#!/usr/bin/env bash
set -euo pipefail

export NCCL_ASYNC_ERROR_HANDLING=1
export NCCL_P2P_LEVEL=PHB
export CUDA_DEVICE_MAX_CONNECTIONS=1
export OMP_NUM_THREADS=8

torchrun --standalone --nproc-per-node=7 -m nayara.train \
  --config configs/nayara_350m.json \
  --data /mnt/nvme/nayara/packed \
  --output /mnt/nvme/nayara/checkpoints \
  --micro-batch 32 \
  --grad-accum 2 \
  --max-tokens 3800000000 \
  --stop-at 15:15