Spaces:

HOLOGRAMTECH
/

bonsai-kappa-stream

Running

App Files Files Community

bonsai-kappa-stream / q

7.2 MB

Ctrl+K

Ctrl+K

4 contributors

History: 8 commits

humuhumu33

Claude Opus 4.8

Self-heal poisoned weight cache: purge + re-fetch on κ MISMATCH

d9f5555 4 months ago

core
Keep multi-turn conversations fast: snapshot the conversation's KV before the follow-up-suggestion pass evicts it, and restore it before the next turn so only the new turn prefills instead of re-processing the whole thread. Measured 20x faster re-prefill on a 459-token thread (12s to 0.6s, same GPU back-to-back), byte-identical by determinism, gated to long threads on capable desktops. 4 months ago
forge
WARP W1: M=k batched prefill (2.6 -> 11.5 tok/s at M=16, bit-exact) + verifyBatch primitive 4 months ago
pkg
Bonsai kappa-stream: self-contained showcase (app + engine closure) 4 months ago
holo-delta.mjs

10.8 kB
Bonsai kappa-stream: self-contained showcase (app + engine closure) 4 months ago
holo-load-delta.mjs

3.81 kB
Bonsai kappa-stream: self-contained showcase (app + engine closure) 4 months ago
holo-load2bit.mjs

12.6 kB
Self-heal poisoned weight cache: purge + re-fetch on κ MISMATCH 4 months ago
holo-model-frame.mjs

6.48 kB
Bonsai 27B is now the default: the full 27B mind runs in your browser (4-5 tok/s, watch it think), with a one-tap toggle to the fast 8B. Verified q1 hybrid engine + app bridge. 4 months ago
qvac-2bit.mjs

20 kB
Bonsai kappa-stream: self-contained showcase (app + engine closure) 4 months ago
qvac-gpu.js

184 kB
WARP W0: 27B cross-turn state carry (TTFT no longer grows with history) + incremental detok + specDecode qk_norm lift (byte-identity proven, spec stays opt-in) 4 months ago
qvac-ingest.mjs

28.6 kB
Bonsai kappa-stream: self-contained showcase (app + engine closure) 4 months ago