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

No EMA buffer at beta1==0 #32

Merged
merged 3 commits into from
Dec 18, 2024
Merged

No EMA buffer at beta1==0 #32

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
63ab03f
initial
dxqbYD Dec 16, 2024
6b9e2d6
Merge branch 'konstmish:main' into beta0
dxqbYD Dec 16, 2024
a416d0e
fix formatting
konstmish Dec 18, 2024
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
21 changes: 14 additions & 7 deletions prodigyopt/prodigy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,12 +177,13 @@ def step(self, closure=None):
state['p0'] = torch.tensor(0, device=p.device, dtype=p.dtype)

# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p.data).detach()
if beta1 > 0:
state['exp_avg'] = torch.zeros_like(p.data).detach()
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p.data).detach()

exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
exp_avg_sq = state['exp_avg_sq']

s = state['s']
p0 = state['p0']

Expand All @@ -192,7 +193,9 @@ def step(self, closure=None):
d_numerator += (d / d0) * dlr * torch.dot(sliced_grad, p0.data - p.data.flatten()[::slice_p]).item()

# Adam EMA updates
exp_avg.mul_(beta1).add_(grad, alpha=d * (1-beta1))
if beta1 > 0:
exp_avg = state['exp_avg']
exp_avg.mul_(beta1).add_(grad, alpha=d * (1-beta1))
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=d * d * (1-beta2))

if safeguard_warmup:
Expand Down Expand Up @@ -246,7 +249,7 @@ def step(self, closure=None):

state = self.state[p]

exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
exp_avg_sq = state['exp_avg_sq']

state['step'] += 1

Expand All @@ -256,9 +259,13 @@ def step(self, closure=None):
if decay != 0 and decouple:
p.data.add_(p.data, alpha=-decay * dlr)


### Take step
p.data.addcdiv_(exp_avg, denom, value=-dlr)
if beta1 > 0:
exp_avg = state['exp_avg']
p.data.addcdiv_(exp_avg,denom, value=-dlr)
else:
p.data.addcdiv_(grad,denom, value=-dlr * d)
Copy link
Owner

Choose a reason for hiding this comment

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

Could you please fix the formatting? (missing spaces after ,)

Copy link
Contributor Author

Choose a reason for hiding this comment

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

I'm an a business trip - I have checked the "Allow edits by maintainers" checkbox - not sure how this works exactly, but I suppose there is an easy way for you to directly make this change on the github website, no?

Copy link
Contributor

Choose a reason for hiding this comment

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

Yeah that setting makes it possible for repo admins to edit PR code via GitHub.

Copy link
Owner

Choose a reason for hiding this comment

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

Indeed, I could do it myself, thanks



group['k'] = k + 1

Expand Down