feat: support iluvatar backend qwen3 0.6b run through

Author: zuolve.lai

CMakeLists.txt

@@ -3,6 +3,7 @@ set_property(GLOBAL PROPERTY USE_FOLDERS ON)
 
 option(USE_NPU "Enable NPU support" OFF)
 option(USE_MLU "Enable MLU support" OFF)
+option(USE_ILU "Enable ILU support" OFF)
 option(USE_CUDA "Enable CUDA support" OFF)
 add_compile_definitions(YLT_ENABLE_IBV)
 add_definitions(-DYLT_ENABLE_IBV)
@@ -105,7 +106,7 @@ set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS ON)
 
-if(USE_NPU OR USE_CUDA)
+if(USE_NPU OR USE_CUDA OR USE_ILU)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D_GLIBCXX_USE_CXX11_ABI=0")
   add_definitions(-D_GLIBCXX_USE_CXX11_ABI=0)
 elseif(USE_MLU)
@@ -208,6 +209,19 @@ if(USE_CUDA)
   message(STATUS "TORCH_CUDA_ARCH_LIST: ${TORCH_CUDA_ARCH_LIST}")
 endif()
 
+if(USE_ILU)
+    set(CMAKE_MODULE_PATH "${CMAKE_CURRENT_LIST_DIR}/cmake/Modules;${CMAKE_MODULE_PATH}")
+    set(CMAKE_EXPORT_COMPILE_COMMANDS   ON)
+    set(CMAKE_CUDA_ARCHITECTURES "ivcore11")
+    set(WARNINGS_AS_ERRORS OFF)
+    if (CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+      add_definitions(
+        -Wno-c++11-narrowing
+        -Wno-thread-safety-analysis
+      )
+  endif()
+endif()
+
 # configure vcpkg
 # have to set CMAKE_TOOLCHAIN_FILE before first project call.
 # if (DEFINED ENV{VCPKG_ROOT} AND NOT DEFINED CMAKE_TOOLCHAIN_FILE)
@@ -425,6 +439,23 @@ if(USE_CUDA)
   )
 endif()
 
+if(USE_ILU)
+  add_definitions(-DUSE_ILU)
+  set(CMAKE_VERBOSE_MAKEFILE ON)
+  include_directories(
+      $ENV{PYTHON_INCLUDE_PATH}
+      $ENV{PYTORCH_INSTALL_PATH}/include
+      $ENV{PYTORCH_INSTALL_PATH}/include/torch/csrc/api/include
+      $ENV{IXFORMER_INSTALL_PATH}/csrc/include/ixformer
+  )
+
+  link_directories(
+    $ENV{PYTHON_LIB_PATH}
+    $ENV{PYTORCH_INSTALL_PATH}/lib
+    $ENV{IXFORMER_INSTALL_PATH}
+  )
+endif()
+
 # check if USE_CXX11_ABI is set correctly
 # if (DEFINED USE_CXX11_ABI)
 #   parse_make_options(${TORCH_CXX_FLAGS} "TORCH_CXX_FLAGS")

setup.py

@@ -109,11 +109,11 @@ def get_torch_root_path():
     except ImportError:
         return None
 
-def get_torch_mlu_root_path():
+def get_ixformer_root_path():
     try:
-        import torch_mlu
+        import ixformer
         import os
-        return os.path.dirname(os.path.abspath(torch_mlu.__file__))
+        return os.path.dirname(os.path.abspath(ixformer.__file__))
     except ImportError:
         return None
 
@@ -229,7 +229,15 @@ def set_cuda_envs():
     os.environ["CUDA_TOOLKIT_ROOT_DIR"] = "/usr/local/cuda"
     os.environ["NCCL_ROOT"] = get_nccl_root_path()
     os.environ["NCCL_VERSION"] = "2"
-    
+
+def set_ilu_envs():
+    os.environ["PYTHON_INCLUDE_PATH"] = get_python_include_path()
+    os.environ["PYTHON_LIB_PATH"] =  get_torch_root_path()
+    os.environ["LIBTORCH_ROOT"] = get_torch_root_path()
+    os.environ["PYTORCH_INSTALL_PATH"] = get_torch_root_path()
+    os.environ["CUDA_TOOLKIT_ROOT_DIR"] = "/usr/local/corex"
+    os.environ["IXFORMER_INSTALL_PATH"] = get_ixformer_root_path()
+
 class CMakeExtension(Extension):
     def __init__(self, name: str, path: str, sourcedir: str = "") -> None:
         super().__init__(name, sources=[])
@@ -280,8 +288,7 @@ def run(self):
             for ext in self.extensions:
                 self.build_extension(ext)
         except Exception as e:
-            print("ERROR: Build failed.")
-            print(f"Details: {e}")
+            print("Build failed.")
             exit(1)
 
     def build_extension(self, ext: CMakeExtension):
@@ -313,7 +320,7 @@ def build_extension(self, ext: CMakeExtension):
             f"-DDEVICE_ARCH={self.arch.upper()}",
             f"-DINSTALL_XLLM_KERNELS={'ON' if self.install_xllm_kernels else 'OFF'}",
         ]
-        
+
         if self.device == "a2" or self.device == "a3":
             cmake_args += ["-DUSE_NPU=ON"]
             # set npu environment variables
@@ -328,6 +335,9 @@ def build_extension(self, ext: CMakeExtension):
                            f"-DCMAKE_CUDA_ARCHITECTURES={cuda_architectures}"]
             # set cuda environment variables
             set_cuda_envs()
+        elif self.device == "ilu":
+            cmake_args += ["-DUSE_ILU=ON"]
+            set_ilu_envs()
         else:
             raise ValueError("Please set --device to a2 or a3 or mlu or cuda.")
 
@@ -351,6 +361,7 @@ def build_extension(self, ext: CMakeExtension):
 
         build_args = ["--config", build_type]
         max_jobs = os.getenv("MAX_JOBS", str(os.cpu_count()))
+        # max_jobs="2"
         build_args += ["-j" + max_jobs]
 
         env = os.environ.copy()
@@ -572,7 +583,7 @@ def pre_build():
         idx = sys.argv.index('--device')
         if idx + 1 < len(sys.argv):
             device = sys.argv[idx+1].lower()
-            if device not in ('a2', 'a3', 'mlu', 'cuda'):
+            if device not in ('a2', 'a3', 'mlu', 'cuda', 'ilu'):
                 print("Error: --device must be a2 or a3 or mlu (case-insensitive)")
                 sys.exit(1)
             # Remove the arguments so setup() doesn't see them

third_party/CMakeLists.txt

@@ -19,4 +19,32 @@ target_include_directories(mooncake_store PUBLIC
   ${CMAKE_CURRENT_SOURCE_DIR}/Mooncake/mooncake-transfer-engine/include
 )
 
+if(USE_ILU)
+  if(TARGET cpprest)
+      set_target_properties(cpprest PROPERTIES
+          CXX_STANDARD 20
+          CXX_STANDARD_REQUIRED ON
+          CXX_EXTENSIONS OFF
+      )
+  endif()
+  if(TARGET transfer_engine)
+    target_compile_options(transfer_engine PRIVATE -std=c++20)
+    set_target_properties(transfer_engine PROPERTIES
+        CXX_STANDARD 20
+        CXX_STANDARD_REQUIRED ON
+    )
+    message(STATUS "Set C++20 for transfer_engine target")
+  endif()
+  if(TARGET SMHasherSupport)
+      set_target_properties(SMHasherSupport PROPERTIES
+          CXX_STANDARD 11
+          CXX_STANDARD_REQUIRED ON
+          CXX_EXTENSIONS OFF
+      )
+      message(STATUS "SMHasherSupport target found and configured")
+  else()
+      message(WARNING "SMHasherSupport target not found after adding smhasher")
+  endif()
+endif()
+
 target_link_libraries(mooncake_store PUBLIC transfer_engine cachelib_memory_allocator)

xllm/core/distributed_runtime/worker_server.cpp

@@ -1,4 +1,5 @@
 /* Copyright 2025 The xLLM Authors. All Rights Reserved.
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.

xllm/core/framework/batch/batch_input_builder.cpp

@@ -207,7 +207,7 @@ void BatchInputBuilder::process_sequences_multithreaded() {
     state_.q_seq_lens.insert(state_.q_seq_lens.end(),
                              state.q_seq_lens.begin(),
                              state.q_seq_lens.end());
-#elif defined(USE_MLU) || defined(USE_CUDA)
+#elif defined(USE_MLU) || defined(USE_CUDA) || defined(USE_ILU)
     int32_t seq_len_offset = state_.seq_lens.back();
     // skip the first element which is 0
     for (size_t i = 1; i < state.seq_lens.size(); ++i) {
@@ -293,7 +293,7 @@ void BatchInputBuilder::process_single_sequence(
 #if defined(USE_NPU)
   state.seq_lens.push_back(seq_len + offset);
   state.q_seq_lens.push_back(q_seq_len);
-#elif defined(USE_MLU) || defined(USE_CUDA)
+#elif defined(USE_MLU) || defined(USE_CUDA) || defined(USE_ILU)
   state.seq_lens.push_back(state.seq_lens.back() + seq_len + offset);
   state.q_seq_lens.push_back(state.q_seq_lens.back() + q_seq_len);
 #endif
@@ -527,7 +527,7 @@ void BatchInputBuilder::padding_decode_batch_size(
 #if defined(USE_NPU)
         state_.seq_lens.push_back(num_decoding_tokens);
         state_.q_seq_lens.push_back(num_decoding_tokens);
-#elif defined(USE_MLU) || defined(USE_CUDA)
+#elif defined(USE_MLU) || defined(USE_CUDA) || defined(USE_ILU)
         state_.seq_lens.push_back(state_.seq_lens.back() + num_decoding_tokens);
         state_.q_seq_lens.push_back(state_.q_seq_lens.back() +
                                     num_decoding_tokens);

xllm/core/framework/batch/batch_input_builder.h

@@ -85,7 +85,7 @@ class BatchInputBuilder {
 #if defined(USE_NPU)
     std::vector<int32_t> seq_lens;
     std::vector<int32_t> q_seq_lens;
-#elif defined(USE_MLU) || defined(USE_CUDA)
+#elif defined(USE_MLU) || defined(USE_CUDA) || defined(USE_ILU)
     std::vector<int32_t> seq_lens = {0};    // cu_seq_lens
     std::vector<int32_t> q_seq_lens = {0};  // q_cu_seq_len
 #endif

xllm/core/framework/parallel_state/CMakeLists.txt

@@ -12,6 +12,7 @@ cc_library(
     $<$<BOOL:${USE_NPU}>:npu_process_group.h>
     $<$<BOOL:${USE_MLU}>:mlu_process_group.h>
     $<$<BOOL:${USE_CUDA}>:cuda_process_group.h>
+    $<$<BOOL:${USE_ILU}>:ilu_process_group.h>
     collective_communicator.h
   SRCS
     mapping_npu.cpp

xllm/core/framework/parallel_state/collective_communicator.cpp

@@ -25,6 +25,8 @@ limitations under the License.
 #include "mlu_process_group.h"
 #elif defined(USE_CUDA)
 #include "cuda_process_group.h"
+#elif defined(USE_ILU)
+#include "ilu_process_group.h"
 #endif
 #include "common/global_flags.h"
 #include "parallel_args.h"

xllm/core/framework/parallel_state/ilu_process_group.h

@@ -0,0 +1,55 @@
+/* Copyright 2025 The xLLM Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    https://github.com/jd-opensource/xllm/blob/main/LICENSE
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#pragma once
+
+#include <torch/csrc/distributed/c10d/ProcessGroupNCCL.hpp>
+
+#include "process_group.h"
+
+namespace xllm {
+
+class ProcessGroupImpl : public ProcessGroup {
+ public:
+  ProcessGroupImpl(int32_t global_rank,
+                   int32_t world_size,
+                   int32_t rank_size,
+                   int32_t port,
+                   bool trans,
+                   const std::string& host,
+                   const std::string& group_name,
+                   const torch::Device& device)
+      : ProcessGroup(device) {
+    c10::intrusive_ptr<c10d::ProcessGroupNCCL::Options> pg_options =
+        c10d::ProcessGroupNCCL::Options::create();
+#if TORCH_VERSION_MAJOR > 2 || \
+    (TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR >= 7)
+    pg_options->group_name = group_name;
+#endif
+    int32_t rank = global_rank;
+    if (world_size != rank_size) {
+      auto [local_rank, group_ranks] =
+          get_group_rank(world_size, global_rank, rank_size, trans);
+      pg_options->global_ranks_in_group = group_ranks;
+      rank = local_rank;
+    }
+
+    auto store = create_tcp_store(host, port, rank);
+    pg_ = std::make_unique<c10d::ProcessGroupNCCL>(
+        store, rank, rank_size, pg_options);
+  }
+};
+
+}  // namespace xllm

xllm/core/framework/parallel_state/process_group.cpp

@@ -21,6 +21,8 @@ limitations under the License.
 #include "mlu_process_group.h"
 #elif defined(USE_CUDA)
 #include "cuda_process_group.h"
+#elif defined(USE_ILU)
+#include "ilu_process_group.h"
 #endif
 
 namespace {