Skip to content

Fix Lion decoupled weight decay in the CUDA 32-bit kernel #3

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
eaglstun merged 1 commit into from
Jul 15, 2026
Merged

Fix Lion decoupled weight decay in the CUDA 32-bit kernel #3

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
10 changes: 9 additions & 1 deletion csrc/kernels.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -858,7 +858,11 @@ __launch_bounds__(TH, 1) __global__ void kOptimizer32bit1State(
#pragma unroll 4
for (unsigned int j = 0; j < NUM_PER_THREAD; j++) {
g_vals[j] = gnorm_scale * ((float)g_vals[j]);
if (weight_decay > 0.0f)
// Coupled (L2) weight decay folds decay into the gradient. Correct for
// MOMENTUM/RMSPROP/ADAGRAD, but NOT for LION, which uses decoupled
// (AdamW-style) decay applied to the param directly (see the LION branch
// below and Chen et al. 2023). This matches the CUDA 8-bit blockwise kernel.
if (weight_decay > 0.0f && OPTIMIZER != LION)
g_vals[j] = (float)g_vals[j] + (((float)p_vals[j]) * weight_decay);
}

Expand All @@ -875,6 +879,10 @@ __launch_bounds__(TH, 1) __global__ void kOptimizer32bit1State(
p_vals[j] = ((float)p_vals[j]) + update_scale * (-lr * (s1_vals[j]));
break;
case LION:
// Decoupled weight decay: shrink the param directly, outside the
// sign update (Chen et al. 2023), matching the 8-bit blockwise kernel.
if (weight_decay > 0.0f)
p_vals[j] = ((float)p_vals[j]) * (1.0f - lr * weight_decay);
p_vals[j] =
((float)p_vals[j]) -
update_scale * (lr * sgn(((float)s1_vals[j]) * beta1 + ((1.0f - beta1) * ((float)g_vals[j]))));
Expand Down