Upload run_sdxl_fast_benchmark.slurm with huggingface_hub
Browse files
run_sdxl_fast_benchmark.slurm
ADDED
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/bash
|
2 |
+
#SBATCH --job-name=benchmarking
|
3 |
+
#SBATCH --nodes=1
|
4 |
+
#SBATCH --qos=normal
|
5 |
+
# set 2h for job wall time limit
|
6 |
+
#SBATCH --time=02:00:00
|
7 |
+
# activate the requeue option
|
8 |
+
#SBATCH --requeue
|
9 |
+
#SBATCH --ntasks-per-node=1
|
10 |
+
#SBATCH --cpus-per-task=12
|
11 |
+
#SBATCH --gres=gpu:1
|
12 |
+
#SBATCH --mem-per-cpu=10G
|
13 |
+
#SBATCH --partition=production-cluster
|
14 |
+
#SBATCH -o /fsx/sayak/logs/benchmarking-job/sdxl-fast-%x-%j.out
|
15 |
+
|
16 |
+
|
17 |
+
set -x -e
|
18 |
+
|
19 |
+
source /admin/home/sayak/.bashrc
|
20 |
+
source /admin/home/sayak/miniconda3/etc/profile.d/conda.sh
|
21 |
+
conda activate /fsx/sayak
|
22 |
+
|
23 |
+
echo "START TIME: $(date)"
|
24 |
+
|
25 |
+
REPO=/fsx/sayak/sdxl-fast
|
26 |
+
OUTPUT_DIR=/fsx/sayak/benchmarking
|
27 |
+
LOG_PATH=$OUTPUT_DIR/main_log.txt
|
28 |
+
|
29 |
+
mkdir -p $OUTPUT_DIR
|
30 |
+
touch $LOG_PATH
|
31 |
+
pushd $REPO
|
32 |
+
|
33 |
+
PROGRAM="run_specific.sh"
|
34 |
+
|
35 |
+
# Note: it is important to escape `$SLURM_PROCID` since we want the srun on each node to evaluate this variable
|
36 |
+
export LAUNCHER="CUDA_VISIBLE_DEVICES=0 bash"
|
37 |
+
|
38 |
+
|
39 |
+
export CMD="$LAUNCHER $PROGRAM"
|
40 |
+
echo $CMD
|
41 |
+
|
42 |
+
|
43 |
+
# AWS specific
|
44 |
+
export NCCL_PROTO=simple
|
45 |
+
export RDMAV_FORK_SAFE=1
|
46 |
+
export FI_EFA_FORK_SAFE=1
|
47 |
+
export FI_EFA_USE_DEVICE_RDMA=1
|
48 |
+
export FI_PROVIDER=efa
|
49 |
+
export FI_LOG_LEVEL=1
|
50 |
+
export NCCL_IB_DISABLE=1
|
51 |
+
export NCCL_SOCKET_IFNAME=ens
|
52 |
+
|
53 |
+
|
54 |
+
# srun error handling:
|
55 |
+
# --wait=60: wait 60 sec after the first task terminates before terminating all remaining tasks
|
56 |
+
# --kill-on-bad-exit=1: terminate a step if any task exits with a non-zero exit code
|
57 |
+
SRUN_ARGS=" \
|
58 |
+
--wait=60 \
|
59 |
+
--kill-on-bad-exit=1 \
|
60 |
+
"
|
61 |
+
|
62 |
+
clear; srun $SRUN_ARGS --jobid $SLURM_JOB_ID bash -c "$CMD" 2>&1 | tee $LOG_PATH
|
63 |
+
|
64 |
+
echo "END TIME: $(date)"
|