Repository navigation

All pull requests

Labels (34)Milestones (0)New pull request

server: learn a past example: /learn's messages form serves its prompt through the ordinary completion path first, then steps
#89joelteply opened Nov 2, 2026
server: the lost-reply binding: a learn names its execution, base step and engine; a repeat is answered from its receipt, a stale one is fenced
#88joelteply opened Nov 1, 2026
2
server: one learning ledger (draft): every learn step debits the grant that admitted it
#87joelteply opened Nov 1, 2026
4
server: POST /lora-adapters/snapshot writes a mind's shadow to an immutable GGUF between steps, and /learn reports the shadow's committed step
#86joelteply opened Nov 1, 2026
server: one shadow per mind on the serving context: each shadow keeps its own optimizer, /learn trains any loaded shadow, and unload releases it
#85joelteply opened Nov 1, 2026
server: create dormant dream adapters and save immutable snapshots
#81joelteply opened Oct 28, 2026
2
server: a page carries the draft's KV once; checkpoints stop copying it
#79joelteply opened Oct 28, 2026
1
server: a disk tier for the prompt cache (--cache-disk)
#78joelteply opened Oct 27, 2026
server: the checkpoint budget's eviction gives up what a saved page gives up first, never oldest-first
#74joelteply opened Oct 25, 2026
server: /reconfigure changes the context geometry without reloading the model (S2)
#73joelteply opened Oct 24, 2026
server: genes (LoRA adapters) page between the device and host RAM (stacked on #70)
#72joelteply opened Oct 24, 2026
server: a slot pages to named host RAM and back; checkpoints keep a byte budget; a page carries its draft KV
#70joelteply opened Oct 24, 2026
1
server: rotate admitted prefill service without moving KV
#69joelteply opened Oct 23, 2026
2
server: checkpoint at a client-named user message (checkpoint_user_message)
#68joelteply opened Oct 20, 2026
server: expose realized context allocation geometry
#67joelteply opened Oct 20, 2026
grammar: preserve tagged unions through singleton allOf
#66joelteply opened Oct 17, 2026
train in place: labels ride the batch, the reply learns on the slot that served it (POST /learn)
#65joelteply opened Oct 15, 2026
3
llama: connect labelled decode to shared physical cache rows
#64joelteply opened Oct 14, 2026
4
cuda: implement shared FLASH_ATTN_BACK with direct quantized cache reads
#63joelteply opened Oct 14, 2026
5
metal: OUT_PROD runs batched, broadcast and narrow shapes on the GPU (training 1.69x on the M5)
#62joelteply opened Oct 14, 2026
train: one rule for the recurrent states a graph keeps, so training runs in a context with rollback slots
#61joelteply opened Oct 13, 2026
2
train: the start's claim is released on every exit that doesn't hand it to the worker
#58joelteply opened Oct 13, 2026
server /train: a per-turn slowdown receipt beside the cycle average
#39joelteply opened Oct 6, 2026
7
continuum: explicit Metal device selection — avoid hang in AMD Polaris lazy init (card 346b356f)
#1joelteply opened Jun 7, 2026