Missing python3.10-dev fixed | Generated by AI
Fixed. The problem was missing python3.10-dev — the triton CUDA module needs Python.h to compile the driver module. Installing it resolved the issue.
You can re-run the training now from your tmux session:
source .venv/bin/activate
export NANOCHAT_DATA_DIR=/workspace/zz/datasets/fineweb-edu/
python -m scripts.base_train \
--depth=24 \
--device-batch-size=16 \
--total-batch-size=524288 \
--max-seq-len=2048 \
--window-pattern SSSL \
--fp8 \
--num-iterations=29000 \
--eval-every=500 \
--save-every=5000 \
--run=h200-d24-760m \
--tracker=wandb