From d44c836cfdcfa6367158df16f3ab5858bc755d37 Mon Sep 17 00:00:00 2001 From: ashwini rathi Date: Thu, 13 Aug 2026 08:49:39 +0530 Subject: [PATCH] [XPU][CI] disable SYCL_CACHE_PERSISTENT to fix topk segfault (#34567) Co-authored-by: Claude Opus 4.7 (1M context) --- scripts/ci/xpu/xpu_ci_start_container.sh | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/scripts/ci/xpu/xpu_ci_start_container.sh b/scripts/ci/xpu/xpu_ci_start_container.sh index 11314a792..f4e49bead 100755 --- a/scripts/ci/xpu/xpu_ci_start_container.sh +++ b/scripts/ci/xpu/xpu_ci_start_container.sh @@ -128,6 +128,9 @@ echo "Launching container: ${CONTAINER_NAME} from ${IMAGE}" # SGLANG_SERVER_LAUNCH_TIMEOUT=36000 matches /data/pgirijal/scripts/setup_upstream_env.sh: # 4-GPU MoE loads (Qwen3.5-35B-A3B, gemma-4-26B-A4B, ...) on Arc Pro B60 can # take >1h from a cold HF cache, so give sglang server startup a 10h ceiling. +# SYCL_CACHE_PERSISTENT=0: the SYCL runtime's persistent kernel cache mishandles +# torch 2.13's XPU aten.topk kernel on compute-runtime 26.05 / IGC 2.28 -- reload +# segfaults inside libsycl. Keep off until Intel ships a fix in newer runtimes. docker run -dt \ --shm-size 8g \ --group-add 992 \ @@ -145,7 +148,7 @@ docker run -dt \ -e NEO_CACHE_DIR=/root/.cache/sglang-xpu/neo \ -e NEO_CACHE_PERSISTENT=1 \ -e SYCL_CACHE_DIR=/root/.cache/sglang-xpu/sycl \ - -e SYCL_CACHE_PERSISTENT=1 \ + -e SYCL_CACHE_PERSISTENT=0 \ "${GPU_AFFINITY_ARGS[@]}" \ --name "${CONTAINER_NAME}" \ "${IMAGE}"