v0.2.1 (#116)

* Fix a header installation bug * Check the Python version during build * Resolve compilation warnings
microsoft · Sep 7, 2023 · 335da67 · 335da67
1 parent 693fe01
commit 335da67
Show file tree

Hide file tree

Showing 28 changed files with 54 additions and 80 deletions.
diff --git a/CITATION.cff b/CITATION.cff
@@ -1,6 +1,6 @@
 cff-version: 1.2.0
 title: "ARK: A GPU-driven system framework for scalable AI applications"
-version: 0.2.0
+version: 0.2.1
 message: >-
   If you use this project in your research, please cite it as below.
 authors:

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -3,7 +3,7 @@
 
 set(ARK_MAJOR "0")
 set(ARK_MINOR "2")
-set(ARK_PATCH "0")
+set(ARK_PATCH "1")
 
 set(ARK_VERSION "${ARK_MAJOR}.${ARK_MINOR}.${ARK_PATCH}")
 set(ARK_SOVERSION "${ARK_MAJOR}.${ARK_MINOR}")
@@ -45,11 +45,11 @@ if(CUDAToolkit_FOUND)
     endif()
 
     if(CUDAToolkit_VERSION_MAJOR GREATER_EQUAL 11)
-        set(CMAKE_CUDA_ARCHITECTURES 70 80)
+        set(CMAKE_CUDA_ARCHITECTURES 60 70 80)
     endif()
 
     if(CUDAToolkit_VERSION_MAJOR GREATER_EQUAL 12)
-        set(CMAKE_CUDA_ARCHITECTURES ${CMAKE_CUDA_ARCHITECTURES} 70 80 90)
+        set(CMAKE_CUDA_ARCHITECTURES ${CMAKE_CUDA_ARCHITECTURES} 90)
     endif()
 endif()
 
@@ -100,50 +100,21 @@ add_custom_target(ut)
 # Details
 add_subdirectory(ark)
 
-# Add header files to library targets
-file(GLOB_RECURSE ARK_HEADERS CONFIGURE_DEPENDS
-    ${CMAKE_SOURCE_DIR}/ark/include/*.h)
-file(GLOB_RECURSE CUTLASS_HEADERS CONFIGURE_DEPENDS
-    ${CMAKE_SOURCE_DIR}/third_party/cutlass/include/*.h
-    ${CMAKE_SOURCE_DIR}/third_party/cutlass/include/*.hpp
-    ${CMAKE_SOURCE_DIR}/third_party/cutlass/include/*.inl)
-target_sources(ark PUBLIC
-    FILE_SET install_headers
-    TYPE HEADERS
-    BASE_DIRS ${CMAKE_SOURCE_DIR}/ark/include
-    FILES ${ARK_HEADERS}
-)
-target_sources(ark_static PUBLIC
-    FILE_SET install_headers
-    TYPE HEADERS
-    BASE_DIRS ${CMAKE_SOURCE_DIR}/ark/include
-    FILES ${ARK_HEADERS}
-)
-target_sources(ark PUBLIC
-    FILE_SET install_cutlass_headers
-    TYPE HEADERS
-    BASE_DIRS ${CMAKE_SOURCE_DIR}/third_party/cutlass/include
-    FILES ${CUTLASS_HEADERS}
-)
-target_sources(ark_static PUBLIC
-    FILE_SET install_cutlass_headers
-    TYPE HEADERS
-    BASE_DIRS ${CMAKE_SOURCE_DIR}/third_party/cutlass/include
-    FILES ${CUTLASS_HEADERS}
-)
-
-# Install libraries
-install(TARGETS ark ark_static
-    LIBRARY DESTINATION ark/lib
-    ARCHIVE DESTINATION ark/lib
-    FILE_SET install_headers DESTINATION ark/include
-    FILE_SET install_cutlass_headers DESTINATION ark/include/kernels
-)
-
-# Install Python module
 if(BUILD_PYTHON)
+    # Install Python module
     add_subdirectory(python)
+    add_dependencies(ark_py build)
+    install(TARGETS ark_py LIBRARY DESTINATION ark)
+else()
+    # Install libraries
+    install(TARGETS ark ark_static
+        LIBRARY DESTINATION ark/lib
+        ARCHIVE DESTINATION ark/lib
+    )
 endif()
 
+# Install header files
+install(DIRECTORY ${BUILD_DIR}/include DESTINATION ark)
+
 # Utils
 include(${PROJECT_SOURCE_DIR}/cmake/Utils.cmake)
diff --git a/ark/include/ark.h b/ark/include/ark.h
@@ -11,7 +11,7 @@
 
 #define ARK_MAJOR 0
 #define ARK_MINOR 2
-#define ARK_PATCH 0
+#define ARK_PATCH 1
 #define ARK_VERSION (ARK_MAJOR * 10000 + ARK_MINOR * 100 + ARK_PATCH)
 
 namespace ark {

diff --git a/ark/ops/ops_add.cc b/ark/ops/ops_add.cc
@@ -52,7 +52,7 @@ Tensor *Model::add(Tensor *input, Tensor *other, Tensor *output,
 {
     CHECK(input != nullptr);
     CHECK(other != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_div.cc b/ark/ops/ops_div.cc
@@ -53,7 +53,7 @@ Tensor *Model::div(Tensor *input, Tensor *other, Tensor *output,
 {
     assert(input != nullptr);
     assert(other != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_exp.cc b/ark/ops/ops_exp.cc
@@ -47,7 +47,7 @@ std::string ExpOp::function_name(const OpConfig &cfg) const
 Tensor *Model::exp(Tensor *input, Tensor *output, const string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_gelu.cc b/ark/ops/ops_gelu.cc
@@ -45,7 +45,7 @@ std::string GeluOp::function_name(const OpConfig &cfg) const
 Tensor *Model::gelu(Tensor *input, Tensor *output, const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_im2col.cc b/ark/ops/ops_im2col.cc
@@ -85,7 +85,7 @@ Tensor *Model::im2col(Tensor *input, int kernel_height, int kernel_width,
                       Tensor *output, const std::string &name)
 {
     assert(input != nullptr);
-    DimType n, c, h, w;
+    DimType n = 1, c = 1, h = 1, w = 1;
     int input_ndims = input->ndims();
     if (input_ndims == 2) {
         n = 1;
@@ -107,7 +107,7 @@ Tensor *Model::im2col(Tensor *input, int kernel_height, int kernel_width,
             "invalid # of input dimensions. Expected 2, 3, or 4, but given ",
             input_ndims);
     }
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_layernorm.cc b/ark/ops/ops_layernorm.cc
@@ -46,7 +46,7 @@ std::string LayernormOp::function_name(const OpConfig &cfg) const
 Tensor *Model::layernorm(Tensor *input, Tensor *output, const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_matmul.cc b/ark/ops/ops_matmul.cc
@@ -136,7 +136,7 @@ Tensor *Model::matmul(Tensor *mat_a, Tensor *mat_b, Tensor *mat_y,
         LOG(ERROR, "inner dimensions mismatch: ", k, " and ", k2);
     }
 
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (mat_a->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (mat_a->type == FP32) {

diff --git a/ark/ops/ops_max_pool.cc b/ark/ops/ops_max_pool.cc
@@ -20,7 +20,7 @@ Tensor *Model::max_pool(Tensor *input, DimType kernel_size, DimType stride,
                         Tensor *output, const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_mul.cc b/ark/ops/ops_mul.cc
@@ -51,7 +51,7 @@ Tensor *Model::mul(Tensor *input, Tensor *other, Tensor *output,
 {
     assert(input != nullptr);
     assert(other != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_reduce.cc b/ark/ops/ops_reduce.cc
@@ -140,7 +140,7 @@ Tensor *Model::reduce_sum(Tensor *input, int axis, Tensor *output,
                           const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {
@@ -174,7 +174,7 @@ Tensor *Model::reduce_mean(Tensor *input, int axis, Tensor *output,
                            const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {
@@ -208,7 +208,7 @@ Tensor *Model::reduce_max(Tensor *input, int axis, Tensor *output,
                           const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_relu.cc b/ark/ops/ops_relu.cc
@@ -45,7 +45,7 @@ std::string ReluOp::function_name(const OpConfig &cfg) const
 Tensor *Model::relu(Tensor *input, Tensor *output, const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_rmsnorm.cc b/ark/ops/ops_rmsnorm.cc
@@ -46,7 +46,7 @@ std::string RMSnormOp::function_name(const OpConfig &cfg) const
 Tensor *Model::rmsnorm(Tensor *input, Tensor *output, const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_rope.cc b/ark/ops/ops_rope.cc
@@ -51,7 +51,7 @@ Tensor *Model::rope(Tensor *input, Tensor *other, Tensor *output,
 {
     assert(input != nullptr);
     assert(other != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_rope_test.cc b/ark/ops/ops_rope_test.cc
@@ -19,8 +19,6 @@ void baseline_rope(std::vector<void *> &outputs, const std::vector<ark::Dims> &,
     T *other = static_cast<T *>(inputs[1]);
 
     ark::Dims ish = input_shapes[0].dims4();
-    ark::Dims other_sh = input_shapes[1].dims4();
-    assert(ish == other_sh);
 
     for (ark::DimType n = 0; n < ish[0]; ++n) {
         for (ark::DimType c = 0; c < ish[1]; ++c) {

diff --git a/ark/ops/ops_scale.cc b/ark/ops/ops_scale.cc
@@ -62,7 +62,7 @@ Tensor *Model::scale(Tensor *input, float val, Tensor *output,
                      const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_sigmoid.cc b/ark/ops/ops_sigmoid.cc
@@ -46,7 +46,7 @@ std::string SigmoidOp::function_name(const OpConfig &cfg) const
 Tensor *Model::sigmoid(Tensor *input, Tensor *output, const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_softmax.cc b/ark/ops/ops_softmax.cc
@@ -39,7 +39,7 @@ std::string SoftmaxOp::function_name(const OpConfig &cfg) const
 Tensor *Model::softmax(Tensor *input, Tensor *output, const std::string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_sqrt.cc b/ark/ops/ops_sqrt.cc
@@ -47,7 +47,7 @@ std::string SqrtOp::function_name(const OpConfig &cfg) const
 Tensor *Model::sqrt(Tensor *input, Tensor *output, const string &name)
 {
     assert(input != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_sub.cc b/ark/ops/ops_sub.cc
@@ -53,7 +53,7 @@ Tensor *Model::sub(Tensor *input, Tensor *other, Tensor *output,
 {
     assert(input != nullptr);
     assert(other != nullptr);
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/ops/ops_transpose.cc b/ark/ops/ops_transpose.cc
@@ -48,7 +48,7 @@ std::string TransposeOp::function_name(const OpConfig &cfg) const
 Tensor *Model::transpose(Tensor *input, Dims perm, Tensor *output,
                          const std::string &name)
 {
-    OpPrecType pt;
+    OpPrecType pt = OP_PREC_NONE;
     if (input->type == FP16) {
         pt = OP_PREC_FP16;
     } else if (input->type == FP32) {

diff --git a/ark/sched/sched.cc b/ark/sched/sched.cc
@@ -75,7 +75,7 @@ const OpConfig *BaseScheduler::sched_op_config(const Op *op)
         return nullptr;
     }
     const GpuInfo &gpu_info = this->gpu_mgr->get_gpu_info();
-    OpArchType arch_type;
+    OpArchType arch_type = OP_ARCH_CUDA_60;
     if (gpu_info.arch == GPU_ARCH_CUDA_60) {
         arch_type = OP_ARCH_CUDA_60;
     } else if (gpu_info.arch == GPU_ARCH_CUDA_70) {

diff --git a/docs/sphinx/source/conf.py b/docs/sphinx/source/conf.py
@@ -20,8 +20,8 @@
 project = "ARK"
 copyright = "2023, ARK Team"
 author = "ARK Team"
-version = "0.2.0"
-release = "0.2.0"
+version = "0.2.1"
+release = "0.2.1"
 
 # -- General configuration ---------------------------------------------------
 

diff --git a/pyproject.toml b/pyproject.toml
@@ -4,14 +4,17 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "ark"
-version = "0.2.0"
+version = "0.2.1"
 
 [tool.scikit-build]
 cmake.minimum-version = "3.25"
 cmake.args = []
 cmake.verbose = false
 cmake.build-type = "Release"
+cmake.targets = ["ark_py"]
 wheel.packages = ["python/ark"]
+wheel.license-files = ["LICENSE", "CITATION.cff", "CODE_OF_CONDUCT.md", "SECURITY.md", "SUPPORT.md"]
+install.strip = true
 build-dir = "build/{wheel_tag}"
 
 [tool.scikit-build.cmake.define]

diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt
@@ -1,7 +1,13 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
 
-find_package(Python3 COMPONENTS Interpreter Development.Module REQUIRED)
+find_package(Python3 REQUIRED COMPONENTS Interpreter Development.Module)
+if(Python3_FOUND)
+  if(${Python3_VERSION} VERSION_LESS 3.8)
+    message(FATAL_ERROR "Python version must be at least 3.8")
+  endif()
+endif()
+
 include(FetchContent)
 FetchContent_Declare(
     pybind11
@@ -13,7 +19,3 @@ FetchContent_MakeAvailable(pybind11)
 pybind11_add_module(ark_py ${CMAKE_CURRENT_SOURCE_DIR}/bindings.cpp)
 set_target_properties(ark_py PROPERTIES OUTPUT_NAME _ark_core)
 target_link_libraries(ark_py PRIVATE ark_static)
-
-if (SKBUILD)
-    install(TARGETS ark_py LIBRARY DESTINATION ark)
-endif()
diff --git a/third_party/CMakeLists.txt b/third_party/CMakeLists.txt
@@ -12,7 +12,7 @@ add_custom_target(tp-update
 
 # Patch CUTLASS
 add_custom_target(tp-cutlass
-    COMMAND cd ${CMAKE_CURRENT_SOURCE_DIR}/cutlass && patch -sN -p1 < ${CMAKE_CURRENT_SOURCE_DIR}/patches/cutlass/cutlass.patch | true
+    COMMAND cd ${CMAKE_CURRENT_SOURCE_DIR}/cutlass && patch -sN -p1 < ${CMAKE_CURRENT_SOURCE_DIR}/patches/cutlass/cutlass.patch || true
 )
 add_dependencies(tp-cutlass tp-update)