Make ServerState announcements work better (#93)

- Before this PR, `ServerState.JOINING` was announced only once. This announcement quickly expires in case of the full-size BLOOM, since loading blocks takes several minutes. This PR fixes it, so `ServerState.JOINING` is announced periodically in a thread until blocks are loaded.

- This PR also makes the `Server` class a non-thread, so it runs in the main thread and can catch `KeyboardInterrupt`. This is important, since if we are downloading blocks right now, we need to stop it and send the `ServerState.OFFLINE` message. Note that `ModuleContainer` is still a thread.

- (minor) For the sake of readability, I moved the `ModuleContainer.create()` definition, so it is now defined before `Server.__init__()` (this is because `.create()` is invoked first).

cli/run_server.py

@@ -124,10 +124,9 @@ def main():
     use_auth_token = args.pop("use_auth_token")
     args["use_auth_token"] = True if use_auth_token in ("True", "true", "") else use_auth_token
 
-    server = Server(**args, compression=compression, attn_cache_size=attn_cache_size, start=True)
-
+    server = Server(**args, compression=compression, attn_cache_size=attn_cache_size)
     try:
-        server.join()
+        server.run()
     except KeyboardInterrupt:
         logger.info("Caught KeyboardInterrupt, shutting down")
     finally:

src/server/server.py

@@ -32,7 +32,7 @@ use_hivemind_log_handler("in_root_logger")
 logger = get_logger(__file__)
 
 
-class Server(threading.Thread):
+class Server:
     """
     Runs ModuleContainer, periodically checks that the network is balanced,
     restarts the ModuleContainer with other layers if the imbalance is significant
@@ -68,13 +68,10 @@ class Server(threading.Thread):
         mean_block_selection_delay: float = 0.5,
         use_auth_token: Optional[str] = None,
         load_in_8bit: bool = False,
-        start: bool,
         **kwargs,
     ):
         """Create a server with one or more bloom blocks. See run_server.py for documentation."""
 
-        super().__init__()
-
         self.converted_model_name_or_path = converted_model_name_or_path
         self.num_handlers = num_handlers
         self.min_batch_size, self.max_batch_size = min_batch_size, max_batch_size
@@ -147,8 +144,6 @@ class Server(threading.Thread):
         self.mean_block_selection_delay = mean_block_selection_delay
 
         self.stop = threading.Event()
-        if start:
-            self.start()
 
     def run(self):
         while True:
@@ -231,65 +226,6 @@ class Server(threading.Thread):
 class ModuleContainer(threading.Thread):
     """Serves a set of specific Bloom layers for inference, forward, and backward. Announces itself over the DHT."""
 
-    def __init__(
-        self,
-        dht: DHT,
-        module_backends: Dict[str, TransformerBackend],
-        *,
-        inference_max_length: int,
-        num_connection_handlers: int,
-        throughput: float,
-        update_period: float,
-        expiration: Optional[float] = None,
-        start: bool,
-        **kwargs,
-    ):
-        super().__init__()
-
-        self.dht, self.module_backends = dht, module_backends
-        self.throughput, self.update_period, self.expiration = throughput, update_period, expiration
-        self.conn_handlers = [
-            TransformerConnectionHandler(dht, self.module_backends, inference_max_length)
-            for _ in range(num_connection_handlers)
-        ]
-        self.runtime = Runtime(self.module_backends, **kwargs)
-        self.dht_handler_thread = ModuleAnnouncerThread(
-            self.module_backends,
-            dht,
-            throughput=throughput,
-            update_period=update_period,
-            expiration=expiration,
-            daemon=True,
-        )
-        self.checkpoint_saver = None  # no need to save checkpoints since we do not change model state
-
-        if start:
-            self.run_in_background(await_ready=True)
-
-    def run(self):
-        """
-        Runs ModuleContainer in the current thread. Initializes dht if necessary, starts connection handlers,
-        runs Runtime (self.runtime) to process incoming requests.
-        """
-        logger.info(f"Serving {len(self.module_backends)} blocks:")
-        for expert_name, backend in self.module_backends.items():
-            num_parameters = sum(p.numel() for p in backend.module.parameters() if p.requires_grad)
-            logger.info(f"{expert_name}: {backend.module.__class__.__name__}, {num_parameters} parameters")
-
-        if not self.dht.is_alive():
-            self.dht.run_in_background(await_ready=True)
-
-        if self.module_backends:
-            self.dht_handler_thread.start()
-
-        if self.checkpoint_saver is not None:
-            self.checkpoint_saver.start()
-
-        for handler in self.conn_handlers:
-            handler.run_in_background()
-
-        self.runtime.run()
-
     # noinspection PyMethodOverriding
     @classmethod
     def create(
@@ -320,15 +256,19 @@ class ModuleContainer(threading.Thread):
         start: bool,
     ) -> ModuleContainer:
         module_uids = [f"{prefix}.{block_index}" for block_index in block_indices]
-        declare_active_modules(
-            dht,
+        joining_announcer = ModuleAnnouncerThread(
             module_uids,
-            expiration_time=get_dht_time() + expiration,
-            state=ServerState.JOINING,
+            dht,
+            ServerState.JOINING,
             throughput=throughput,
+            update_period=update_period,
+            expiration=expiration,
+            daemon=True,
         )
+        joining_announcer.start()
         logger.info(f"Announced that blocks {block_indices} are joining")
 
+        try:
             blocks = {}
             for module_uid, block_index in zip(module_uids, block_indices):
                 block = load_pretrained_block(
@@ -367,6 +307,21 @@ class ModuleContainer(threading.Thread):
                     min_batch_size=min_batch_size,
                     max_batch_size=max_batch_size,
                 )
+        except:
+            joining_announcer.stop.set()
+            joining_announcer.join()
+            declare_active_modules(
+                dht,
+                module_uids,
+                expiration_time=get_dht_time() + expiration,
+                state=ServerState.OFFLINE,
+                throughput=throughput,
+            )
+            logger.info(f"Announced that blocks {module_uids} are offline")
+            raise
+        else:
+            joining_announcer.stop.set()
+            joining_announcer.join()
 
         return cls(
             dht,
@@ -383,6 +338,65 @@ class ModuleContainer(threading.Thread):
             start=start,
         )
 
+    def __init__(
+        self,
+        dht: DHT,
+        module_backends: Dict[str, TransformerBackend],
+        *,
+        inference_max_length: int,
+        num_connection_handlers: int,
+        throughput: float,
+        update_period: float,
+        expiration: Optional[float] = None,
+        start: bool,
+        **kwargs,
+    ):
+        super().__init__()
+
+        self.dht, self.module_backends = dht, module_backends
+        self.throughput, self.update_period, self.expiration = throughput, update_period, expiration
+        self.conn_handlers = [
+            TransformerConnectionHandler(dht, self.module_backends, inference_max_length)
+            for _ in range(num_connection_handlers)
+        ]
+        self.runtime = Runtime(self.module_backends, **kwargs)
+        self.online_announcer = ModuleAnnouncerThread(
+            list(self.module_backends.keys()),
+            dht,
+            ServerState.ONLINE,
+            throughput=throughput,
+            update_period=update_period,
+            expiration=expiration,
+            daemon=True,
+        )
+        self.checkpoint_saver = None  # no need to save checkpoints since we do not change model state
+
+        if start:
+            self.run_in_background(await_ready=True)
+
+    def run(self):
+        """
+        Runs ModuleContainer in the current thread. Initializes dht if necessary, starts connection handlers,
+        runs Runtime (self.runtime) to process incoming requests.
+        """
+        logger.info(f"Serving {len(self.module_backends)} blocks:")
+        for expert_name, backend in self.module_backends.items():
+            num_parameters = sum(p.numel() for p in backend.module.parameters() if p.requires_grad)
+            logger.info(f"{expert_name}: {backend.module.__class__.__name__}, {num_parameters} parameters")
+
+        if not self.dht.is_alive():
+            self.dht.run_in_background(await_ready=True)
+
+        self.online_announcer.start()
+
+        if self.checkpoint_saver is not None:
+            self.checkpoint_saver.start()
+
+        for handler in self.conn_handlers:
+            handler.run_in_background()
+
+        self.runtime.run()
+
     def run_in_background(self, await_ready=True, timeout=None):
         """
         Starts ModuleContainer in a background thread. if await_ready, this method will wait until the container
@@ -411,9 +425,8 @@ class ModuleContainer(threading.Thread):
         Please note that terminating container otherwise (e.g. by killing processes) may result in zombie processes.
         If you did already cause a zombie outbreak, your only option is to kill them with -9 (SIGKILL).
         """
-        if self.module_backends:
-            self.dht_handler_thread.stop.set()
-            self.dht_handler_thread.join()
+        self.online_announcer.stop.set()
+        self.online_announcer.join()
 
         declare_active_modules(
             self.dht,
@@ -450,8 +463,9 @@ class ModuleAnnouncerThread(threading.Thread):
 
     def __init__(
         self,
-        module_backends: Dict[str, TransformerBackend],
+        module_uids: List[str],
         dht: DHT,
+        state: ServerState,
         *,
         throughput: float,
         update_period: float = 30,
@@ -459,8 +473,9 @@ class ModuleAnnouncerThread(threading.Thread):
         **kwargs,
     ):
         super().__init__(**kwargs)
-        self.module_backends = module_backends
+        self.module_uids = module_uids
         self.dht = dht
+        self.state = state
         self.throughput = throughput
         self.update_period = update_period
         self.expiration = expiration
@@ -470,9 +485,9 @@ class ModuleAnnouncerThread(threading.Thread):
         while True:
             declare_active_modules(
                 self.dht,
-                self.module_backends.keys(),
+                self.module_uids,
                 expiration_time=get_dht_time() + self.expiration,
-                state=ServerState.ONLINE,
+                state=self.state,
                 throughput=self.throughput,
             )
             if self.stop.wait(self.update_period):