Migrate to C++20 and drop CUDA 11 support

.azure-pipelines/codecov.yml

@@ -38,6 +38,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)
@@ -59,6 +61,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)

.azure-pipelines/integration-test.yml

@@ -30,10 +30,10 @@ jobs:
   displayName: Integration test A100
   strategy:
     matrix:
-      cuda11:
-        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda11.8
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   pool:
     name: msccl-ci
@@ -53,6 +53,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   pool:
     name: msccl-ci-h100

.azure-pipelines/multi-nodes-test.yml

@@ -31,6 +31,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
   pool:
     name: mscclpp-multi-node
   container:

.azure-pipelines/nccl-api-test.yml

@@ -35,6 +35,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)
@@ -56,6 +58,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)

.azure-pipelines/sglang-multi-node-test.yml

@@ -67,6 +67,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
   pool:
     name: mscclpp-multi-node
   container:

.azure-pipelines/sglang-test.yml

@@ -48,6 +48,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
   pool:
     name: msccl-ci-h100
   container:

.azure-pipelines/ut.yml

@@ -34,10 +34,10 @@ jobs:
     name: msccl-ci
   strategy:
     matrix:
-      cuda11:
-        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda11.8
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)
@@ -55,10 +55,10 @@ jobs:
     name: msccl-ci
   strategy:
     matrix:
-      cuda11:
-        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda11.8
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)
@@ -78,6 +78,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)
@@ -97,6 +99,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)
@@ -118,6 +122,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)
@@ -161,6 +167,8 @@ jobs:
     matrix:
       cuda12:
         containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda12.9
+      cuda13:
+        containerImage: ghcr.io/microsoft/mscclpp/mscclpp:base-dev-cuda13.0
 
   container:
     image: $(containerImage)

.devcontainer/devcontainer.json

@@ -32,7 +32,7 @@
                     "/usr/include"
                 ],
                 "C_Cpp.default.cStandard": "c17",
-                "C_Cpp.default.cppStandard": "c++17"
+                "C_Cpp.default.cppStandard": "c++20"
             }
         }
     },

.devcontainer/devcontainer_amd.json

@@ -32,7 +32,7 @@
                     "/usr/include"
                 ],
                 "C_Cpp.default.cStandard": "c17",
-                "C_Cpp.default.cppStandard": "c++17"
+                "C_Cpp.default.cppStandard": "c++20"
             }
         }
     },

.github/workflows/codeql-analysis.yml

@@ -40,7 +40,7 @@ jobs:
       fail-fast: false
       matrix:
         language: [ 'cpp', 'python' ]
-        version: [ 'cuda11.8', 'cuda12.9' ]
+        version: [ 'cuda12.9', 'cuda13.0' ]
 
     steps:
     - name: Checkout repository

.github/workflows/mscclpp-lang.yml

@@ -15,7 +15,7 @@ jobs:
     strategy:
         fail-fast: false
         matrix:
-          version: [ 'cuda11.8', 'cuda12.9' ]
+          version: [ 'cuda12.9', 'cuda13.0' ]
 
     steps:
     - uses: actions/checkout@v4

CMakeLists.txt

@@ -169,8 +169,8 @@ elseif(MSCCLPP_USE_CUDA)
     if(NVIDIA_FOUND)
         set(MSCCLPP_GPU_ARCHS "native")
     else()
-	    if(CUDAToolkit_VERSION VERSION_LESS "11.8")
-	        message(FATAL_ERROR "CUDA 11.8 or higher required, found ${CUDAToolkit_VERSION}")
+	    if(CUDAToolkit_VERSION VERSION_LESS "12.0")
+	        message(FATAL_ERROR "CUDA 12.0 or higher required (C++20 build), found ${CUDAToolkit_VERSION}")
 	    endif()
 	    set(MSCCLPP_GPU_ARCHS 80)
 	    if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.0")
@@ -199,10 +199,11 @@ if(MSCCLPP_USE_ROCM)
 endif()
 
 # Declare project
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra")
 if(MSCCLPP_USE_CUDA)
-    set(CMAKE_CUDA_STANDARD 17)
+    set(CMAKE_CUDA_STANDARD 20)
+    set(CMAKE_CUDA_EXTENSIONS OFF)
     set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Xcompiler -Wall,-Wextra")
     enable_language(CUDA)
 
@@ -217,7 +218,7 @@ if(MSCCLPP_USE_CUDA)
         set(GPU_LIBRARIES CUDA::cudart CUDA::cuda_driver)
     endif()
 else()
-    set(CMAKE_HIP_STANDARD 17)
+    set(CMAKE_HIP_STANDARD 20)
     set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -Wall -Wextra")
 
     set(CMAKE_HIP_ARCHITECTURES ${MSCCLPP_GPU_ARCHS})
@@ -226,6 +227,14 @@ else()
     set(GPU_INCLUDE_DIRS ${hip_INCLUDE_DIRS})
 endif()
 
+message(STATUS "C++ standard: C++${CMAKE_CXX_STANDARD}")
+if(MSCCLPP_USE_CUDA)
+    message(STATUS "CUDA toolkit version: ${CUDAToolkit_VERSION}")
+    message(STATUS "CUDA language standard: C++${CMAKE_CUDA_STANDARD}")
+else()
+    message(STATUS "HIP language standard: C++${CMAKE_HIP_STANDARD}")
+endif()
+
 if(CMAKE_BUILD_TYPE STREQUAL "Debug")
     add_compile_definitions(DEBUG_BUILD)
 endif()

docker/build.sh

@@ -4,7 +4,6 @@ set -e
 
 declare -A baseImageTable
 baseImageTable=(
-    ["cuda11.8"]="nvidia/cuda:11.8.0-devel-ubuntu22.04"
     ["cuda12.4"]="nvidia/cuda:12.4.1-devel-ubuntu22.04"
     ["cuda12.8"]="nvidia/cuda:12.8.1-devel-ubuntu22.04"
     ["cuda12.9"]="nvidia/cuda:12.9.1-devel-ubuntu24.04"
@@ -21,7 +20,6 @@ extraLdPathTable=(
 
 declare -A ofedVersionTable
 ofedVersionTable=(
-    ["cuda11.8"]="23.07-0.5.1.2"
     ["cuda12.4"]="23.07-0.5.1.2"
     ["cuda12.8"]="24.10-1.1.4.0"
     ["cuda12.9"]="24.10-1.1.4.0"
@@ -34,7 +32,7 @@ TARGET=${1}
 OS_ARCH=$(uname -m)
 
 print_usage() {
-    echo "Usage: $0 [cuda11.8|cuda12.4|cuda12.8|cuda12.9|cuda13.0|rocm6.2|rocm7.2]"
+    echo "Usage: $0 [cuda12.4|cuda12.8|cuda12.9|cuda13.0|rocm6.2|rocm7.2]"
 }
 
 if [[ ! -v "baseImageTable[${TARGET}]" ]]; then

docs/quickstart.md

@@ -14,10 +14,15 @@
             * [NDm_A100_v4](https://learn.microsoft.com/en-us/azure/virtual-machines/ndm-a100-v4-series)
             * [ND_H100_v5](https://learn.microsoft.com/en-us/azure/virtual-machines/nd-h100-v5-series)
         * Non-Azure Systems
-            * NVIDIA A100 GPUs + CUDA >= 11.8
+            * NVIDIA A100 GPUs + CUDA >= 12.0
             * NVIDIA H100 GPUs + CUDA >= 12.0
             * AMD MI250X GPUs + ROCm >= 5.7
             * AMD MI300X GPUs + ROCm >= 6.0
+* Toolchain
+    * MSCCL++ is built as **C++20** (both host and device code), so a C++20-capable toolchain is required.
+    * [CMake](https://cmake.org/) >= 3.25
+    * A C++20-capable host compiler, e.g., GCC >= 11 or Clang >= 14
+    * On NVIDIA platforms, **CUDA Toolkit >= 12.0** is required. `nvcc` first added `-std=c++20` support in CUDA 12.0, so earlier toolkits (11.x and below) cannot build the project.
 * OS
     * Tested on Ubuntu 20.04 and later
 * Libraries
@@ -109,12 +114,12 @@ $ python -m pip install ".[cuda12]"
 $ CXX=/opt/rocm/bin/hipcc python -m pip install ".[rocm7]"
 ```
 
-> **Note:** A platform extra (`cuda11`, `cuda12`, `cuda13`, `rocm6`, or `rocm7`) is required to install CuPy.
+> **Note:** A platform extra (`cuda12`, `cuda13`, `rocm6`, or `rocm7`) is required to install CuPy.
 > The CUDA extras install pre-built CuPy wheels and CUDA Python bindings. The ROCm extras install CuPy from source
 > and HIP Python for the matching ROCm major version, which require ROCm and may take longer. Running `pip install .` without an extra will not install CuPy.
 
 Optional extras can be installed by specifying them in brackets. Available extras:
-- **`cuda11`**, **`cuda12`**, **`cuda13`**: Install a pre-built CuPy package and CUDA Python bindings for your CUDA version.
+- **`cuda12`**, **`cuda13`**: Install a pre-built CuPy package and CUDA Python bindings for your CUDA version.
 - **`rocm6`**, **`rocm7`**: Install CuPy from source and HIP Python for AMD ROCm platforms.
 - **`benchmark`**: Install benchmark dependencies (mpi4py, prettytable, netifaces, matplotlib).
 - **`test`**: Install test dependencies (pytest, mpi4py, netifaces).
@@ -215,7 +220,7 @@ $ mpirun -np 16 -npernode 8 -hostfile hostfile ./bin/mp_unit_tests -ip_port 10.0
 
 ```bash
 # Install with benchmark dependencies and the appropriate CUDA/ROCm extras.
-# Replace `cuda12` with your platform: cuda11, cuda12, cuda13, rocm6, or rocm7.
+# Replace `cuda12` with your platform: cuda12, cuda13, rocm6, or rocm7.
 $ python3 -m pip install ".[cuda12,benchmark,test]"
 
 ```

pyproject.toml

@@ -21,10 +21,6 @@ dependencies = [
 ]
 
 [project.optional-dependencies]
-cuda11 = [
-    "cupy-cuda11x",
-    "cuda-bindings>=11.8,<12",
-]
 cuda12 = [
     "cupy-cuda12x",
     "cuda-bindings>=12,<13",

python/mscclpp/_core/compiler.py

@@ -198,7 +198,7 @@ def _do_init(self):
         self._is_hip = cp.cuda.runtime.is_hip
         self._device_arch = get_device_arch()
         self._compiler = self._get_compiler()
-        self._default_options = ["-std=c++17", "-O3", "--shared"]
+        self._default_options = ["-std=c++20", "-O3", "--shared"]
         python_include = sysconfig.get_path("include")
         pybind11_include = pybind11.get_include()
         self._default_options += [f"-I{python_include}", f"-I{pybind11_include}"]

python/mscclpp/utils.py

@@ -99,7 +99,7 @@ def __init__(self, file: str, kernel_name: str, file_dir: str = None, macro_dict
         self._kernel = Kernel(cubin, kernel_name)
         self.kernel_map[kernel_key] = self._kernel
 
-    def _compile_cuda(self, source_file, output_file, std_version="c++17"):
+    def _compile_cuda(self, source_file, output_file, std_version="c++20"):
         mscclpp_home = os.environ.get("MSCCLPP_HOME", "/usr/local/mscclpp")
         include_dir = os.path.join(mscclpp_home, "include")
         if not cp.cuda.runtime.is_hip:

src/core/core.cc

@@ -75,7 +75,7 @@ MSCCLPP_API_CPP bool TransportFlags::operator==(TransportFlags other) const {
 }
 
 MSCCLPP_API_CPP bool TransportFlags::operator!=(TransportFlags other) const {
-  return detail::TransportFlagsBase::operator!=(other);
+  return !detail::TransportFlagsBase::operator==(other);
 }
 
 MSCCLPP_API_CPP detail::TransportFlagsBase TransportFlags::toBitset() const { return *this; }

src/core/executor/execution_plan.cc

@@ -232,7 +232,7 @@ void ExecutionPlan::Impl::loadExecutionPlan(size_t inputSize, size_t outputSize,
                                             size_t constDstOffset) {
   std::ifstream file(this->planPath);
   json obj = json::parse(file);
-  if (this->name != obj["name"]) {
+  if (this->name != obj["name"].get<std::string>()) {
     throw Error("Plan name does not match", ErrorCode::ExecutorError);
   }
   this->collective = obj["collective"];
@@ -268,7 +268,7 @@ void ExecutionPlan::Impl::lightLoadExecutionPlan(size_t inputSize, size_t output
                                                  size_t constDstOffset) {
   std::ifstream file(this->planPath);
   json obj = json::parse(file);
-  if (this->name != obj["name"]) {
+  if (this->name != obj["name"].get<std::string>()) {
     throw Error("Plan name does not match", ErrorCode::ExecutorError);
   }
   std::string protocol = obj["protocol"];

test/deploy/setup.sh

@@ -61,9 +61,7 @@ if [ -f "${PIP_CMAKE_ARGS_FILE}" ]; then
 fi
 
 cd /root/mscclpp
-if [[ "${CUDA_VERSION}" == *"11."* ]]; then
-    pip3 install ".[cuda11,benchmark,test]"
-elif [[ "${CUDA_VERSION}" == *"12."* ]]; then
+if [[ "${CUDA_VERSION}" == *"12."* ]]; then
     pip3 install ".[cuda12,benchmark,test]"
 elif [[ "${CUDA_VERSION}" == *"13."* ]]; then
     pip3 install ".[cuda13,benchmark,test]"