Price trainer-rank memory profiles by packed tokens, not sharing ratio - #950

Merged

bradhilton merged 13 commits into

mainfrom

dalinar/planner-sharing-ratio

Sep 24, 2026

Commits on Sep 24, 2026

Price trainer-rank memory profiles by packed tokens, not sharing ratio

bradhilton
and
claude
committed
8d70537
Price logical-sized head copies and output gradients; drop stale callers

bradhilton
and
claude
committed
1cac203
Scope packed-token pricing to single-target training requests

bradhilton
and
claude
committed
bc5a1b6
Tighten packed-pricing scope and charge head rows beyond profile sharing

bradhilton
and
claude
committed
d243137
Charge retained head rows and keep packed pricing monotone

bradhilton
and
claude
committed
73db1f9
Charge GDN branch states and keep packed pricing monotone at the floor

bradhilton
and
claude
committed
597061d
Limit packed pricing to one-layer full recompute

bradhilton
and
claude
committed
c4a9427
Detect one-layer recompute from exact config and training mode

bradhilton
and
claude
committed
81e0eb1
Read the decoder's own mode and recompute config when pricing

bradhilton
and
claude
committed
d2d4d81
Charge allocator blocks for shared rows; read the decoder's own mode

bradhilton
and
claude
committed
960fae3
Size the shared-row charge from a measured head peak

bradhilton
and
claude
committed
740242c
Charge head and caller memory per logical row; clamp sharing to 8x

bradhilton
and
claude
committed
058547e
Keep short-request batches on the calibrated profile

bradhilton
and
claude
committed
ccca0dc