On March 18, 2025, Moore Threads unveiled two groundbreaking AI frameworks, MT-MegatronLM and MT-TransformerEngine, designed to revolutionize mixed parallel training and inference capabilities on domestic, full-featured GPUs. MT-MegatronLM is tailored for open-source mixed parallel training on these GPUs, whereas MT-TransformerEngine is specifically engineered to enhance the training and inference processes of Transformer models. These frameworks embody a harmonious fusion of hardware adaptability and algorithmic ingenuity, highlighted by their advanced mixed parallel training capabilities. In real-world applications, they have yielded impressive results, complemented by robust exception handling and broad compatibility. Moore Threads pledged to continuously refine these frameworks and integrate even more advanced features. Concurrently, the company also announced the open-source repositories for these frameworks and Moore Threads' Simumax.
