feat: add InfiniOps as optional kernel provider

Wire InfiniOps in as a pluggable kernel provider keyed at the GEMM level:
Dispatcher consults a per-key whitelist hook and routes registered ops to
InfiniOps, falling back to the default CUDA kernel otherwise. linear, matmul
and outer now invoke Gemm via Dispatcher rather than calling the cuBLAS
wrapper directly, so InfiniOps Gemm transparently covers all three.

Author: chen2021673

.gitmodules

@@ -10,3 +10,6 @@
 [submodule "third_party/googletest"]
 	path = third_party/googletest
 	url = git@github.com:google/googletest.git
+[submodule "third_party/InfiniOps"]
+	path = third_party/InfiniOps
+	url = git@github.com:InfiniTensor/InfiniOps.git

CMakeLists.txt

@@ -4,6 +4,7 @@ option(USE_CUDA "Support NVIDIA CUDA" OFF)
 option(PROFILE_MODE "ENABLE PROFILE MODE" OFF)
 option(USE_OMP "Use OpenMP as backend for Eigen" ON)
 option(USE_NCCL "Build project for distributed running" ON)
+option(USE_INFINIOPS "Use InfiniOps as an optional kernel provider" OFF)
 option(BUILD_TEST "Build InfiniTrain tests" OFF)
 
 project(infini_train VERSION 0.5.0 LANGUAGES CXX)
@@ -51,6 +52,32 @@ include_directories(${PROJECT_SOURCE_DIR}/third_party/eigen)
 
 include_directories(${PROJECT_SOURCE_DIR})
 
+if(USE_INFINIOPS)
+  add_compile_definitions(USE_INFINIOPS=1)
+
+  set(INFINIOPS_SOURCE_DIR "${PROJECT_SOURCE_DIR}/third_party/InfiniOps")
+  if(NOT EXISTS "${INFINIOPS_SOURCE_DIR}/CMakeLists.txt")
+    message(FATAL_ERROR
+      "USE_INFINIOPS=ON requires InfiniOps under third_party/InfiniOps. "
+      "Run: git submodule update --init third_party/InfiniOps")
+  endif()
+
+  set(INFINIOPS_WITH_CPU OFF)
+  if(NOT USE_CUDA)
+    set(INFINIOPS_WITH_CPU ON)
+  endif()
+
+  set(WITH_CPU ${INFINIOPS_WITH_CPU} CACHE BOOL "Enable InfiniOps CPU backend" FORCE)
+  set(WITH_NVIDIA ${USE_CUDA} CACHE BOOL "Enable InfiniOps NVIDIA backend" FORCE)
+  add_subdirectory(${INFINIOPS_SOURCE_DIR} ${CMAKE_BINARY_DIR}/third_party/InfiniOps EXCLUDE_FROM_ALL)
+  if(NOT TARGET infiniops)
+    message(FATAL_ERROR "InfiniOps third-party project did not define target `infiniops`")
+  endif()
+  if(NOT TARGET InfiniOps::infiniops)
+    add_library(InfiniOps::infiniops ALIAS infiniops)
+  endif()
+endif()
+
 if(PROFILE_MODE)
   add_compile_definitions(PROFILE_MODE=1)
 endif()
@@ -62,9 +89,13 @@ endif()
 # Framework core sources (*.cc), excluding cpu kernels (they are built separately)
 file(GLOB_RECURSE SRC ${PROJECT_SOURCE_DIR}/infini_train/src/*.cc)
 list(FILTER SRC EXCLUDE REGEX ".*kernels/cpu/.*")
+if(NOT USE_INFINIOPS)
+  list(FILTER SRC EXCLUDE REGEX ".*infini_train/src/core/kernel_provider/infiniops/.*\.cc$")
+endif()
 if(NOT USE_CUDA)
   list(FILTER SRC EXCLUDE REGEX ".*runtime/cuda/.*")
   list(FILTER SRC EXCLUDE REGEX ".*ccl/cuda/.*")
+  list(FILTER SRC EXCLUDE REGEX ".*infini_train/src/core/kernel_provider/infiniops/cuda/.*")
 endif()
 if(NOT USE_NCCL)
   list(FILTER SRC EXCLUDE REGEX ".*infini_train/src/core/ccl/cuda/.*")
@@ -100,8 +131,14 @@ if(USE_CUDA)
 
   # Only compile CUDA kernels / cuda sources here (your original used src/*.cu)
   file(GLOB_RECURSE CUDA_KERNELS ${PROJECT_SOURCE_DIR}/infini_train/src/*.cu)
+  if(NOT USE_INFINIOPS)
+    list(FILTER CUDA_KERNELS EXCLUDE REGEX ".*infini_train/src/core/kernel_provider/infiniops/.*\.cu$")
+  endif()
 
   add_library(infini_train_cuda_kernels STATIC ${CUDA_KERNELS})
+  if(USE_INFINIOPS)
+    target_link_libraries(infini_train_cuda_kernels PUBLIC InfiniOps::infiniops)
+  endif()
   set_target_properties(infini_train_cuda_kernels PROPERTIES CUDA_ARCHITECTURES "75;80;90")
 
   target_link_libraries(infini_train_cuda_kernels
@@ -126,6 +163,9 @@ endif()
 # ------------------------------------------------------------------------------
 
 add_library(infini_train STATIC ${SRC})
+if(USE_INFINIOPS)
+  target_link_libraries(infini_train PUBLIC InfiniOps::infiniops)
+endif()
 target_link_libraries(infini_train
   PUBLIC
     glog

infini_train/include/core/kernel_provider/infiniops/adapter.h

@@ -0,0 +1,45 @@
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <mutex>
+#include <vector>
+
+#include <handle.h>
+
+#include "data_type.h"
+#include "tensor.h"
+
+#include "infini_train/include/datatype.h"
+#include "infini_train/include/device.h"
+
+namespace infini_train {
+class Tensor;
+} // namespace infini_train
+
+namespace infini_train::core {
+class Stream;
+} // namespace infini_train::core
+
+namespace infini_train::kernel_provider::infiniops {
+
+infini::ops::DataType ToOpsDataType(DataType dtype);
+
+infini::ops::Device ToOpsDevice(const Device &device);
+
+std::mutex &InfiniOpsCallMutex();
+
+using HandleFactory = infini::ops::Handle (*)(const Device &device, core::Stream *stream);
+
+void RegisterHandleFactory(Device::DeviceType type, HandleFactory factory);
+
+infini::ops::Handle GetHandle(const Device &device);
+
+infini::ops::Tensor ToOpsTensor(const std::shared_ptr<Tensor> &tensor);
+
+infini::ops::Tensor ToOpsTensor(void *data, const std::vector<int64_t> &dims, DataType dtype, const Device &device);
+
+infini::ops::Tensor ToOpsTensor(void *data, const std::vector<int64_t> &dims, DataType dtype, const Device &device,
+                                const std::vector<int64_t> &strides);
+
+} // namespace infini_train::kernel_provider::infiniops

infini_train/include/core/kernel_provider/infiniops_registry.h

@@ -0,0 +1,50 @@
+#pragma once
+
+#include <map>
+#include <string>
+#include <utility>
+
+#include "glog/logging.h"
+
+#include "infini_train/include/device.h"
+#include "infini_train/include/dispatcher.h"
+
+namespace infini_train::kernel_provider {
+
+using KeyT = std::pair<Device::DeviceType, std::string>;
+
+class InfiniOpsRegistry {
+public:
+    static InfiniOpsRegistry &Instance() {
+        static InfiniOpsRegistry instance;
+        return instance;
+    }
+
+    const KernelFunction *Lookup(const std::string &kernel_name) const {
+        auto it = name_to_kernel_map_.find(kernel_name);
+        return it == name_to_kernel_map_.end() ? nullptr : &it->second;
+    }
+
+    template <typename FuncT> void Register(const std::string &kernel_name, FuncT &&kernel) {
+        CHECK(!name_to_kernel_map_.contains(kernel_name)) << "InfiniOps kernel already registered: " << kernel_name;
+        name_to_kernel_map_.emplace(kernel_name, kernel);
+    }
+
+private:
+    std::map<std::string, KernelFunction> name_to_kernel_map_;
+};
+
+// Bridge functions used by Dispatcher::GetKernel. Implemented in
+// infiniops_registry.cc; declared here for users that already include
+// the full registry header (e.g. unit tests).
+bool InfiniOpsEnabled();
+bool InfiniOpsEnabled(const KeyT &key);
+const KernelFunction *LookupInfiniOpsKernel(const KeyT &key);
+
+} // namespace infini_train::kernel_provider
+
+#define REGISTER_INFINIOPS_KERNEL(kernel_name, kernel_func)                                                            \
+    static const bool _##kernel_name##_infiniops_registered##__COUNTER__ = []() {                                      \
+        infini_train::kernel_provider::InfiniOpsRegistry::Instance().Register(#kernel_name, kernel_func);              \
+        return true;                                                                                                   \
+    }();

infini_train/include/dispatcher.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include <map>
+#include <string>
 #include <type_traits>
 #include <utility>
 
@@ -47,6 +48,11 @@ class KernelFunction {
     void *func_ptr_ = nullptr;
 };
 
+namespace kernel_provider {
+bool InfiniOpsEnabled(const std::pair<Device::DeviceType, std::string> &key);
+const KernelFunction *LookupInfiniOpsKernel(const std::pair<Device::DeviceType, std::string> &key);
+} // namespace kernel_provider
+
 class Dispatcher {
 public:
     using KeyT = std::pair<Device::DeviceType, std::string>;
@@ -57,6 +63,17 @@ class Dispatcher {
     }
 
     const KernelFunction &GetKernel(KeyT key) const {
+        if (kernel_provider::InfiniOpsEnabled(key)) {
+            if (const auto *kernel = kernel_provider::LookupInfiniOpsKernel(key)) {
+#ifdef PROFILE_MODE
+                SetProfileContext(key.second, key.first);
+#endif
+                return *kernel;
+            }
+            LOG(WARNING) << "InfiniOps kernel enabled but not registered: " << key.second
+                         << " on device: " << static_cast<int>(key.first) << "; falling back to default kernel";
+        }
+
         CHECK(key_to_kernel_map_.contains(key))
             << "Kernel not found: " << key.second << " on device: " << static_cast<int>(key.first);
 #ifdef PROFILE_MODE

infini_train/include/kernels/common/gemm.h

@@ -0,0 +1,48 @@
+#pragma once
+
+#include "infini_train/include/datatype.h"
+#include "infini_train/include/device.h"
+
+namespace infini_train::kernels {
+
+enum class GemmTranspose : int {
+    kNoTranspose = 0,
+    kTranspose = 1,
+};
+
+/**
+ * Parameter bundle for a single GEMM call:
+ *   C = alpha * op(A) * op(B) + beta * C
+ *
+ * batch_count == 1 describes a non-batched GEMM. batch_count > 1 describes a
+ * strided-batched GEMM. When batch_count == 1, stride_a/b/c are unused and must
+ * be left at 0.
+ */
+struct GemmParams {
+    GemmTranspose trans_a = GemmTranspose::kNoTranspose;
+    GemmTranspose trans_b = GemmTranspose::kNoTranspose;
+
+    int m = 0; // rows of op(A) and C
+    int n = 0; // cols of op(B) and C
+    int k = 0; // cols of op(A) == rows of op(B)
+
+    const void *A = nullptr;
+    int lda = 0;
+    const void *B = nullptr;
+    int ldb = 0;
+    void *C = nullptr;
+    int ldc = 0;
+
+    float alpha = 1.0f;
+    float beta = 0.0f;
+
+    int batch_count = 1;
+    long long stride_a = 0;
+    long long stride_b = 0;
+    long long stride_c = 0;
+
+    DataType input_dtype;
+    DataType output_dtype;
+};
+
+} // namespace infini_train::kernels

infini_train/include/tensor.h

@@ -139,7 +139,7 @@ class Tensor : public std::enable_shared_from_this<Tensor> {
     std::shared_ptr<Tensor> View(const std::vector<int64_t> &dims);
     std::shared_ptr<Tensor> Contiguous();
     // FIXME: Currently returns true unconditionally. Requires stride tracking in the Tensor
-    // class before this can be implemented correctly. The guard in elementwise.cu ensures
+    // class before this can be implemented correctly. The elementwise broadcast guard ensures
     // non-contiguous tensors fall back to the broadcast path until this is resolved.
     bool IsContiguous() const;
     std::shared_ptr<Tensor> Flatten(int64_t start = 0, int64_t end = -1);