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 14, 05:30 UTC
[CI] Nixl integration tests DP-EP (#27199)
Signed-off-by: NickLucche <nlucches@redhat.com>
Nicolò Lucchesi committed
Oct 22, 2025 at 05:17 UTC
bfa59be8f18512ac1b6d2764e17db25f29c75933
2 files changed
+10
-1
tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh
+9
-1
@@ -178,8 +178,16 @@ run_tests_for_model() {
178
--port $PORT \
179
--enforce-eager \
180
--gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
181
- --tensor-parallel-size $DECODER_TP_SIZE \
181
--kv-transfer-config '$KV_CONFIG'"
182
+
183
+ # DP-EP attention mode
184
+ if [[ -z "$DP_EP" ]]; then
185
+ BASE_CMD="${BASE_CMD} --tensor-parallel-size $DECODER_TP_SIZE"
186
+ else
187
+ echo "DP-EP Attention enabled, deploying with dp=DECODER_TP_SIZE and tp=1"
188
+ BASE_CMD="${BASE_CMD} --data-parallel-size $DECODER_TP_SIZE \
189
+ --tensor-parallel-size 1 --enable-expert-parallel"
190
+ fi
191
192
if [ -n "$model_args" ]; then
193
FULL_CMD="$BASE_CMD $model_args"
tests/v1/kv_connector/nixl_integration/tp_config_sweep_accuracy_test.sh
+1
@@ -10,6 +10,7 @@ configs=(
10
"GPU_MEMORY_UTILIZATION=0.6 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2"
11
"GPU_MEMORY_UTILIZATION=0.8 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # MLA case
12
"GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny"
13
+ "DP_EP=1 GPU_MEMORY_UTILIZATION=0.8 PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 MODEL_NAMES=deepseek-ai/deepseek-vl2-tiny" # MLA+P-TP1, D-DPEP=2 (TP=1)
14
)
15
16
run_tests() {