File size: 3,740 Bytes
2e3ebcb 95b4916 2e3ebcb 76fc218 2e3ebcb 76fc218 814cbbb 76fc218 65e9690 76fc218 7c7eafb 76fc218 362ef00 76fc218 2e3ebcb ab85772 2e3ebcb 76fc218 814cbbb 76fc218 7c7eafb 76fc218 362ef00 76fc218 362ef00 2e3ebcb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 |
# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/embedding.py
# Commit id: f1a73d074002226c42ce65a1df170ecff9f022c0
# Copyright (c) 2022, Tri Dao.
import torch
import torch.nn as nn
from einops import rearrange
from torch import Tensor
from transformers.models.xlm_roberta.modeling_xlm_roberta import create_position_ids_from_input_ids
class XLMRobertaEmbeddings(nn.Module):
def __init__(
self,
embed_dim,
vocab_size,
max_position_embeddings,
type_vocab_size,
padding_idx=None,
device=None,
dtype=None,
):
"""
If max_position_embeddings <= 0, there's no position embeddings
If type_vocab_size <= 0, there's no token type embeddings
"""
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.word_embeddings = nn.Embedding(
vocab_size, embed_dim, padding_idx=padding_idx, **factory_kwargs
)
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
if self.max_position_embeddings > 0:
self.position_embeddings = nn.Embedding(
max_position_embeddings, embed_dim, **factory_kwargs
)
if self.type_vocab_size > 0:
self.token_type_embeddings = nn.Embedding(type_vocab_size, embed_dim, **factory_kwargs)
def forward(self, input_ids, position_ids=None, token_type_ids=None, adapter_mask=None):
"""
input_ids: (batch, seqlen)
position_ids: (batch, seqlen)
token_type_ids: (batch, seqlen)
"""
batch_size, seqlen = input_ids.shape
if adapter_mask is not None:
unique_tasks = torch.unique(adapter_mask)
embedding_dtype = next(self.word_embeddings.parameters()).dtype
embeddings = torch.empty(*input_ids.shape, self.word_embeddings.embedding_dim,
dtype=embedding_dtype, device=input_ids.device)
for task_id in unique_tasks:
task_indices = (adapter_mask == task_id).nonzero(as_tuple=True)[0]
task_input_ids = input_ids[task_indices]
task_embeddings = self.word_embeddings(task_input_ids, task_id=task_id)
embeddings[task_indices] = task_embeddings
else:
embeddings = self.word_embeddings(input_ids)
if self.max_position_embeddings > 0:
if position_ids is None:
position_ids = create_position_ids_from_input_ids(input_ids, padding_idx=self.word_embeddings.padding_idx).to(input_ids.device)
# position_ids = torch.arange(seqlen, dtype=torch.long, device=input_ids.device)
position_embeddings = self.position_embeddings(position_ids)
embeddings = embeddings + position_embeddings
if self.type_vocab_size > 0:
if token_type_ids is None:
token_type_ids = torch.zeros(seqlen, dtype=torch.long, device=input_ids.device)
if adapter_mask is not None:
unique_tasks = torch.unique(adapter_mask)
for task_id in unique_tasks:
task_token_type_embeddings = self.token_type_embeddings(token_type_ids, task_id=task_id)
task_indices = (adapter_mask == task_id).nonzero(as_tuple=True)[0]
embeddings[task_indices] = embeddings[task_indices] + task_token_type_embeddings
else:
token_type_embeddings = self.token_type_embeddings(token_type_ids)
embeddings = embeddings + token_type_embeddings
return embeddings
|