start_training.sh

#!/bin/bash
set -e
export PATH="/usr/local/bin:/usr/bin:/bin:/usr/sbin:/sbin:$HOME/.local/bin:$PATH"

# Load RunPod env vars (Docker injects them into PID 1 but SSH doesn't inherit them)
if [ -f /proc/1/environ ]; then
    while IFS= read -r -d '' line; do
        export "$line"
    done < /proc/1/environ
fi

exec > >(tee -a /workspace/training_boot.log) 2>&1
echo "=== Training startup $(date) ==="
echo "MODEL_NAME=$MODEL_NAME"
nvidia-smi -L

echo "=== Installing dependencies ==="
pip install "transformers>=4.45,<5" "peft>=0.13,<0.15" trl datasets accelerate bitsandbytes huggingface_hub sentencepiece protobuf 2>&1 | tail -5
pip install flash-attn --no-build-isolation 2>/dev/null || echo "flash-attn skip"

echo "=== HF Login ==="
python3 -c "from huggingface_hub import login; login(token='$HF_TOKEN')"

echo "=== Downloading training script ==="
python3 -c "
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id='$DATASET_REPO', filename='train_lora.py', repo_type='dataset', local_dir='/workspace')
print('Downloaded train_lora.py')
"

mkdir -p /workspace/output /workspace/hf_cache

echo "=== Starting training at $(date) ==="
python3 /workspace/train_lora.py 2>&1 | tee /workspace/training.log
echo "DONE" > /workspace/DONE
echo "=== Training complete at $(date) ==="