Run DeepSeek-V2-Lite prefetch-offload eval eager on ROCm (#46386)
Signed-off-by: aarushjain29 <Aarushi.Jain2@amd.com>
Aarushi Jain committed
Jun 24, 2026 at 12:25 UTC
cf5731118756858337689d1980ca898f3b9d5936
1 file changed
+8
.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh
+8
@@ -33,6 +33,14 @@ if [[ -n "${ATTENTION_BACKEND:-}" ]]; then
33
EXTRA_ARGS+=(--attention-backend "${ATTENTION_BACKEND}")
34
fi
35
36
+# ROCm: run eager to avoid intermittent HIP-graph decode corruption.
37
+# See https://github.com/ROCm/clr/issues/279
38
+# TODO(aarushjain29): Revert after TheRock 7.14
39
+if command -v rocm-smi &> /dev/null || command -v amd-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
40
+ echo "ROCm platform detected: adding --enforce-eager to avoid HIP-graph decode corruption"
41
+ EXTRA_ARGS+=(--enforce-eager)
42
+fi
43
+
44
cleanup() {
45
if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
46
kill "${SERVER_PID}" 2>/dev/null || true