[CI][BugFix] ShellCheck cleanup to remove baseline and preserve runtime behavior (#34514)

Signed-off-by: junuxyz <216036880+junuxyz@users.noreply.github.com>

tests/standalone_tests/python_only_compile.sh

@@ -6,7 +6,7 @@ set -e
 
 merge_base_commit=$(git merge-base HEAD origin/main)
 echo "INFO: current merge base commit with main: $merge_base_commit"
-git show --oneline -s $merge_base_commit
+git show --oneline -s "$merge_base_commit"
 
 # test whether the metadata.json url is valid, retry each 3 minutes up to 5 times
 # this avoids cumbersome error messages & manual retries in case the precompiled wheel
@@ -40,7 +40,7 @@ for i in {1..5}; do
         fi
     fi
     # failure handling & retry logic
-    if [ $i -eq 5 ]; then
+    if [ "$i" -eq 5 ]; then
         echo "ERROR: metadata is still not available after 5 attempts."
         echo "ERROR: Please check whether the precompiled wheel for commit $merge_base_commit is available."
         echo " NOTE: If $merge_base_commit is a new commit on main, maybe try again after its release pipeline finishes."

tests/v1/ec_connector/integration/run_epd_correctness_test.sh

@@ -24,7 +24,7 @@ MODEL="${MODEL:-Qwen/Qwen2.5-VL-3B-Instruct}"
 # Set 1 to use multimodal prompts; else to use text-only
 USE_MM_PROMPTS="${USE_MM_PROMPTS:-1}"
 MM_FLAG=""
-if [ $USE_MM_PROMPTS = "1" ]; then
+if [ "$USE_MM_PROMPTS" = "1" ]; then
     MM_FLAG="--use_mm_prompts"
 fi
 
@@ -51,7 +51,7 @@ LOG_PATH="${LOG_PATH:-/tmp}"
 BASELINE_FILE="${BASELINE_FILE:-/tmp/vllm_baseline.txt}"
 BASELINE_PD_FILE="${BASELINE_PD_FILE:-/tmp/vllm_epd_baseline.txt}"
 
-mkdir -p $LOG_PATH
+mkdir -p "$LOG_PATH"
 
 # Trap the SIGINT signal (triggered by Ctrl+C)
 trap 'kill $(jobs -pr)' SIGINT SIGTERM EXIT
@@ -87,20 +87,20 @@ run_baseline() {
     # Start baseline instance
     echo "Starting baseline instance on GPU $GPU_SINGLE, port $PORT"
     CUDA_VISIBLE_DEVICES="$GPU_SINGLE" vllm serve "$MODEL" \
-        --port $PORT \
+        --port "$PORT" \
         --enforce-eager \
         --gpu-memory-utilization 0.7 \
         --max-num-seqs 128 \
-        --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
-        > $LOG_PATH/baseline.log 2>&1 &
+        --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
+        > "$LOG_PATH"/baseline.log 2>&1 &
     
     local BASELINE_PID=$!
     
     # Wait for baseline to start
     echo "Waiting for baseline instance to start..."
-    wait_for_server $PORT
+    wait_for_server "$PORT"
 
-    curl http://127.0.0.1:$PORT/v1/models
+    curl http://127.0.0.1:"$PORT"/v1/models
     echo ""
     
     # Run test in baseline mode
@@ -139,14 +139,14 @@ run_epd_1e_1pd() {
     # Start encoder instance
     echo "Starting encoder instance on GPU $GPU_E, port $ENCODE_PORT"
     CUDA_VISIBLE_DEVICES="$GPU_E" vllm serve "$MODEL" \
-        --port $ENCODE_PORT \
+        --port "$ENCODE_PORT" \
         --enforce-eager \
         --gpu-memory-utilization 0.01 \
         --enable-request-id-headers \
         --no-enable-prefix-caching \
         --max-num-batched-tokens 114688 \
         --max-num-seqs 128 \
-        --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+        --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
         --ec-transfer-config '{
             "ec_connector": "ECExampleConnector",
             "ec_role": "ec_producer",
@@ -154,18 +154,18 @@ run_epd_1e_1pd() {
                 "shared_storage_path": "'"$EC_SHARED_STORAGE_PATH"'"
             }
         }' \
-        > $LOG_PATH/1e1pd_encoder.log 2>&1 &
+        > "$LOG_PATH"/1e1pd_encoder.log 2>&1 &
     PIDS+=($!)
     
     # Start prefill+decode instance
     echo "Starting PD instance on GPU $GPU_PD, port $PREFILL_DECODE_PORT"
     CUDA_VISIBLE_DEVICES="$GPU_PD" vllm serve "$MODEL" \
-        --port $PREFILL_DECODE_PORT \
+        --port "$PREFILL_DECODE_PORT" \
         --enforce-eager \
         --gpu-memory-utilization 0.7 \
         --enable-request-id-headers \
         --max-num-seqs 128 \
-        --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+        --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
         --ec-transfer-config '{
             "ec_connector": "ECExampleConnector",
             "ec_role": "ec_consumer",
@@ -173,32 +173,32 @@ run_epd_1e_1pd() {
                 "shared_storage_path": "'"$EC_SHARED_STORAGE_PATH"'"
             }
         }' \
-        > $LOG_PATH/1e1pd_pd.log 2>&1 &
+        > "$LOG_PATH"/1e1pd_pd.log 2>&1 &
     PIDS+=($!)
     
     # Wait for instances to start
     echo "Waiting for encoder instance..."
-    wait_for_server $ENCODE_PORT
+    wait_for_server "$ENCODE_PORT"
     echo "Waiting for PD instance..."
-    wait_for_server $PREFILL_DECODE_PORT
+    wait_for_server "$PREFILL_DECODE_PORT"
 
     # Start proxy
     echo "Starting EPD proxy on port $PROXY_PORT"
     python "${GIT_ROOT}/examples/online_serving/disaggregated_encoder/disagg_epd_proxy.py" \
         --host "0.0.0.0" \
-        --port $PROXY_PORT \
+        --port "$PROXY_PORT" \
         --encode-servers-urls "http://localhost:$ENCODE_PORT" \
         --prefill-servers-urls "disable" \
         --decode-servers-urls "http://localhost:$PREFILL_DECODE_PORT" \
-        > $LOG_PATH/1e1pd_proxy.log 2>&1 &
+        > "$LOG_PATH"/1e1pd_proxy.log 2>&1 &
     PIDS+=($!)
     
     # Wait for proxy
     echo "Waiting for proxy..."
-    wait_for_server $PROXY_PORT
+    wait_for_server "$PROXY_PORT"
 
-    curl http://127.0.0.1:$PROXY_PORT/v1/models
-    curl http://127.0.0.1:$PROXY_PORT/health
+    curl http://127.0.0.1:"$PROXY_PORT"/v1/models
+    curl http://127.0.0.1:"$PROXY_PORT"/health
     echo ""
 
     echo "All EPD (1E+1PD) services are up!"
@@ -217,7 +217,7 @@ run_epd_1e_1pd() {
     echo "✓✓ 1E+1PD Correctness Test finished"
     echo "Stopping EPD (1E+1PD) instances..."
     for pid in "${PIDS[@]}"; do
-        kill $pid 2>/dev/null || true
+        kill "$pid" 2>/dev/null || true
     done
     sleep 2
     cleanup_instances
@@ -244,17 +244,17 @@ run_baseline_1p_1d() {
     CUDA_VISIBLE_DEVICES="$GPU_P" \
     VLLM_NIXL_SIDE_CHANNEL_PORT=5559 \
     vllm serve "$MODEL" \
-        --port $PREFILL_PORT \
+        --port "$PREFILL_PORT" \
         --enforce-eager \
         --gpu-memory-utilization 0.7 \
         --enable-request-id-headers \
         --max-num-seqs 128 \
-        --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+        --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
         --kv-transfer-config '{
             "kv_connector": "NixlConnector",
             "kv_role": "kv_producer"
         }' \
-        > $LOG_PATH/1p1d_prefill.log 2>&1 &
+        > "$LOG_PATH"/1p1d_prefill.log 2>&1 &
     PIDS+=($!)
     
     # Start decode instance
@@ -262,40 +262,40 @@ run_baseline_1p_1d() {
     CUDA_VISIBLE_DEVICES="$GPU_D" \
     VLLM_NIXL_SIDE_CHANNEL_PORT=6000 \
     vllm serve "$MODEL" \
-        --port $DECODE_PORT \
+        --port "$DECODE_PORT" \
         --enforce-eager \
         --gpu-memory-utilization 0.7 \
         --enable-request-id-headers \
         --max-num-seqs 128 \
-        --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+        --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
         --kv-transfer-config '{
             "kv_connector": "NixlConnector",
             "kv_role": "kv_consumer"
         }' \
-        > $LOG_PATH/1p1d_decode.log 2>&1 &
+        > "$LOG_PATH"/1p1d_decode.log 2>&1 &
     PIDS+=($!)
     
     # Wait for instances to start
     echo "Waiting for prefill instance..."
-    wait_for_server $PREFILL_PORT
+    wait_for_server "$PREFILL_PORT"
     echo "Waiting for decode instance..."
-    wait_for_server $DECODE_PORT
+    wait_for_server "$DECODE_PORT"
     
     # Start proxy
     echo "Starting EPD proxy on port $PROXY_PORT"
     python "${GIT_ROOT}/tests/v1/kv_connector/nixl_integration/toy_proxy_server.py" \
         --host "0.0.0.0" \
-        --port $PROXY_PORT \
-        --prefiller-ports $PREFILL_PORT \
-        --decoder-ports $DECODE_PORT \
-        > $LOG_PATH/1p1d_proxy.log 2>&1 &
+        --port "$PROXY_PORT" \
+        --prefiller-ports "$PREFILL_PORT" \
+        --decoder-ports "$DECODE_PORT" \
+        > "$LOG_PATH"/1p1d_proxy.log 2>&1 &
     PIDS+=($!)
     
     # Wait for proxy
     echo "Waiting for proxy..."
-    wait_for_server $PROXY_PORT
+    wait_for_server "$PROXY_PORT"
 
-    curl http://127.0.0.1:$PROXY_PORT/healthcheck
+    curl http://127.0.0.1:"$PROXY_PORT"/healthcheck
     echo ""
 
     echo "All PD (1P+1D) services are up!"
@@ -313,7 +313,7 @@ run_baseline_1p_1d() {
     # Cleanup
     echo "Stopping PD (1P+1D) instances..."
     for pid in "${PIDS[@]}"; do
-        kill $pid 2>/dev/null || true
+        kill "$pid" 2>/dev/null || true
     done
     sleep 2
     cleanup_instances
@@ -339,14 +339,14 @@ run_epd_1e_1p_1d() {
     # Start encoder instance
     echo "Starting encoder instance on GPU $GPU_E, port $ENCODE_PORT"
     CUDA_VISIBLE_DEVICES="$GPU_E" vllm serve "$MODEL" \
-        --port $ENCODE_PORT \
+        --port "$ENCODE_PORT" \
         --enforce-eager \
         --gpu-memory-utilization 0.01 \
         --enable-request-id-headers \
         --no-enable-prefix-caching \
         --max-num-batched-tokens 114688 \
         --max-num-seqs 128 \
-        --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+        --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
         --ec-transfer-config '{
             "ec_connector": "ECExampleConnector",
             "ec_role": "ec_producer",
@@ -354,7 +354,7 @@ run_epd_1e_1p_1d() {
                 "shared_storage_path": "'"$EC_SHARED_STORAGE_PATH"'"
             }
         }' \
-        > $LOG_PATH/1e1p1d_encoder.log 2>&1 &
+        > "$LOG_PATH"/1e1p1d_encoder.log 2>&1 &
     PIDS+=($!)
     
     # Start prefill instance
@@ -362,12 +362,12 @@ run_epd_1e_1p_1d() {
     CUDA_VISIBLE_DEVICES="$GPU_P" \
     VLLM_NIXL_SIDE_CHANNEL_PORT=5559 \
     vllm serve "$MODEL" \
-        --port $PREFILL_PORT \
+        --port "$PREFILL_PORT" \
         --enforce-eager \
         --gpu-memory-utilization 0.7 \
         --enable-request-id-headers \
         --max-num-seqs 128 \
-        --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+        --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
         --ec-transfer-config '{
             "ec_connector": "ECExampleConnector",
             "ec_role": "ec_consumer",
@@ -379,7 +379,7 @@ run_epd_1e_1p_1d() {
             "kv_connector": "NixlConnector",
             "kv_role": "kv_producer"
         }' \
-        > $LOG_PATH/1e1p1d_prefill.log 2>&1 &
+        > "$LOG_PATH"/1e1p1d_prefill.log 2>&1 &
     PIDS+=($!)
     
     # Start decode instance
@@ -387,44 +387,44 @@ run_epd_1e_1p_1d() {
     CUDA_VISIBLE_DEVICES="$GPU_D" \
     VLLM_NIXL_SIDE_CHANNEL_PORT=6000 \
     vllm serve "$MODEL" \
-        --port $DECODE_PORT \
+        --port "$DECODE_PORT" \
         --enforce-eager \
         --gpu-memory-utilization 0.7 \
         --enable-request-id-headers \
         --max-num-seqs 128 \
-        --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+        --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
         --kv-transfer-config '{
             "kv_connector": "NixlConnector",
             "kv_role": "kv_consumer"
         }' \
-        > $LOG_PATH/1e1p1d_decode.log 2>&1 &
+        > "$LOG_PATH"/1e1p1d_decode.log 2>&1 &
     PIDS+=($!)
     
     # Wait for instances to start
     echo "Waiting for encoder instance..."
-    wait_for_server $ENCODE_PORT
+    wait_for_server "$ENCODE_PORT"
     echo "Waiting for prefill instance..."
-    wait_for_server $PREFILL_PORT
+    wait_for_server "$PREFILL_PORT"
     echo "Waiting for decode instance..."
-    wait_for_server $DECODE_PORT
+    wait_for_server "$DECODE_PORT"
     
     # Start proxy
     echo "Starting EPD proxy on port $PROXY_PORT"
     python "${GIT_ROOT}/examples/online_serving/disaggregated_encoder/disagg_epd_proxy.py" \
         --host "0.0.0.0" \
-        --port $PROXY_PORT \
+        --port "$PROXY_PORT" \
         --encode-servers-urls "http://localhost:$ENCODE_PORT" \
         --prefill-servers-urls "http://localhost:$PREFILL_PORT" \
         --decode-servers-urls "http://localhost:$DECODE_PORT" \
-        > $LOG_PATH/1e1p1d_proxy.log 2>&1 &
+        > "$LOG_PATH"/1e1p1d_proxy.log 2>&1 &
     PIDS+=($!)
     
     # Wait for proxy
     echo "Waiting for proxy..."
-    wait_for_server $PROXY_PORT
+    wait_for_server "$PROXY_PORT"
 
-    curl http://127.0.0.1:$PROXY_PORT/v1/models
-    curl http://127.0.0.1:$PROXY_PORT/health
+    curl http://127.0.0.1:"$PROXY_PORT"/v1/models
+    curl http://127.0.0.1:"$PROXY_PORT"/health
     echo ""
 
     echo "All EPD (1E+1P+1D) services are up!"
@@ -443,7 +443,7 @@ run_epd_1e_1p_1d() {
     echo "✓✓ 1E+1P+1D Correctness Test finished"
     echo "Stopping EPD (1E+1P+1D) instances..."
     for pid in "${PIDS[@]}"; do
-        kill $pid 2>/dev/null || true
+        kill "$pid" 2>/dev/null || true
     done
     sleep 2
     cleanup_instances

tests/v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh

@@ -32,9 +32,14 @@ run_tests() {
 
   echo "=== Running tests (${label}) ==="
   for cfg in "${configs[@]}"; do
+    local -a cfg_parts extra_args_parts
+    read -r -a cfg_parts <<< "$cfg"
+    read -r -a extra_args_parts <<< "$extra_args"
+
     echo "-> Running with ${cfg} ${extra_args:+and ${extra_args}}"
     # Use 'env' to safely set variables without eval
-    if ! env ${cfg} bash "${SCRIPT}" ${extra_args}; then
+    # keep argv splitting safe and SC2086-clean via arrays.
+    if ! env "${cfg_parts[@]}" bash "${SCRIPT}" "${extra_args_parts[@]}"; then
       echo "❌ Test failed for config: ${cfg} ${extra_args:+(${extra_args})}"
       exit 1
     fi

tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh

@@ -109,9 +109,9 @@ get_model_args() {
 
 get_num_gpus() {
   if [[ "$SMI_BIN" == *"nvidia"* ]]; then
-    echo "$($SMI_BIN --query-gpu=name --format=csv,noheader | wc -l)"
+    $SMI_BIN --query-gpu=name --format=csv,noheader | wc -l
   elif [[ "$SMI_BIN" == *"rocm"* ]]; then
-    echo "$($SMI_BIN -l | grep GPU | wc -l)"
+    $SMI_BIN -l | grep -c GPU
   else
     # works for non-cuda platforms,
     # assuming at least 1 device and
@@ -182,7 +182,7 @@ run_tests_for_model() {
 
     # Store host and port for proxy configuration
     PREFILL_HOSTS+=("localhost")
-    PREFILL_PORTS+=($PORT)
+    PREFILL_PORTS+=("$PORT")
   done
 
   # Start decode instances
@@ -237,30 +237,30 @@ run_tests_for_model() {
 
     # Store host and port for proxy configuration
     DECODE_HOSTS+=("localhost")
-    DECODE_PORTS+=($PORT)
+    DECODE_PORTS+=("$PORT")
   done
 
   # Wait for all instances to start
   for PORT in "${PREFILL_PORTS[@]}"; do
     echo "Waiting for prefill instance on port $PORT to start..."
-    wait_for_server $PORT
+    wait_for_server "$PORT"
   done
 
   for PORT in "${DECODE_PORTS[@]}"; do
     echo "Waiting for decode instance on port $PORT to start..."
-    wait_for_server $PORT
+    wait_for_server "$PORT"
   done
 
   # Build the command for the proxy server with all the hosts and ports
   PROXY_CMD="python3 ${GIT_ROOT}/tests/v1/kv_connector/nixl_integration/toy_proxy_server.py --port 8192"
 
   # Add all prefill hosts and ports
-  PROXY_CMD+=" --prefiller-hosts ${PREFILL_HOSTS[@]}"
-  PROXY_CMD+=" --prefiller-ports ${PREFILL_PORTS[@]}"
+  PROXY_CMD+=" --prefiller-hosts ${PREFILL_HOSTS[*]}"
+  PROXY_CMD+=" --prefiller-ports ${PREFILL_PORTS[*]}"
 
   # Add all decode hosts and ports
-  PROXY_CMD+=" --decoder-hosts ${DECODE_HOSTS[@]}"
-  PROXY_CMD+=" --decoder-ports ${DECODE_PORTS[@]}"
+  PROXY_CMD+=" --decoder-hosts ${DECODE_HOSTS[*]}"
+  PROXY_CMD+=" --decoder-ports ${DECODE_PORTS[*]}"
 
   # Start the proxy server
   echo "Starting proxy server with command: $PROXY_CMD"
@@ -271,7 +271,7 @@ run_tests_for_model() {
 
   # Run lm eval for this model
   echo "Running tests for $model_name"
-  TEST_MODEL=$model_name python3 -m pytest -s -x ${GIT_ROOT}/tests/v1/kv_connector/nixl_integration/test_accuracy.py
+  TEST_MODEL=$model_name python3 -m pytest -s -x "${GIT_ROOT}"/tests/v1/kv_connector/nixl_integration/test_accuracy.py
 
   # Clean up before running next model
   cleanup_instances

tests/v1/kv_connector/nixl_integration/run_edge_case_test.sh

@@ -114,10 +114,10 @@ run_tests_for_model() {
   eval "$FULL_CMD &"
 
   # Wait for all instances to start
-  echo "Waiting for prefill instance on port $PORT to start..."
-  wait_for_server $PREFILL_PORT
-  echo "Waiting for decode instance on port $PORT to start..."
-  wait_for_server $DECODE_PORT
+  echo "Waiting for prefill instance on port $PREFILL_PORT to start..."
+  wait_for_server "$PREFILL_PORT"
+  echo "Waiting for decode instance on port $DECODE_PORT to start..."
+  wait_for_server "$DECODE_PORT"
 
   # Build the command for the proxy server with all the hosts and ports
   PROXY_PORT=8192
@@ -133,7 +133,7 @@ run_tests_for_model() {
 
   # Run lm eval for this model
   echo "Running tests for $model_name"
-  PREFILL_PORT=$PREFILL_PORT DECODE_PORT=$DECODE_PORT PROXY_PORT=$PROXY_PORT python -m pytest -s -v ${GIT_ROOT}/tests/v1/kv_connector/nixl_integration/test_edge_cases.py
+  PREFILL_PORT=$PREFILL_PORT DECODE_PORT=$DECODE_PORT PROXY_PORT=$PROXY_PORT python -m pytest -s -v "${GIT_ROOT}"/tests/v1/kv_connector/nixl_integration/test_edge_cases.py
 
   # Clean up before running next model
   cleanup_instances

tests/v1/kv_connector/nixl_integration/run_tpu_disagg_accuracy_test.sh

@@ -63,8 +63,8 @@ launch_baseline() {
       --block-size ${BLOCK_SIZE} \
       --gpu-memory-utilization 0.5 \
       --enforce-eager"
-  echo ${BASELINE_BASE_CMD}
-  ssh -tt ${BASELINE_HOST} "${BASELINE_BASE_CMD}" &
+  echo "${BASELINE_BASE_CMD}"
+  ssh -tt "${BASELINE_HOST}" "${BASELINE_BASE_CMD}" &
 }
 
 launch_pd() {
@@ -103,17 +103,17 @@ launch_pd() {
       --gpu-memory-utilization 0.5 \
       --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\",\"kv_buffer_device\":\"cpu\"}'"
 
-  echo ${PREFILL_BASE_CMD}
-  echo ${DECODE_BASE_CMD}
+  echo "${PREFILL_BASE_CMD}"
+  echo "${DECODE_BASE_CMD}"
   sleep 2
 
   # execute on hosts
-  ssh -tt ${PREFILL_HOST} "${PREFILL_BASE_CMD}" &
-  ssh -tt ${DECODE_HOST} "${DECODE_BASE_CMD}" &
+  ssh -tt "${PREFILL_HOST}" "${PREFILL_BASE_CMD}" &
+  ssh -tt "${DECODE_HOST}" "${DECODE_BASE_CMD}" &
   sleep 1
-  wait_for_server ${PREFILL_HOST} ${PREFILL_PORT}
+  wait_for_server "${PREFILL_HOST}" "${PREFILL_PORT}"
   sleep 1
-  wait_for_server ${DECODE_HOST} ${DECODE_PORT}
+  wait_for_server "${DECODE_HOST}" "${DECODE_PORT}"
   sleep 1
 }
 
@@ -123,21 +123,21 @@ launch_pd_proxy(){
   --prefiller-host ${PREFILL_HOST} --prefiller-port ${PREFILL_PORT} \
   --decoder-host ${DECODE_HOST} --decoder-port ${DECODE_PORT} \
   --host=${PROXY_HOST} --port ${PROXY_PORT}"
-  echo ${PROXY_BASE_CMD}
-  ssh -tt ${PROXY_HOST} "${PROXY_BASE_CMD}" &
+  echo "${PROXY_BASE_CMD}"
+  ssh -tt "${PROXY_HOST}" "${PROXY_BASE_CMD}" &
 }
 
 run_tests(){
   local service_url=$1
   local mode=$2
-  python3 ${EXP_ROOT}/test_disagg_accuracy.py --service_url=${service_url} --model_name=${MODEL_NAME} --mode=${mode} --file_name=${OUTPUT_FILE}
+  python3 "${EXP_ROOT}"/test_disagg_accuracy.py --service_url="${service_url}" --model_name="${MODEL_NAME}" --mode="${mode}" --file_name="${OUTPUT_FILE}"
 }
 
 
 # run non-disagg. baseline & save outputs
 launch_baseline
 sleep 2
-wait_for_server ${BASELINE_HOST} ${BASELINE_PORT}
+wait_for_server "${BASELINE_HOST}" "${BASELINE_PORT}"
 run_tests "http://${BASELINE_HOST}:${BASELINE_PORT}" "baseline"
 cleanup
 sleep 10
@@ -150,7 +150,7 @@ sleep 10
 run_tests "http://${PROXY_HOST}:${PROXY_PORT}" "disagg"
 echo "-----P/D success----"
 
-rm ${OUTPUT_FILE}
+rm "${OUTPUT_FILE}"
 cleanup
 
 exit 0
\ No newline at end of file

tests/v1/kv_connector/nixl_integration/run_tpu_edge_case_test.sh

@@ -86,17 +86,17 @@ launch_pd() {
       --gpu-memory-utilization 0.5 \
       --kv-transfer-config '{\"kv_connector\":\"NixlConnector\",\"kv_role\":\"kv_both\",\"kv_buffer_device\":\"cpu\"}'"
 
-  echo ${PREFILL_BASE_CMD}
-  echo ${DECODE_BASE_CMD}
+  echo "${PREFILL_BASE_CMD}"
+  echo "${DECODE_BASE_CMD}"
   sleep 2
 
   # execute on hosts
-  ssh -tt ${PREFILL_HOST} "${PREFILL_BASE_CMD}" &
-  ssh -tt ${DECODE_HOST} "${DECODE_BASE_CMD}" &
+  ssh -tt "${PREFILL_HOST}" "${PREFILL_BASE_CMD}" &
+  ssh -tt "${DECODE_HOST}" "${DECODE_BASE_CMD}" &
   sleep 1
-  wait_for_server ${PREFILL_HOST} ${PREFILL_PORT}
+  wait_for_server "${PREFILL_HOST}" "${PREFILL_PORT}"
   sleep 1
-  wait_for_server ${DECODE_HOST} ${DECODE_PORT}
+  wait_for_server "${DECODE_HOST}" "${DECODE_PORT}"
   sleep 1
 }
 
@@ -106,8 +106,8 @@ launch_pd_proxy(){
   --prefiller-host ${PREFILL_HOST} --prefiller-port ${PREFILL_PORT} \
   --decoder-host ${DECODE_HOST} --decoder-port ${DECODE_PORT} \
   --host=${PROXY_HOST} --port ${PROXY_PORT}"
-  echo ${PROXY_BASE_CMD}
-  ssh -tt ${PROXY_HOST} "${PROXY_BASE_CMD}" &
+  echo "${PROXY_BASE_CMD}"
+  ssh -tt "${PROXY_HOST}" "${PROXY_BASE_CMD}" &
 }
 
 
@@ -121,4 +121,4 @@ PREFILL_PORT=${PREFILL_PORT} \
 DECODE_HOST=${DECODE_HOST} \
 DECODE_PORT=${DECODE_PORT} \
 PROXY_HOST=${PROXY_HOST} \
-PROXY_PORT=${PROXY_PORT} python -m pytest -s -v ${GIT_ROOT}/tests/v1/kv_connector/nixl_integration/test_edge_cases.py
\ No newline at end of file
+PROXY_PORT=${PROXY_PORT} python -m pytest -s -v "${GIT_ROOT}"/tests/v1/kv_connector/nixl_integration/test_edge_cases.py

tools/ep_kernels/elastic_ep/install_eep_libraries.sh

@@ -23,7 +23,7 @@ while getopts "w:n" opt; do
 done
 
 if [ ! -d "$WORKSPACE" ]; then
-    mkdir -p $WORKSPACE
+    mkdir -p "$WORKSPACE"
 fi
 
 
@@ -31,7 +31,7 @@ fi
 pip3 install cmake torch ninja
 
 # build nvshmem
-pushd $WORKSPACE
+pushd "$WORKSPACE"
 # Reset NVSHMEM build if requested
 if [ "$INSTALL_NVSHMEM" = true ]; then
     mkdir -p nvshmem_src
@@ -69,15 +69,15 @@ export NVSHMEM_BUILD_HYDRA_LAUNCHER=0
 export NVSHMEM_BUILD_TXZ_PACKAGE=0
 export NVSHMEM_TIMEOUT_DEVICE_POLLING=0
 
-cmake -G Ninja -S . -B $WORKSPACE/nvshmem_build/ -DCMAKE_INSTALL_PREFIX=$WORKSPACE/nvshmem_install
-cmake --build $WORKSPACE/nvshmem_build/ --target install
+cmake -G Ninja -S . -B "$WORKSPACE"/nvshmem_build/ -DCMAKE_INSTALL_PREFIX="$WORKSPACE"/nvshmem_install
+cmake --build "$WORKSPACE"/nvshmem_build/ --target install
 
 popd
 
 export CMAKE_PREFIX_PATH=$WORKSPACE/nvshmem_install:$CMAKE_PREFIX_PATH
 
 # build and install pplx, require pytorch installed
-pushd $WORKSPACE
+pushd "$WORKSPACE"
 git clone https://github.com/ppl-ai/pplx-kernels
 cd pplx-kernels
 # see https://github.com/pypa/pip/issues/9955#issuecomment-838065925

tools/ep_kernels/install_python_libraries.sh

@@ -14,7 +14,7 @@ DEEPEP_COMMIT_HASH=${DEEPEP_COMMIT_HASH:-"73b6ea4"}
 NVSHMEM_VER=${NVSHMEM_VER:-"3.3.24"}  # Default supports both CUDA 12 and 13
 WORKSPACE=${WORKSPACE:-$(pwd)/ep_kernels_workspace}
 MODE=${MODE:-install}
-CUDA_VERSION_MAJOR=$(${CUDA_HOME}/bin/nvcc --version | egrep -o "release [0-9]+" | cut -d ' ' -f 2)
+CUDA_VERSION_MAJOR=$("${CUDA_HOME}"/bin/nvcc --version | grep -E -o "release [0-9]+" | cut -d ' ' -f 2)
 
 # Parse arguments
 while [[ $# -gt 0 ]]; do

tools/flashinfer-build.sh

@@ -5,8 +5,6 @@ set -ex
 
 # FlashInfer configuration
 FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
-FLASHINFER_GIT_REF="${FLASHINFER_GIT_REF}"
-CUDA_VERSION="${CUDA_VERSION}"
 BUILD_WHEEL="${BUILD_WHEEL:-true}"
 
 if [[ -z "${FLASHINFER_GIT_REF}" ]]; then
@@ -23,7 +21,7 @@ echo "🏗️  Building FlashInfer ${FLASHINFER_GIT_REF} for CUDA ${CUDA_VERSION
 
 # Clone FlashInfer
 git clone --depth 1 --recursive --shallow-submodules \
-    --branch ${FLASHINFER_GIT_REF} \
+    --branch "${FLASHINFER_GIT_REF}" \
     ${FLASHINFER_GIT_REPO} flashinfer
 
 # Set CUDA arch list based on CUDA version
@@ -44,7 +42,7 @@ echo "🏗️ Building FlashInfer AOT for arches: ${FI_TORCH_CUDA_ARCH_LIST}"
 
 pushd flashinfer
     # Make sure the wheel is built for the correct CUDA version
-    export UV_TORCH_BACKEND=cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
+    export UV_TORCH_BACKEND=cu$(echo "$CUDA_VERSION" | cut -d. -f1,2 | tr -d '.')
 
     # Build AOT kernels
     export TORCH_CUDA_ARCH_LIST="${FI_TORCH_CUDA_ARCH_LIST}"
@@ -63,4 +61,4 @@ pushd flashinfer
 popd
 
 # Cleanup
-rm -rf flashinfer
\ No newline at end of file
+rm -rf flashinfer

tools/install_deepgemm.sh

@@ -65,7 +65,7 @@ fi
 
 # Extract major and minor version numbers
 CUDA_MAJOR="${CUDA_VERSION%%.*}"
-CUDA_MINOR="${CUDA_VERSION#${CUDA_MAJOR}.}"
+CUDA_MINOR="${CUDA_VERSION#"${CUDA_MAJOR}".}"
 CUDA_MINOR="${CUDA_MINOR%%.*}"
 echo "CUDA version: $CUDA_VERSION (major: $CUDA_MAJOR, minor: $CUDA_MINOR)"
 
@@ -92,7 +92,7 @@ git checkout "$DEEPGEMM_GIT_REF"
 
 # Clean previous build artifacts
 # (Based on https://github.com/deepseek-ai/DeepGEMM/blob/main/install.sh)
-rm -rf build dist *.egg-info
+rm -rf -- build dist *.egg-info 2>/dev/null || true
 
 # Build wheel
 echo "🏗️  Building DeepGEMM wheel..."

tools/pre_commit/shellcheck.baseline

-benchmarks/auto_tune/auto_tune.sh:SC2034
-benchmarks/auto_tune/auto_tune.sh:SC2086
-benchmarks/auto_tune/batch_auto_tune.sh:SC2086
-benchmarks/run_structured_output_benchmark.sh:SC2028
-benchmarks/run_structured_output_benchmark.sh:SC2034
-benchmarks/run_structured_output_benchmark.sh:SC2086
-.buildkite/image_build/image_build_cpu_arm64.sh:SC2086
-.buildkite/image_build/image_build_cpu.sh:SC2086
-.buildkite/image_build/image_build_hpu.sh:SC2086
-.buildkite/lm-eval-harness/run-lm-eval-chartqa-vllm-vlm-baseline.sh:SC2086
-.buildkite/lm-eval-harness/run-lm-eval-mmlupro-vllm-baseline.sh:SC2034
-.buildkite/performance-benchmarks/scripts/run-performance-benchmarks.sh:SC2027
-.buildkite/performance-benchmarks/scripts/run-performance-benchmarks.sh:SC2086
-.buildkite/performance-benchmarks/scripts/run-performance-benchmarks.sh:SC2126
-.buildkite/scripts/annotate-rocm-release.sh:SC2086
-.buildkite/scripts/cache-rocm-base-wheels.sh:SC2012
-.buildkite/scripts/cherry-pick-from-milestone.sh:SC2064
-.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh:SC2086
-.buildkite/scripts/hardware_ci/run-cpu-test.sh:SC2086
-.buildkite/scripts/hardware_ci/run-hpu-test.sh:SC2086
-.buildkite/scripts/hardware_ci/run-npu-test.sh:SC1090
-.buildkite/scripts/hardware_ci/run-npu-test.sh:SC2006
-.buildkite/scripts/hardware_ci/run-npu-test.sh:SC2086
-.buildkite/scripts/hardware_ci/run-npu-test.sh:SC2181
-.buildkite/scripts/hardware_ci/run-xpu-test.sh:SC2086
-.buildkite/scripts/push-nightly-builds.sh:SC2086
-.buildkite/scripts/run-multi-node-test.sh:SC2086
-.buildkite/scripts/run-multi-node-test.sh:SC2089
-.buildkite/scripts/run-multi-node-test.sh:SC2090
-.buildkite/scripts/run-prime-rl-test.sh:SC2086
-.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh:SC2086
-.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh:SC2086
-.buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh:SC2086
-.buildkite/scripts/tpu/docker_run_bm.sh:SC1090
-.buildkite/scripts/tpu/docker_run_bm.sh:SC2086
-.buildkite/scripts/tpu/run_bm.sh:SC2034
-.buildkite/scripts/tpu/run_bm.sh:SC2086
-.buildkite/scripts/upload-nightly-wheels.sh:SC2086
-.buildkite/scripts/upload-nightly-wheels.sh:SC2115
-.buildkite/scripts/upload-nightly-wheels.sh:SC2236
-.buildkite/scripts/upload-release-wheels-pypi.sh:SC2086
-.buildkite/scripts/upload-rocm-wheels.sh:SC2012
-examples/online_serving/disaggregated_encoder/disagg_1e1p1d_example.sh:SC2086
-examples/online_serving/disaggregated_encoder/disagg_1e1pd_example.sh:SC2086
-examples/online_serving/disaggregated_prefill.sh:SC2086
-examples/online_serving/disaggregated_serving/kv_events.sh:SC2086
-examples/online_serving/disaggregated_serving/mooncake_connector/run_mooncake_connector.sh:SC2046
-examples/online_serving/disaggregated_serving/mooncake_connector/run_mooncake_connector.sh:SC2086
-examples/online_serving/disaggregated_serving/mooncake_connector/run_mooncake_connector.sh:SC2317
-examples/online_serving/disaggregated_serving_p2p_nccl_xpyd/disagg_example_p2p_nccl_xpyd.sh:SC2046
-examples/online_serving/disaggregated_serving_p2p_nccl_xpyd/disagg_example_p2p_nccl_xpyd.sh:SC2086
-examples/online_serving/disaggregated_serving_p2p_nccl_xpyd/disagg_example_p2p_nccl_xpyd.sh:SC2317
-examples/online_serving/elastic_ep/bench.sh:SC2086
-examples/online_serving/elastic_ep/serve_deepseek_v2.sh:SC2086
-examples/online_serving/multi-node-serving.sh:SC2006
-examples/online_serving/multi-node-serving.sh:SC2086
-examples/online_serving/multi-node-serving.sh:SC2181
-examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh:SC2046
-examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh:SC2126
-examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh:SC2181
-examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh:SC2206
-examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_vllm_launcher.sh:SC2086
-examples/pooling/embed/openai_embedding_long_text/service.sh:SC2086
-tests/standalone_tests/python_only_compile.sh:SC2086
-tests/v1/ec_connector/integration/run_epd_correctness_test.sh:SC2086
-tests/v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh:SC2086
-tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh:SC2005
-tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh:SC2086
-tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh:SC2124
-tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh:SC2126
-tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh:SC2206
-tests/v1/kv_connector/nixl_integration/run_edge_case_test.sh:SC2086
-tests/v1/kv_connector/nixl_integration/run_edge_case_test.sh:SC2153
-tests/v1/kv_connector/nixl_integration/run_tpu_disagg_accuracy_test.sh:SC2086
-tests/v1/kv_connector/nixl_integration/run_tpu_disagg_accuracy_test.sh:SC2089
-tests/v1/kv_connector/nixl_integration/run_tpu_disagg_accuracy_test.sh:SC2090
-tests/v1/kv_connector/nixl_integration/run_tpu_edge_case_test.sh:SC2086
-tests/v1/kv_connector/nixl_integration/run_tpu_edge_case_test.sh:SC2089
-tests/v1/kv_connector/nixl_integration/run_tpu_edge_case_test.sh:SC2090
-tools/ep_kernels/elastic_ep/install_eep_libraries.sh:SC2086
-tools/ep_kernels/install_python_libraries.sh:SC2086
-tools/ep_kernels/install_python_libraries.sh:SC2196
-tools/flashinfer-build.sh:SC2086
-tools/flashinfer-build.sh:SC2269
-tools/install_deepgemm.sh:SC2035
-tools/install_deepgemm.sh:SC2295
-tools/pre_commit/shellcheck.sh:SC2016
-tools/vllm-rocm/generate-rocm-wheels-root-index.sh:SC2295
-tools/vllm-tpu/build.sh:SC2145

tools/pre_commit/shellcheck.sh

@@ -2,7 +2,6 @@
 set -euo pipefail
 
 scversion="stable"
-baseline="tools/pre_commit/shellcheck.baseline"
 
 if [ -d "shellcheck-${scversion}" ]; then
     export PATH="$PATH:$(pwd)/shellcheck-${scversion}"
@@ -20,38 +19,6 @@ if ! [ -x "$(command -v shellcheck)" ]; then
 fi
 
 # TODO - fix warnings in .buildkite/scripts/hardware_ci/run-amd-test.sh
-# collects warnings as "file:SCcode" pairs for baseline comparison.
-collect() {
-  find . -path ./.git -prune -o -name "*.sh" \
-    -not -path "./.buildkite/scripts/hardware_ci/run-amd-test.sh" -print0 | \
-    xargs -0 sh -c 'for f in "$@"; do git check-ignore -q "$f" || shellcheck -s bash -f gcc "$f" || true; done' -- | \
-    sed -nE 's|^\./||; s|^([^:]+):[0-9]+:[0-9]+:.*\[(SC[0-9]+)\]$|\1:\2|p' | \
-    sort -u
-}
-
-if [[ "${1:-}" == "--generate-baseline" ]]; then
-  collect > "$baseline"
-  echo "Wrote baseline to $baseline"
-  exit 0
-fi
-
-if [[ ! -f "$baseline" ]]; then
-  echo "Baseline not found: $baseline (run: $0 --generate-baseline)"
-  exit 1
-fi
-
-current="$(mktemp)"
-trap 'rm -f "$current"' EXIT
-collect > "$current"
-
-# finds new warnings not in baseline
-new_errors="$(comm -23 "$current" <(sort -u "$baseline") || true)"
-if [ -n "$new_errors" ]; then
-  echo "$new_errors" | cut -d: -f1 | sort -u | while IFS= read -r file; do
-    if [[ -f "$file" ]]; then
-      codes=$(echo "$new_errors" | awk -F: -v f="$file" '$1==f {print $2}' | paste -sd ',' -)
-      shellcheck -s bash --include="$codes" "$file" 2>&1 || true
-    fi
-  done
-  exit 1
-fi
+find . -path ./.git -prune -o -name "*.sh" \
+  -not -path "./.buildkite/scripts/hardware_ci/run-amd-test.sh" -print0 | \
+  xargs -0 sh -c "for f in \"\$@\"; do git check-ignore -q \"\$f\" || shellcheck -s bash \"\$f\"; done" --

tools/vllm-rocm/generate-rocm-wheels-root-index.sh

@@ -190,7 +190,7 @@ echo ""
 # List what would be uploaded
 echo "Files to upload:"
 find "$WORK_DIR/output" -name "*.html" -type f | while read -r file; do
-    rel_path="${file#$WORK_DIR/output/}"
+    rel_path="${file#"$WORK_DIR"/output/}"
     echo "  rocm/$rel_path"
 done
 echo ""

tools/vllm-tpu/build.sh

@@ -38,7 +38,7 @@ if ! grep -q "name = \"vllm-tpu\"" "$PYPROJECT_FILE"; then
     cp "$PYPROJECT_FILE" "${PYPROJECT_FILE}.bak"
     sed -i '0,/^name = "vllm"/s//name = "vllm-tpu"/' "$PYPROJECT_FILE"
 
-    echo "Patching ${CHANGE_FILE_LIST[@]} vllm to vllm-tpu..."
+    echo "Patching ${CHANGE_FILE_LIST[*]} vllm to vllm-tpu..."
     # patching
     #   importlib.metadata.version('vllm') -> importlib.metadata.version('vllm-tpu')
     #   importlib.metadata.version("vllm") -> importlib.metadata.version("vllm-tpu")