Skip to content

Add accelerate support for LongT5 models #20341

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
sgugger merged 4 commits into from
Dec 12, 2022
Merged

Add accelerate support for LongT5 models #20341

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
1269fb7
✨ add accelerate support for LongT5 models
pszemraj Nov 20, 2022
a3e8eb4
fix `accelerate` tests
younesbelkada Dec 8, 2022
b83bb25
Trigger CI test
younesbelkada Dec 11, 2022
ec1f0dc
Merge remote-tracking branch 'upstream/main' into HEAD
younesbelkada Dec 11, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 14 additions & 5 deletions src/transformers/models/longt5/modeling_longt5.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -648,9 +648,12 @@ def _relative_position_bucket(relative_position, bidirectional=True, num_buckets

def compute_bias(self, block_length: int):
"""Compute binned relative position bias"""
memory_position = torch.arange(
3 * block_length, dtype=torch.long, device=self.relative_attention_bias.weight.device
target_device = (
self.relative_attention_bias.weight.device
if self.relative_attention_bias.weight.device.type != "meta"
else None
)
memory_position = torch.arange(3 * block_length, dtype=torch.long, device=target_device)
context_position = memory_position[block_length:-block_length]

# (block_length, 3 * block_length)
Expand Down Expand Up @@ -843,9 +846,12 @@ def _relative_position_bucket(relative_position, bidirectional=True, num_buckets

def compute_bias(self, block_length: int):
"""Compute binned relative position bias"""
memory_position = torch.arange(
3 * block_length, dtype=torch.long, device=self.relative_attention_bias.weight.device
target_device = (
self.relative_attention_bias.weight.device
if self.relative_attention_bias.weight.device.type != "meta"
else None
)
memory_position = torch.arange(3 * block_length, dtype=torch.long, device=target_device)
context_position = memory_position[block_length:-block_length]

# (block_length, 3 * block_length)
Expand Down Expand Up @@ -1271,6 +1277,7 @@ class LongT5PreTrainedModel(PreTrainedModel):
config_class = LongT5Config
base_model_prefix = "transformer"
supports_gradient_checkpointing = True
_no_split_modules = ["LongT5Block"]

@property
# Copied from transformers.models.t5.modeling_t5.T5PreTrainedModel.dummy_inputs
Expand Down Expand Up @@ -1366,7 +1373,9 @@ class LongT5Stack(LongT5PreTrainedModel):
def __init__(self, config, embed_tokens=None):
super().__init__(config)

self.embed_tokens = embed_tokens
self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model)
if embed_tokens is not None:
self.embed_tokens.weight = embed_tokens.weight
self.is_decoder = config.is_decoder

self.local_radius = config.local_radius
Expand Down