scrapegoat-lora/train_unsloth.sh

#!/bin/bash
set -euo pipefail
export PATH=$PATH:/home/ubuntu/.local/bin

export CUDA_VISIBLE_DEVICES=0
export TOKENIZERS_PARALLELISM=false
export NCCL_DEBUG=WARN
export CUDA_DEVICE_MAX_CONNECTIONS=1

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
OUTPUT_DIR=${OUTPUT_DIR:-"/ephemeral/scrapegoat-lora"}

mkdir -p /ephemeral/ds_offload
mkdir -p "${OUTPUT_DIR}"

echo "============================================"
echo "  ScrapeGoat Unsloth SFT (H100)"
echo "  Output: ${OUTPUT_DIR}"
echo "  GPUs: $(nvidia-smi -L 2>/dev/null | wc -l)"
echo "  DeepSpeed offload: /ephemeral/ds_offload"
echo "============================================"

torchrun \
    --nproc_per_node=1 \
    --master_port="${MASTER_PORT:-29501}" \
    "${SCRIPT_DIR}/train_unsloth.py" \
    --output_dir "${OUTPUT_DIR}" \
    --max_steps 200 \
    --save_steps 50 \
    --learning_rate 2e-4 \
    --warmup_steps 10 \
    --max_seq_length 4096 \
    "$@"

echo "[done] Training complete. Adapter saved to ${OUTPUT_DIR}"