start_torch.sh

#!/usr/bin/env bash

set -euo pipefail

export NCCL_ALGO="Ring"
export NCCL_PROTO="Simple"
export MP=8
export CONFIG="config.json"
export CKPT_PATH="deepseek-ai/DeepSeek-V4-Flash-bf16-mp8"

torchrun \
  --nproc-per-node "${MP}" \
  generate.py \
  --ckpt-path "${CKPT_PATH}" \
  --config "${CONFIG}" \
  --interactive