feat: add multi stream concurrent execution for rec_model

Author: zhang-minchao

xllm/core/common/global_flags.cpp

@@ -478,6 +478,17 @@ DEFINE_bool(enable_constrained_decoding,
             "that the output meets specific format or structural requirements "
             "through pre-defined rules.");
 
+// --- concurrent llm worker config ---
+DEFINE_uint32(llm_worker_max_concurrency,
+              1,
+              "Concurrency for llm worker parallel execution. Less than or "
+              "equal to 1 means disable concurrent llm worker.");
+
+// --- fixedsteps scheduler config ---
+DEFINE_bool(enable_fixedsteps_scheduler,
+            false,
+            "Whether to use fixsteps scheduler.");
+
 #if defined(USE_NPU)
 DEFINE_string(
     npu_kernel_backend,

xllm/core/common/global_flags.h

@@ -235,6 +235,9 @@ DECLARE_double(dit_cache_residual_diff_threshold);
 
 DECLARE_bool(enable_constrained_decoding);
 
+DECLARE_uint32(llm_worker_max_concurrency);
+
+DECLARE_bool(enable_fixedsteps_scheduler);
 #if defined(USE_NPU)
 DECLARE_string(npu_kernel_backend);
 #endif

xllm/core/runtime/CMakeLists.txt

@@ -18,6 +18,7 @@ cc_library(
     worker.h
     worker_impl.h
     llm_worker_impl.h
+    concurrent_llm_worker_impl.h
     vlm_worker_impl.h
     dit_worker.h
     embed_worker_impl.h
@@ -34,6 +35,7 @@ cc_library(
     worker.cpp
     worker_impl.cpp
     llm_worker_impl.cpp
+    concurrent_llm_worker_impl.cpp
     vlm_worker_impl.cpp
     dit_worker.cpp
     embed_worker_impl.cpp