout_rate self.inner_dim = self.n_heads * self.key_value_proj_dim self.enable_recompute = False # Mesh TensorFlow initialization to avoid scaling before softmax self.q = nn.Linear(self.d_model, self.inner_dim, bias_attr=False) self.k = nn.Linear(self.d_model, self.inner_dim, bias_attr=False) self.v = nn.Linear(self.d_model, self.inner_dim, bias_attr=False) self.o = nn.Linear(self.inner_dim, self.d_model, bias_attr=False) if self.has_relative_attention_bias: self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads) @staticmethod def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128): """ Adapted from Mesh Tensorflow: https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593 Translate relative position to a bucket number for relative attention. The relative position is defined as memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for small absolute relative_position and larger buckets for larger absolute relative_positions. All relative positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket. This should allow for more graceful generalization to longer sequences than the model has been trained on Args: relative_position: an int64 Tensor bidirectional: a boolean - whether the attention is bidirectional num_buckets: an integer max_distance: an integer Returns: a Tensor with the same shape as relative_position, containing int64 values in the range [0, num_buckets) """ relative_buckets = 0 if bidirectional: num_buckets //= 2 relative_buckets += (relative_position > 0).astype(paddle.int64) * num_buckets relative_position = paddle.abs(relative_position) else: relative_position = -paddle.minimum(relative_position, paddle.zeros_like(relative_position)) # now relative_position is in the range [0, inf) # half of the buckets are for exact increments in positions max_exact = num_buckets // 2 is_small = relative_position < max_exact # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance relative_postion_if_large = max_exact + ( paddle.log(relative_position.astype(paddle.get_default_dtype()) / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact) ).astype(paddle.int64) relative_postion_if_large = paddle.minimum( relative_postion_if_large, paddle.full_like(relative_postion_if_large, num_buckets - 1), ) relative_buckets += paddle.where(is_small, relative_position, relative_postion_if_large) return relative_buckets def compute_bias(self, query_length, key_length): """Compute binned relative position bias""" context_position = paddle.arange(query_length).unsqueeze(-1) memory_position = paddle.arange(key_length).unsqueeze(0) relative_position = memory_position - context_position # shape (query_length, key_length) relative_position_bucket = self._relative_position_bucket( relative_position, # shape (query_length, key_length) bidirectional=(not self.is_decoder), num_buckets=self.relative_attention_num_buckets, ) values = self.relative_attention_bias(relative_position_bucket) # shape (query_length, key_length, num_heads) values = values.transpose(perm=[2, 0, 1]).unsqueeze(0) # shape (1, num_heads, query_length, key_length) return values def forward( self, hidden_states, mask=None, key_value_states=None, position_bias=None, cache=None, query_length=None, use_cache=False, output_attentions=False, ): """ Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states). """ # Input is (batch_size, seq_length, dim) # Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length) # cache[0] is (batch_size, n_heads, q_len - 1, dim_per_head) batch_size, seq_length = paddle.shape(hidden_states)[:2] real_seq_length = seq_length if cache is not None: assert len(cache) == 2, f"cache should have 2 past states: keys and values. Got { len(cache)} past states" real_seq_length += paddle.shape(cache[0])[2] if query_length is None else query_length key_length = real_seq_length if key_value_states is None else paddle.shape(key_value_states)[1] def shape(states): """projection""" return states.reshape(shape=[batch_size, -1, self.n_heads, self.key_value_proj_dim]).transpose( perm=[0, 2, 1, 3] ) def unshape(states): """reshape""" return states.transpose(perm=[0, 2, 1, 3]).reshape(shape=[batch_size, -1, self.inner_dim]) def project(hidden_states, proj_layer, key_value_states, cache): """projects hidden states correctly to key/query states""" if key_value_states is None: # self-attn # (batch_size, n_heads, seq_length, dim_per_head) hidden_states = shape(proj_layer(hidden_states)) elif cache is None: # cross-attn # (batch_size, n_heads, seq_length, dim_per_head) hidden_states = shape(proj_layer(key_value_states)) if cache is not None: if key_value_states is None: # self-attn # (batch_size, n_heads, key_length, dim_per_head) hidden_states = paddle.concat([cache, hidden_states], axis=2) else: # cross-attn hidden_states = cache return hidden_states # get query states query_states = shape(self.q(hidden_states)) # (batch_size, n_heads, seq_length, dim_per_head) # get key/value states key_states = project( hidden_states, self.k, key_value_states, cache[0] if cache is not None else None, ) value_states = project( hidden_states, self.v, key_value_states, cache[1] if cache is not None else None, ) # compute scores scores = paddle.matmul(query_states, key_states, transpose_y=True) if position_bias is None: if not self.has_relative_attention_bias: position_bias = paddle.zeros( shape=(1, self.n_heads, real_seq_length, key_length), dtype=scores.dtype, ) if self.training and self.enable_recompute: position_bias.stop_gradient = False else: position_bias = self.compute_bias(real_seq_length, key_length) # if key and values are already calculated # we want only the last query position bias if cache is not None: position_bias = position_bias[:, :, -paddle.shape(hidden_states)[1] :, :] if mask is not None: position_bias = position_bias + mask # (batch_size, n_heads, seq_length, key_length) scores += position_bias attn_weights = F.softmax(scores.astype(paddle.float32), axis=-1).astype( scores.dtype ) # (batch_size, n_heads, seq_length, key_length) attn_weights = F.dropout( attn_weights, p=self.dropout, training=self.training ) # (batch_size, n_heads, seq_length, key_length) attn_output = unshape(paddle.matmul(attn_weights, value_states)) # (batch_size, seq_length, dim) attn_output = self.o(attn_output) present_key_value_state = (key_states, value_states) if (self.is_decoder and use_cache) else None outputs = (attn_output,) + (present_key_value_state,) + (position_bias,) if output_attentions: outputs = outputs + (attn_weights,) return outputs class T5LayerSelfAttention(nn.Layer): def __init__(self, config: T5Config, has_relative_attention_bias: bool = False): super().__init__() self.SelfAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias) self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) def forward( self, hidden_states, attention_mask=None, position_bias=None, cache=None, use_cache=False, output_attentions=False, ): normed_hidden_states = self.layer_norm(hidden_states) attention_output = self.SelfAttention( normed_hidden_states, mask=attention_mask, position_bias=position_bias, cache=cache, use_cache=use_cache, output_attentions=output_attentions, ) hidden_states = hidden_states + self.dropout(attention_output[0]) outputs = (hidden_states,) + attention_output[1:] # add attentions if we output them return outputs class T5LayerCrossAttention(nn.Layer): def __init__(self, config: T5Config): super().__init__() self.EncDecAttention = T5Attention(config, has_relative_attention_bias=False) self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) def forward( self, hidden_states, key_value_states, attention_mask=None, position_bias=None, cache=None, use_cache=False, query_length=None, output_attentions=False, ): normed_hidden_states = self.layer_norm(hidden_states) attention_output = self.EncDecAttention( normed_hidden_states, mask=attention_mask, key_value_states=key_value_states, position_bias=position_bias, cache=cache, use_cache=use_cache, query_length=query_length, output_attentions=output_attentions, ) layer_output = hidden_states + self.dropout(attention_output[0]) outputs = (layer_output,) + attention_output[1:] # add attentions if we output them return outputs class T5Block(nn.Layer): def __init__(self, config: T5Config, has_relative_attention_bias: bool = False): super().__init__() self.is_decoder = config.is_decoder self.layer = nn.LayerList() self.layer.append(T5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias)) if self.is_decoder: self.layer.append(T5LayerCrossAttention(config)) self.layer.append(T5LayerFF(config)) def forward( self, hidden_states, attention_mask=None, position_bias=None, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None, cache=None, use_cache=False, output_attentions=False, ): if cache is not None: assert self.is_decoder, "Only decoder can use `caches`" expected_num_caches = 2 if encoder_hidden_states is None else 4 if len(cache) != expected_num_caches: raise ValueError( f"There should be {expected_num_caches} past states. " f"{'2 (past / key) for cross attention. ' if expected_num_caches == 4 else ''}" f"Got {len(cache)} past key / value states" ) self_attn_cache = cache[:2] cross_attn_cache = cache[2:] else: self_attn_cache, cross_attn_cache = None, None self_attention_outputs = self.layer[0]( hidden_states, attention_mask=attention_mask, position_bias=position_bias, cache=self_attn_cache, use_cache=use_cache, output_attentions=output_attentions, ) hidden_states, present_key_value_state = self_attention_outputs[:2] attention_outputs = self_attention_outputs[2:] # Keep self-attention outputs and relative position weights # clamp inf values to enable fp16 training if (amp_state() or hidden_states.dtype == paddle.float16) and paddle.isinf(hidden_states).any(): # TODO finfo clamp_value = finfo(hidden_states.dtype).max - 1000 hidden_states = paddle.clip(hidden_states, min=-clamp_value, max=clamp_value) do_cross_attention = self.is_decoder and encoder_hidden_states is not None if do_cross_attention: # the actual query length is unknown for cross attention # if using past key value states. Need to inject it here if present_key_value_state is not None: query_length = paddle.shape(present_key_value_state[0])[2] else: query_length = None cross_attention_outputs = self.layer[1]( hidden_states, key_value_states=encoder_hidden_states, attention_mask=encoder_attention_mask, position_bias=encoder_decoder_position_bias, cache=cross_attn_cache, query_length=query_length, use_cache=use_cache, output_attentions=output_attentions, ) hidden_states = cross_attention_outputs[0] # clamp inf values to enable fp16 training if (amp_state() or hidden_states.dtype == paddle.float16) and paddle.isinf(hidden_states).any(): clamp_value = finfo(hidden_states.dtype).max - 1000 hidden_states = paddle.clip(hidden_states, min=-clamp_value, max=clamp_value) # Combine self attn and cross attn key value states if present_key_value_state is not None: present_key_value_state = present_key_value_state + cross_attention_outputs[1] # Keep cross-attention outputs and relative position weights attention_outputs = attention_outputs + cross_attention_outputs[2:] # Apply Feed Forward layer hidden_states = self.layer[-1](hidden_states) # clamp inf values to enable fp16 training if (amp_state() or hidden_states.dtype == paddle.float16) and paddle.isinf(hidden_states).any(): clamp_value = finfo(hidden_states.dtype).max - 1000 hidden_states = paddle.clip(hidden_states, min=-clamp_value, max=clamp_value) outputs = (hidden_states,) if use_cache: outputs = outputs + (present_key_value_state,) + attention_outputs else: outputs = outputs + attention_outputs return outputs # hidden-states, present_key_value_states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights) class T5PretrainedModel(PretrainedModel): """ An abstract class for pretrained T5 models. It provides T5 related `model_config_file`, `resource_files_names`, `pretrained_resource_files_map`, `pretrained_init_configuration`, `base_model_prefix` for downloading and loading pretrained models. See `PretrainedModel` for more details. """ base_model_prefix = "t5" config_class = T5Config pretrained_init_configuration = T5_PRETRAINED_INIT_CONFIGURATION pretrained_resource_files_map = T5_PRETRAINED_RESOURCE_FILES_MAP @classmethod def _get_name_mappings(cls, config: T5Config) -> list[StateDictNameMapping]: mappings: list[StateDictNameMapping] = [] model_mappings = [ ["shared.weight", "shared.weight"], ["encoder.embed_tokens.weight", "encoder.embed_tokens.weight"], ["encoder.final_layer_norm.weight", "encoder.final_layer_norm.weight"], ["decoder.embed_tokens.weight", "decoder.embed_tokens.weight"], ["decoder.final_layer_norm.weight", "decoder.final_layer_norm.weight"], [ "encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight", "encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight", ], [ "decoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight", "decoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight", ], ] for layer_index in range(config.num_hidden_layers): for att_head in ["q", "k", "v", "o"]: model_mappings.extend( [ [ f"encoder.block.{layer_index}.layer.0.SelfAttention.{att_head}.weight", f"encoder.block.{layer_index}.layer.0.SelfAttention.{att_head}.weight", "transpose", ], [ f"decoder.block.{layer_index}.layer.0.SelfAttention.{att_head}.weight", f"decoder.block.{layer_index}.layer.0.SelfAttention.{att_head}.weight", "transpose", ], [ f"decoder.block.{layer_index}.layer.1.EncDecAttention.{att_head}.weight", f"decoder.block.{layer_index}.layer.1.EncDecAttention.{att_head}.weight", "transpose", ], ] ) layer_mappings = [ [ f"encoder.block.{layer_index}.layer.1.DenseReluDense.wo.weight", f"encoder.block.{layer_index}.layer.1.DenseReluDense.wo.weight", "transpose", ], [ f"decoder.block.{layer_index}.layer.2.DenseReluDense.wo.weight", f"decoder.block.{layer_index}.layer.2.DenseReluDense.wo.weight", "transpose", ], [ f"encoder.block.{layer_index}.layer.0.layer_norm.weight", f"encoder.block.{layer_index}.layer.0.layer_norm.weight", ], [ f"encoder.block.{layer_index}.layer.1.layer_norm.weight", f"encoder.block.{layer_index}.layer.1.layer_norm.weight", ], [ f"decoder.block.{layer_index}.layer.0.layer_norm.weight", f"decoder.block.{layer_index}.layer.0.layer_norm.weight", ], [ f"decoder.block.{layer_index}.layer.1.layer_norm.weight", f"decoder.block.{layer_index}.layer.1.layer_norm.weight", ], [ f"decoder.block.{layer_index}.layer.2.layer_norm.weight", f"decoder.block.{layer_index}.layer.2.layer_norm.weight", ], ] if config.feed_forward_proj == "relu": layer_mappings.extend( [ [ f"encoder.block.{layer_index}.layer.1.DenseReluDense.wi.weight", f"encoder.block.{layer_index}.layer.1.DenseReluDense.wi.weight", "transpose", ], [ f"decoder.block.{layer_index}.layer.2.DenseReluDense.wi.weight", f"decoder.block.{layer_index}.layer.2.DenseReluDense.wi.weight", "transpose", ], ] ) elif config.feed_forward_proj == "gated-gelu": for i in range(2): layer_mappings.extend( [ [ f"encoder.block.{layer_index}.layer.1.DenseReluDense.wi_{i}.weight", f"encoder.block.{layer_index}.layer.1.DenseReluDense.wi_{i}.weight", "transpose", ], [ f"decoder.block.{layer_index}.layer.2.DenseReluDense.wi_{i}.weight", f"decoder.block.{layer_index}.layer.2.DenseReluDense.wi_{i}.weight", "transpose", ], ] ) model_mappings.extend(layer_mappings) if cls.__name__ != "T5Model": for mapping in model_mappings: mapping[1] = "t5." + mapping[1] if config.architectures is not None and "T5ForConditionalGeneration" in config.architectures: model_mappings.append(["lm_head.weight", "lm_head.weight", "transpose"]) mappings = [StateDictNameMapping(*mapping) for mapping in model_mappings] return mappings @property def dummy_inputs(self): DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]] DUMMY_MASK = [[1, 1, 1, 1, 1], [1, 1, 1, 0, 0], [0, 0, 0, 1, 1]] input_ids = paddle.assign(np.asarray(DUMMY_INPUTS, dtype="int64")) input_mask = paddle.assign(np.asarray(DUMMY_MASK, dtype="int64")) dummy_inputs = { "decoder_input_ids": input_ids, "input_ids": input_ids, "decoder_attention_mask": input_mask, } return dummy_inputs def init_weights(self): """ Initializes and tie weights if needed. """ # Initialize weights self.apply(self._init_weights) def _init_weights(self, layer): """Initialize the weights""" # Used for testing weights initialization factor = self.config.initializer_factor d_model = self.config.d_model d_ff = self.config.d_ff n_heads = self.config.num_heads key_value_proj_dim = self.config.d_kv if isinstance(layer, T5LayerNorm): layer.weight.set_value(paddle.ones_like(layer.weight) * factor) elif isinstance(layer, T5Model): # Mesh TensorFlow embeddings initialization # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624 layer.shared.weight.set_value(paddle.normal(mean=0.0, std=factor * 1.0, shape=layer.shared.weight.shape)) elif isinstance(layer, (T5ForConditionalGeneration,)): layer.t5.shared.weight.set_value( paddle.normal(mean=0.0, std=factor * 1.0, shape=layer.t5.shared.weight.shape) ) elif isinstance(layer, T5DenseReluDense): # Mesh TensorFlow FF initialization # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56 # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89 layer.wi.weight.set_value( paddle.normal(mean=0.0, std=factor * ((d_model) ** -0.5), shape=layer.wi.weight.shape) ) if hasattr(layer.wi, "bias") and layer.wi.bias is not None: layer.wi.bias.set_value(paddle.zeros_like(layer.wi.bias)) layer.wo.weight.set_value( paddle.normal(mean=0.0, std=factor * ((d_ff) ** -0.5), shape=layer.wo.weight.shape) ) if hasattr(layer.wo, "bias") and layer.wo.bias is not None: layer.wo.bias.set_value(paddle.zeros_like(layer.wo.bias)) elif isinstance(layer, T5DenseGatedGeluDense): layer.wi_0.weight.set_value( paddle.normal(mean=0.0, std=factor * ((d_model) ** -0.5), shape=layer.wi_0.weight.shape) ) if hasattr(layer.wi_0, "bias") and layer.wi_0.bias is not None: layer.wi_0.bias.set_value(paddle.zeros_like(layer.wi_0.bias)) layer.wi_1.weight.set_value( paddle.normal(mean=0.0, std=factor * ((d_model) ** -0.5), shape=layer.wi_1.weight.shape) ) if hasattr(layer.wi_1, "bias") and layer.wi_1.bias is not None: layer.wi_1.bias.set_value(paddle.zeros_like(layer.wi_1.bias)) layer.wo.weight.set_value( paddle.normal(mean=0.0, std=factor * ((d_ff) ** -0.5), shape=layer.wo.weight.shape) ) if hasattr(layer.wo, "bias") and layer.wo.bias is not None: layer.wo.bias.set_value(paddle.zeros_like(layer.wo.bias)) elif isinstance(layer, T5Attention): # Mesh TensorFlow attention initialization to avoid scaling before softmax # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136 layer.q.weight.set_value( paddle.normal( mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5), shape=layer.q.weight.shape ) ) layer.k.weight.set_value( paddle.normal(mean=0.0, std=factor * (d_model**-0.5), shape=layer.k.weight.shape) ) layer.v.weight.set_value( paddle.normal(mean=0.0, std=factor * (d_model**-0.5), shape=layer.v.weight.shape) ) layer.o.weight.set_value( paddle.normal( mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5), shape=layer.o.weight.shape ) ) if layer.has_relative_attention_bias: layer.relative_attention_bias.weight.set_value( paddle.normal( mean=0.0, std=factor * ((d_model) ** -0.5), shape=layer.relative_attention_bias.weight.shape ) ) def _shift_right(self, input_ids): bos_token_id = self.config.bos_token_id pad_token_id = self.config.pad_token_id assert ( bos_token_id is not None ), "bos_token_id has to be defined. In T5 it is usually set to the pad_token_id. See T5 docs for more information" # shift inputs to the right shifted_input_ids = paddle.zeros_like(input_ids) shifted_input_ids[:, 1:] = input_ids[:, :-1].clone() shifted_input_ids[:, 0] = bos_token_id assert pad_token_id is not None, "pad_token_id has to be defined." # replace possible -100 values in labels by `pad_token_id` shifted_input_ids = paddle.where( shifted_input_ids == -100, paddle.assign(np.asarray(pad_token_id, dtype=data_type_converter(shifted_input_ids)).reshape([1])), shifted_input_ids, ) assert paddle.all(shifted_input_ids >= 0), "Verify that `shifted_input_ids` has only positive values" return shifted_input_ids class T5Stack(T5PretrainedModel): def __init__(self, config: T5Config, embed_tokens: Optional[nn.Embedding] = None): super().__init__(config) self.is_decoder = config.is_decoder self.embed_tokens = embed_tokens self.block = nn.LayerList( [T5Block(config, has_relative_attention_bias=bool(i == 0)) for i in range(config.num_layers)] ) self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) self.enable_recompute = config.enable_recompute def get_input_embeddings(self): return self.embed_tokens def set_input_embeddings(self, new_embeddings): self.embed_tokens = new_embeddings @property def dtype(self): return self.embed_tokens.weight.dtype @paddle.jit.not_to_static def recompute_training( self, layer_module, hidden_states, extended_attention_mask, position_bias, encoder_hidden_states, encoder_extended_attention_mask, encoder_decoder_position_bias, use_cache, output_attentions, ): def create_custom_forward(module): def custom_forward(*inputs): return tuple(module(*inputs, use_cache, output_attentions)) return custom_forward layer_outputs = recompute( create_custom_forward(layer_module), hidden_states, extended_attention_mask, position_bias, encoder_hidden_states, encoder_extended_attention_mask, encoder_decoder_position_bias, None, ) return layer_outputs def forward( self, input_ids=None, attention_mask=None, encoder_hidden_states=None, encoder_attention_mask=None, inputs_embeds=None, cache=None, use_cache=False, output_attentions=False, output_hidden_states=False, return_dict=False, **model_kwargs ): if input_ids is not None and inputs_embeds is not None: err_msg_prefix = "decoder_" if self.is_decoder else "" raise ValueError( f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time" ) elif input_ids is not None: input_shape = paddle.shape(input_ids) # input_ids = input_ids.reshape(shape=[-1, input_shape[-1]]) elif inputs_embeds is not None: input_shape = paddle.shape(inputs_embeds)[:-1] else: err_msg_prefix = "decoder_" if self.is_decoder else "" raise ValueError(f"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds") if inputs_embeds is None: assert self.embed_tokens is not None, "You have to initialize the model with valid token embeddings" inputs_embeds = self.embed_tokens(input_ids) batch_size, seq_length = input_shape # required mask seq length can be calculated via length of past mask_seq_length = paddle.shape(cache[0][0])[2] + seq_length if cache is not None else seq_length if use_cache is True: assert self.is_decoder, f"`use_cache` can only be set to `True` if {self.__class__} is used as a decoder" if attention_mask is None: attention_mask = paddle.ones(shape=[batch_size, mask_seq_length]) if self.is_decoder and encoder_attention_mask is None and encoder_hidden_states is not None: encoder_seq_length = paddle.shape(encoder_hidden_states)[1] encoder_attention_mask = paddle.ones([batch_size, encoder_seq_length], dtype=paddle.int64) # initialize caches with `None` if past does not exist if cache is None: cache = [None] * len(self.block) # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length] # ourselves in which case we just need to make it broadcastable to all heads. extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape) # If a 2D or 3D attention mask is provided for the cross-attention # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length] if self.is_decoder and encoder_hidden_states is not None: encoder_batch_size, encoder_sequence_length, _ = paddle.shape(encoder_hidden_states) encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length) if encoder_attention_mask is None: encoder_attention_mask = paddle.ones(shape=encoder_hidden_shape) encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask) else: encoder_extended_attention_mask = None present_key_value_states = () if use_cache else None all_hidden_states = () if output_hidden_states else None all_attentions = () if output_attentions else None all_cross_attentions = () if (output_attentions and self.is_decoder) else None position_bias = None encoder_decoder_position_bias = None hidden_states = self.dropout(inputs_embeds) for i, (layer_module, past_key_value) in enumerate(zip(self.block, cache)): if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) if self.enable_recompute and self.training: if use_cache: logger.warning( "`use_cache=True` is incompatible with `config.enable_recompute=True`. Setting " "`use_cache=False`..." ) use_cache = False layer_outputs = self.recompute_training( layer_module, hidden_states, extended_attention_mask, position_bias, encoder_hidden_states, encoder_extended_attention_mask, encoder_decoder_position_bias, use_cache, output_attentions, ) else: layer_outputs = layer_module( hidden_states, attention_mask=extended_attention_mask, position_bias=position_bias, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_extended_attention_mask, encoder_decoder_position_bias=encoder_decoder_position_bias, cache=past_key_value, use_cache=use_cache, output_attentions=output_attentions, ) # layer_outputs is a tuple with: # hidden-states, key-value-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights) if not use_cache: layer_outputs = layer_outputs[:1] + (None,) + layer_outputs[1:] hidden_states, present_key_value_state = layer_outputs[:2] # We share the position biases between the layers - the first layer store them # layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights), # (cross-attention position bias), (cross-attention weights) position_bias = layer_outputs[2] if self.is_decoder and encoder_hidden_states is not None: encoder_decoder_position_bias = layer_outputs[4 if output_attentions else 3] # append next layer key value states if use_cache: present_key_value_states = present_key_value_states + (present_key_value_state,) if output_attentions: all_attentions = all_attentions + (layer_outputs[3],) if self.is_decoder: all_cross_attentions = all_cross_attentions + (layer_outputs[5],) hidden_states = self.final_layer_norm(hidden_states) hidden_states = self.dropout(hidden_states) # Add last layer if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) if not return_dict: return tuple( v for v in [ hidden_states, present_key_value_states, all_hidden_states, all_attentions, all_cross_attentions, ] if v is not None ) return BaseModelOutputWithPastAndCrossAttentions( last_hidden_state=hidden_states, past_key_values=present_key_value_states, hidden_states=all_hidden_states, attentions=all_attentions, cross_attentions=all_cross_attentions, ) def get_extended_attention_mask(self, attention_mask, input_shape): if attention_mask.ndim == 3: extended_attention_mask = attention_mask.unsqueeze(1) elif attention_mask.ndim == 2: # Provided a padding mask of dimensions [batch_size, seq_length] # - if the model is a decoder, apply a causal mask in addition to the padding mask # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length] if self.is_decoder: batch_size, seq_length = input_shape seq_ids = paddle.arange(seq_length) causal_mask = paddle.tile( seq_ids.unsqueeze(axis=[0, 1]), [batch_size, seq_length, 1] ) <= seq_ids.unsqueeze(axis=[0, 2]) causal_mask = causal_mask.astype(attention_mask.dtype) if causal_mask.shape[1] < attention_mask.shape[1]: prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1] causal_mask = paddle.concat( [ paddle.ones( [batch_size, seq_length, prefix_seq_len], dtype=causal_mask.dtype, ), causal_mask, ], axis=-1, ) extended_attention_mask = causal_mask.unsqueeze(1) * attention_mask.unsqueeze([1, 2]) else: extended_attention_mask = attention_mask.unsqueeze([1, 2]) elif attention_mask.ndim == 4: if self.is_decoder: batch_size, seq_length = input_shape seq_ids = paddle.arange(seq_length) causal_mask = paddle.tile( seq_ids.unsqueeze(axis=[0, 1]), [batch_size, seq_length, 1] ) <= seq_ids.unsqueeze(axis=[0, 2]) # in case cache are used we need to add a prefix ones mask to the causal mask # causal and attention masks must have same type causal_mask = causal_mask.astype(attention_mask.dtype) if causal_mask.shape[1] < attention_mask.shape[-1]: prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1] causal_mask = paddle.concat( [ paddle.ones( [batch_size, seq_length, prefix_seq_len], dtype=causal_mask.dtype, ), causal_mask, ], axis=-1, ) extended_attention_mask = causal_mask.unsqueeze(1) * attention_mask else: extended_attention_mask = attention_mask else: raise ValueError( f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})" ) extended_attention_mask = extended_attention_mask.astype(self.dtype) extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0 return extended_attention_mask def invert_attention_mask(self, encoder_attention_mask): if encoder_attention_mask.ndim == 4: encoder_extended_attention_mask = encoder_attention_mask elif encoder_attention_mask.ndim == 3: encoder_extended_attention_mask = encoder_attention_mask.unsqueeze(1) elif encoder_attention_mask.ndim == 2: encoder_extended_attention_mask = encoder_attention_mask.unsqueeze([1, 2]) encoder_extended_attention_mask = encoder_extended_attention_mask.astype(self.dtype) # fp16 compatibility if amp_state() or self.dtype == paddle.float16: encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e4 elif self.dtype == paddle.float32: encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e4 else: encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e4 # raise ValueError( # f"{self.dtype} not recognized. `dtype` should be set to either `paddle.float32` or `paddle.float16`" # ) return encoder_extended_attention_mask @register_base_model class T5Model(T5PretrainedModel): """ The bare T5 Model transformer outputting raw hidden-states without any specific head on top. This model inherits from :class:`~paddlenlp.transformers.model_utils.PretrainedModel`. Refer to the superclass documentation for the generic methods. This model is also a Paddle `paddle.nn.Layer `__ subclass. Use it as a regular Paddle Layer and refer to the Paddle documentation for all matter related to general usage and behavior. Args: config (class:`T5Config`): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. """ def __init__(self, config: T5Config): super().__init__(config) self.bos_token_id = config.bos_token_id self.pad_token_id = config.pad_token_id self.initializer_factor = config.initializer_factor self.d_model = config.d_model self.num_heads = config.num_heads self.d_kv = config.d_kv self.d_ff = config.d_ff self.tie_word_embeddings = config.tie_word_embeddings self.shared = nn.Embedding(config.vocab_size, config.d_model) encoder_config = copy.deepcopy(config) encoder_config.is_decoder = False encoder_config.use_cache = False encoder_config.is_encoder_decoder = False self.encoder = T5Stack(encoder_config, self.shared) decoder_config = copy.deepcopy(config) decoder_config.is_decoder = True decoder_config.is_encoder_decoder = False decoder_config.num_layers = config.num_decoder_layers self.decoder = T5Stack(decoder_config, self.shared) self.init_weights() def get_input_embeddings(self): return self.shared def set_input_embeddings(self, new_embeddings): self.shared = new_embeddings self.encoder.set_input_embeddings(new_embeddings) self.decoder.set_input_embeddings(new_embeddings) def get_encoder(self): return self.encoder def get_decoder(self): return self.decoder def forward( self, input_ids=None, attention_mask=None, decoder_input_ids=None, decoder_attention_mask=None, encoder_output=None, cache=None, inputs_embeds=None, decoder_inputs_embeds=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): r""" The T5Model forward method, overrides the `__call__()` special method. Args: input_ids (Tensor): Indices of input sequence tokens in the vocabulary. They are numerical representations of tokens that build the input sequence. Its data type should be `int64` and it has a shape of [batch_size, sequence_length]. attention_mask (Tensor, optional): Mask used in multi-head attention to avoid performing attention on to some unwanted positions, usually the paddings or the subsequent positions. Its data type can be int, float. When the data type is int, the `masked` tokens have `0` values and the others have `1` values. When the data type is float, the `masked` tokens have `0.0` values and the others have `1.0` values. It is a tensor with shape broadcasted to [batch_size, num_attention_heads, sequence_length, sequence_length]. Defaults to `None`, which means nothing needed to be prevented attention to. decoder_input_ids (Tensor, optional): Indices of decoder input sequence tokens in the vocabulary. Its data type should be `int64` and it has a shape of [batch_size, sequence_length]. Defaults to `None`, which means no `decoder_input_ids` is provided, the model will create the tensor by shifting the `input_ids` to the right. decoder_attention_mask (Tensor, optional): Mask used in multi-head attention to avoid performing attention to some unwanted positions in `decoder_input_ids`. Its data type and shape is the same as `attention_mask`. Defaults to `None`. encoder_output (tuple, optional): The output of the encoder, a tuple consists `last_hidden_state`, `hidden_states`(optional), `attentions`(optional). The data type of `last_hidden_state` is float32 and its shape is [batch_size, sequence_length, hidden_size]. `hidden_states` is hidden_states of all layers in the Transformer encoder. The length of `hidden_states` is `num_hidden_layers + 1`. For all element in the tuple, its data type should be float32 and its shape is [batch_size, sequence_length, hidden_size]. `attentions` is attentions of all layers of in the Transformer encoder. The length of `attentions` is `num_hidden_layers`. For all element in the tuple, its data type should be float32 and its shape is [batch_size, num_attention_heads, sequence_length, sequence_length]. cache (Tuple[Tuple[Tensor]], optional): Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model. Can be used to speed up sequential decoding. The `input_ids` which have their past given to this model should not be passed as input ids as they have already been computed. Defaults to `None`. inputs_embeds (Tensor, optional): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation of shape `(batch_size, sequence_length, hidden_size)`. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. Default to None. decoder_inputs_embeds (Tensor, optional): Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded representation of shape `(batch_size, target_sequence_length, hidden_size)`. If `cache` is used, optionally only the last `decoder_inputs_embeds` have to be input (see `past_key_values`). This is useful if you want more control over how to convert `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix. Default to None. If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value of `inputs_embeds`. use_cache (bool, optional): Whether or not to use cache. If set to `True`, `past_buckets_states` states are returned and can be used to speed up decoding. Defaults to `False`. output_attentions (bool, optional): Whether or not to return the attentions tensors of all attention layers. Defaults to `False`. output_hidden_states (bool, optional): Whether or not to return the output of all hidden layers. Defaults to `False`. return_dict (bool, optional): Whether or not to return a class:`~paddlenlp.transformers.model_outputs.Seq2SeqModelOutput`. If `False`, the output will be a tuple of tensors. Defaults to `False`. Returns: An instance of :class:`~paddlenlp.transformers.model_outputs.Seq2SeqModelOutput` if `return_dict=True`. Otherwise it returns a tuple of tensors corresponding to ordered and not None (depending on the input arguments) fields of :class:`~paddlenlp.transformers.model_outputs.Seq2SeqModelOutput`. tuple: Returns tuple (`last_hidden_state`, `cache`, `decoder_hidden_states`, `decoder_attentions`, `cross_attentions`, `encoder_last_hidden_state`, `encoder_hidden_states`, `encoder_attentions`) With the fields: - `last_hidden_state` (Tensor): Sequence of hidden-states at the last layer of the decoder of the model. It's data type should be float32 and its shape is [batch_size, sequence_length, hidden_size]. - `cache` (List[tuple(Tensor, Tensor)], optional): returned when `use_cache=True` is passed. List of `tuple(Tensor, Tensor)` of length `config["num_layers"]`, with the first element being the previous `buckets` of shape `[batch_size, num_heads, num_hashes, sequence_length]` and the second being the previous `hidden_states` of shape `[batch_size, sequence_length, hidden_size]`. - `decoder_hidden_states` (tuple(Tensor), optional) returned when ``output_hidden_states=True`` is passed. Tuple of `Tensor` (one for the output of the embeddings + one for the output of decoder each layer) of shape `(batch_size, sequence_length, hidden_size)`. - `decoder_attentions` (tuple(Tensor), optional): returned when `output_attentions=True` is passed. tuple of `Tensor` (one for each layer) of shape. Each Tensor has a data type of float32 and its shape is [batch_size, num_heads, sequence_length, sequence_length]. - `cross_attentions` (tuple(Tensor), optional): returned when `output_attentions=True` is passed. tuple of `Tensor` (one for each layer) of shape. Each Tensor has a data type of float32 and its shape is [batch_size, num_heads, sequence_length, sequence_length]. - `encoder_last_hidden_state` (Tensor): Sequence of hidden-states at the last layer of the encoder of the model. It's data type should be float32 and its shape is [batch_size, sequence_length, hidden_size]. - `encoder_hidden_states` (tuple(Tensor), optional): returned when `output_hidden_states=True` is passed. tuple of `Tensor` (one for the output of the embeddings + one for the output of encoder each layer). Each Tensor has a data type of float32 and its shape is [batch_size, sequence_length, hidden_size]. - `encoder_attentions` (tuple(Tensor), optional): returned when `output_attentions=True` is passed. tuple of `Tensor` (one for each layer) of shape. Each Tensor has a data type of float32 and its shape is [batch_size, num_heads, sequence_length, sequence_length]. Example: .. code-block:: import paddle from paddlenlp.transformers import T5Model, T5Tokenizer tokenizer = T5Tokenizer.from_pretrained('t5-base') model = T5Model.from_pretrained('t5-base') inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!") input_ids = paddle.to_tensor([inputs["input_ids"]], dtype="int64") decoder_inputs = tokenizer("It means you can") decoder_input_ids = paddle.to_tensor([decoder_inputs["input_ids"]], dtype="int64") outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids) last_hidden_state = outputs[0] print(last_hidden_state.shape) # [1, 5, 768] """ use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict # Encode if needed (training, first prediction pass) if encoder_output is None: encoder_output = self.encoder( input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) elif return_dict and not isinstance(encoder_output, BaseModelOutput): encoder_output = convert_encoder_output(encoder_output) hidden_states = encoder_output[0] # Decode decoder_outputs = self.decoder( input_ids=decoder_input_ids, attention_mask=decoder_attention_mask, inputs_embeds=decoder_inputs_embeds, cache=cache, encoder_hidden_states=hidden_states, encoder_attention_mask=attention_mask, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) if not return_dict: return decoder_outputs + encoder_output return Seq2SeqModelOutput( last_hidden_state=decoder_outputs.last_hidden_state, past_key_values=decoder_outputs.past_key_values, decoder_hidden_states=decoder_outputs.hidden_states, decoder_attentions=decoder_outputs.attentions, cross_attentions=decoder_outputs.cross_attentions, encoder_last_hidden_state=encoder_output.last_hidden_state, encoder_hidden_states=encoder_output.hidden_states, encoder_attentions=encoder_output.attentions, ) class T5ForConditionalGeneration(T5PretrainedModel): """ The T5 Model transformer with a language modeling head on top. Args: config (:class:`T5Config`): An instance of T5Config used to construct T5ForConditionalGeneration. """ def __init__(self, config: T5Config): super().__init__(config) self.t5 = T5Model(config) if not self.t5.config["tie_word_embeddings"]: self.lm_head = nn.Linear(self.t5.config["d_model"], self.t5.config["vocab_size"], bias_attr=False) self.init_weights() def get_input_embeddings(self): return self.t5.shared def set_input_embeddings(self, new_embeddings): self.t5.shared = new_embeddings self.t5.encoder.set_input_embeddings(new_embeddings) self.t5.decoder.set_input_embeddings(new_embeddings) def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings def get_output_embeddings(self): if self.t5.config["tie_word_embeddings"]: return self.t5.shared else: return self.lm_head def get_encoder(self): return self.t5.encoder def get_decoder(self): return self.t5.decoder def forward( self, input_ids=None, attention_mask=None, decoder_input_ids=None, decoder_attention_mask=None, encoder_output=None, cache=None, labels=None, inputs_embeds=None, decoder_inputs_embeds=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): r""" Args: input_ids (Tensor, optional): See :class:`T5Model`. attention_mask (Tensor, optional): See :class:`T5Model`. decoder_input_ids (Tensor, optional): See :class:`T5Model`. decoder_attention_mask (Tensor, optional): See :class:`T5Model`. encoder_output (tuple(Tensor), optional): See :class:`T5Model`. cache (List[tuple(Tensor, Tensor)], optional): See :class:`T5Model`. labels (Tensor, optional): Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set `labels = input_ids` Indices are selected in `[-100, 0, ..., vocab_size]` All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., vocab_size]`. Shape is [batch_size, sequence_length] and dtype is int64. inputs_embeds (Tensor, optional): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation of shape `(batch_size, sequence_length, hidden_size)`. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. Default to None. decoder_inputs_embeds (Tensor , optional): Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded representation of shape `(batch_size, target_sequence_length, hidden_size)`. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be input (see `past_key_values`). This is useful if you want more control over how to convert `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix. Default to None. If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value of `inputs_embeds`. use_cache (bool, optional): See :class:`T5Model`. output_attentions (bool, optional): See :class:`T5Model`. output_hidden_states (bool, optional): See :class:`T5Model`. return_dict (bool, optional): Whether or not to return a class:`~paddlenlp.transformers.model_outputs.Seq2SeqLMOutput`. If `False`, the output will be a tuple of tensors. Defaults to `False`. Returns: An instance of :class:`~paddlenlp.transformers.model_outputs.Seq2SeqLMOutput` if `return_dict=True`. Otherwise it returns a tuple of tensors corresponding to ordered and not None (depending on the input arguments) fields of :class:`~paddlenlp.transformers.model_outputs.Seq2SeqLMOutput`. tuple: Returns tuple (`loss`, `logits`, `cache`, `decoder_hidden_states`, `decoder_attentions`, `cross_attentions`, `encoder_last_hidden_state`, `encoder_hidden_states`, `encoder_attentions`) With the fields: - `loss` (Tensor): returned when `labels` is provided. Language modeling loss. It's data type should be float32 and its shape is [1,]. - `logits` (Tensor): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). It's data type should be float32 and its shape is [batch_size, sequence_length, vocab_size]. - `cache` (List[tuple(Tensor, Tensor)], optional): See :class:`T5Model`. - `decoder_hidden_states` (tuple(Tensor), optional) See :class:`T5Model`. - `decoder_attentions` (tuple(Tensor), optional): See :class:`T5Model`. - `cross_attentions` (tuple(Tensor), optional): See :class:`T5Model`. - `encoder_last_hidden_state` (Tensor): See :class:`T5Model`. - `encoder_hidden_states` (tuple(Tensor), optional): See :class:`T5Model`. - `encoder_attentions` (tuple(Tensor), optional): See :class:`T5Model`. Example: .. code-block:: import paddle from paddlenlp.transformers import T5ForConditionalGeneration, T5Tokenizer tokenizer = T5Tokenizer.from_pretrained('t5-base') model = T5ForConditionalGeneration.from_pretrained('t5-base') inputs = tokenizer("Welcome to use PaddlePaddle and PaddleNLP!") inputs = {k:paddle.to_tensor([v]) for (k, v) in inputs.items()} output = model(**inputs, labels=inputs["input_ids"]) loss = output[0] logits = output[1] """ input_type = type(decoder_input_ids) if decoder_input_ids is not None else type(decoder_inputs_embeds) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict # Encode if needed (training, first prediction pass) if encoder_output is None: # Convert encoder inputs in embeddings if needed encoder_output = self.t5.encoder( input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) else: if isinstance(encoder_output, input_type): encoder_output = (encoder_output,) if return_dict and not isinstance(encoder_output, BaseModelOutput): encoder_output = convert_encoder_output(encoder_output) hidden_states = encoder_output[0] if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None: # get decoder inputs from shifting lm labels to the right decoder_input_ids = self._shift_right(labels) # If decoding with past key value states, only the last tokens # should be given as an input if cache is not None: assert labels is None, "Decoder should not use cached key value states when training." if decoder_input_ids is not None: decoder_input_ids = decoder_input_ids[:, -1:] encoder_attention_mask = attention_mask if attention_mask is not None: if attention_mask.ndim == 4: encoder_attention_mask = attention_mask[:, :, -1:, :] elif attention_mask.ndim == 3: encoder_attention_mask = attention_mask[:, -1:, :].unsqueeze([1]) elif attention_mask.ndim == 2: encoder_attention_mask = attention_mask.unsqueeze([1, 2]) else: raise ValueError("Invalid attention mask shape. ") # Decode decoder_outputs = self.t5.decoder( input_ids=decoder_input_ids, attention_mask=decoder_attention_mask, inputs_embeds=decoder_inputs_embeds, cache=cache, encoder_hidden_states=hidden_states, encoder_attention_mask=encoder_attention_mask, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) sequence_output = decoder_outputs[0] if self.t5.config["tie_word_embeddings"]: # Rescale output before projecting on vocab # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586 sequence_output = sequence_output * (self.t5.config["d_model"] ** -0.5) lm_logits = paddle.matmul(sequence_output, self.t5.shared.weight, transpose_y=True) else: lm_logits = self.lm_head(sequence_output) loss = None if labels is not None: loss_fct = nn.CrossEntropyLoss(ignore_index=-100) loss = loss_fct(lm_logits.reshape(shape=[-1, lm_logits.shape[-1]]).astype("float32"), labels.flatten()) if not return_dict: output = (lm_logits,) + decoder_outputs[1:] + encoder_output return ((loss,) + output) if loss is not None else output return Seq2SeqLMOutput( loss=loss, logits=lm_logits, past_key_values=decoder_outputs.past_key_values, decoder_hidden_states=decoder_outputs.hidden_states, decoder_attentions=decoder_outputs.attentions, cross_attentions=decoder_outputs.cross_attentions, encoder_last_hidden_state=encoder_output.last_hidden_state, encoder_hidden_states=encoder_output.hidden_states, encoder_attentions=encoder_output.attentions, ) def prepare_fast_entry(self, kwargs): from paddlenlp.ops import FasterT5 use_fp16_decoding = kwargs.get("use_fp16_decoding", False) decode_strategy = kwargs.get("decode_strategy") if decode_strategy == "sampling" and kwargs.get("top_k") != 0 and kwargs.get("top_p") != 1: raise AttributeError( "Only topk sampling or topp sampling are supported. " "Topk sampling and topp sampling cannot be both applied in the fast version." ) if kwargs["repetition_penalty"] != 1.0: # not support for repetition_penalty yet in the fast version raise AttributeError("'repetition_penalty != 1' is not supported yet in the fast version") if kwargs["forced_bos_token_id"] is not None: # not support for min_length yet in the fast version raise AttributeError("'forced_bos_token_id != None' is not supported yet in the fast version") self._fast_entry = FasterT5(self, use_fp16_decoding=use_fp16_decoding).forward return self._fast_entry @staticmethod def prepare_input_ids_for_generation(bos_token_id, encoder_output=None): batch_size = 1 if bos_token_id is None: raise ValueError("`bos_token_id` should be defined when no " "`input_ids` are provided.") if encoder_output is not None: if isinstance(encoder_output, tuple): encoder_output = encoder_output[0] batch_size = encoder_output.shape[0] return paddle.ones([batch_size, 1], dtype="int64") * bos_token_id def prepare_inputs_for_generation( self, input_ids, cache=None, attention_mask=None, use_cache=None, encoder_output=None, **kwargs ): # cut decoder_input_ids if past is used if cache is not None: input_ids = input_ids[:, -1:] return { "decoder_input_ids": input_ids, "cache": cache, "encoder_output": encoder_output, "attention_mask": attention_mask, "use_cache": use_cache, } def prepare_decoder_input_ids_from_labels(self, labels: paddle.Tensor): return self._shift_right(labels) @staticmethod def expand_inputs_for_generation(input_ids, expand_size, attention_mask=None, **model_kwargs): index = paddle.tile(paddle.arange(input_ids.shape[0]).unsqueeze(-1), [1, expand_size]).reshape([-1]) input_ids = paddle.index_select(input_ids, index) if attention_mask is not None: model_kwargs["attention_mask"] = paddle.index_select(attention_mask, index) if "token_type_ids" in model_kwargs: token_type_ids = model_kwargs["token_type_ids"] model_kwargs["token_type_ids"] = paddle.index_select(token_type_ids, index) if "position_ids" in model_kwargs: position_ids = model_kwargs["position_ids"] model_kwargs["position_ids"] = paddle.index_select(position_ids, index) if "seq_len" in model_kwargs: seq_len = model_kwargs["seq_len"] model_kwargs["seq_len"] = paddle.index_select(seq_len, index) if "encoder_output" in model_kwargs: encoder_output = model_kwargs["encoder_output"] if isinstance(encoder_output, tuple): model_kwargs["encoder_output"] = (paddle.index_select(encoder_output[0], index),) + encoder_output[1:] else: model_kwargs["encoder_output"] = paddle.index_select(encoder_output, index) return input_ids, model_kwargs @staticmethod def prepare_attention_mask_for_generation(input_ids, pad_token_id, eos_token_id): is_pad_token_in_inputs_ids = (pad_token_id is not None) and paddle.any(input_ids == pad_token_id).item() is_pad_token_not_equal_to_eos_token_id = (eos_token_id is None) or ( (eos_token_id is not None) and (pad_token_id != eos_token_id) ) if is_pad_token_in_inputs_ids and is_pad_token_not_equal_to_eos_token_id: attention_mask = (input_ids != pad_token_id).astype("int64") return attention_mask else: attention_mask = paddle.ones_like(input_ids) return attention_mask def __getattr__(self, name): try: return super().__getattr__(name) except AttributeError as e: try: return getattr(getattr(self, self.base_model_prefix), name) except AttributeError: try: return getattr(self, self.base_model_prefix).config[name] except KeyError: raise e class T5EncoderModel(T5PretrainedModel): base_model_class = None def __init__(self, config: T5Config): super().__init__(config) encoder_config = copy.deepcopy(config) encoder_config.use_cache = False encoder_config.is_encoder_decoder = False self.shared = nn.Embedding(encoder_config.vocab_size, encoder_config.d_model) self.encoder = T5Stack(encoder_config, embed_tokens=self.shared) # Initialize weights and apply final processing self.init_weights() @property def t5(self): return self def get_input_embeddings(self) -> nn.Embedding: return self.shared def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None: self.shared = new_embeddings self.encoder.set_input_embeddings(new_embeddings) def get_encoder(self) -> T5Stack: return self.encoder def forward( self, input_ids: Tensor = None, attention_mask: Optional[Tensor] = None, encoder_hidden_states: Optional[Tuple[Tensor]] = None, encoder_attention_mask: Optional[Tensor] = None, cache=None, inputs_embeds: Optional[Tensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ): return_dict = return_dict if return_dict is not None else self.config.use_return_dict encoder_outputs = self.encoder( input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_attention_mask, cache=cache, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) return encoder_outputs T5EncoderModel.base_model_class = T5EncoderModel