Update Rapids version to 22.10 (#617)

Closes #486
Closes #251

Authors:
  - Eli Fajardo (https://github.com/efajardo-nv)
  - Michael Demoret (https://github.com/mdemoret-nv)

Approvers:
  - Devin Robison (https://github.com/drobison00)
  - David Gardner (https://github.com/dagardner-nv)
  - Michael Demoret (https://github.com/mdemoret-nv)

URL: #617

.devcontainer/docker-compose.yml

@@ -20,7 +20,7 @@ services:
   triton:
     container_name: morpheus-triton
     runtime: nvidia
-    image: nvcr.io/nvidia/tritonserver:22.08-py3
+    image: nvcr.io/nvidia/tritonserver:22.10-py3
     command: tritonserver --model-repository=/models --exit-on-error=false ${TRITON_MODEL_ARGS}
     ports:
       - 8000:8000

.gitmodules

@@ -4,5 +4,5 @@
 	branch = branch-23.01
 [submodule "morpheus_utils"]
 	path = external/utilities
-	url = ../../nv-morpheus/utilities.git
-	branch = branch-23.01 
+	url = https://github.com/nv-morpheus/utilities.git
+	branch = branch-23.01

CMakeLists.txt

@@ -35,7 +35,7 @@ option(MORPHEUS_USE_IWYU "Enable running include-what-you-use as part of the bui
 option(MORPHEUS_BUILD_DOCS "Enable building of API documentation" OFF)
 set(MORPHEUS_PY_INSTALL_DIR "${CMAKE_CURRENT_BINARY_DIR}/wheel" CACHE STRING "Location to install the python directory")
 
-set(MORPHEUS_RAPIDS_CMAKE_VERSION "22.08" CACHE STRING "Sets default versions for RAPIDS CMake.")
+set(MORPHEUS_RAPIDS_VERSION "22.10" CACHE STRING "Sets default versions for RAPIDS libraries.")
 set(MORPHEUS_CACHE_DIR "${CMAKE_SOURCE_DIR}/.cache" CACHE PATH "Directory to contain all CPM and CCache data")
 mark_as_advanced(MORPHEUS_CACHE_DIR)
 
@@ -68,8 +68,10 @@ set(MORPHEUS_CMAKE_PREFIX_PATH_EXTENSIONS
 list(PREPEND CMAKE_MODULE_PATH "${MORPHEUS_CMAKE_MODULE_PATH_EXTENSIONS}")
 list(PREPEND CMAKE_PREFIX_PATH "${MORPHEUS_CMAKE_PREFIX_PATH_EXTENSIONS}")
 
+# Force the MORPHEUS_UTILS_RAPIDS_VERSION to match our value
+set(MORPHEUS_UTILS_RAPIDS_VERSION ${MORPHEUS_RAPIDS_VERSION} CACHE STRING "" FORCE)
+
 # Load morpheus utils and update CMake paths
-set(MORPHEUS_UTILS_RAPIDS_CMAKE_VERSION ${MORPHEUS_RAPIDS_CMAKE_VERSION})
 include(morpheus_utils/load)
 
 # Configure project package manager

ci/conda/recipes/morpheus/meta.yaml

@@ -18,7 +18,7 @@
 {% set py_version=environ.get('CONDA_PY', '3.8') %}
 {% set cuda_version='.'.join(environ.get('CUDA', '11.5').split('.')[:2]) %}
 {% set cuda_major=cuda_version.split('.')[0] %}
-{% set rapids_version = "22.08" %}
+{% set rapids_version = "22.10" %}
 
 package:
   name: morpheus-split
@@ -55,7 +55,6 @@ outputs:
         - ccache
         - ninja
       host:
-        - cuda-python <=11.7.0 # Remove when Issue #251 is closed
         - cudf {{ rapids_version }}
         - cython >=0.29,<0.30
         - libcudf {{ rapids_version }}
@@ -71,7 +70,6 @@ outputs:
         # Runtime only requirements. This + setup.y is the definitive runtime requirement list
         - click >=8
         - configargparse 1.5
-        - cuda-python <=11.7.0 # Remove when Issue #251 is closed
         - cudf {{ rapids_version }}
         - cudf_kafka {{ rapids_version }}
         - cupy # Version determined from cudf

docker/Dockerfile

@@ -29,7 +29,7 @@ FROM ${FROM_IMAGE}:${CUDA_VER}-devel-${LINUX_DISTRO}${LINUX_VER} AS base
 # Required arguments
 ARG IMAGE_TYPE=base
 ARG RAPIDS_CHANNEL=rapidsai-nightly
-ARG RAPIDS_VER=22.08
+ARG RAPIDS_VER=22.10
 ARG PYTHON_VER=3.8
 ARG CONDA_CHANNEL=rapidsai
 

docker/build_container.sh

@@ -26,7 +26,7 @@ FROM_IMAGE=${FROM_IMAGE:-"gpuci/miniforge-cuda"}
 CUDA_VER=${CUDA_VER:-11.5}
 LINUX_DISTRO=${LINUX_DISTRO:-ubuntu}
 LINUX_VER=${LINUX_VER:-20.04}
-RAPIDS_VER=${RAPIDS_VER:-22.08}
+RAPIDS_VER=${RAPIDS_VER:-22.10}
 PYTHON_VER=${PYTHON_VER:-3.8}
 TENSORRT_VERSION=${TENSORRT_VERSION:-8.2.1.3}
 

docker/conda/environments/cuda11.5_dev.yml

@@ -32,9 +32,8 @@ dependencies:
     - cmake=3.24
     - configargparse=1.5
     - cuda-nvml-dev=11.5
-    - cuda-python<=11.5.0 # Remove when Issue #251 is closed
     - cudatoolkit=11.5
-    - cudf 22.08
+    - cudf 22.10
     - cupy=9.5.0
     - cython=0.29.24
     - datacompy=0.8
@@ -63,7 +62,7 @@ dependencies:
     - networkx=2.8
     - ninja=1.10
     - nodejs=17.4.0
-    - numba==0.55
+    - numba>=0.56.2
     - numpydoc=1.4
     - nvcc_linux-64=11.5
     - pandas=1.3

morpheus/CMakeLists.txt

@@ -17,6 +17,8 @@ list(APPEND CMAKE_MESSAGE_CONTEXT "morpheus")
 
 option(MORPHEUS_PYTHON_INPLACE_BUILD
         "Whether or not to copy built python modules back to the source tree for debug purposes." OFF)
+option(MORPHEUS_PYTHON_PERFORM_INSTALL
+        "Whether or not to automatically `pip install` any built python library. WARNING: This may overwrite any existing installation of the same name." OFF)
 option(MORPHEUS_BUILD_PYTHON_STUBS
         "Whether or not to generate stubs for python files." ON)
 

morpheus/_lib/cudf_helpers.pyx

@@ -25,6 +25,7 @@ from libcpp.vector cimport vector
 
 from cudf._lib.column cimport Column
 from cudf._lib.cpp.column.column_view cimport column_view
+from cudf._lib.cpp.io.types cimport column_name_info
 from cudf._lib.cpp.io.types cimport table_metadata
 from cudf._lib.cpp.io.types cimport table_with_metadata
 from cudf._lib.cpp.table.table_view cimport table_view
@@ -60,9 +61,23 @@ cdef public api:
         index_names = None
 
         if (index_col_count > 0):
-            index_names = [x.decode() for x in table.metadata.column_names[0:index_col_count]]
-
-        column_names = [x.decode() for x in table.metadata.column_names[index_col_count:]]
+            index_names = []
+
+            # Need to support both column_names and schema_info
+            if (table.metadata.column_names.size() > 0):
+                index_names = [x.decode() for x in table.metadata.column_names[0:index_col_count]]
+            elif (table.metadata.schema_info.size() > 0):
+                for i in range(min(index_col_count, table.metadata.schema_info.size())):
+                    index_names.append(table.metadata.schema_info[i].name.decode())
+
+        column_names = []
+
+        # Need to support both column_names and schema_info
+        if (table.metadata.column_names.size() > 0):
+            column_names = [x.decode() for x in table.metadata.column_names[index_col_count:]]
+        elif (table.metadata.schema_info.size() > 0):
+            for i in range(index_col_count, table.metadata.schema_info.size()):
+                column_names.append(table.metadata.schema_info[i].name.decode())
 
         data, index = data_from_unique_ptr(move(table.tbl), column_names=column_names, index_names=index_names)
 

morpheus/_lib/include/morpheus/io/deserializers.hpp

@@ -30,6 +30,15 @@ namespace morpheus {
  * @file
  */
 
+/**
+ * @brief Get the column names from table object. Looks at both column_names as well as schema_info and returns the
+ * correct one.
+ *
+ * @param table The table to pull the columns from
+ * @return std::vector<std::string>
+ */
+std::vector<std::string> get_column_names_from_table(const cudf::io::table_with_metadata& table);
+
 /**
  * @brief Loads a cudf table from either CSV or JSON file
  *

morpheus/_lib/include/morpheus/utilities/type_util.hpp

@@ -61,11 +61,5 @@ struct DType : DataType  // NOLINT
     static DType from_triton(const std::string& type_str);
 };
 
-template <typename T>
-DType type_to_dtype()
-{
-    return DType::from_triton(cudf::type_to_id<T>);
-}
-
 /** @} */  // end of group
 }  // namespace morpheus

morpheus/_lib/src/io/deserializers.cpp

@@ -17,6 +17,8 @@
 
 #include "morpheus/io/deserializers.hpp"
 
+#include "morpheus/utilities/stage_util.hpp"
+
 #include <cudf/io/csv.hpp>
 #include <cudf/io/json.hpp>
 #include <cudf/scalar/scalar.hpp>  // for string_scalar
@@ -25,27 +27,51 @@
 #include <cudf/types.hpp>        // for cudf::type_id
 #include <glog/logging.h>
 
+#include <algorithm>
 #include <filesystem>
 #include <ostream>  // needed for logging
 #include <regex>
 
 namespace morpheus {
 
+std::vector<std::string> get_column_names_from_table(const cudf::io::table_with_metadata& table)
+{
+    DCHECK(!(!table.metadata.column_names.empty() && !table.metadata.schema_info.empty()))
+        << "Both column_names and schema_info were set on the table_with_metadata object. Defaulting to column_names";
+
+    // If column_names is populated, use that
+    if (!table.metadata.column_names.empty())
+    {
+        return table.metadata.column_names;
+    }
+
+    // Otherwise, use schema_info
+    if (!table.metadata.schema_info.empty())
+    {
+        return foreach_map(table.metadata.schema_info, [](auto schema) { return schema.name; });
+    }
+
+    // Return empty
+    return {};
+}
+
 cudf::io::table_with_metadata load_json_table(cudf::io::json_reader_options&& json_options)
 {
     auto tbl = cudf::io::read_json(json_options);
 
-    auto found = std::find(tbl.metadata.column_names.begin(), tbl.metadata.column_names.end(), "data");
+    auto column_names = get_column_names_from_table(tbl);
+
+    auto found = std::find(column_names.begin(), column_names.end(), "data");
 
-    if (found == tbl.metadata.column_names.end())
+    if (found == column_names.end())
         return tbl;
 
     // Super ugly but cudf cant handle newlines and add extra escapes. So we need to convert
     // \\n -> \n
     // \\/ -> \/
     auto columns = tbl.tbl->release();
 
-    size_t idx = found - tbl.metadata.column_names.begin();
+    size_t idx = found - column_names.begin();
 
     auto updated_data = cudf::strings::replace(
         cudf::strings_column_view{columns[idx]->view()}, cudf::string_scalar("\\n"), cudf::string_scalar("\n"));
@@ -87,21 +113,21 @@ int get_index_col_count(cudf::io::table_with_metadata& data_table)
     int index_col_count = 0;
 
     // Check if we have a first column with INT64 data type
-    if (data_table.metadata.column_names.size() >= 1 &&
+    if (data_table.metadata.schema_info.size() >= 1 &&
         data_table.tbl->get_column(0).type().id() == cudf::type_id::INT64)
     {
         std::regex index_regex(R"((unnamed: 0|id))", std::regex_constants::ECMAScript | std::regex_constants::icase);
 
         // Get the column name
-        auto col_name = data_table.metadata.column_names[0];
+        auto col_name = data_table.metadata.schema_info[0].name;
 
         // Check it against some common terms
         if (std::regex_search(col_name, index_regex))
         {
             // Also, if its the hideous 'Unnamed: 0', then just use an empty string
             if (col_name == "Unnamed: 0")
             {
-                data_table.metadata.column_names[0] = "";
+                data_table.metadata.schema_info[0].name = "";
             }
 
             index_col_count = 1;

morpheus/_lib/src/io/serializers.cpp

@@ -115,10 +115,20 @@ void df_to_csv(const TableInfo& tbl, std::ostream& out_stream, bool include_head
                                .false_value("False"s);
 
     cudf::io::table_metadata metadata{};
+
     if (include_header)
     {
         metadata.column_names = column_names;
-        options_builder       = options_builder.metadata(&metadata);
+
+        // After cuDF PR #11364, use schema_info instead of column_names (actually just set both)
+        metadata.schema_info = std::vector<cudf::io::column_name_info>();
+
+        for (auto& name : column_names)
+        {
+            metadata.schema_info.emplace_back(cudf::io::column_name_info{name});
+        }
+
+        options_builder = options_builder.metadata(&metadata);
     }
 
     cudf::io::write_csv(options_builder.build(), rmm::mr::get_current_device_resource());

morpheus/_lib/src/messages/meta.cpp

@@ -106,14 +106,14 @@ TableInfo MessageMeta::get_info() const
     return this->m_data->get_info();
 }
 
-std::shared_ptr<MessageMeta> MessageMeta::create_from_python(pybind11::object &&data_table)
+std::shared_ptr<MessageMeta> MessageMeta::create_from_python(pybind11::object&& data_table)
 {
     auto data = std::make_unique<PyDataTable>(std::move(data_table));
 
     return std::shared_ptr<MessageMeta>(new MessageMeta(std::move(data)));
 }
 
-std::shared_ptr<MessageMeta> MessageMeta::create_from_cpp(cudf::io::table_with_metadata &&data_table,
+std::shared_ptr<MessageMeta> MessageMeta::create_from_cpp(cudf::io::table_with_metadata&& data_table,
                                                           int index_col_count)
 {
     // Convert to py first
@@ -126,7 +126,7 @@ std::shared_ptr<MessageMeta> MessageMeta::create_from_cpp(cudf::io::table_with_m
 
 MessageMeta::MessageMeta(std::shared_ptr<IDataTable> data) : m_data(std::move(data)) {}
 
-pybind11::object MessageMeta::cpp_to_py(cudf::io::table_with_metadata &&table, int index_col_count)
+pybind11::object MessageMeta::cpp_to_py(cudf::io::table_with_metadata&& table, int index_col_count)
 {
     pybind11::gil_scoped_acquire gil;
 
@@ -145,17 +145,17 @@ pybind11::object MessageMeta::cpp_to_py(cudf::io::table_with_metadata &&table, i
 }
 
 /********** MessageMetaInterfaceProxy **********/
-std::shared_ptr<MessageMeta> MessageMetaInterfaceProxy::init_python(pybind11::object &&data_frame)
+std::shared_ptr<MessageMeta> MessageMetaInterfaceProxy::init_python(pybind11::object&& data_frame)
 {
     return MessageMeta::create_from_python(std::move(data_frame));
 }
 
-cudf::size_type MessageMetaInterfaceProxy::count(MessageMeta &self)
+cudf::size_type MessageMetaInterfaceProxy::count(MessageMeta& self)
 {
     return self.count();
 }
 
-pybind11::object MessageMetaInterfaceProxy::get_data_frame(MessageMeta &self)
+pybind11::object MessageMetaInterfaceProxy::get_data_frame(MessageMeta& self)
 {
     // // Get the column and convert to cudf
     // auto py_table_struct = make_table_from_view_and_meta(self.m_pydf;.tbl->view(),
@@ -168,7 +168,7 @@ pybind11::object MessageMetaInterfaceProxy::get_data_frame(MessageMeta &self)
     return self.get_py_table();
 }
 
-std::shared_ptr<MessageMeta> MessageMetaInterfaceProxy::init_cpp(const std::string &filename)
+std::shared_ptr<MessageMeta> MessageMetaInterfaceProxy::init_cpp(const std::string& filename)
 {
     // Load the file
     auto df_with_meta = CuDFTableUtil::load_table(filename);