Skip to content
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

Jump to bottom

multi-gpu support for prefix-tuning #359

Merged
merged 7 commits into from
Jul 5, 2022
Merged

multi-gpu support for prefix-tuning #359

Changes from 3 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
07bd025
modify prefix-tuning forward for nn.DataParallel compatibility in PyT…
Jun 3, 2022
ce6e637
ModuleUtilsMixin
Jun 3, 2022
6c906a9
do not initialize input_tokens to avoid device detection issue
Jun 3, 2022
dec0cc9
update python import
Jun 28, 2022
459b1cc
Merge remote-tracking branch 'upstream/master'
Jun 28, 2022
c041a4e
modify FlatPrefixTuning to inherit ModeleUtilsMixin
Jun 28, 2022
3ae90d5
isort import
Jun 28, 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
15 changes: 7 additions & 8 deletions src/transformers/adapters/prefix_tuning.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,9 +8,10 @@
from .context import AdapterSetup, ForwardContext
from .layer import AdapterLayerBase
from .modeling import Activation_Function_Class
from ..modeling_utils import ModuleUtilsMixin


class PrefixTuning(nn.Module):
class PrefixTuning(nn.Module, ModuleUtilsMixin):
def __init__(
self,
n_layers: int,
Expand All @@ -25,7 +26,6 @@ def __init__(
self.n_embd_per_head = self.input_size // self.n_heads
self.config = config

self.input_tokens = torch.arange(self.config.prefix_length).long()
self.wte = nn.Embedding(self.config.prefix_length, self.input_size)
self.control_trans = nn.Sequential(
nn.Linear(self.input_size, self.config.bottleneck_size),
Expand All @@ -35,8 +35,8 @@ def __init__(
self.dropout = nn.Dropout(self.config.dropout)

def eject(self):
device = next(self.parameters()).device
input_tokens = self.input_tokens.unsqueeze(0).expand(1, -1).to(device)
input_tokens = torch.arange(self.config.prefix_length).long()
input_tokens = input_tokens.unsqueeze(0).expand(1, -1).to(self.device)
embs = self.wte(input_tokens)
key_values = self.control_trans(embs) # batch_size x prefix_length x n_layers*2*input_size
key_values = key_values.view(
Expand All @@ -46,8 +46,8 @@ def eject(self):
return key_values

def forward(self, batch_size):
device = next(self.parameters()).device
input_tokens = self.input_tokens.unsqueeze(0).expand(batch_size, -1).to(device)
input_tokens = torch.arange(self.config.prefix_length).long()
input_tokens = input_tokens.unsqueeze(0).expand(batch_size, -1).to(self.device)
embs = self.wte(input_tokens)
key_values = self.control_trans(embs) # batch_size x prefix_length x n_layers*2*input_size
key_values = key_values.view(
Expand Down Expand Up @@ -80,12 +80,11 @@ def __init__(
self.dropout = nn.Dropout(self.config.dropout)

def forward(self, batch_size):
device = next(self.parameters()).device
key_values = (
self.control_trans.unsqueeze(0)
.expand(batch_size, -1)
.view(batch_size, self.config.prefix_length, self.n_layers * 2, self.n_heads, self.n_embd_per_head)
.to(device)
.to(self.device)
alexanderhanboli marked this conversation as resolved.
Show resolved Hide resolved
) # *2 for key and value
key_values = self.dropout(key_values)
# n_layers * (2 x batch_size x n_heads x prefix_length x n_embd_per_head)
Expand Down