diff options
| author | jdj <jdj@kunix.org> | 2026-11-18 14:47:33 -0800 |
|---|---|---|
| committer | jdj <jdj@kunix.org> | 2026-11-18 14:47:33 -0800 |
| commit | 5d2e4252e695644c34be605b44a14bd80a98a0a6 (patch) | |
| tree | 311a72eda4459af7547b9cd05c0956eafb235f2d /finalize_0920.sh | |
| download | boris-2-5d2e4252e695644c34be605b44a14bd80a98a0a6.tar.gz boris-2-5d2e4252e695644c34be605b44a14bd80a98a0a6.zip | |
Boris-2 training code
The code, configuration and data filters used to train Boris-2 (144M,
200B tokens, single V100): model, training loop, data preparation,
n-gram table, V100 kernels, final stages and the exact run configs.
Diffstat (limited to 'finalize_0920.sh')
| -rwxr-xr-x | finalize_0920.sh | 22 |
1 files changed, 22 insertions, 0 deletions
diff --git a/finalize_0920.sh b/finalize_0920.sh new file mode 100755 index 0000000..64ba388 --- /dev/null +++ b/finalize_0920.sh @@ -0,0 +1,22 @@ +#!/usr/bin/env bash +set -euo pipefail +repo=/home/joseph/boris-2-D112M-n15M +run=/media/joseph/Media/Boris-2-0920 +exec >>"$run/finalize_0920.log" 2>&1 + +echo "[$(date -Is)] building 90/10 n-gram vocabulary" +nice -n 5 ionice -c2 -n5 "$repo/.venv/bin/python" "$repo/ngram.py" build \ + --config "$run/config.json" --out "$run/ngram_vocab.bin" + +echo "[$(date -Is)] verification" +"$repo/.venv/bin/python" "$repo/verify.py" --quick --device cuda:1 + +echo "[$(date -Is)] V100 micro-batch confirmation" +"$repo/.venv/bin/python" "$repo/preflight.py" --device cuda:0 \ + --block 2048 --sweep 8,16 --iters 4 + +setsid "$repo/.venv/bin/python" "$repo/eval.py" --watch --device cuda:1 \ + >"$run/eval_watch.log" 2>&1 < /dev/null & + +echo "[$(date -Is)] launching Boris-2-0920" +exec "$repo/.venv/bin/python" "$repo/train.py" --config "$run/config.json" |
