.

mm-zero3.sh

@@ -1,6 +1,13 @@
 #!/usr/bin/env bash
 set -euo pipefail
 
+# ==== 0) 日志文件 ====
+LOG_DIR="/home/test/logs"
+mkdir -p "$LOG_DIR"
+LOG_FILE="${LOG_DIR}/train_$(date +%F_%H%M%S)_$$.log"
+export PYTHONUNBUFFERED=1
+echo ">> logging to ${LOG_FILE}"
+
 # 1) 可选：若含 API Key，保护权限
 [ -f .deepspeed_env ] && chmod 600 .deepspeed_env
 
@@ -9,12 +16,6 @@ set -a
 . ./.deepspeed_env
 set +a
 
-# unset PYTHONNOUSERSITE
-# USER_SITE=$(python3 -c 'import site;print(site.getusersitepackages())')
-# export PATH="$HOME/.local/bin:$PATH"
-# export PYTHONPATH="$USER_SITE:/home/test/jd_train:${PYTHONPATH:-}"
-
-
 # 统一环境（会被 deepspeed 的 ssh 继承到各节点）
 unset DS_BUILD_OPS DS_SKIP_CUDA_BUILD PYTHONNOUSERSITE
 export TORCH_EXTENSIONS_DIR=/tmp/$USER/torch_ext
@@ -25,6 +26,8 @@ export OMP_NUM_THREADS=8
 export MKL_NUM_THREADS=8
 export OPENBLAS_NUM_THREADS=8
 
+# ==== 3) 运行并同步到文件 + 控制台（保留退出码）====
+set +e
 FORCE_COLOR=1 deepspeed --hostfile hostfile \
   --num_nodes 6 --num_gpus 4 \
   /home/test/jd_train/train_sft_ds.py \
@@ -48,5 +51,8 @@ FORCE_COLOR=1 deepspeed --hostfile hostfile \
     --early_stopping_patience 5 \
     --early_stopping_threshold 0.0 \
     --metric_for_best_model eval_loss \
-    --eval_data_glob "/home/test/datasets/my_corpus/test.jsonl"
+    --eval_data_glob "/home/test/datasets/my_corpus/test.jsonl" \
-
+  2>&1 | tee -a "$LOG_FILE"
+DS_RC=${PIPESTATUS[0]}
+set -e
+exit "$DS_RC"