SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD
Researchers from multiple institutions have developed a framework called SLAI T-Rex for efficiently training large-scale deep learning models on the Ascend NPU SuperPOD. The system achieves significant improvements in model FLOPs utilization and training stability compared to an open-source baseline recipe. Building on this optimized infrastructure, the team also established a workflow for complex Operations Research tasks using a specialized model called DeepSeek-V4-Flash.
Researchers from multiple institutions have developed a framework called SLAI T-Rex for efficiently training large-scale deep learning models on the Ascend NPU SuperPOD. The system achieves significant improvements in model FLOPs utilization and training stability compared to an open-source baseline recipe. Building on this optimized infrastructure, the team also established a workflow for complex Operations Research tasks using a specialized model called DeepSeek-V4-Flash.
---
Why it matters: This work matters because it provides a full-stack pathway for efficiently training trillion-parameter-scale models on Ascend infra, which can be used to improve performance in various domains such as mathematical modeling and solver-grounded reasoning.
Source: https://arxiv.org/abs/2607.20145
This article was originally published at: https://arxiv.org/abs/2607.20145