Jump to content

Connect SuperML | Leeroopedia MCP: Equip your AI agents with best practices, code verification, and debugging knowledge. Powered by Leeroo — building Organizational Superintelligence. Contact us at founders@leeroo.com.

Implementation:Microsoft Onnxruntime CUDA NcclCommon

From Leeroopedia


Knowledge Sources
Domains Training, CUDA_Kernels
Last Updated 2026-02-10 04:00 GMT

Overview

Concrete tool providing common NCCL context management and type conversion utilities in the ONNX Runtime CUDA training framework.

Description

Implements the NcclContext class and shared NCCL utilities. GetNcclDataType maps ORT data types (uint8, int8, int32, int64, MLFloat16, float, double) to NCCL data types. NcclContext manages NCCL communicators for five worker group types: GlobalParallel, DataParallel, HorizontalParallel, NodeLocalDataParallel, and CrossNodeDataParallel. During construction, it initializes MPI (if not already initialized), creates MPI sub-groups and sub-communicators for each worker group, broadcasts NCCL unique IDs, and initializes NCCL communicators via ncclCommInitRank. The destructor properly destroys all NCCL communicators and finalizes MPI. NcclKernel is the base class for NCCL-based operators that holds a pointer to the static NcclContext singleton and reads the group_type attribute.

Usage

Used as the foundation for all NCCL-based collective operations. NcclContext is initialized once and shared across all NCCL kernel instances throughout the training session.

Code Reference

Source Location

Signature

ncclDataType_t GetNcclDataType(onnxruntime::MLDataType type);

class NcclContext {
  NcclContext();
  ~NcclContext();
  ncclComm_t Comm(training::WorkerGroupType group_type);
};

class NcclKernel : public CudaKernel {
  NcclKernel(const OpKernelInfo& info);
protected:
  NcclContext* nccl_;
  training::WorkerGroupType group_type_;
};

Import

#include "orttraining/training_ops/cuda/collective/nccl_common.h"

I/O Contract

Inputs

Name Type Required Description
group_type int64_t (attribute) No Worker group type (default 0 = GlobalParallel)
type MLDataType Yes Data type to convert to NCCL type (for GetNcclDataType)

Outputs

Name Type Description
nccl_type ncclDataType_t Corresponding NCCL data type
comm ncclComm_t NCCL communicator for the specified group type

Usage Examples

// NcclKernel base class usage
NcclKernel::NcclKernel(const OpKernelInfo& info) : CudaKernel(info) {
  static NcclContext context;
  nccl_ = &context;
  int64_t group_type;
  info.GetAttrOrDefault("group_type", &group_type, static_cast<int64_t>(0));
  group_type_ = static_cast<training::WorkerGroupType>(group_type);
}

// Type conversion
ncclDataType_t dtype = GetNcclDataType(onnx_type);

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.
Principle
Implementation
Heuristic
Environment