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

add resx experimental mode #173

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

add resx experimental mode #173

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
34 changes: 23 additions & 11 deletions RWKV-v4neo/src/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -632,18 +632,30 @@ def forward(self, idx):

if args.dropout > 0:
x = self.drop0(x)
if args.tiny_att_dim > 0:
for block in self.blocks:
if args.grad_cp == 1:
x = deepspeed.checkpointing.checkpoint(block, x, x_emb)
else:
x = block(x, x_emb)
else:
for block in self.blocks:
if args.grad_cp == 1:
x = deepspeed.checkpointing.checkpoint(block, x)


# resx is an exponential residual connection between layers.
# It cuts early training speed in half
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
# It cuts early training speed in half
# It cuts early training time in half

resx = "x" in os.environ["RWKV_MY_TESTING"]
resxstack = None

usex_emb = args.tiny_att_dim > 0

for block in self.blocks:

if resx:
if resxstack is None:
resxstack = x
else:
x = block(x)
resxstack = resxstack*2 + x

x = resxstack + x

if args.grad_cp == 1:
x = deepspeed.checkpointing.checkpoint(block, *([x, x_emb] if usex_emb else [x]))
else:
x = block(*([x, x_emb] if usex_emb else [x]))


x = self.ln_out(x)

Expand Down