fix tcp port in model_service_worker.py (#2377)

* fix tcp port in model_service_worker.py * update torchrun parameter name * fmt
pytorch · Jun 1, 2023 · a81fc52 · a81fc52
1 parent c685b5c
commit a81fc52
Show file tree

Hide file tree

Showing 3 changed files with 9 additions and 32 deletions.
diff --git a/frontend/archive/src/main/java/org/pytorch/serve/archive/model/ModelConfig.java b/frontend/archive/src/main/java/org/pytorch/serve/archive/model/ModelConfig.java
@@ -339,7 +339,6 @@ public static class TorchRun {
         private String rdzvEndpoint;
         private String rdzvBackend = "c10d";
         private String rdzvConf;
-        private int maxRestarts = 3;
         private int monitorInterval = 5;
         private int nodeRank;
         private String masterAddr;
@@ -388,13 +387,6 @@ public static TorchRun build(Map<?, ?> torchRunMap) {
                                     logger.warn("Invalid torchrun.rdzv-conf:{}", v);
                                 }
                                 break;
-                            case "max-restarts":
-                                if (v instanceof Integer) {
-                                    torchRun.setMaxRestarts((Integer) v);
-                                } else {
-                                    logger.warn("Invalid torchrun.max-restarts:{}, reset to 3", v);
-                                }
-                                break;
                             case "monitor-interval":
                                 if (v instanceof Integer) {
                                     torchRun.setMonitorInterval((Integer) v);
@@ -417,6 +409,7 @@ public static TorchRun build(Map<?, ?> torchRunMap) {
                                 }
                                 break;
                             default:
+                                logger.warn("unsupported parameter {}", k);
                                 break;
                         }
                     });
@@ -479,18 +472,6 @@ public void setRdzvConf(String rdzvConf) {
             this.rdzvConf = rdzvConf;
         }
 
-        public int getMaxRestarts() {
-            return maxRestarts;
-        }
-
-        public void setMaxRestarts(int maxRestarts) {
-            if (maxRestarts <= 0) {
-                logger.warn("Invalid torchrun.max-restarts:{}, reset to 3", maxRestarts);
-                return;
-            }
-            this.maxRestarts = maxRestarts;
-        }
-
         public int getMonitorInterval() {
             return monitorInterval;
         }

diff --git a/frontend/server/src/main/java/org/pytorch/serve/wlm/WorkerLifeCycle.java b/frontend/server/src/main/java/org/pytorch/serve/wlm/WorkerLifeCycle.java
@@ -195,21 +195,17 @@ private void attachRunner(
         argl.add("torchrun");
         argl.add("--nnodes");
         argl.add(String.valueOf(torchRun.getNnodes()));
-        argl.add("--nproc_per_node");
+        argl.add("--nproc-per-node");
         argl.add(String.valueOf(torchRun.getNprocPerNode()));
-        argl.add("--max_restarts");
-        argl.add(String.valueOf(torchRun.getMaxRestarts()));
-        argl.add("--log_dir");
+        argl.add("--log-dir");
         argl.add(ConfigManager.getInstance().getTorchRunLogDir());
-        argl.add("--rdzv_backend");
+        argl.add("--rdzv-backend");
         argl.add(torchRun.getRdzvBackend());
-        argl.add("--rdzv_endpoint");
         if (torchRun.getRdzvEndpoint() != null) {
+            argl.add("--rdzv-endpoint");
             argl.add(torchRun.getRdzvEndpoint());
-        } else {
-            argl.add(String.format("localhost:%d", port));
         }
-        argl.add("--rdzv_id");
+        argl.add("--rdzv-id");
         argl.add(String.format("%s_%d", model.getModelName(), port));
         if (torchRun.getMasterAddr() != null) {
             argl.add("--master-addr");

diff --git a/ts/model_service_worker.py b/ts/model_service_worker.py
@@ -61,16 +61,16 @@ def __init__(
                     raise RuntimeError(
                         "socket already in use: {}.".format(s_name_new)
                     ) from e
-
+            logging.info("Listening on port: %s", s_name_new)
         elif s_type == "tcp":
             self.sock_name = host_addr if host_addr is not None else "127.0.0.1"
             if port_num is None:
                 raise ValueError("Wrong arguments passed. No socket port given.")
-            self.port = port_num + LOCAL_RANK
+            self.port = int(port_num) + LOCAL_RANK
+            logging.info("Listening on addr:port: %s:%d", self.sock_name, self.port)
         else:
             raise ValueError("Incomplete data provided")
 
-        logging.info("Listening on port: %s", s_name)
         socket_family = socket.AF_INET if s_type == "tcp" else socket.AF_UNIX
         self.sock = socket.socket(socket_family, socket.SOCK_STREAM)
         self.metrics_cache = MetricsCacheYamlImpl(config_file_path=metrics_config)