---- the output's and the backward's planes ----
Each phase's tensors, placed in the workspace where the phase's inputs
are not: the output from the workspace's start (the final norm where a
layer's first norm is), the backward's scratch in the scores item (dead
once the recompute has its probabilities). Which item a plane reuses is
the placement; that no plane is overwritten while a later phase needs it
is what Proof.CoppeliusArenaCertificate decides.
528def cgLogitsBytes : Nat =
529 (naturalMultiply cgSeq cgRowBytes)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.