GPT-JX-3b / modeling_gptjx.py

Upload 11 files

7f92913 5 months ago

No virus

53.7 kB

	""" PyTorch GPT-JX model."""
	import math
	import os
	import warnings
	from typing import Optional, Tuple, Union

	import torch
	from torch import Tensor, nn
	from torch.cuda.amp import autocast
	from torch.nn import CrossEntropyLoss

	from transformers.activations import ACT2FN
	from transformers.modeling_outputs import (
	BaseModelOutputWithPastAndCrossAttentions,
	CausalLMOutputWithCrossAttentions
	)
	from transformers.modeling_utils import PreTrainedModel
	from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
	from transformers.utils import (
	add_code_sample_docstrings,
	add_start_docstrings,
	add_start_docstrings_to_model_forward,
	logging,
	)
	from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
	from .configuration_gptjx import GPTJXConfig


	logger = logging.get_logger(__name__)

	_CHECKPOINT_FOR_DOC = "alien-ai/gpt-jx-3b"
	_CONFIG_FOR_DOC = "GPTJXConfig"

	class SwiGLUActivation(nn.Module):
	def forward(self, x1: Tensor, x2: Tensor) -> Tensor:
	return x1 * nn.functional.silu(x2)


	class AlibiPositionEmbeddingLayer(nn.Module):
	def __init__(self, num_heads):
	super(AlibiPositionEmbeddingLayer, self).__init__()

	self.num_heads = num_heads
	slopes = torch.tensor(AlibiPositionEmbeddingLayer._get_alibi_slopes(num_heads)).unsqueeze(-1)
	self.slopes = nn.parameter.Parameter(slopes, requires_grad=False)

	def forward(
	self,
	seq_length,
	key_length,
	cached_qk_len,
	):
	context_position = torch.arange(
	cached_qk_len, cached_qk_len + seq_length, device=self.slopes.device
	)[:, None]
	memory_position = torch.arange(
	key_length + cached_qk_len, device=self.slopes.device
	)[None, :]
	relative_position = memory_position - context_position
	relative_position = torch.abs(relative_position).unsqueeze(0).expand(self.num_heads, -1, -1)
	alibi = (self.slopes * -1.0).unsqueeze(1) * relative_position
	return alibi

	@staticmethod
	def _get_alibi_slopes(n):
	def get_slopes_power_of_2(n):
	start = 2 (-(2 -(math.log2(n) - 3)))
	ratio = start
	return [start * ratio**i for i in range(n)]

	if math.log2(n).is_integer():
	return get_slopes_power_of_2(
	n
	) # In the paper, we only train models that have 2^a heads for some a. This function has
	else: # some good properties that only occur when the input is a power of 2. To maintain that even
	closest_power_of_2 = 2 ** math.floor(
	math.log2(n)
	) # when the number of heads is not a power of 2, we use this workaround.
	return (
	get_slopes_power_of_2(closest_power_of_2)
	+ AlibiPositionEmbeddingLayer._get_alibi_slopes(2 * closest_power_of_2)[0::2][: n - closest_power_of_2]
	)


	def load_tf_weights_in_model(model, config, checkpoint_path):
	"""Load tf checkpoints in a pytorch model"""
	try:
	import re

	import tensorflow as tf
	except ImportError:
	logger.error(
	"Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
	"https://www.tensorflow.org/install/ for installation instructions."
	)
	raise
	tf_path = os.path.abspath(checkpoint_path)
	logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
	# Load weights from TF model
	init_vars = tf.train.list_variables(tf_path)
	names = []
	arrays = []
	for name, shape in init_vars:
	logger.info(f"Loading TF weight {name} with shape {shape}")
	array = tf.train.load_variable(tf_path, name)
	names.append(name)
	arrays.append(array.squeeze())

	for name, array in zip(names, arrays):
	name = name[6:] # skip "model/"
	name = name.split("/")
	pointer = model
	for m_name in name:
	if re.fullmatch(r"[A-Za-z]+\d+", m_name):
	scope_names = re.split(r"(\d+)", m_name)
	else:
	scope_names = [m_name]
	if scope_names[0] == "w" or scope_names[0] == "g":
	pointer = getattr(pointer, "weight")
	elif scope_names[0] == "b":
	pointer = getattr(pointer, "bias")
	elif scope_names[0] == "wpe" or scope_names[0] == "wte":
	pointer = getattr(pointer, scope_names[0])
	pointer = getattr(pointer, "weight")
	else:
	pointer = getattr(pointer, scope_names[0])
	if len(scope_names) >= 2:
	num = int(scope_names[1])
	pointer = pointer[num]
	try:
	assert (
	pointer.shape == array.shape
	), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
	except AssertionError as e:
	e.args += (pointer.shape, array.shape)
	raise
	logger.info(f"Initialize PyTorch weight {name}")
	pointer.data = torch.from_numpy(array)
	return model


	class GPTJXAttention(nn.Module):
	def __init__(self, config, is_cross_attention=False, layer_idx=None):
	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
	),
	persistent=False,
	)
	self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)

	self.embed_dim = config.hidden_size
	self.num_heads = config.num_attention_heads
	self.head_dim = self.embed_dim // self.num_heads
	self.split_size = self.embed_dim
	if self.head_dim * self.num_heads != self.embed_dim:
	raise ValueError(
	f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
	f" {self.num_heads})."
	)

	self.scale_attn_weights = config.scale_attn_weights
	self.is_cross_attention = is_cross_attention

	# Layer-wise attention scaling, reordering, and upcasting
	self.scale_attn_by_inverse_layer_idx = config.scale_attn_by_inverse_layer_idx
	self.layer_idx = layer_idx
	self.reorder_and_upcast_attn = config.reorder_and_upcast_attn

	if self.is_cross_attention:
	self.c_attn = Conv1D(2 * self.embed_dim, self.embed_dim)
	self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
	else:
	self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
	self.c_proj = Conv1D(self.embed_dim, self.embed_dim)

	self.attn_dropout = nn.Dropout(config.attn_pdrop)
	self.resid_dropout = nn.Dropout(config.resid_pdrop)

	self.pruned_heads = set()

	self.attn_scale_power = 1.0 if config.mup_scale_qk_dot_by_d else 0.5

	def prune_heads(self, heads):
	if len(heads) == 0:
	return
	heads, index = find_pruneable_heads_and_indices(heads, self.num_heads, self.head_dim, self.pruned_heads)
	index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])

	# Prune conv1d layers
	self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
	self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)

	# Update hyper params
	self.split_size = (self.split_size // self.num_heads) * (self.num_heads - len(heads))
	self.num_heads = self.num_heads - len(heads)
	self.pruned_heads = self.pruned_heads.union(heads)

	def _attn(self, query, key, value, attention_mask=None, head_mask=None, position_bias=None):
	attn_weights = torch.matmul(query, key.transpose(-1, -2))

	if self.scale_attn_weights:
	attn_weights = attn_weights / torch.full(
	[], value.size(-1) ** self.attn_scale_power, dtype=attn_weights.dtype, device=attn_weights.device
	)

	# Layer-wise attention scaling
	if self.scale_attn_by_inverse_layer_idx:
	attn_weights = attn_weights / float(self.layer_idx + 1)

	if not self.is_cross_attention:
	# if only "normal" attention layer implements causal mask
	query_length, key_length = query.size(-2), key.size(-2)
	causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
	mask_value = torch.finfo(attn_weights.dtype).min
	# Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
	# Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
	mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
	attn_weights = torch.where(causal_mask, attn_weights.to(attn_weights.dtype), mask_value)

	if attention_mask is not None:
	# Apply the attention mask
	attn_weights = attn_weights + attention_mask

	if position_bias is not None:
	attn_weights += position_bias.type_as(attn_weights).unsqueeze(0)
	attn_weights = nn.functional.softmax(attn_weights, dim=-1)

	# Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
	attn_weights = attn_weights.type(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 _upcast_and_reordered_attn(self, query, key, value, attention_mask=None, head_mask=None, position_bias=None):
	# Use `torch.baddbmm` (a bit more efficient w/ alpha param for scaling -- from Megatron-LM)
	bsz, num_heads, q_seq_len, dk = query.size()
	_, _, k_seq_len, _ = key.size()

	# Preallocate attn_weights for `baddbmm`
	attn_weights = torch.empty(bsz * num_heads, q_seq_len, k_seq_len, dtype=torch.float32, device=query.device)

	# Compute Scale Factor
	scale_factor = 1.0
	if self.scale_attn_weights:
	scale_factor /= float(value.size(-1)) ** self.attn_scale_power

	if self.scale_attn_by_inverse_layer_idx:
	scale_factor /= float(self.layer_idx + 1)

	# Upcast (turn off autocast) and reorder (Scale K by 1 / root(dk))
	with autocast(enabled=False):
	q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(-1, dk, k_seq_len)
	attn_weights = torch.baddbmm(attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor)
	attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)

	if not self.is_cross_attention:
	# if only "normal" attention layer implements causal mask
	query_length, key_length = query.size(-2), key.size(-2)
	causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
	mask_value = torch.finfo(attn_weights.dtype).min
	# Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
	# Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
	mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
	attn_weights = torch.where(causal_mask, attn_weights, mask_value)

	if attention_mask is not None:
	# Apply the attention mask
	attn_weights = attn_weights + attention_mask

	if position_bias is not None:
	attn_weights += position_bias.type_as(attn_weights).unsqueeze(0)
	attn_weights = nn.functional.softmax(attn_weights, dim=-1)

	# Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
	if attn_weights.dtype != torch.float32:
	raise RuntimeError("Error with upcasting, attn_weights does not have dtype torch.float32")
	attn_weights = attn_weights.type(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 _split_heads(self, tensor, num_heads, attn_head_size):
	"""
	Splits hidden_size dim into attn_head_size and num_heads
	"""
	new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
	tensor = tensor.view(new_shape)
	return tensor.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features)

	def _merge_heads(self, tensor, num_heads, attn_head_size):
	"""
	Merges attn_head_size dim and num_attn_heads dim into hidden_size
	"""
	tensor = tensor.permute(0, 2, 1, 3).contiguous()
	new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
	return tensor.view(new_shape)

	def forward(
	self,
	hidden_states: Optional[Tuple[torch.FloatTensor]],
	layer_past: Optional[Tuple[torch.Tensor]] = None,
	attention_mask: Optional[torch.FloatTensor] = None,
	head_mask: Optional[torch.FloatTensor] = None,
	encoder_hidden_states: Optional[torch.Tensor] = None,
	encoder_attention_mask: Optional[torch.FloatTensor] = None,
	use_cache: Optional[bool] = False,
	output_attentions: Optional[bool] = False,
	position_bias: Optional[torch.FloatTensor] = None,
	) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
	if encoder_hidden_states is not None:
	if not hasattr(self, "q_attn"):
	raise ValueError(
	"If class is used as cross attention, the weights `q_attn` have to be defined. "
	"Please make sure to instantiate class with `GPTJXAttention(..., is_cross_attention=True)`."
	)

	query = self.q_attn(hidden_states)
	key, value = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
	attention_mask = encoder_attention_mask
	else:
	query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)

	query = self._split_heads(query, self.num_heads, self.head_dim)
	key = self._split_heads(key, self.num_heads, self.head_dim)
	value = self._split_heads(value, self.num_heads, self.head_dim)

	if layer_past is not None:
	past_key, past_value = layer_past
	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

	if self.reorder_and_upcast_attn:
	attn_output, attn_weights = self._upcast_and_reordered_attn(
	query, key, value, attention_mask, head_mask, position_bias
	)
	else:
	attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask, position_bias)

	attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
	attn_output = self.c_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 GPTJXMLP(nn.Module):
	def __init__(self, intermediate_size, config):
	super().__init__()
	embed_dim = config.hidden_size
	self.swiglu = config.activation_function == "swiglu"
	self.c_fc = Conv1D(intermediate_size, embed_dim)
	self.c_fc2 = Conv1D(intermediate_size, embed_dim) if self.swiglu else None
	self.c_proj = Conv1D(embed_dim, intermediate_size)
	self.act = SwiGLUActivation() if self.swiglu else ACT2FN[config.activation_function]
	self.dropout = nn.Dropout(config.resid_pdrop)

	def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor:
	if self.swiglu:
	hidden_states2 = self.c_fc2(hidden_states)
	hidden_states = self.c_fc(hidden_states)
	hidden_states = self.act(hidden_states, hidden_states2) if self.swiglu else self.act(hidden_states)
	hidden_states = self.c_proj(hidden_states)
	hidden_states = self.dropout(hidden_states)
	return hidden_states


	class DecoderBlock(nn.Module):
	def __init__(self, config, layer_idx=None):
	super().__init__()
	hidden_size = config.hidden_size
	inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size

	self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
	self.attn = GPTJXAttention(config, layer_idx=layer_idx)
	self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)

	if config.add_cross_attention:
	self.crossattention = GPTJXAttention(config, is_cross_attention=True, layer_idx=layer_idx)
	self.ln_cross_attn = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)

	self.mlp = GPTJXMLP(inner_dim, config)

	def forward(
	self,
	hidden_states: Optional[Tuple[torch.FloatTensor]],
	layer_past: Optional[Tuple[torch.Tensor]] = None,
	attention_mask: Optional[torch.FloatTensor] = None,
	head_mask: Optional[torch.FloatTensor] = None,
	encoder_hidden_states: Optional[torch.Tensor] = None,
	encoder_attention_mask: Optional[torch.FloatTensor] = None,
	use_cache: Optional[bool] = False,
	output_attentions: Optional[bool] = False,
	position_bias: Optional[torch.FloatTensor] = None,
	) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
	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,
	position_bias=position_bias,
	)
	attn_output = attn_outputs[0] # output_attn: a, present, (attentions)
	outputs = attn_outputs[1:]
	# residual connection
	hidden_states = attn_output + residual

	if encoder_hidden_states is not None:
	# add one self-attention block for cross-attention
	if not hasattr(self, "crossattention"):
	raise ValueError(
	f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
	"cross-attention layers by setting `config.add_cross_attention=True`"
	)
	residual = hidden_states
	hidden_states = self.ln_cross_attn(hidden_states)
	cross_attn_outputs = self.crossattention(
	hidden_states,
	attention_mask=attention_mask,
	head_mask=head_mask,
	encoder_hidden_states=encoder_hidden_states,
	encoder_attention_mask=encoder_attention_mask,
	output_attentions=output_attentions,
	position_bias=position_bias,
	)
	attn_output = cross_attn_outputs[0]
	# residual connection
	hidden_states = residual + attn_output
	outputs = outputs + cross_attn_outputs[2:] # add cross attentions if we output attention weights

	residual = hidden_states
	hidden_states = self.ln_2(hidden_states)
	feed_forward_hidden_states = self.mlp(hidden_states)
	# residual connection
	hidden_states = residual + feed_forward_hidden_states

	if use_cache:
	outputs = (hidden_states,) + outputs
	else:
	outputs = (hidden_states,) + outputs[1:]

	return outputs # hidden_states, present, (attentions, cross_attentions)


	class GPTJXPreTrainedModel(PreTrainedModel):
	"""
	An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
	models.
	"""

	config_class = GPTJXConfig
	load_tf_weights = load_tf_weights_in_model
	base_model_prefix = "transformer"
	is_parallelizable = True
	supports_gradient_checkpointing = True
	_no_split_modules = ["DecoderBlock"]
	_skip_keys_device_placement = "past_key_values"

	def __init__(self, inputs, *kwargs):
	super().__init__(inputs, *kwargs)

	def _init_weights(self, module):
	"""Initialize the weights."""
	mup_init_scale = math.sqrt(self.config.mup_width_scale)
	if isinstance(module, (nn.Linear, Conv1D)):
	# Slightly different from the TF version 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 * mup_init_scale))
	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)

	for name, p in module.named_parameters():
	if name == "c_proj.weight":
	# Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
	stddev = self.config.initializer_range * mup_init_scale / math.sqrt(2 * self.config.n_layer)
	p.data.normal_(mean=0.0, std=stddev)

	def _set_gradient_checkpointing(self, module, value=False):
	if isinstance(module, GPTJXModel):
	module.gradient_checkpointing = value

	def get_mup_param_groups(self, lr, weight_decay=0.0, decoupled_wd=True):
	"""
	Returns list of dicts defining parameter groups for muP:
	group 0: most model params get scaled learning rate and weight decay.
	group 1: embedding layer gets non-scaled learning rate and weight decay.
	group 2: normalization layers and biases get non-scaled learning rate only.

	The output can be passed to Adam-base optimizers
	e.g.
	param_groups = model.get_mup_param_groups(lr=1e-3, weight_decay=0.1)
	torch.optim.AdamW(param_groups, betas=(0.9, 0.95), eps=1e-8)
	"""
	norm_modules = (
	torch.nn.LayerNorm,
	torch.nn.BatchNorm1d,
	torch.nn.BatchNorm2d,
	torch.nn.BatchNorm3d,
	torch.nn.InstanceNorm1d,
	torch.nn.InstanceNorm2d,
	torch.nn.InstanceNorm3d,
	torch.nn.GroupNorm,
	torch.nn.SyncBatchNorm,
	torch.nn.LocalResponseNorm,
	)

	def get_group_index(param_name):
	for name, module in self.named_modules():
	if name in param_name:
	if isinstance(module, norm_modules):
	return 2
	elif isinstance(module, torch.nn.Embedding):
	return 1
	return 0

	width_scale = self.config.mup_width_scale
	new_param_groups = []
	new_param_groups.append({"params": [], "lr": lr * width_scale, "weight_decay": weight_decay})
	if not decoupled_wd:
	new_param_groups[0]["weight_decay"] /= width_scale
	new_param_groups.append({"params": [], "lr": lr, "weight_decay": weight_decay})
	new_param_groups.append({"params": [], "lr": lr, "weight_decay": 0.0})

	for name, param in self.named_parameters():
	if not param.requires_grad:
	continue

	if name.endswith("bias"):
	new_param_groups[2]["params"].append(param)
	else:
	new_param_groups[get_group_index(name)]["params"].append(param)

	for idx, param_group in enumerate(new_param_groups):
	if len(param_group["params"]) == 0:
	del new_param_groups[idx]

	return new_param_groups


	START_DOCSTRING = r"""

	This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
	library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
	etc.)

	This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
	Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
	and behavior.

	Parameters:
	config ([`GPTJXConfig`]): 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.
	"""

	INPUTS_DOCSTRING = r"""
	Args:
	input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
	`input_ids_length` = `sequence_length` if `past_key_values` is `None` else
	`past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
	sequence tokens in the vocabulary.

	If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
	`input_ids`.

	Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
	[`PreTrainedTokenizer.__call__`] for details.

	[What are input IDs?](../glossary#input-ids)
	past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.n_layers`):
	Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
	`past_key_values` output below). 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.
	attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, 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.

	If `past_key_values` is used, `attention_mask` needs to contain the masking strategy that was used for
	`past_key_values`. In other words, the `attention_mask` always has to have the length:
	`len(past_key_values) + len(input_ids)`

	[What are attention masks?](../glossary#attention-mask)
	token_type_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, 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#token-type-ids)
	position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, optional):
	Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
	config.max_position_embeddings - 1]`.

	[What are position IDs?](../glossary#position-ids)
	head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_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 (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, optional):
	Optionally, instead of passing `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.

	If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
	`past_key_values`).
	use_cache (`bool`, optional):
	If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
	`past_key_values`).
	output_attentions (`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 (`bool`, optional):
	Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
	more detail.
	return_dict (`bool`, optional):
	Whether or not to return a [`~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 (`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-JX model have the
	following number of attention modules:
	-gpt-jx-3b: 32
	"""
	DEPARALLELIZE_DOCSTRING = r"""
	Moves the model to cpu from a model parallel state.

	Example:

	```python
	# On a 4 GPU machine with gptjx:
	model = GPTJXForCausalLM.from_pretrained("alien-ai/gpt-jx-3b")
	device_map = {
	0: [0, 1, 2, 3, 4, 5, 6, 7],
	1: [8, 9, 10, 11, 12, 13, 14, 15],
	2: [16, 17, 18, 19, 20, 21, 22, 23],
	3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35],
	}
	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()
	```
	"""


	@add_start_docstrings(
	"The bare GPT-JX Model transformer outputting raw hidden-states without any specific head on top.",
	START_DOCSTRING,
	)
	class GPTJXModel(GPTJXPreTrainedModel):
	_keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
	_keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias"]

	def __init__(self, config):
	super().__init__(config)

	self.embed_dim = config.hidden_size

	self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
	self.wpe = (
	nn.Embedding(config.max_position_embeddings, self.embed_dim)
	if config.position_embedding_type != "alibi"
	else None
	)
	self.embeddings_scale = config.mup_embeddings_scale

	self.drop = nn.Dropout(config.embd_pdrop)
	self.h = nn.ModuleList([DecoderBlock(config, layer_idx=i) for i in range(config.num_hidden_layers)])
	self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

	self.relative_pe = (
	AlibiPositionEmbeddingLayer(config.num_attention_heads)
	if config.position_embedding_type == "alibi"
	else None
	)

	# Model parallel
	self.model_parallel = False
	self.device_map = None
	self.gradient_checkpointing = False

	# Initialize weights and apply final processing
	self.post_init()

	@add_start_docstrings(PARALLELIZE_DOCSTRING)
	def parallelize(self, device_map=None):
	# Check validity of device_map
	warnings.warn(
	"`GPTJXModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your"
	" model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
	" `device_map` but it needs to be a dictionary module_name to device, so for instance {'h.0': 0, 'h.1': 1,"
	" ...}",
	FutureWarning,
	)
	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)
	if self.wpe is not None:
	self.wpe = self.wpe.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):
	warnings.warn(
	"Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
	FutureWarning,
	)
	self.model_parallel = False
	self.device_map = None
	self.first_device = "cpu"
	self.last_device = "cpu"
	self.wte = self.wte.to("cpu")
	if self.wpe is not None:
	self.wpe = self.wpe.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

	def _prune_heads(self, heads_to_prune):
	"""
	Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
	"""
	for layer, heads in heads_to_prune.items():
	self.h[layer].attn.prune_heads(heads)

	@add_start_docstrings_to_model_forward(INPUTS_DOCSTRING)
	@add_code_sample_docstrings(
	checkpoint=_CHECKPOINT_FOR_DOC,
	output_type=BaseModelOutputWithPastAndCrossAttentions,
	config_class=_CONFIG_FOR_DOC,
	)
	def forward(
	self,
	input_ids: Optional[torch.LongTensor] = None,
	past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
	attention_mask: Optional[torch.FloatTensor] = None,
	token_type_ids: Optional[torch.LongTensor] = None,
	position_ids: Optional[torch.LongTensor] = None,
	head_mask: Optional[torch.FloatTensor] = None,
	inputs_embeds: Optional[torch.FloatTensor] = None,
	encoder_hidden_states: Optional[torch.Tensor] = None,
	encoder_attention_mask: Optional[torch.FloatTensor] = None,
	use_cache: Optional[bool] = None,
	output_attentions: Optional[bool] = None,
	output_hidden_states: Optional[bool] = None,
	return_dict: Optional[bool] = None,
	) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
	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])

	# GPTJXAttention mask.
	if attention_mask is not None:
	if batch_size <= 0:
	raise ValueError("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 the dtype's smallest value 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) * torch.finfo(self.dtype).min

	# 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.config.add_cross_attention and encoder_hidden_states is not None:
	encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
	encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
	if encoder_attention_mask is None:
	encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
	encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
	else:
	encoder_attention_mask = None

	# Prepare head mask if needed
	# 1.0 in head_mask indicate we keep the head
	# attention_probs has shape bsz x n_heads x N x N
	# head_mask has shape n_layer x batch x n_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)
	if self.wpe is not None:
	position_embeds = self.wpe(position_ids)
	hidden_states = inputs_embeds + position_embeds
	else:
	hidden_states = inputs_embeds
	hidden_states *= torch.tensor(
	float(self.embeddings_scale), dtype=hidden_states.dtype, device=hidden_states.device
	)

	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)

	if self.relative_pe is not None:
	length = input_ids.shape[1]
	cached_kv_length = 0
	cached_kv = past_key_values[0]
	if cached_kv is not None:
	cached_kv_length = cached_kv[0].shape[-2]
	position_bias = self.relative_pe(length, length, cached_kv_length)
	else:
	position_bias = None

	output_shape = input_shape + (hidden_states.size(-1),)

	if self.gradient_checkpointing and self.training:
	if use_cache:
	logger.warning_once(
	"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
	)
	use_cache = False

	presents = () if use_cache else None
	all_self_attentions = () if output_attentions else None
	all_cross_attentions = () if output_attentions and self.config.add_cross_attention 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:

	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],
	encoder_hidden_states,
	encoder_attention_mask,
	)
	else:
	outputs = block(
	hidden_states,
	layer_past=layer_past,
	attention_mask=attention_mask,
	head_mask=head_mask[i],
	encoder_hidden_states=encoder_hidden_states,
	encoder_attention_mask=encoder_attention_mask,
	use_cache=use_cache,
	output_attentions=output_attentions,
	position_bias=position_bias,
	)

	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],)
	if self.config.add_cross_attention:
	all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)

	# 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, all_cross_attentions]
	if v is not None
	)

	return BaseModelOutputWithPastAndCrossAttentions(
	last_hidden_state=hidden_states,
	past_key_values=presents,
	hidden_states=all_hidden_states,
	attentions=all_self_attentions,
	cross_attentions=all_cross_attentions,
	)


	@add_start_docstrings(
	"""
	The GPTJX Model transformer with a language modeling head on top (linear layer with weights tied to the input
	embeddings).
	""",
	START_DOCSTRING,
	)
	class GPTJXForCausalLM(GPTJXPreTrainedModel):
	_keys_to_ignore_on_load_missing = [r"lm_head.weight"]
	_keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias"]

	def __init__(self, config):
	super().__init__(config)
	self.transformer = GPTJXModel(config)
	self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
	self.output_logits_scale = config.mup_output_alpha * config.mup_width_scale

	# Model parallel
	self.model_parallel = False
	self.device_map = None

	# Initialize weights and apply final processing
	self.post_init()

	@add_start_docstrings(PARALLELIZE_DOCSTRING)
	def parallelize(self, device_map=None):
	warnings.warn(
	"`GPTJXForCausalLM.parallelize` is deprecated and will be removed in v5 of Transformers, you should load"
	" your model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
	" `device_map` but it needs to be a dictionary module_name to device, so for instance {'transformer.h.0':"
	" 0, 'transformer.h.1': 1, ...}",
	FutureWarning,
	)
	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):
	warnings.warn(
	"Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
	FutureWarning,
	)
	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 self.lm_head

	def set_output_embeddings(self, new_embeddings):
	self.lm_head = new_embeddings

	def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=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_key_values:
	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_key_values:
	position_ids = position_ids[:, -1].unsqueeze(-1)
	else:
	position_ids = None

	# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
	if inputs_embeds is not None and past_key_values is None:
	model_inputs = {"inputs_embeds": inputs_embeds}
	else:
	model_inputs = {"input_ids": input_ids}

	model_inputs.update(
	{
	"past_key_values": past_key_values,
	"use_cache": kwargs.get("use_cache"),
	"position_ids": position_ids,
	"attention_mask": attention_mask,
	"token_type_ids": token_type_ids,
	}
	)
	return model_inputs

	@add_start_docstrings_to_model_forward(INPUTS_DOCSTRING)
	@add_code_sample_docstrings(
	checkpoint=_CHECKPOINT_FOR_DOC,
	output_type=CausalLMOutputWithCrossAttentions,
	config_class=_CONFIG_FOR_DOC,
	)
	def forward(
	self,
	input_ids: Optional[torch.LongTensor] = None,
	past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
	attention_mask: Optional[torch.FloatTensor] = None,
	token_type_ids: Optional[torch.LongTensor] = None,
	position_ids: Optional[torch.LongTensor] = None,
	head_mask: Optional[torch.FloatTensor] = None,
	inputs_embeds: Optional[torch.FloatTensor] = None,
	encoder_hidden_states: Optional[torch.Tensor] = None,
	encoder_attention_mask: Optional[torch.FloatTensor] = None,
	labels: Optional[torch.LongTensor] = None,
	use_cache: Optional[bool] = None,
	output_attentions: Optional[bool] = None,
	output_hidden_states: Optional[bool] = None,
	return_dict: Optional[bool] = None,
	) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
	r"""
	labels (`torch.LongTensor` of shape `(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,
	encoder_hidden_states=encoder_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,
	)
	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)

	lm_logits = self.lm_head(hidden_states)
	lm_logits *= torch.tensor(float(self.output_logits_scale), dtype=lm_logits.dtype, device=lm_logits.device)

	loss = None
	if labels is not None:
	# move labels to correct device to enable model parallelism
	labels = labels.to(lm_logits.device)
	# 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))

	if not return_dict:
	output = (lm_logits,) + transformer_outputs[1:]
	return ((loss,) + output) if loss is not None else output

	return CausalLMOutputWithCrossAttentions(
	loss=loss,
	logits=lm_logits,
	past_key_values=transformer_outputs.past_key_values,
	hidden_states=transformer_outputs.hidden_states,
	attentions=transformer_outputs.attentions,
	cross_attentions=transformer_outputs.cross_attentions,
	)

	@staticmethod
	def _reorder_cache(
	past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
	) -> Tuple[Tuple[torch.Tensor]]:
	"""
	This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
	[`~PreTrainedModel.beam_sample`] is called. This is required to match `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_key_values
	)