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

# Select GPUs 0 and 1 for tensor-parallel sharding
export CUDA_VISIBLE_DEVICES=0,1

# Unbuffered stdout for real-time logs
export PYTHONUNBUFFERED=1

# Help PyTorch allocator avoid fragmentation (see OOM hint)
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

# Start the local web app
exec python3 web_app_local.py "$@"