Files
gpt-2-inference-engine/verify_ncu.slurm
T
2026-06-07 13:43:20 -05:00

47 lines
995 B
Bash

#!/bin/bash
#SBATCH --job-name="verify"
#SBATCH --partition=gpuA40x4
#SBATCH --mem=12G
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=1
#SBATCH --constraint="projects,perf,nvperf"
#SBATCH --gpus-per-node=1
#SBATCH --gpu-bind=closest
#SBATCH --account=bche-delta-gpu
#SBATCH -t 00:15:00
#SBATCH --exclude=gpub029,gpub034,gpub038,gpub043,gpub074,gpub054,gpub012
module reset
module load cuda
SUFFIX=${1:-op9}
BIN=./build/test_gpt2_${SUFFIX}
# -------------------------
# output dirs
# -------------------------
OUTDIR=output/perf/ncu
REPORTDIR=output/perf/reports
mkdir -p $OUTDIR
mkdir -p $REPORTDIR
OUTFILE=$OUTDIR/verify_${SUFFIX}.out
ERRFILE=$OUTDIR/verify_${SUFFIX}.err
REPORT=$REPORTDIR/analysis_${SUFFIX}
echo "Binary: $BIN"
echo "NCU report: $REPORT"
exec > $OUTFILE 2> $ERRFILE
echo -e "job $SLURM_JOBID starting on $(hostname)\n\n"
# -------------------------
# NCU profiling
# -------------------------
srun ncu --set full -f \
-o $REPORT \
$BIN