Skip to content

Fix FSDP2 offload param+adamw8bit #3116

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

Open
wants to merge 6 commits into
base: main
Choose a base branch
from
Open

Fix FSDP2 offload param+adamw8bit #3116

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
85b24cf
Fix FSDP offload+Adamw8bits compatible
nguyen599 Sep 30, 2025
5936312
fix
nguyen599 Sep 30, 2025
64020e7
Merge branch 'pytorch:main' into fsdp2-offload-adamw8bit
nguyen599 Oct 1, 2025
deced26
Merge branch 'pytorch:main' into fsdp2-offload-adamw8bit
nguyen599 Oct 1, 2025
3b8f3a7
Merge branch 'pytorch:main' into fsdp2-offload-adamw8bit
nguyen599 Oct 2, 2025
b05ed3a
Merge branch 'pytorch:main' into fsdp2-offload-adamw8bit
nguyen599 Oct 5, 2025
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
12 changes: 7 additions & 5 deletions torchao/optim/adam.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -179,10 +179,11 @@ def single_param_adam(
p_f32 = p.float()
grad_f32 = grad.float()

if IS_ADAMW:
p_f32 = p_f32 - lr * weight_decay * p_f32
else:
grad_f32 = grad_f32 + weight_decay * p_f32
if weight_decay != 0:
if IS_ADAMW:
p_f32.mul_(1 - lr * weight_decay)
else:
grad_f32 = grad_f32.add(p_f32, alpha=weight_decay)

bias_correction1 = 1 - beta1**step
bias_correction2 = 1 - beta2**step
Expand All @@ -201,7 +202,8 @@ def single_param_adam(
else:
denom = (exp_avg_sq_f32.sqrt() / bias_correction2.sqrt()) + eps

p_f32 = p_f32 - lr * (exp_avg_f32 / bias_correction1) / denom
step_size = lr / bias_correction1
p_f32.addcdiv_(exp_avg_f32, denom, value=-step_size.to(p_f32.device))

if BF16_STOCHASTIC_ROUND:
p.copy_(_fp32_to_bf16_sr(p_f32))
Expand Down