SLAI introduces T-Rex, an end-to-end optimization framework for the full-parameter post-training of trillion-parameter-scale MoE models on Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, the system addresses memory pressure and communication overhead through hierarchical parallelism and kernel execution optimizations.

  • The infrastructure achieves 34.22% Model FLOPs Utilization (MFU), representing a 2.93x improvement over the open-source baseline while maintaining training stability.
  • A combined CPT and SFT workflow is established for complex Operations Research tasks, utilizing solver-verified synthetic optimization documents.
  • The resulting dataset contains 10K high-quality SFT samples spanning four task categories and three problem representations.
  • The specialized model achieves a 71.81% average zero-shot Pass@1 score, outperforming GPT-5.4-Mini and the base DeepSeek-V4-Flash by 3.98 and 11.27 percentage points respectively.

This work demonstrates a full-stack pathway for efficient post-training on Ascend infrastructure, advancing frontier-model systems for complex reasoning.