[ROCm][CI] Keep native datasets cache off shared NFS (#49516)
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Andreas Karatzas committed
Jul 26, 2026 at 22:31 UTC
da99ffcc13362ab446a22a9bc6ed36c4c14942e3
1 file changed
+6
-2
.buildkite/scripts/hardware_ci/run-amd-test.sh
+6
-2
@@ -405,11 +405,14 @@ initialize_native_environment() {
405
VLLM_CACHE_ROOT="${native_root}/cache/vllm"
406
XDG_CACHE_HOME="${native_root}/cache/xdg"
407
: "${HF_HOME:=/home/buildkite-agent/huggingface}"
408
+ # datasets uses POSIX locks that are unsupported by the shared HF NFS cache.
409
+ # Keep processed datasets job-local while retaining the persistent Hub cache.
410
+ HF_DATASETS_CACHE="${native_root}/cache/huggingface/datasets"
411
: "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
412
: "${HF_HUB_ETAG_TIMEOUT:=60}"
413
export TMPDIR VLLM_RPC_BASE_PATH
414
export TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT XDG_CACHE_HOME
412
- export HF_HOME HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
415
+ export HF_HOME HF_DATASETS_CACHE HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
416
export PYTORCH_ROCM_ARCH=""
417
418
mkdir -p "${TMPDIR}" \
@@ -417,7 +420,8 @@ initialize_native_environment() {
420
"${TRITON_CACHE_DIR}" \
421
"${VLLM_CACHE_ROOT}" \
422
"${XDG_CACHE_HOME}" \
420
- "${HF_HOME}" || return 1
423
+ "${HF_HOME}" \
424
+ "${HF_DATASETS_CACHE}" || return 1
425
426
echo "Native compile caches: VLLM_CACHE_ROOT=${VLLM_CACHE_ROOT} TORCHINDUCTOR_CACHE_DIR=${TORCHINDUCTOR_CACHE_DIR}"
427