diff --git a/merge_zero3_safetensors.sh b/merge_zero3_safetensors.sh index bddea24..80bb614 100755 --- a/merge_zero3_safetensors.sh +++ b/merge_zero3_safetensors.sh @@ -3,7 +3,7 @@ set -euo pipefail # ===== 可调参数 ===== CKPT_ROOT="/home/test/checkpoints/q3-32b-ds4" # 若实际是 .../checkpoint-62/global_step62,请把 CKPT_ROOT 改成 .../checkpoint-62 -TAG="global_step62" +TAG="global_step20" HOSTS=(tn01 tn02 tn03 tn04 tn05 tn06) AGGREGATOR_HOST="tn06" # 本脚本运行/汇总所在机器 EXPECTED_SHARDS_PER_HOST=4 # 每机应写出分片数(按你的并行布局)