WEBVTT

00:00:00.000 --> 00:00:02.528
Same outputs. Same gradients.

00:00:02.528 --> 00:00:05.056
But half the optimizer step.

00:00:05.056 --> 00:00:07.584
A reproducible result in Apple’s MLX.

00:00:08.100 --> 00:00:10.505
A 1×1 convolution and an equivalent

00:00:10.505 --> 00:00:12.910
linear layer match before the update.

00:00:13.433 --> 00:00:16.116
Eight output channels. Two input channels.

00:00:16.116 --> 00:00:18.799
The convolution weight step is half as large.

00:00:19.300 --> 00:00:22.279
Muon restores the weight shape too early.

00:00:22.279 --> 00:00:25.259
The scale uses the wrong dimensions.

00:00:25.767 --> 00:00:28.740
Move scaling before the reshape.

00:00:28.740 --> 00:00:31.714
31 focused tests pass. 19 failed before.

00:00:32.233 --> 00:00:36.064
Local CPU test. Model effects unmeasured.

00:00:36.064 --> 00:00:39.894
Code, patch and logs on GERO.