feat: Parallelise Model Loading

exo/inference/inference_engine.py

@@ -8,6 +8,9 @@
 
 class InferenceEngine(ABC):
   @abstractmethod
+  async def preload_model(self, shard: Shard) -> None:
+    pass
+  @abstractmethod
   async def infer_prompt(self, request_id: str, shard: Shard, prompt: str, image_str: Optional[str] = None, inference_state: Optional[str] = None) -> (np.ndarray, str, bool):
     pass
 

exo/inference/mlx/sharded_inference_engine.py

@@ -9,6 +9,7 @@
 import asyncio
 from concurrent.futures import ThreadPoolExecutor
 from functools import partial
+import pathlib as Path
 
 class MLXDynamicShardInferenceEngine(InferenceEngine):
   def __init__(self, shard_downloader: ShardDownloader):
@@ -43,8 +44,15 @@ async def ensure_shard(self, shard: Shard):
     model_path = await self.shard_downloader.ensure_shard(shard)
 
     if self.shard != shard:
-      loop = asyncio.get_running_loop()
-      def load_shard_wrapper(): return asyncio.run(load_shard(model_path, shard))
-      model_shard, self.tokenizer = await loop.run_in_executor(self.executor, load_shard_wrapper)
-      self.stateful_sharded_model = await loop.run_in_executor(self.executor, StatefulShardedModel, shard, model_shard)
-      self.shard = shard
+      await self.load_model(model_path, shard)
+
+  async def load_model(self, model_path: Path, shard: Shard):
+    loop = asyncio.get_running_loop()
+    def load_shard_wrapper(): return asyncio.run(load_shard(model_path, shard))
+    model_shard, self.tokenizer = await loop.run_in_executor(self.executor, load_shard_wrapper)
+    self.stateful_sharded_model = await loop.run_in_executor(self.executor, StatefulShardedModel, shard, model_shard)
+    self.shard = shard
+
+    # already loaded into memory by ensure_shard,
+  async def preload_model(self, shard: Shard) -> None:
+    await self.ensure_shard(shard)

exo/main.py

@@ -47,6 +47,7 @@
 parser.add_argument("--prompt", type=str, help="Prompt for the model when using --run-model", default="Who are you?")
 parser.add_argument("--tailscale-api-key", type=str, default=None, help="Tailscale API key")
 parser.add_argument("--tailnet-name", type=str, default=None, help="Tailnet name")
+parser.add_argument("--preload-models", type=str, help="Comma-separated list of models to preload")
 args = parser.parse_args()
 
 print_yellow_exo()
@@ -100,16 +101,18 @@
   lambda req_id, tokens, __: topology_viz.update_prompt_output(req_id, inference_engine.tokenizer.decode(tokens)) if topology_viz and hasattr(inference_engine, "tokenizer") else None
 )
 def preemptively_start_download(request_id: str, opaque_status: str):
-  try:
-    status = json.loads(opaque_status)
-    if status.get("type") == "node_status" and status.get("status") == "start_process_prompt":
-      current_shard = node.get_current_shard(Shard.from_dict(status.get("shard")))
-      if DEBUG >= 2: print(f"Preemptively starting download for {current_shard}")
-      asyncio.create_task(shard_downloader.ensure_shard(current_shard))
-  except Exception as e:
-    if DEBUG >= 2:
-      print(f"Failed to preemptively start download: {e}")
-      traceback.print_exc()
+    try:
+        status = json.loads(opaque_status)
+        if status.get("type") == "node_status" and status.get("status") == "start_process_prompt":
+            current_shard = node.get_current_shard(Shard.from_dict(status.get("shard")))
+            if DEBUG >= 2: print(f"Preemptively starting download for {current_shard}")
+            asyncio.create_task(shard_downloader.ensure_shard(current_shard))
+            return current_shard
+    except Exception as e:
+        if DEBUG >= 2:
+            print(f"Failed to preemptively start download: {e}")
+            traceback.print_exc()
+    return None
 node.on_opaque_status.register("start_download").on_next(preemptively_start_download)
 
 if args.prometheus_client_port:
@@ -176,7 +179,6 @@ async def run_model_cli(node: Node, inference_engine: InferenceEngine, model_nam
 async def main():
   loop = asyncio.get_running_loop()
 
-  # Use a more direct approach to handle signals
   def handle_exit():
     asyncio.ensure_future(shutdown(signal.SIGTERM, loop))
 
@@ -195,6 +197,18 @@ def handle_exit():
     asyncio.create_task(api.run(port=args.chatgpt_api_port))  # Start the API server as a non-blocking task
     await asyncio.Event().wait()
 
+  if args.preload_models:
+    models_to_preload = [Shard.from_dict(model_base_shards[model_name][inference_engine.__class__.__name__]) 
+                         for model_name in args.preload_models.split(',')]
+    for shard in models_to_preload:
+      current_shard = preemptively_start_download(str(uuid.uuid4()), json.dumps({
+        "type": "node_status",
+        "status": "start_process_prompt",
+        "shard": shard.to_dict()
+      }))
+      if current_shard:
+        await node.preload_models([current_shard])
+
 
 def run():
   loop = asyncio.new_event_loop()

exo/orchestration/node.py

@@ -31,6 +31,10 @@ async def get_inference_result(self, request_id: str) -> Tuple[Optional[np.ndarr
   async def collect_topology(self, visited: set[str] = set(), max_depth: int = 2) -> Topology:
     pass
 
+  @abstractmethod
+  async def preload_models(self, shards: List[Shard]) -> None:
+    pass
+
   @property
   @abstractmethod
   def current_topology(self) -> Topology:
@@ -45,3 +49,4 @@ def on_token(self) -> AsyncCallbackSystem[str, Tuple[str, List[int], bool]]:
   @abstractmethod
   def on_opaque_status(self) -> AsyncCallbackSystem[str, Tuple[str, str]]:
     pass
+

exo/orchestration/standard_node.py

@@ -432,3 +432,9 @@ async def send_status_to_peer(peer):
   @property
   def current_topology(self) -> Topology:
     return self.topology
+
+  async def preload_models(self, shards: List[Shard]) -> None:
+    preload_tasks = []
+    for shard in shards:
+        preload_tasks.append(asyncio.create_task(self.inference_engine.preload_model(shard)))
+    await asyncio.gather(*preload_tasks)