SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[Attention Mask] Refactor all encoder-decoder attention mask (#27086)

Browse Source 
* [FA2 Bart] Add FA2 to all Bart-like

* better

* Refactor attention mask

* remove all customized atteniton logic

* format

* mass rename

* replace _expand_mask

* replace _expand_mask

* mass rename

* add pt files

* mass replace & rename

* mass replace & rename

* mass replace & rename

* mass replace & rename

* Update src/transformers/models/idefics/modeling_idefics.py

* fix more

* clean more

* fix more

* make style

* fix again

* finish

* finish

* finish

* finish

* finish

* finish

* finish

* finish

* finish

* finish

* Apply suggestions from code review

* Apply suggestions from code review

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* small fix mistral

* finish

* finish

* finish

* finish

---------

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
This commit is contained in:
Patrick von Platen
2023-10-27 16:42:01 +02:00
committed by GitHub
parent 29c74f58ae
commit ac5893756b
55 changed files with 647 additions and 2879 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py
View File

@@ -1567,6 +1567,7 @@ from ...utils import (
add_start_docstrings_to_model_forward,
replace_return_docstrings,
)
from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
from ...modeling_outputs import (
BaseModelOutput,
BaseModelOutputWithPastAndCrossAttentions,
@@ -1608,37 +1609,6 @@ def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start
return shifted_input_ids
def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0):
"""
Make causal mask used for bi-directional self-attention.
"""
bsz, tgt_len = input_ids_shape
mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min)
mask_cond = torch.arange(mask.size(-1))
mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
mask = mask.to(dtype)
if past_key_values_length > 0:
mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1)
return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
def _expand_mask(
mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None
):
"""
Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
"""
bsz, src_len = mask.size()
tgt_len = tgt_len if tgt_len is not None else src_len
expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
inverted_mask = 1.0 - expanded_mask
return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
class {{cookiecutter.camelcase_modelname}}LearnedPositionalEmbedding(nn.Embedding):
"""
This module learns positional embeddings up to a fixed maximum size.
@@ -2280,7 +2250,7 @@ class {{cookiecutter.camelcase_modelname}}Encoder({{cookiecutter.camelcase_model
# expand attention_mask
if attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)
encoder_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
@@ -2369,25 +2339,6 @@ class {{cookiecutter.camelcase_modelname}}Decoder({{cookiecutter.camelcase_model
def set_input_embeddings(self, value):
self.embed_tokens = value
# Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
# create causal mask
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
combined_attention_mask = None
if input_shape[-1] > 1:
combined_attention_mask = _make_causal_mask(
input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
).to(self.device)
if attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
combined_attention_mask = (
expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
)
return combined_attention_mask
def forward(
self,
input_ids=None,
@@ -2491,12 +2442,12 @@ class {{cookiecutter.camelcase_modelname}}Decoder({{cookiecutter.camelcase_model
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
attention_mask = self._prepare_decoder_attention_mask(attention_mask, input_shape, inputs_embeds, past_key_values_length)
attention_mask = _prepare_4d_causal_attention_mask(attention_mask, input_shape, inputs_embeds, past_key_values_length)
# expand encoder attention mask
if encoder_hidden_states is not None and encoder_attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
encoder_attention_mask = _prepare_4d_attention_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
# embed positions
positions = self.embed_positions(input_shape, past_key_values_length)