fix(server): better handling of inference mode (#57)

launcher/src/main.rs

@@ -38,9 +38,9 @@ struct Args {
     port: u16,
     #[clap(default_value = "/tmp/text-generation-server", long, env)]
     shard_uds_path: String,
-    #[clap(default_value = "localhost", long, env)]
+    #[clap(default_value = "0.0.0.0", long, env)]
     master_addr: String,
-    #[clap(default_value = "29500", long, env)]
+    #[clap(default_value = "6000", long, env)]
     master_port: usize,
     #[clap(long, env)]
     json_output: bool,

server/text_generation/models/__init__.py

@@ -28,6 +28,9 @@ torch.backends.cuda.matmul.allow_tf32 = True
 # The flag below controls whether to allow TF32 on cuDNN. This flag defaults to True.
 torch.backends.cudnn.allow_tf32 = True
 
+# Disable gradients
+torch.set_grad_enabled(False)
+
 
 def get_model(
     model_id: str, revision: Optional[str], sharded: bool, quantize: bool

server/text_generation/models/causal_lm.py

@@ -289,11 +289,6 @@ class CausalLM(Model):
     def generate_token(
         self, batch: CausalLMBatch
     ) -> Tuple[List[Generation], Optional[CausalLMBatch]]:
-        # For some reason, inference_mode does not work well with GLOO which we use on CPU
-        context_manager = (
-            torch.no_grad if self.device.type == "cpu" else torch.inference_mode
-        )
-        with context_manager():
         logits, past = self.forward(
             batch.input_ids,
             batch.attention_mask,

server/text_generation/models/seq2seq_lm.py

@@ -364,11 +364,6 @@ class Seq2SeqLM(Model):
     def generate_token(
         self, batch: Seq2SeqLMBatch
     ) -> Tuple[List[Generation], Optional[Seq2SeqLMBatch]]:
-        # For some reason, inference_mode does not work well with GLOO which we use on CPU
-        context_manager = (
-            torch.no_grad if self.device.type == "cpu" else torch.inference_mode
-        )
-        with context_manager():
         logits, encoder_last_hidden_state, past = self.forward(
             batch.input_ids,
             batch.attention_mask,

server/text_generation/server.py

@@ -1,5 +1,6 @@
 import asyncio
 import os
+import torch
 
 from grpc import aio
 from loguru import logger
@@ -19,6 +20,10 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
         self.cache = cache
         self.model = model
         self.server_urls = server_urls
+        # For some reason, inference_mode does not work well with GLOO which we use on CPU
+        if model.device.type == "cuda":
+            # Force inference mode for the lifetime of TextGenerationService
+            self._inference_mode_raii_guard = torch._C._InferenceMode(True)
 
     async def ServiceDiscovery(self, request, context):
         return generate_pb2.ServiceDiscoveryResponse(urls=self.server_urls)
@@ -89,7 +94,11 @@ def serve(
             local_url = unix_socket_template.format(uds_path, 0)
             server_urls = [local_url]
 
+        try:
             model = get_model(model_id, revision, sharded, quantize)
+        except Exception:
+            logger.exception("Error when initializing model")
+            raise
 
         server = aio.server(interceptors=[ExceptionInterceptor()])
         generate_pb2_grpc.add_TextGenerationServiceServicer_to_server(
@@ -101,8 +110,11 @@ def serve(
         )
         reflection.enable_server_reflection(SERVICE_NAMES, server)
         server.add_insecure_port(local_url)
+
         await server.start()
+
         logger.info("Server started at {}".format(local_url))
+
         try:
             await server.wait_for_termination()
         except KeyboardInterrupt:

server/text_generation/utils.py

@@ -171,9 +171,14 @@ def initialize_torch_distributed():
     else:
         backend = "gloo"
 
+    master_ip = os.getenv("MASTER_ADDR", "0.0.0.0")
+    master_port = os.getenv("MASTER_PORT", "6000")
+    init_method = f"tcp://{master_ip}:{master_port}"
+
     # Call the init process.
     torch.distributed.init_process_group(
         backend=backend,
+        init_method=init_method,
         world_size=world_size,
         rank=rank,
         timeout=timedelta(seconds=60),