---- the invocation ----
one update's shared part: gradients zeroed, forward, backward, and the
embedding gradient's first scalars to window 1
(the gradients are zero: the initialization clears them, and every
AdamW clears each after its update)
2573def cgUpdate : (family NvidiaLaunchSchedule) =
2574 (cgThen
2575 cgForward
2576 (cgThen
2577 cgBackward
2578 (cgPhase
2579 cgPhaseDiagnostics
2580 (cgCopy (cgWindow 1) (cgParameter cgBankG cgEmbeddingEntry) cgWindowWords))))The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.