# coding=utf-8
# Copyright 2021 The EleutherAI and HuggingFace Teams. All rights reserved.
#
# 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.
""" PyTorch GPT-J model. """
from typing import Tuple
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss, MSELoss
from ...activations import ACT2FN
from ...file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
from ...modeling_utils import PreTrainedModel
from ...utils import logging
from ...utils.model_parallel_utils import assert_device_map, get_device_map
from .configuration_gptj import GPTJConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "EleutherAI/gpt-j-6B"
_CONFIG_FOR_DOC = "GPTJConfig"
_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
GPTJ_PRETRAINED_MODEL_ARCHIVE_LIST = [
"EleutherAI/gpt-j-6B",
# See all GPT-J models at https://huggingface.co/models?filter=gptj
]
def fixed_pos_embedding(x, seq_dim=1, seq_len=None):
dim = x.shape[-1]
if seq_len is None:
seq_len = x.shape[seq_dim]
inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2) / dim))
sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(seq_len), inv_freq).to(x.device).float()
return torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)
def rotate_every_two(x):
x1 = x[:, :, :, ::2]
x2 = x[:, :, :, 1::2]
x = torch.stack((-x2, x1), axis=-1)
return x.flatten(-2) # in einsum notation: rearrange(x, '... d j -> ... (d j)')
def apply_rotary_pos_emb(x, sincos, offset=0):
sin, cos = map(lambda t: t[None, offset : x.shape[1] + offset, None, :].repeat_interleave(2, 3), sincos)
# einsum notation for lambda t: repeat(t[offset:x.shape[1]+offset,:], "n d -> () n () (d j)", j=2)
return (x * cos) + (rotate_every_two(x) * sin)
class GPTJAttention(nn.Module):
def __init__(self, config):
super().__init__()
max_positions = config.max_position_embeddings
self.register_buffer(
"bias",
torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
1, 1, max_positions, max_positions
),
)
self.register_buffer("masked_bias", torch.tensor(-1e9))
self.attn_dropout = nn.Dropout(config.attn_pdrop)
self.resid_dropout = nn.Dropout(config.resid_pdrop)
self.embed_dim = config.hidden_size
self.num_attention_heads = config.num_attention_heads
self.head_dim = self.embed_dim // self.num_attention_heads
if self.head_dim * self.num_attention_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and `num_attention_heads`: {self.num_attention_heads})."
)
self.scale_attn = torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype())
self.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
self.rotary_dim = None
if config.rotary_dim is not None:
self.rotary_dim = config.rotary_dim
def _split_heads(self, tensor, num_attention_heads, attn_head_size, rotary):
"""
Splits n_ctx dim into attn_head_size and num_attention_heads
"""
new_shape = tensor.size()[:-1] + (num_attention_heads, attn_head_size)
tensor = tensor.view(*new_shape)
if rotary:
return tensor
if len(tensor.shape) == 5:
return tensor.permute(0, 1, 3, 2, 4) # (batch, blocks, head, block_length, head_features)
elif len(tensor.shape) == 4:
return tensor.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features)
else:
raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")
def _merge_heads(self, tensor, num_attention_heads, attn_head_size):
"""
Merges attn_head_size dim and num_attn_heads dim into n_ctx
"""
if len(tensor.shape) == 5:
tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()
elif len(tensor.shape) == 4:
tensor = tensor.permute(0, 2, 1, 3).contiguous()
else:
raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")
new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)
return tensor.view(new_shape)
def _attn(
self,
query,
key,
value,
attention_mask=None,
head_mask=None,
):
# compute causal mask from causal mask buffer
query_length, key_length = query.size(-2), key.size(-2)
causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
# Keep the attention weights computation in fp32 to avoid overflow issues
query = query.to(torch.float32)
key = key.to(torch.float32)
attn_weights = torch.matmul(query, key.transpose(-1, -2))
attn_weights = torch.where(causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype))
attn_weights = attn_weights / self.scale_attn
if attention_mask is not None:
# Apply the attention mask
attn_weights = attn_weights + attention_mask
attn_weights = nn.Softmax(dim=-1)(attn_weights)
attn_weights = attn_weights.to(value.dtype)
attn_weights = self.attn_dropout(attn_weights)
# Mask heads if we want to
if head_mask is not None:
attn_weights = attn_weights * head_mask
attn_output = torch.matmul(attn_weights, value)
return attn_output, attn_weights
def forward(
self,
hidden_states,
attention_mask=None,
layer_past=None,
head_mask=None,
use_cache=False,
output_attentions=False,
):
query = self.q_proj(hidden_states)
key = self.k_proj(hidden_states)
value = self.v_proj(hidden_states)
query = self._split_heads(query, self.num_attention_heads, self.head_dim, True)
key = self._split_heads(key, self.num_attention_heads, self.head_dim, True)
value = self._split_heads(value, self.num_attention_heads, self.head_dim, False)
seq_len = key.shape[1]
offset = 0
if layer_past is not None:
offset = layer_past[0].shape[-2]
seq_len += offset
if self.rotary_dim is not None:
k_rot = key[:, :, :, : self.rotary_dim]
k_pass = key[:, :, :, self.rotary_dim :]
q_rot = query[:, :, :, : self.rotary_dim]
q_pass = query[:, :, :, self.rotary_dim :]
sincos = fixed_pos_embedding(k_rot, 1, seq_len=seq_len)
k_rot = apply_rotary_pos_emb(k_rot, sincos, offset=offset)
q_rot = apply_rotary_pos_emb(q_rot, sincos, offset=offset)
key = torch.cat([k_rot, k_pass], dim=-1)
query = torch.cat([q_rot, q_pass], dim=-1)
else:
sincos = fixed_pos_embedding(key, 1, seq_len=seq_len)
key = apply_rotary_pos_emb(key, sincos, offset=offset)
query = apply_rotary_pos_emb(query, sincos, offset=offset)
key = key.permute(0, 2, 1, 3)
query = query.permute(0, 2, 1, 3)
if layer_past is not None:
past_key = layer_past[0]
past_value = layer_past[1]
key = torch.cat((past_key, key), dim=-2)
value = torch.cat((past_value, value), dim=-2)
if use_cache is True:
present = (key, value)
else:
present = None
# compute self-attention: V x Softmax(QK^T)
attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)
attn_output = self.out_proj(attn_output)
attn_output = self.resid_dropout(attn_output)
outputs = (attn_output, present)
if output_attentions:
outputs += (attn_weights,)
return outputs # a, present, (attentions)
class GPTJMLP(nn.Module):
def __init__(self, intermediate_size, config): # in MLP: intermediate_size= 4 * embed_dim
super().__init__()
embed_dim = config.n_embd
self.fc_in = nn.Linear(embed_dim, intermediate_size)
self.fc_out = nn.Linear(intermediate_size, embed_dim)
self.act = ACT2FN[config.activation_function]
self.dropout = nn.Dropout(config.resid_pdrop)
def forward(self, hidden_states):
hidden_states = self.fc_in(hidden_states)
hidden_states = self.act(hidden_states)
hidden_states = self.fc_out(hidden_states)
hidden_states = self.dropout(hidden_states)
return hidden_states
class GPTJBlock(nn.Module):
def __init__(self, config):
super().__init__()
inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd
self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
self.attn = GPTJAttention(config)
self.mlp = GPTJMLP(inner_dim, config)
def forward(
self,
hidden_states,
layer_past=None,
attention_mask=None,
head_mask=None,
use_cache=False,
output_attentions=False,
):
residual = hidden_states
hidden_states = self.ln_1(hidden_states)
attn_outputs = self.attn(
hidden_states,
layer_past=layer_past,
attention_mask=attention_mask,
head_mask=head_mask,
use_cache=use_cache,
output_attentions=output_attentions,
)
attn_output = attn_outputs[0] # output_attn: a, present, (attentions)
outputs = attn_outputs[1:]
feed_forward_hidden_states = self.mlp(hidden_states)
hidden_states = attn_output + feed_forward_hidden_states + residual
if use_cache:
outputs = (hidden_states,) + outputs
else:
outputs = (hidden_states,) + outputs[1:]
return outputs # hidden_states, present, (attentions)
class GPTJPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = GPTJConfig
base_model_prefix = "transformer"
is_parallelizable = True
supports_gradient_checkpointing = True
def __init__(self, *inputs, **kwargs):
super().__init__(*inputs, **kwargs)
def _init_weights(self, module):
"""Initialize the weights."""
if isinstance(module, (nn.Linear,)):
# Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization
# cf https://github.com/pytorch/pytorch/pull/5617
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
def _set_gradient_checkpointing(self, module, value=False):
if isinstance(module, GPTJModel):
module.gradient_checkpointing = value
GPTJ_START_DOCSTRING = r"""
This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`_ sub-class. Use
it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
behavior.
Parameters:
config (:class:`~transformers.GPTJConfig`): 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 :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
weights.
"""
GPTJ_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using :class:`transformers.GPTJTokenizer`. See
:func:`transformers.PreTrainedTokenizer.encode` and :func:`transformers.PreTrainedTokenizer.__call__` for
details.
`What are input IDs? <../glossary.html#input-ids>`__
attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({0})`, `optional`):
Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
`What are attention masks? <../glossary.html#attention-mask>`__
token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
1]``:
- 0 corresponds to a `sentence A` token,
- 1 corresponds to a `sentence B` token.
`What are token type IDs? <../glossary.html#token-type-ids>`_
position_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
config.n_positions - 1]``.
`What are position IDs? <../glossary.html#position-ids>`_
head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_attention_heads,)` or :obj:`(n_layer, num_attention_heads)`, `optional`):
Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`({0}, n_ctx)`, `optional`):
Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
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.
output_attentions (:obj:`bool`, `optional`):
Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
tensors for more detail.
output_hidden_states (:obj:`bool`, `optional`):
Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
more detail.
return_dict (:obj:`bool`, `optional`):
Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""
PARALLELIZE_DOCSTRING = r"""
This is an experimental feature and is a subject to change at a moment's notice. Uses a device map to distribute
attention modules of the model across several devices. If no device map is given, it will evenly distribute blocks
across all devices.
Args:
device_map (:obj:`Dict[int, list]`, optional, defaults to None):
A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
automatically mapped to the first device (for esoteric reasons). That means that the first device should
have fewer attention modules mapped to it than other devices. For reference, the GPT-J models have the
following number of attention modules:
- gpt-j-6B: 28
Example::
# Here is an example of a device map on a machine with 4 GPUs using gpt-j-6B, which has a total of 28 attention modules:
model = GPTJForCausalLM.from_pretrained('EleutherAI/gpt-j-6B')
device_map = {0: [0, 1, 2, 3, 4, 5, 6],
1: [7, 8, 9, 10, 11, 12, 13],
2: [14, 15, 16, 17, 18, 19, 20],
3: [21, 22, 23, 24, 25, 26, 27]}
model.parallelize(device_map)
"""
DEPARALLELIZE_DOCSTRING = r"""
Moves the model to CPU from a model parallel state.
Example::
# On a 4 GPU machine with gpt-j-6B:
model = GPTJForCausalLM.from_pretrained('EleutherAI/gpt-j-6B')
device_map = {0: [0, 1, 2, 3, 4, 5, 6],
1: [7, 8, 9, 10, 11, 12, 13],
2: [14, 15, 16, 17, 18, 19, 20],
3: [21, 22, 23, 24, 25, 26, 27]}
model.parallelize(device_map) # Splits the model across several devices
model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
"""
[docs]@add_start_docstrings(
"The bare GPT-J Model transformer outputting raw hidden-states without any specific head on top.",
GPTJ_START_DOCSTRING,
)
class GPTJModel(GPTJPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.embed_dim = config.n_embd
self.vocab_size = config.vocab_size
self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
self.drop = nn.Dropout(config.embd_pdrop)
self.h = nn.ModuleList([GPTJBlock(config) for _ in range(config.n_layer)])
self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
self.rotary_dim = min(config.rotary_dim, config.n_ctx // config.num_attention_heads)
self.init_weights()
# Model parallel
self.model_parallel = False
self.device_map = None
self.gradient_checkpointing = False
@add_start_docstrings(PARALLELIZE_DOCSTRING)
def parallelize(self, device_map=None):
# Check validity of device_map
self.device_map = (
get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
)
assert_device_map(self.device_map, len(self.h))
self.model_parallel = True
self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
self.last_device = "cuda:" + str(max(self.device_map.keys()))
self.wte = self.wte.to(self.first_device)
# Load onto devices
for k, v in self.device_map.items():
for block in v:
cuda_device = "cuda:" + str(k)
self.h[block] = self.h[block].to(cuda_device)
# ln_f to last
self.ln_f = self.ln_f.to(self.last_device)
@add_start_docstrings(DEPARALLELIZE_DOCSTRING)
def deparallelize(self):
self.model_parallel = False
self.device_map = None
self.first_device = "cpu"
self.last_device = "cpu"
self.wte = self.wte.to("cpu")
for index in range(len(self.h)):
self.h[index] = self.h[index].to("cpu")
self.ln_f = self.ln_f.to("cpu")
torch.cuda.empty_cache()
def get_input_embeddings(self):
return self.wte
def set_input_embeddings(self, new_embeddings):
self.wte = new_embeddings
[docs] @add_start_docstrings_to_model_forward(GPTJ_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=BaseModelOutputWithPast,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
past_key_values=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
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
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
input_shape = input_ids.size()
input_ids = input_ids.view(-1, input_shape[-1])
batch_size = input_ids.shape[0]
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
batch_size = inputs_embeds.shape[0]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
device = input_ids.device if input_ids is not None else inputs_embeds.device
if token_type_ids is not None:
token_type_ids = token_type_ids.view(-1, input_shape[-1])
if position_ids is not None:
position_ids = position_ids.view(-1, input_shape[-1])
if past_key_values is None:
past_length = 0
past_key_values = tuple([None] * len(self.h))
else:
past_length = past_key_values[0][0].size(-2)
if position_ids is None:
position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
# Attention mask.
if attention_mask is not None:
assert batch_size > 0, "batch_size has to be defined and > 0"
attention_mask = attention_mask.view(batch_size, -1)
# We create a 3D attention mask from a 2D tensor mask.
# Sizes are [batch_size, 1, 1, to_seq_length]
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
# this attention mask is more simple than the triangular masking of causal attention
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
attention_mask = attention_mask[:, None, None, :]
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility
attention_mask = (1.0 - attention_mask) * -10000.0
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape bsz x num_attention_heads x N x N
# head_mask has shape n_layer x batch x num_attention_heads x N x N
head_mask = self.get_head_mask(head_mask, self.config.n_layer)
if inputs_embeds is None:
inputs_embeds = self.wte(input_ids)
hidden_states = inputs_embeds
if token_type_ids is not None:
token_type_embeds = self.wte(token_type_ids)
hidden_states = hidden_states + token_type_embeds
hidden_states = self.drop(hidden_states)
output_shape = input_shape + (hidden_states.size(-1),)
presents = () if use_cache else None
all_self_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
# Model parallel
if self.model_parallel:
torch.cuda.set_device(hidden_states.device)
# Ensure layer_past is on same device as hidden_states (might not be correct)
if layer_past is not None:
layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
# Ensure that attention_mask is always on the same device as hidden_states
if attention_mask is not None:
attention_mask = attention_mask.to(hidden_states.device)
if isinstance(head_mask, torch.Tensor):
head_mask = head_mask.to(hidden_states.device)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if self.gradient_checkpointing and self.training:
if use_cache:
logger.warning(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
)
use_cache = False
def create_custom_forward(module):
def custom_forward(*inputs):
# None for past_key_value
return module(*inputs, use_cache, output_attentions)
return custom_forward
outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(block),
hidden_states,
None,
attention_mask,
head_mask[i],
)
else:
outputs = block(
hidden_states,
layer_past=layer_past,
attention_mask=attention_mask,
head_mask=head_mask[i],
use_cache=use_cache,
output_attentions=output_attentions,
)
hidden_states = outputs[0]
if use_cache is True:
presents = presents + (outputs[1],)
if output_attentions:
all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
# Model Parallel: If it's the last layer for that device, put things on the next device
if self.model_parallel:
for k, v in self.device_map.items():
if i == v[-1] and "cuda:" + str(k) != self.last_device:
hidden_states = hidden_states.to("cuda:" + str(k + 1))
hidden_states = self.ln_f(hidden_states)
hidden_states = hidden_states.view(*output_shape)
# Add last hidden state
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
return BaseModelOutputWithPast(
last_hidden_state=hidden_states,
past_key_values=presents,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
)
[docs]@add_start_docstrings(
"""
The GPT-J Model transformer with a language modeling head on top (linear layer with weights tied to the input
embeddings).
""",
GPTJ_START_DOCSTRING,
)
class GPTJForCausalLM(GPTJPreTrainedModel):
_keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias", r"lm_head\.weight"]
def __init__(self, config):
super().__init__(config)
self.transformer = GPTJModel(config)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
self.init_weights()
# Model parallel
self.model_parallel = False
self.device_map = None
@add_start_docstrings(PARALLELIZE_DOCSTRING)
def parallelize(self, device_map=None):
self.device_map = (
get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
if device_map is None
else device_map
)
assert_device_map(self.device_map, len(self.transformer.h))
self.transformer.parallelize(self.device_map)
self.lm_head = self.lm_head.to(self.transformer.first_device)
self.model_parallel = True
@add_start_docstrings(DEPARALLELIZE_DOCSTRING)
def deparallelize(self):
self.transformer.deparallelize()
self.transformer = self.transformer.to("cpu")
self.lm_head = self.lm_head.to("cpu")
self.model_parallel = False
torch.cuda.empty_cache()
def get_output_embeddings(self):
return None
def set_output_embeddings(self, new_embeddings):
return
def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
token_type_ids = kwargs.get("token_type_ids", None)
# only last token for inputs_ids if past is defined in kwargs
if past:
input_ids = input_ids[:, -1].unsqueeze(-1)
if token_type_ids is not None:
token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
attention_mask = kwargs.get("attention_mask", None)
position_ids = kwargs.get("position_ids", None)
if attention_mask is not None and position_ids is None:
# create position_ids on the fly for batch generation
position_ids = attention_mask.long().cumsum(-1) - 1
position_ids.masked_fill_(attention_mask == 0, 1)
if past:
position_ids = position_ids[:, -1].unsqueeze(-1)
else:
position_ids = None
return {
"input_ids": input_ids,
"past_key_values": past,
"use_cache": kwargs.get("use_cache"),
"position_ids": position_ids,
"attention_mask": attention_mask,
"token_type_ids": token_type_ids,
}
[docs] @add_start_docstrings_to_model_forward(GPTJ_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=CausalLMOutputWithPast,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
past_key_values=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `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, ..., config.vocab_size]`` All labels set to
``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
transformer_outputs = self.transformer(
input_ids,
past_key_values=past_key_values,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
hidden_states = transformer_outputs[0]
# Set device for model parallelism
if self.model_parallel:
torch.cuda.set_device(self.transformer.first_device)
hidden_states = hidden_states.to(self.lm_head.weight.device)
# make sure sampling in fp16 works correctly and
# compute loss in fp32 to match with mesh-tf version
# https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179
lm_logits = self.lm_head(hidden_states).to(torch.float32)
loss = None
if labels is not None:
# Shift so that tokens < n predict n
shift_logits = lm_logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten the tokens
loss_fct = CrossEntropyLoss()
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
loss = loss.to(hidden_states.dtype)
if not return_dict:
output = (lm_logits,) + transformer_outputs[1:]
return ((loss,) + output) if loss is not None else output
return CausalLMOutputWithPast(
loss=loss,
logits=lm_logits,
past_key_values=transformer_outputs.past_key_values,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
)
@staticmethod
def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
"""
This function is used to re-order the :obj:`past_key_values` cache if
:meth:`~transformers.PretrainedModel.beam_search` or :meth:`~transformers.PretrainedModel.beam_sample` is
called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
"""
return tuple(
tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
for layer_past in past
)
[docs]@add_start_docstrings(
"""
The GPT-J Model transformer with a sequence classification head on top (linear layer).
:class:`~transformers.GPTJForSequenceClassification` uses the last token in order to do the classification, as
other causal models (e.g. GPT, GPT-2, GPT-Neo) do.
Since it does classification on the last token, it requires to know the position of the last token. If a
:obj:`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each
row. If no :obj:`pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot
guess the padding tokens when :obj:`inputs_embeds` are passed instead of :obj:`input_ids`, it does the same (take
the last value in each row of the batch).
""",
GPTJ_START_DOCSTRING,
)
class GPTJForSequenceClassification(GPTJPreTrainedModel):
_keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head\.weight"]
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.transformer = GPTJModel(config)
self.score = nn.Linear(config.n_ctx, self.num_labels, bias=False)
self.init_weights()
# Model parallel
self.model_parallel = False
self.device_map = None
[docs] @add_start_docstrings_to_model_forward(GPTJ_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=SequenceClassifierOutputWithPast,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
past_key_values=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
transformer_outputs = self.transformer(
input_ids,
past_key_values=past_key_values,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
hidden_states = transformer_outputs[0]
logits = self.score(hidden_states)
if input_ids is not None:
batch_size = input_ids.shape[0]
else:
batch_size = inputs_embeds.shape[0]
assert (
self.config.pad_token_id is not None or batch_size == 1
), "Cannot handle batch sizes > 1 if no padding token is defined."
if self.config.pad_token_id is None:
sequence_lengths = -1
else:
if input_ids is not None:
sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
else:
sequence_lengths = -1
logger.warning(
f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
f"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
)
pooled_logits = logits[range(batch_size), sequence_lengths]
loss = None
if labels is not None:
if self.num_labels == 1:
# We are doing regression
loss_fct = MSELoss()
loss = loss_fct(pooled_logits.view(-1), labels.to(self.dtype).view(-1))
else:
loss_fct = CrossEntropyLoss()
loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
if not return_dict:
output = (pooled_logits,) + transformer_outputs[1:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutputWithPast(
loss=loss,
logits=pooled_logits,
past_key_values=transformer_outputs.past_key_values,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
)