Correct the model card (pooling is at the LABELS suffix's last token); add code/classifier from project commit 4ced7554f
3d7c238 verified | # Pod setup, run on the pod after sync.sh up: venv, packages, the base checkpoint, then the environment | |
| # check (classifier.check_env) and the tiny-model tests (classifier.selftest). Rerun after a pod restart. | |
| # bootstrap.sh Qwen3.5-9B-Base | |
| # bootstrap.sh instruct also Qwen3.5-9B (only for the optional extra arm E1) | |
| set -euo pipefail | |
| W=/workspace | |
| cd "$W/code" | |
| # The base checkpoint and the HF cache also live on the container disk: the network volume (FUSE) reads and | |
| # writes about 200 MB/s, and every stage reloads the 19 GB model. Runs stay on /workspace, which survives the pod. | |
| M=${MODELS:-/root/models} | |
| export HF_HOME=/root/hf HF_XET_HIGH_PERFORMANCE=1 | |
| export PATH=/usr/local/cuda/bin:$PATH CUDA_HOME=${CUDA_HOME:-/usr/local/cuda} | |
| # The venv lives on the container disk: the network volume writes small files very slowly (reference lesson). | |
| V=${VENV:-/root/venv} | |
| [ -d "$V" ] || python3 -m venv --system-site-packages "$V" | |
| . "$V/bin/activate" | |
| pip install -q --upgrade pip | |
| # Keep the image's torch and Triton: nothing may replace them (the tilelang pin and selftest target them). | |
| python -c "import torch, triton; print('torch==' + torch.__version__.split('+')[0]); print('triton==' + triton.__version__)" > "$W/constraints.txt" | |
| before=$(python -c "import torch; print(torch.__version__)") | |
| pip install -q -c "$W/constraints.txt" -r classifier/requirements-train.txt | |
| after=$(python -c "import torch; print(torch.__version__)") | |
| [ "$before" = "$after" ] || { echo "FATAL: pip replaced the image's torch ($before -> $after)"; exit 1; } | |
| # Compiles CUDA code (devel image); without it transformers falls back to PyTorch and check_env fails. | |
| MAX_JOBS=${MAX_JOBS:-16} pip install -q --no-build-isolation -c "$W/constraints.txt" causal-conv1d==1.7.0 \ | |
| || echo "WARN: causal-conv1d did not build" | |
| models=(Qwen3.5-9B-Base) | |
| [ "${1:-}" = instruct ] && models+=(Qwen3.5-9B) | |
| for m in "${models[@]}"; do | |
| [ -f "$M/$m/.complete" ] || { hf download "Qwen/$m" --local-dir "$M/$m" && touch "$M/$m/.complete"; } | |
| done | |
| nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv | tee "$W/nvidia-smi.txt" | |
| status=0 | |
| python -m classifier.check_env --model "$M/Qwen3.5-9B-Base" --out "$W/check_env.json" || status=1 | |
| python -m classifier.selftest --model "$M/Qwen3.5-9B-Base" --out "$W/selftest.json" || status=1 | |
| [ $status = 0 ] && echo "BOOTSTRAP OK" || echo "BOOTSTRAP CHECKS FAILED: read check_env.json and selftest.json before any training" | |
| exit $status | |