A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 18:15 UTC
Enabling cooperative multi-gpu tests on multi-gpu nodes (#27986)
Signed-off-by: Alexei V. Ivanov <alexei.ivanov@amd.com>
Alexei-V-Ivanov-AMD committed
Nov 5, 2025 at 09:35 UTC
80c9275348c1c0be21f03938c918d2ed5d98574b
1 file changed
+10
-1
.buildkite/scripts/hardware_ci/run-amd-test.sh
+10
-1
@@ -173,6 +173,14 @@ fi
173
PARALLEL_JOB_COUNT=8
174
MYPYTHONPATH=".."
175
176
+# Test that we're launching on the machine that has
177
+# proper access to GPUs
178
+render_gid=$(getent group render | cut -d: -f3)
179
+if [[ -z "$render_gid" ]]; then
180
+ echo "Error: 'render' group not found. This is required for GPU access." >&2
181
+ exit 1
182
+fi
183
+
184
# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs.
185
if [[ $commands == *"--shard-id="* ]]; then
186
# assign job count as the number of shards used
@@ -186,6 +194,7 @@ if [[ $commands == *"--shard-id="* ]]; then
194
--device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
195
--network=host \
196
--shm-size=16gb \
197
+ --group-add "$render_gid" \
198
--rm \
199
-e HIP_VISIBLE_DEVICES="${GPU}" \
200
-e HF_TOKEN \
@@ -217,8 +226,8 @@ else
226
--device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
227
--network=host \
228
--shm-size=16gb \
229
+ --group-add "$render_gid" \
230
--rm \
221
- -e HIP_VISIBLE_DEVICES=0 \
231
-e HF_TOKEN \
232
-e AWS_ACCESS_KEY_ID \
233
-e AWS_SECRET_ACCESS_KEY \