tacit/nanochat
1
0
Fork 0
mirror of https://github.com/karpathy/nanochat.git synced 2025-12-06 04:12:13 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fix OOM in Japanese tokenizer training by reducing max_chars

Browse Source 
Reduce --max_chars from 2B to 500M characters to prevent OOM during
BPE training. Japanese text generates significantly more unique
sequences than English (92M+ unique sequences observed), causing
memory exhaustion during heap construction.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
karaage0703 2025-12-02 12:09:24 +09:00
parent 86a6cf6668
commit 55f8d4acf2
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
speedrun_spark_ja.sh
View File

@ -8,6 +8,7 @@ DEPTH=20
DEVICE_BATCH_SIZE=16
DATA_SHARDS=30
NUM_ITERATIONS=1000
#NUM_ITERATIONS=10
CACHE_DIR="$HOME/.cache/nanochat"
# ========================
@ -34,7 +35,7 @@ echo "== 1) 日本語データ準備 =="
python -m nanochat.dataset -n "${DATA_SHARDS}" --lang ja
echo "== 2) 日本語トークナイザ学習 =="
python -m scripts.tok_train --max_chars=2000000000
python -m scripts.tok_train --max_chars=500000000
python -m scripts.tok_eval || true
ls -l "${CACHE_DIR}/tokenizer" || true