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
[Bugfix] Fix GPU_ID issue in test script (#26442)
Signed-off-by: Chendi Xue <chendi.xue@intel.com>
Chendi.Xue committed
Oct 12, 2025 at 06:39 UTC
9bb38130cb19eb084d39f269cbeae2952789fafd
1 file changed
+12
-1
tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh
+12
-1
@@ -101,6 +101,12 @@ run_tests_for_model() {
101
for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
102
# Calculate GPU ID - we'll distribute across available GPUs
103
GPU_ID=$((i % $(get_num_gpus)))
104
+ NEXT_GPU=${GPU_ID}
105
+ # If PREFILLER_TP_SIZE is more than 1
106
+ for (( j=1; j < PREFILLER_TP_SIZE; j++ )); do
107
+ NEXT_GPU=$(((GPU_ID + j) % $(get_num_gpus)))
108
+ GPU_ID="${GPU_ID},${NEXT_GPU}"
109
+ done
110
111
# Calculate port number (base port + instance number)
112
PORT=$((8100 + i))
@@ -136,7 +142,12 @@ run_tests_for_model() {
142
# Start decode instances
143
for i in $(seq 0 $((NUM_DECODE_INSTANCES-1))); do
144
# Calculate GPU ID - we'll distribute across available GPUs, starting from after prefill GPUs
139
- GPU_ID=$(((i + NUM_PREFILL_INSTANCES) % $(get_num_gpus)))
145
+ GPU_ID=$(((i + NEXT_GPU + 1) % $(get_num_gpus)))
146
+ # If DECODER_TP_SIZE is more than 1
147
+ for (( j=1; j < DECODER_TP_SIZE; j++ )); do
148
+ NEXT_GPU=$(((GPU_ID + j) % $(get_num_gpus)))
149
+ GPU_ID="${GPU_ID},${NEXT_GPU}"
150
+ done
151
# Calculate port number (base port + instance number)
152
PORT=$((8200 + i))
153
# Calculate side channel port