Add ucx test

Signed-off-by: Daniel Huang <daniel1.huang@intel.com>

Author: pi314ever

tests/full_tests/ci_gsm8k_tests.sh

@@ -290,6 +290,15 @@ run_pd_disaggregate_nixl_libfabric_test() {
     echo "✅ PD disaggregate through NIXL libfabric."
 }
 
+run_pd_disaggregate_nixl_ucx_test() {
+    echo "➡️ Testing PD disaggregate through NIXL UCX."
+    git clone https://github.com/intel-staging/ucx.git -b intel_gaudi_gdr_enabling_0
+    bash ucx/setup_nixl_ucx.sh
+    rm -rf ucx
+    cd ${VLLM_GAUDI_PREFIX}/tests/unit_tests; DECODER_TP_SIZE=1 NIXL_BUFFER_DEVICE=hpu VLLM_NIXL_BACKEND=UCX bash run_accuracy_test.sh
+    echo "✅ PD disaggregate through NIXL UCX."
+}
+
 # sleep mode
 run_sleep_mode_test() {
     echo "Testing basic model with sleep mode / wake up functionality"

tests/unit_tests/run_accuracy_test.sh

@@ -29,8 +29,12 @@ export PT_HPU_LAZY_MODE=1
 NIXL_BUFFER_DEVICE=${NIXL_BUFFER_DEVICE:-"cpu"}
 VLLM_NIXL_BACKEND=${VLLM_NIXL_BACKEND:-"UCX"}
 
+UCX_TLS="tcp"
 if [ "$VLLM_NIXL_BACKEND" == "UCX" ]; then
   export VLLM_NIXL_DEVICE_TO_DEVICE=false
+  if [ "$NIXL_BUFFER_DEVICE" == "hpu" ]; then
+    UCX_TLS="gaudi_gdr,ib,rc,ud"
+  fi
 else
   export VLLM_NIXL_DEVICE_TO_DEVICE=true
 fi
@@ -42,8 +46,7 @@ PREFILLER_TP_SIZE=${PREFILLER_TP_SIZE:-1}
 DECODER_TP_SIZE=${DECODER_TP_SIZE:-2}
 
 # Find the git repository root directory
-#GIT_ROOT=$(git rev-parse --show-toplevel)
-GIT_ROOT="/home/vllm-nixl/vllm"
+GIT_ROOT=$(git rev-parse --show-toplevel)
 
 #SMI_BIN=$(which nvidia-smi || which rocm-smi)
 
@@ -116,7 +119,7 @@ run_tests_for_model() {
     echo "Starting prefill instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=0 UCX_TLS=tcp VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="RANK=0 UCX_TLS=$UCX_TLS VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --enforce-eager \
     --max_num_batched_tokens 8192 \
@@ -149,7 +152,7 @@ run_tests_for_model() {
     echo "Starting decode instance $i on GPU $GPU_ID, port $PORT"
 
     # Build the command with or without model-specific args
-    BASE_CMD="RANK=1 UCX_TLS=tcp VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
+    BASE_CMD="RANK=1 UCX_TLS=$UCX_TLS VLLM_NIXL_SIDE_CHANNEL_PORT=$SIDE_CHANNEL_PORT vllm serve $model_name \
     --port $PORT \
     --enforce-eager \
     --max_num_batched_tokens 8192 \