Advanced Development Overview

MindSpore Transformers advanced development targets scenarios such as model migration, tuning, and accuracy verification, helping users go beyond basic training and inference to perform development migration, debugging, optimization, and accuracy comparison. This section summarizes all advanced development documentation by category: Diagnostics and Optimization, Model Development and Configuration, Accuracy Comparison, and API Reference, for quick reference and navigation.

Diagnostics and Optimization

We provide systematic methods for identifying and resolving precision and performance issues during training and inference.

Document	Description	Architecture Support
Precision Optimization	Common precision issues in large model training and general troubleshooting methods, including checklists, parameter alignment, and long-run validation.	Mcore/Legacy
Performance Optimization	Large model performance tuning approach and tools, covering data loading, forward/backward computation, communication, scheduling, and Profile usage.	Mcore/Legacy

Model Development and Configuration

Support for building or migrating models from scratch and quickly launching training and inference via configuration templates.

Document	Description	Architecture Support
Development Migration	End-to-end workflow for building large models with MindSpore Transformers, including config, model, tokenizer, and YAML configuration.	Legacy
Guide to Using the Inference Configuration Template	How to use the YAML configuration template for inference, including quick setup with Hugging Face/ModelScope model directories.	Mcore
Training Template Instruction	Overview of general configuration templates for pre-training and fine-tuning (dense/MoE, etc.) and quick start for custom or unsupported model sizes.	Mcore
Weight Transfer	Adapting new models to MindSpore Transformers weight conversion for unified Hugging Face to MindSpore weight conversion and loading.	Mcore

Accuracy Comparison

Validation of alignment with reference implementations or GPU environments for both training and inference.

Document	Description	Architecture Support
Compare Training Accuracy with Megatron-LM	Training accuracy alignment with Megatron-LM at the model level, including equivalent structure setup and comparison of forward, loss, and gradients.	Mcore
Comparison of Inference Precision	Inference precision acceptance workflow and troubleshooting, including online inference checks, dataset evaluation, and common issue resolution.	Mcore

API Reference

Entry point to API documentation for MindSpore Transformers modules.

Document	Description
API	API index and detailed interface documentation for MindSpore Transformers and submodules.