WEBVTT

00:00:00.000 --> 00:00:10.900
Four finite numbers become zeros in recorded MLX GroupNorm tests. The expected float sixteen output is minus one, one, minus one, one.

00:00:10.900 --> 00:00:20.333
The variance fits in float sixteen. But the sum of squares overflows before division. A representable answer is lost inside the calculation.

00:00:20.333 --> 00:00:30.100
Switching compatibility mode can change which channels belong together. The research patch preserves the original groups and uses float thirty two statistics.

00:00:30.100 --> 00:00:39.567
Across three hundred thirty two forward cases and twenty four gradient cases, the main suite records forty five failures before repair and none after.

00:00:39.567 --> 00:00:48.800
The first patch stalled on an empty array within a bounded test. A guard restored the original behavior. The failed prototype remains in the archive.

00:00:48.800 --> 00:01:00.100
Float thirty two range limits remain. GPU behavior and model performance are untested. A separate float sixty four observation still needs its own validated fix.