Revert "feat: nanobind bindings (#5961)"

cpp/CMakeLists.txt

@@ -198,7 +198,7 @@ set(TRT_LIB TensorRT::NvInfer)
 get_filename_component(TRT_LLM_ROOT_DIR ${CMAKE_CURRENT_SOURCE_DIR} PATH)
 
 set(3RDPARTY_DIR ${TRT_LLM_ROOT_DIR}/3rdparty)
-if(BINDING_TYPE STREQUAL "pybind" OR BUILD_DEEP_EP)
+if(BINDING_TYPE STREQUAL "pybind")
   add_subdirectory(${3RDPARTY_DIR}/pybind11
                    ${CMAKE_CURRENT_BINARY_DIR}/pybind11)
 endif()
@@ -217,7 +217,7 @@ include_directories(
   ${3RDPARTY_DIR}/cutlass/tools/util/include
   ${3RDPARTY_DIR}/NVTX/include
   ${3RDPARTY_DIR}/json/include)
-if(BINDING_TYPE STREQUAL "pybind" OR BUILD_DEEP_EP)
+if(BINDING_TYPE STREQUAL "pybind")
   include_directories(${3RDPARTY_DIR}/pybind11/include)
 endif()
 if(BINDING_TYPE STREQUAL "nanobind")

cpp/include/tensorrt_llm/batch_manager/runtimeBuffers.h

@@ -168,7 +168,7 @@ class RuntimeBuffers
 
 public:
     //! Additional buffers depending on model type
-    std::shared_ptr<TransformerBuffers> transformerBuffers;
+    std::unique_ptr<TransformerBuffers> transformerBuffers;
     std::unique_ptr<RnnStateBuffers> rnnStateBuffers;
 
     //! Encoder-Decoder

cpp/tensorrt_llm/batch_manager/runtimeBuffers.cpp

@@ -84,7 +84,7 @@ void RuntimeBuffers::create(SizeType32 maxBatchSize, SizeType32 maxBeamWidth,
 
     if (modelConfig.isTransformerBased())
     {
-        transformerBuffers = std::make_shared<TransformerBuffers>(maxBatchSize, maxBeamWidth, maxAttentionWindowVec,
+        transformerBuffers = std::make_unique<TransformerBuffers>(maxBatchSize, maxBeamWidth, maxAttentionWindowVec,
             maxAttentionWindow, sinkTokenLen, runtime, modelConfig, worldConfig);
     }
     if (modelConfig.isRnnBased())

cpp/tensorrt_llm/nanobind/CMakeLists.txt

@@ -3,23 +3,7 @@ set(TRTLLM_NB_MODULE
     ${TRTLLM_NB_MODULE}
     PARENT_SCOPE)
 
-set(SRCS
-    batch_manager/algorithms.cpp
-    batch_manager/bindings.cpp
-    batch_manager/buffers.cpp
-    batch_manager/cacheTransceiver.cpp
-    batch_manager/kvCacheManager.cpp
-    batch_manager/llmRequest.cpp
-    executor/bindings.cpp
-    executor/executor.cpp
-    executor/executorConfig.cpp
-    executor/request.cpp
-    runtime/bindings.cpp
-    testing/modelSpecBinding.cpp
-    runtime/moeBindings.cpp
-    userbuffers/bindings.cpp
-    ../runtime/ipcNvlsMemory.cu
-    bindings.cpp)
+set(SRCS ../runtime/ipcNvlsMemory.cu bindings.cpp)
 
 include_directories(${PROJECT_SOURCE_DIR}/include)
 
@@ -30,29 +14,20 @@ set_property(TARGET ${TRTLLM_NB_MODULE} PROPERTY POSITION_INDEPENDENT_CODE ON)
 target_link_directories(${TRTLLM_NB_MODULE} PUBLIC
                         "${TORCH_INSTALL_PREFIX}/lib")
 
-if(ENABLE_NVSHMEM)
-  target_link_libraries(${TRTLLM_NB_MODULE} PUBLIC nvshmem::nvshmem_host
-                                                   nvshmem::nvshmem_device)
-endif()
-
 target_link_libraries(
   ${TRTLLM_NB_MODULE}
-  PUBLIC ${SHARED_TARGET}
-         ${UNDEFINED_FLAG}
-         ${NO_AS_NEEDED_FLAG}
-         ${Python3_LIBRARIES}
-         ${TORCH_LIBRARIES}
-         torch_python
-         ${CUDA_NVML_LIB})
+  PUBLIC ${SHARED_TARGET} ${UNDEFINED_FLAG} ${NO_AS_NEEDED_FLAG}
+         ${Python3_LIBRARIES} ${TORCH_LIBRARIES} torch_python)
+
 target_compile_definitions(
   ${TRTLLM_NB_MODULE} PUBLIC TRTLLM_NB_MODULE=${TRTLLM_NB_MODULE}
-                             PYBIND11_DETAILED_ERROR_MESSAGES=1)
+                             NB_DETAILED_ERROR_MESSAGES=1)
 
 if(NOT WIN32)
   set_target_properties(
     ${TRTLLM_NB_MODULE}
     PROPERTIES
       LINK_FLAGS
-      "-Wl,-rpath,'$ORIGIN/libs' -Wl,-rpath,'$ORIGIN/../nvidia/nccl/lib' -Wl,-rpath,'${CUDA_TOOLKIT_ROOT_DIR}/targets/x86_64-linux/lib/stubs' ${AS_NEEDED_FLAG} ${UNDEFINED_FLAG}"
+      "-Wl,-rpath,'$ORIGIN/libs' -Wl,-rpath,'$ORIGIN/../nvidia/nccl/lib' ${AS_NEEDED_FLAG} ${UNDEFINED_FLAG}"
   )
 endif()