train: walk the window — decode context into the cache, train each reply chunk against it by joelteply · Pull Request #40 · CambrianTech/llama.cpp

train: walk the window: decode context, train each reply chunk against the cached prefix (WIP)

joelteply

committed

69a2f8f

train: drop the walk's debug prints; a reply a hybrid memory cannot re-decode stops the run loudly

joelteply

and

claude

committed

0540511

train: a recurrent state is snapshotted around each training chunk (Fable), so a reply spans chunks on a hybrid

joelteply

and

claude

committed

321c1fe

train: the cache holds constants at F16; an interim default chunk of 512

joelteply

and

claude

committed

c424b00

train: q8_0 cache constants in a flash-attention context; training graphs build explicit attention

joelteply

and

claude

committed

a8a235d

train: the walk yields at every chunk, not once per example window

joelteply

and

claude

committed

87e5078

train: the walk's chunk step fails its run through #42's backend-failure path (returns false, like a refused graph)

joelteply

and

claude

committed

539c75d

Navigation Menu

Repository navigation

train: walk the window — decode context into the cache, train each reply chunk against it - #40

train: walk the window — decode context into the cache, train each reply chunk against it#40
joelteply merged 7 commits into
feat/props-weight-residencyfrom
train/walk-on-head

Commits on Oct 6, 2026

train: walk the window: decode context, train each reply chunk against the cached prefix (WIP)

train: drop the walk's debug prints; a reply a hybrid memory cannot re-decode stops the run loudly

train: a recurrent state is snapshotted around each training chunk (Fable), so a reply spans chunks on a hybrid

train: the cache holds constants at F16; an interim default chunk of 512

train: q8_0 cache constants in a flash-attention context; training graphs build explicit attention

train: the walk yields at every chunk, not once per example window

train: the walk's chunk step fails its run through #42's backend-failure path (returns false, like a refused graph)