# Copyright 2022 Cerebras Systems.
#
# 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.
import os
import torch.nn as nn
from cerebras.modelzoo.layers.create_initializer import create_initializer
from cerebras.modelzoo.layers.RotaryPositionEmbeddingHelper import (
RotaryPositionEmbeddingHelper,
)
from .AlibiPositionEmbeddingLayer import AlibiPositionEmbeddingLayer
from .FixedPositionEmbeddingLayer import FixedPositionEmbeddingLayer
from .LearnedPositionEmbeddingLayer import LearnedPositionEmbeddingLayer
from .RelativePositionEmbeddingLayer import RelativePositionEmbeddingLayer
[docs]class EmbeddingLayer(nn.Module):
"""
Creates token and, optionally, position and segment embeddings.
:param int vocab_size: Size of input vocabulary.
:param int embedding_size: Dimension of the embedding space.
:param Optional[int] pad_token_id: If specified, the entries at padding_idx
do not contribute to the gradient; therefore,
the embedding vector at padding_idx is not updated during training.
:param int segment_embedding_size: Dimension of the embedding space for segment
embeddings. Useful when factorized embeddings are used for tokens and
so the size of the embedding space for segments differs from that for
tokens. Defaults to the same value as embedding_size.
:param Optional[str,Callable] embeddings_initializer: Token embeddings
initializer. Defaults to 'uniform'.
:param int max_position_embeddings: Maximum sequence length to train
using model.
:param str position_embedding_type: 'learned', 'fixed' or 'rotary'. Defaults to "learned",
for 'rotary' embeddings, embeddings are not created at bottom but computed with key&query embeddings by RotaryPositionEmbeddingHelper
:param int position_embedding_offset: Offset for position embeddings. Default to 0.
:param Optional[int] min_timescale: The scale of the shortest sinusoid. Default to 1.0. (only need to be specified when position_embedding_type is fixed).
:param Optional[int] max_timescale: The scale of the longest sinusoid. Default to 1.0e4. (only need to be specified when position_embedding_type is fixed).
:param Optional[str,Callable] position_embeddings_initializer: Position
embeddings initializer. Defaults to "uniform".
:param Optional[int] num_segments: Number of segments for the segment
embedding layer. Defaults to None, in which case the segment embedding
layer is not created.
:param Optional[str,Callable] segment_embeddings_initializer: Segment
embeddings initializer. Defaults to "uniform".
:param device (optional): Device to create the model parameters on, can be a cuda device or CS device.
"""
[docs] def __init__(
self,
# Word Embedding Parameters
vocab_size,
embedding_size,
pad_token_id=None,
initializer="xavier_uniform",
embeddings_initializer='uniform',
device=None,
# Positional Embedding Parameters:
position_embedding_type="learned",
max_position_embeddings=None,
positional_embedding_size=None,
position_embedding_offset=0,
position_embeddings_initializer='uniform',
# Fixed PE:
min_timescale=1.0,
max_timescale=1.0e4,
# Learned PE:
mask_padding_in_positional_embed=False,
# Relative PE:
num_heads=None,
relative_attention_bias=None,
num_relative_attention_buckets=32,
bidirectional=False,
# Rotary PE:
rotary_dim=None,
rope_theta=10000,
pad_rope=False,
# Alibi PE:
alibi_slopes=None,
alibi_trainable_slopes=False,
# Rotary & Alibi PE:
pos_scaling_factor=1.0,
# Segment Embedding Parameters:
num_segments=None,
segment_embedding_size=None,
segment_embeddings_initializer='uniform',
):
super(EmbeddingLayer, self).__init__()
if segment_embedding_size is None:
segment_embedding_size = embedding_size
if positional_embedding_size is None:
positional_embedding_size = embedding_size
self.embedding_size = embedding_size
self.positional_embedding_size = positional_embedding_size
self.segment_embedding_size = segment_embedding_size
self.embeddings_initializer = embeddings_initializer
self.position_embeddings_initializer = position_embeddings_initializer
self.segment_embeddings_initializer = segment_embeddings_initializer
self.position_embedding_offset = position_embedding_offset
self.max_position_embeddings = max_position_embeddings
self.position_embedding_type = position_embedding_type
self.pad_token_id = pad_token_id
self.mask_padding_in_positional_embed = mask_padding_in_positional_embed
if self.mask_padding_in_positional_embed:
assert (
self.position_embedding_type == "learned"
), "Masking padding tokens in positional embeddings is only supported in learned position embeddings"
self.word_embeddings = nn.Embedding(
vocab_size,
embedding_size,
padding_idx=self.pad_token_id,
device=device,
)
self.position_embeddings = None
self.position_embed_helper = None
if position_embedding_type is not None:
assert position_embedding_offset >= 0, (
f"Position embedding offset should be non-negative, but it is "
f"{position_embedding_offset}."
)
if position_embedding_type != "learned":
assert position_embedding_offset == 0, (
f"For {position_embedding_type} embeddings, position "
f"embedding offset must be 0, but it is "
f"{position_embedding_offset}."
)
if max_position_embeddings is None:
raise ValueError("max_position_embeddings should be specified.")
if (
not os.environ.get("CEREBRAS_ALLOW_LONG_MSL")
) and max_position_embeddings > 32768:
raise ValueError("max_position_embeddings cannot exceed 32768")
if (
pos_scaling_factor != 1.0
and self.position_embedding_type
not in [
"rotary",
"alibi",
]
):
raise ValueError(
"pos_scaling_factor is only supported for ALiBi and RoPE."
)
if position_embedding_type == "learned":
self.position_embeddings = LearnedPositionEmbeddingLayer(
max_position_embeddings=max_position_embeddings,
positional_embedding_size=positional_embedding_size,
pad_token_id=pad_token_id,
position_embedding_offset=position_embedding_offset,
position_embeddings_initializer=position_embeddings_initializer,
mask_padding_in_positional_embed=mask_padding_in_positional_embed,
device=device,
)
elif position_embedding_type == "fixed":
assert (
max_position_embeddings > 1
), "Max position embeddings of length 1 currently unsupported."
self.position_embeddings = FixedPositionEmbeddingLayer(
max_position_embeddings,
positional_embedding_size,
min_timescale,
max_timescale,
)
elif self.position_embedding_type == "rotary":
assert (
rotary_dim is not None
), "EmbeddingLayer requires rotary_dim when using rotary embeddings"
self.position_embed_helper = RotaryPositionEmbeddingHelper(
self.max_position_embeddings,
rotary_dim,
base=rope_theta,
scaling_factor=pos_scaling_factor,
pad_fixed_pos_emb=pad_rope,
)
elif self.position_embedding_type == "relative":
assert (
num_heads is not None
), "EmbeddingLayer requires num_heads when using relative positional embeddings"
self.position_embed_helper = RelativePositionEmbeddingLayer(
num_heads,
relative_attention_bias=relative_attention_bias,
num_relative_attention_buckets=num_relative_attention_buckets,
bidirectional_relative_attention=bidirectional,
relative_attn_bias_initializer=initializer,
)
elif self.position_embedding_type == "alibi":
assert (
num_heads is not None
), "AlibiPositionEmbeddingLayer requires num_heads when using alibi positional embeddings"
self.position_embed_helper = AlibiPositionEmbeddingLayer(
num_heads,
slopes=alibi_slopes,
alibi_trainable_slopes=alibi_trainable_slopes,
slopes_initializer=initializer,
scaling_factor=pos_scaling_factor,
)
else:
raise ValueError(
f"Unknown position embedding type: {position_embedding_type}"
)
self.segment_embeddings = None
if num_segments:
self.segment_embeddings = nn.Embedding(
num_segments,
segment_embedding_size,
device=device,
)
# Initialize weights
self.__reset_parameters()
def reset_parameters(self):
self.__reset_parameters()
if self.position_embeddings:
self.position_embeddings.reset_parameters()
if self.position_embed_helper:
self.position_embed_helper.reset_parameters()
def __reset_parameters(self):
create_initializer(self.embeddings_initializer)(
self.word_embeddings.weight.data
)
if self.segment_embeddings:
create_initializer(self.segment_embeddings_initializer)(
self.segment_embeddings.weight.data
)
if self.pad_token_id:
self.word_embeddings.weight.data[
self.word_embeddings.padding_idx
].zero_()
if self.segment_embeddings:
self.segment_embeddings.weight.data[
self.segment_embeddings.padding_idx
].zero_()
def get_input_embeddings(self):
return self.word_embeddings
def set_input_embeddings(self, new_embeddings):
self.word_embeddings = new_embeddings
[docs] def forward(
self,
input_ids,
position_ids=None,
segment_ids=None,
past_length=0,
):
"""Convert input_ids to token embeddings according to the embedding type.
Word embeddings (required), segment embeddings (optional) and position embeddings (optional).
Args:
input_ids (Tensor): input token ids with shape ``[batch_size, seq_length]``.
position_ids (Tensor): position ids with shape ``[batch_size, seq_length]``.
segment_ids (Tensor): input segment ids with shape ``[batch_size, seq_length]``.
Returns:
Token embedding output with shape ``[batch_size, seq_length, embedding_size]``.
"""
embeddings = self.compute_token_embeddings(input_ids)
if self.position_embeddings is not None:
assert self.embedding_size == self.positional_embedding_size, (
"Cannot use EmbeddingLayer's forward function since the word "
"embedding size and positional embedding size are different. "
"Manually apply the following functions with the appropriate"
"projections in order to align the sizes: compute_token_embeddings, "
"compute_positional_embeddings, and compute_segment_embeddings"
)
embeddings = embeddings + self.compute_positional_embeddings(
input_ids, position_ids, past_length, embeddings.dtype
)
if segment_ids is not None and self.segment_embeddings is not None:
assert self.embedding_size == self.segment_embedding_size, (
"Cannot use EmbeddingLayer's forward function since the word "
"embedding size and segment embedding size are different. "
"Manually apply the following functions with the appropriate"
"projections in order to align the sizes: compute_token_embeddings, "
"compute_positional_embeddings, and compute_segment_embeddings"
)
embeddings = embeddings + self.compute_segment_embeddings(
segment_ids
)
return embeddings
def compute_token_embeddings(self, input_ids):
input_shape = input_ids.size()
input_ids = input_ids.view(-1, input_shape[-1])
embeddings = self.word_embeddings(input_ids)
return embeddings
def compute_positional_embeddings(
self, input_ids, position_ids=None, past_length=0, dtype=None
):
input_shape = input_ids.size()
embed_dtype = (
self.word_embeddings.weight.dtype if dtype is None else dtype
)
if position_ids is not None:
assert (
position_ids.size() == input_shape
), "position_ids must have shape [batch_size, seq_length]"
position_embeddings = None
if self.position_embedding_type == "learned":
position_embeddings = self.position_embeddings(
input_ids,
position_ids=position_ids,
past_length=past_length,
dtype=dtype,
)
elif self.position_embedding_type == "fixed":
position_embeddings = self.position_embeddings(
input_shape[-1], position_ids=position_ids, dtype=embed_dtype
)
return position_embeddings
def compute_segment_embeddings(self, segment_ids):
segment_embeddings = None
if segment_ids is not None and self.segment_embeddings is not None:
segment_embeddings = self.segment_embeddings(segment_ids.int())
return segment_embeddings
def compute_position_bias(self, seq_length, key_length, past_kv=None):
self_attn_position_bias = None
if self.position_embed_helper and isinstance(
self.position_embed_helper,
(
RelativePositionEmbeddingLayer,
AlibiPositionEmbeddingLayer,
),
):
if (not os.environ.get("CEREBRAS_ALLOW_LONG_MSL")) and (
seq_length > 32768 or key_length > 32768
):
raise ValueError(
f"The tokenized data exceeded the 32768 MSL limit. Make "
f"sure that the tokenized data length and "
f"{self.position_embedding_type}'s max_position_embeddings "
f"<= 32768."
)
self_attn_position_bias = self.position_embed_helper(
seq_length, key_length, past_kv=past_kv
)
return self_attn_position_bias
def get_rope_helper(self):
if self.position_embed_helper and isinstance(
self.position_embed_helper, RotaryPositionEmbeddingHelper
):
return self.position_embed_helper
else:
return None