diff options
Diffstat (limited to 'TRANSFORMER_CROSSOVER.md')
| -rw-r--r-- | TRANSFORMER_CROSSOVER.md | 8 |
1 files changed, 7 insertions, 1 deletions
diff --git a/TRANSFORMER_CROSSOVER.md b/TRANSFORMER_CROSSOVER.md index 2bcff12..df0cf0f 100644 --- a/TRANSFORMER_CROSSOVER.md +++ b/TRANSFORMER_CROSSOVER.md @@ -186,7 +186,13 @@ allocated/reserved memory, forward/feedback parameters, ordinary training and validation tokens, positive/negative or perturbed presentations, candidate-token presentations, relaxation-token passes, local-VJP token evaluations, first nonfinite step, and completed optimizer steps. The -accuracy--cost plot and the within-family depth plot retain timeouts and +hardware-independent ledger additionally enumerates full-forward token +passes, extra local block/head evaluations, block affine MACs per token, +causal-attention MACs per token, and whole-forward MACs per token. These +primitive counts remain separate for DP/EP because a local VJP is not +silently priced as an ordinary forward. + +The accuracy--cost plot and the within-family depth plot retain timeouts and failures. A low-cost method is not declared scalable from depth 4 alone, and an accurate DP/EP point is not declared local in the weight-transport-free sense merely because its VJPs were evaluated at detached edges. |