feat: use shared memory for intra-process communication instead of BRPC. (#276)

Author: Clement-Wang26

xllm/core/common/global_flags.cpp

@@ -275,6 +275,9 @@ DEFINE_int32(sleep_time_second,
              3,
              "The sleep time for worker try to connect to server next time.");
 
+DEFINE_bool(enable_shm,
+            true,
+            "Whether to enable shared memory for executing model.");
 // --- function call config ---
 
 DEFINE_string(tool_call_parser,

xllm/core/common/global_flags.h

@@ -199,4 +199,6 @@ DECLARE_int64(cache_size_per_token);
 
 DECLARE_int64(buffer_size_per_seq);
 
-DECLARE_bool(enable_beam_search_kernel);
+DECLARE_bool(enable_beam_search_kernel);
+
+DECLARE_bool(enable_shm);

xllm/core/common/options.h

@@ -175,6 +175,9 @@ class Options {
   PROPERTY(bool, enable_offline_inference) = false;
   // for offline inference: the path to spawn worker binary
   PROPERTY(std::string, spawn_worker_path) = "";
+
+  // whether the worker and master are on the same machine.
+  PROPERTY(bool, is_local) = false;
 };
 
 }  // namespace xllm

xllm/core/distributed_runtime/CMakeLists.txt

@@ -20,6 +20,8 @@ cc_library(
     remote_worker.h
     worker_server.h
     worker_service.h
+    comm_channel.h
+    shm_channel.h
   SRCS
     disagg_pd_service.cpp
     disagg_pd_service_impl.cpp
@@ -29,6 +31,8 @@ cc_library(
     remote_worker.cpp
     worker_server.cpp
     worker_service.cpp
+    comm_channel.cpp
+    shm_channel.cpp
   DEPS
     :api_service
     :runtime