Add splitwise deployment with using rdma

Author: juncaipeng

.github/workflows/_unit_test_coverage.yml

@@ -105,6 +105,7 @@ jobs:
           FD_CACHE_QUEUE_PORT=$((8098 + DEVICE_PORT * 100))
           FD_ROUTER_PORT=$((8048 + DEVICE_PORT * 100))
           FD_CONNECTOR_PORT=$((8038 + DEVICE_PORT * 100))
+          FD_RDMA_PORT=$((9008 + DEVICE_PORT * 100))
           echo "Test ENV Parameter:"
           echo "========================================================="
           echo "FLASK_PORT=${FLASK_PORT}"
@@ -114,6 +115,7 @@ jobs:
           echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
           echo "FD_ROUTER_PORT=${FD_ROUTER_PORT}"
           echo "FD_CONNECTOR_PORT=${FD_CONNECTOR_PORT}"
+          echo "FD_RDMA_PORT=${FD_RDMA_PORT}"
           echo "DEVICES=${DEVICES}"
           echo "========================================================="
 
@@ -165,6 +167,7 @@ jobs:
           -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
           -e "FD_ROUTER_PORT=${FD_ROUTER_PORT}" \
           -e "FD_CONNECTOR_PORT=${FD_CONNECTOR_PORT}" \
+          -e "FD_RDMA_PORT=${FD_RDMA_PORT}" \
           -e TZ="Asia/Shanghai" \
           -e "fd_wheel_url=${fd_wheel_url}" \
           -e "BASE_REF=${BASE_REF}" \

tests/e2e/test_ernie_03b_pd_router_v1.py renamed to tests/e2e/test_ernie_03b_pd_router_v1_ipc.py

@@ -12,8 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# Test splitwise deployment which uses local_scheduler + router,
-# and ENABLE_V1_KVCACHE_SCHEDULER is 1
+# Test splitwise deployment: use local_scheduler + router,
+# set ENABLE_V1_KVCACHE_SCHEDULER is 1, use ipc to transfer cache.
 
 import json
 import os
@@ -111,7 +111,7 @@ def setup_and_run_server():
     env_prefill["CUDA_VISIBLE_DEVICES"] = "0"
     env_prefill["ENABLE_V1_KVCACHE_SCHEDULER"] = "1"
     env_prefill["FD_LOG_DIR"] = "log_prefill"
-    prefill_log_path = "server.log"
+    prefill_log_path = "prefill.log"
     prefill_cmd = [
         sys.executable,
         "-m",
@@ -161,7 +161,7 @@ def setup_and_run_server():
     env_decode["CUDA_VISIBLE_DEVICES"] = "1"
     env_decode["ENABLE_V1_KVCACHE_SCHEDULER"] = "1"
     env_decode["FD_LOG_DIR"] = "log_decode"
-    decode_log_path = "decode_server.log"
+    decode_log_path = "decode.log"
     decode_cmd = [
         sys.executable,
         "-m",
@@ -216,7 +216,7 @@ def setup_and_run_server():
         try:
             os.killpg(process_prefill.pid, signal.SIGTERM)
             os.killpg(process_decode.pid, signal.SIGTERM)
-            clean_ports()
+            clean_ports(PORTS_TO_CLEAN)
         except Exception as e:
             print(f"Failed to kill process group: {e}")
         raise RuntimeError(f"API server did not start on port {FD_API_PORT}")