[gRPC] Native gRPC server: proto + Rust crate scaffold + server args (#22736)

python/pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools>=61.0", "setuptools-scm>=8.0", "wheel"]
+requires = ["setuptools>=61.0", "setuptools-scm>=8.0", "setuptools-rust>=1.10", "wheel"]
 build-backend = "setuptools.build_meta"
 
 [project]
@@ -217,5 +217,10 @@ git_describe_command = ["python3", "python/tools/get_version_tag.py"]
 # Allow editable installs even when .git metadata is not available.
 fallback_version = "0.0.0.dev0"
 
+[[tool.setuptools-rust.ext-modules]]
+target = "sglang.srt.grpc._core"
+path = "../rust/sglang-grpc/Cargo.toml"
+binding = "PyO3"
+
 [tool.kernels.dependencies]
 "kernels-community/sgl-flash-attn3" = 1

python/sglang/launch_server.py

@@ -27,6 +27,9 @@ def run_server(server_args):
 
             launch_server(server_args)
     elif server_args.grpc_mode:
+        # TODO: Once the native Rust gRPC server starts alongside HTTP in the
+        # default path below (controlled by SGLANG_ENABLE_GRPC / SGLANG_GRPC_PORT),
+        # remove this legacy SMG path and the grpc_mode flag.
         from sglang.srt.entrypoints.grpc_server import serve_grpc
 
         asyncio.run(serve_grpc(server_args))

python/sglang/srt/environ.py

@@ -516,6 +516,10 @@ class Envs:
     # Encoder receiver selection: http|grpc (used by EPD paths).
     SGLANG_ENCODER_MM_RECEIVER_MODE = EnvStr("http")
 
+    # Native gRPC server (internal, not yet user-facing)
+    SGLANG_GRPC_PORT = EnvInt(None)
+    SGLANG_ENABLE_GRPC = EnvBool(False)
+
     # External models
     SGLANG_EXTERNAL_MODEL_PACKAGE = EnvStr("")
     SGLANG_EXTERNAL_MM_MODEL_ARCH = EnvStr("")

python/sglang/srt/server_args.py

@@ -999,6 +999,21 @@ class ServerArgs:
             envs.SGLANG_SPEC_NAN_DETECTION.set(True)
             envs.SGLANG_SPEC_OOB_DETECTION.set(True)
 
+        # Native gRPC flags — env-only for now, not exposed as CLI args.
+        # Set as instance attributes (not dataclass fields) to avoid
+        # argparse namespace lookup in from_cli_args.
+        self.enable_grpc = envs.SGLANG_ENABLE_GRPC.get()
+
+        grpc_port_env = envs.SGLANG_GRPC_PORT.get()
+        self.grpc_port = (
+            grpc_port_env if grpc_port_env is not None else self.port + 10000
+        )
+
+        if not (1 <= self.grpc_port <= 65535):
+            raise ValueError(
+                f"SGLANG_GRPC_PORT ({self.grpc_port}) must be between 1 and 65535"
+            )
+
     def _handle_prefill_delayer_env_compat(self):
         if envs.SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE.get():
             self.enable_prefill_delayer = True
@@ -6622,6 +6637,19 @@ class ServerArgs:
                 "When enabling two batch overlap, moe_a2a_backend cannot be 'none'."
             )
 
+        if (
+            self.enable_grpc
+            and self.grpc_port is not None
+            and self.grpc_port == self.port
+        ):
+            raise ValueError(
+                f"SGLANG_GRPC_PORT ({self.grpc_port}) must differ from --port ({self.port})"
+            )
+
+        # TODO: Also validate grpc_port != metrics_http_port and grpc_port != nccl_port
+        # to avoid opaque bind errors at runtime. Deferred because metrics_http_port
+        # and nccl_port have dynamic defaults that may not be resolved yet here.
+
         if self.gc_threshold:
             if not (1 <= len(self.gc_threshold) <= 3):
                 raise ValueError(