Upload run_cmp_rollout.sh with huggingface_hub
Browse files- run_cmp_rollout.sh +7 -2
run_cmp_rollout.sh
CHANGED
|
@@ -23,8 +23,13 @@ export RENDER_BUNDLE_DIR="${RENDER_BUNDLE_DIR:-/cache/rollout_bundle}"
|
|
| 23 |
source "$(dirname "${BASH_SOURCE[0]}")/bundle_from_hf.sh"
|
| 24 |
|
| 25 |
OUT="${ROLLOUT_OUT_DIR:-/home/ma-user/modelarts/outputs/result_0}"
|
| 26 |
-
|
| 27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
mkdir -p "$OUT" "$WORK" "$STAGE"
|
| 29 |
|
| 30 |
echo "[entry] bundle=$BUNDLE shard=${SHARD_INDEX:-0}/${SHARD_COUNT:-1}"
|
|
|
|
| 23 |
source "$(dirname "${BASH_SOURCE[0]}")/bundle_from_hf.sh"
|
| 24 |
|
| 25 |
OUT="${ROLLOUT_OUT_DIR:-/home/ma-user/modelarts/outputs/result_0}"
|
| 26 |
+
# ⚠️ work/stage 目录要带分片号。`/cache` 是**节点级**的,同一个节点上完全可能
|
| 27 |
+
# 落两个分片(池子跟别人共用,调度不归我们管)。两片共用一个 work 目录时,
|
| 28 |
+
# 预热跑的是**同一批配置**(warmup 从枚举头上取),sid 一模一样,
|
| 29 |
+
# 两个进程同时往同一个 npz 写 —— 写出来的东西谁也说不清。
|
| 30 |
+
# bundle 和 JIT 缓存反过来**必须共享**(只读 / 只增),所以那两个不带分片号。
|
| 31 |
+
WORK="${ROLLOUT_WORK_DIR:-/cache/rollout_work_s${SHARD_INDEX:-0}}"
|
| 32 |
+
STAGE="${ROLLOUT_STAGE_DIR:-/cache/rollout_out_s${SHARD_INDEX:-0}}"
|
| 33 |
mkdir -p "$OUT" "$WORK" "$STAGE"
|
| 34 |
|
| 35 |
echo "[entry] bundle=$BUNDLE shard=${SHARD_INDEX:-0}/${SHARD_COUNT:-1}"
|