← all documents · raw markdown · 3 KB

KAI Native Ternary — Owner checklist (you approve, I build)

Right now (while RunPod trains)

| You | Me (done) |

|-----|-----------|

| Let A100 train finish (GEN at 1k steps can be gibberish — normal) | Rust forward/decode wired for kai-native |

| Watch loss go down over epochs | install_kai_native.ps1 + check_kai_native.ps1 |

| Do not stop for messy GEN samples early | Export improved (tokenizer + MANIFEST) |

| | RSTS sandbox design (lattice path) ready separately |

GEN at step ~1000

[GEN] 'Who are you' -> 'The system are"etime warps...'

= learning, not failure. Early next-token is noisy. Trust loss trend + later epochs + final export.

---

When training finishes (on RunPod)

ls -lah /workspace/Kai_volume/kai-native/
# need: weights.bin  config.json  (+ tokenizer.json if exported)
# also fine: checkpoint_latest.pt  training.log

If only checkpoints exist (export crashed), re-export is automatic at end of train script; or we can add a small export-only command.

Download whole folder kai-native to your PC (Jupyter download or scp).

---

On Windows (you run / I can run when you say)

cd C:\KAI
.\tools\install_kai_native.ps1 -FromDir "PATH\TO\downloaded\kai-native"
cargo build --release --bin kai
.\Start-KAI.ps1
.\tools\check_kai_native.ps1

Boot log should include:

[LanguageWarehouse] KAI native model: LOADED
[LanguageWarehouse] KAI native structured: ...
[Hybrid Router] Routing → KAI native ternary model ...

---

What I already coded for you

1. language_warehouse.rs — parse weights, structure layers, RoPE + ternary attn + SwiGLU + sample, kai_native_decode, router uses it first

2. ternary_math.rs — kai_i2 unpack + matvec matching train export

3. Train export — tokenizer copy + MANIFEST

4. Scripts — install / check

5. RSTS sandbox — lattice-native cousin path (not blocking this ship)

---

What you only need to approve later

[ ] Train finished, weights downloaded
[ ] Install script run
[ ] Engine restart + smoke chat “Who are you”
[ ] If output still weak: more train epochs / more data / smaller temp (I adjust)
[ ] Optional next phase: RSTS lattice spine hybrid gate

---

Pod: keep training healthy

# tail log
tail -f /workspace/Kai_volume/kai-native/training.log

# if dead, resume from latest checkpoint (after script has --resume support)
ls /workspace/Kai_volume/kai-native/checkpoint*.pt

If process dies after export, you’re done — just download.