From 8d9d6a4cf6c6e539a8098eea93615478cbe4d5f1 Mon Sep 17 00:00:00 2001 From: David Rohr Date: Tue, 11 Mar 2025 15:00:18 +0100 Subject: [PATCH] dpl-workflow: Automatically apply MI100 workaround in sync --- prodtests/full-system-test/dpl-workflow.sh | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/prodtests/full-system-test/dpl-workflow.sh b/prodtests/full-system-test/dpl-workflow.sh index 0f5083dbcdebb..f9b0c7accbff9 100755 --- a/prodtests/full-system-test/dpl-workflow.sh +++ b/prodtests/full-system-test/dpl-workflow.sh @@ -259,7 +259,8 @@ if [[ $GPUTYPE == "HIP" ]]; then TIMESLICEOFFSET=$(($GPU_FIRST_ID + ($NUMAGPUIDS != 0 ? ($NGPUS * $NUMAID) : 0))) GPU_CONFIG+=" --environment \"ROCR_VISIBLE_DEVICES={timeslice${TIMESLICEOFFSET}}\"" fi - [[ "${EPN_NODE_MI100:-}" != "1" ]] && export HSA_NO_SCRATCH_RECLAIM=1 + [[ ${EPN_NODE_MI100:-} != "1" ]] && export HSA_NO_SCRATCH_RECLAIM=1 + [[ $EPNSYNCMODE == 1 && ${EPN_NODE_MI100:-} == "1" ]] && GPU_CONFIG_KEY+="GPU_proc.serializeGPU=3;" #export HSA_TOOLS_LIB=/opt/rocm/lib/librocm-debug-agent.so.2 else GPU_CONFIG_KEY+="GPU_proc.deviceNum=-2;"