server: context checkpoints ride with a slot save/restore — a restored slot is warm again by joelteply · Pull Request #2 · CambrianTech/llama.cpp

server: context checkpoints ride with a slot save/restore (sidecar) — a restored slot is warm again

joelteply

and

claude

committed

f8aa8d8

ggml: moe-gguf-source: widen the shard-index buffer — gcc -Werror=format-truncation on the CI (cpu) x64 job

joelteply

and

claude

committed

ae375e6

tools: moe-fit: widen the shard-index buffer too (gcc -Werror=format-truncation)

joelteply

and

claude

committed

45626cf

server tests: the slot-binding proof keeps the long slot busy for the whole restore round-trip (4096 tokens) — fast hosted runners ended the 512-token gen before the assertion

joelteply

and

claude

committed

13bbb16

server tests: the long slot ignores EOS — the tiny model ends a 'long story' in a few tokens on hosted runners, which read as a binding failure

joelteply

and

claude

committed

a02caaa

server tests: the short slot runs 256 tokens so a fast runner can observe both slots busy before the restores queue

joelteply

and

claude

committed

d61a394

server tests: the slot-binding proof is measured in timestamps — the restore returns beside its OWN slot's end, not the long slot's; skips when a runner cannot separate the slots

joelteply

and

claude

committed

1330954

Navigation Menu

Repository navigation

server: context checkpoints ride with a slot save/restore — a restored slot is warm again - #2

server: context checkpoints ride with a slot save/restore — a restored slot is warm again#2
joelteply merged 7 commits into
feat/props-weight-residencyfrom
continuum/kv-checkpoints-survive-slot-save

Commits on Sep 13, 2026

server: context checkpoints ride with a slot save/restore (sidecar) — a restored slot is warm again

ggml: moe-gguf-source: widen the shard-index buffer — gcc -Werror=format-truncation on the CI (cpu) x64 job

tools: moe-fit: widen the shard-index buffer too (gcc -Werror=format-truncation)

server tests: the slot-binding proof keeps the long slot busy for the whole restore round-trip (4096 tokens) — fast hosted runners ended the 512-token gen before the assertion

server tests: the long slot ignores EOS — the tiny model ends a 'long story' in a few tokens on hosted runners, which read as a binding failure

server tests: the short slot runs 256 tokens so a fast runner can observe both slots busy before the restores queue

server tests: the slot-binding proof is measured in timestamps — the restore returns beside its OWN slot's end, not the long slot's; skips when a runner cannot separate the slots