Skip to content

Commit b4154e5

Browse files
authored
ci: submit Phoenix GPU test jobs to a partition list for backfill (#1832)
1 parent 70589fd commit b4154e5

1 file changed

Lines changed: 22 additions & 6 deletions

File tree

.github/scripts/submit-slurm-job.sh

Lines changed: 22 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -126,13 +126,29 @@ elif [ "$device" = "gpu" ]; then
126126
# Determine GPU partition
127127
gpu_partition="batch"
128128
if [ "$gpu_partition_dynamic" = "true" ]; then
129-
# Use pre-selected bench partition if available, otherwise query sinfo
130-
if [ -n "${BENCH_GPU_PARTITION:-}" ]; then
131-
gpu_partition="$BENCH_GPU_PARTITION"
132-
echo "Using pre-selected bench partition: $gpu_partition (PR/master consistency)"
129+
if [ "$job_type" = "bench" ]; then
130+
# Benchmarks compare PR against master, so both jobs must land on the
131+
# SAME GPU type or the comparison is meaningless. That rules out a
132+
# partition list (SLURM could place PR and master on different
133+
# hardware); instead a single partition is picked and pinned across
134+
# both jobs via BENCH_GPU_PARTITION. See run_parallel_benchmarks.sh.
135+
if [ -n "${BENCH_GPU_PARTITION:-}" ]; then
136+
gpu_partition="$BENCH_GPU_PARTITION"
137+
echo "Using pre-selected bench partition: $gpu_partition (PR/master consistency)"
138+
else
139+
source "${SCRIPT_DIR}/select-gpu-partition.sh"
140+
gpu_partition="$SELECTED_GPU_PARTITION"
141+
fi
133142
else
134-
source "${SCRIPT_DIR}/select-gpu-partition.sh"
135-
gpu_partition="$SELECTED_GPU_PARTITION"
143+
# Tests (and build+test) don't compare across hardware, so submit to a
144+
# partition LIST and let SLURM start on whichever frees first instead
145+
# of pinning one partition and queueing behind it. This restores the
146+
# multi-partition backfill that #1299 dropped when it unified test and
147+
# bench onto the single-partition bench selector. gpu-l40s (bad
148+
# hardware) and gpu-rtx6000 (too slow for the test time limit) are
149+
# intentionally omitted.
150+
gpu_partition="gpu-h200,gpu-h100,gpu-a100,gpu-v100"
151+
echo "Using GPU partition list for test job: $gpu_partition"
136152
fi
137153
fi
138154

0 commit comments

Comments
 (0)