SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD

2026-07-22

AI Summary

This work addresses the memory, communication, and operator efficiency bottlenecks encountered when performing full-parameter post-training of trillion-parameter Mixture-of-Experts (MoE) models on non-GPU hardware. Focusing on the Ascend NPU SuperPOD platform, we propose an end-to-end hierarchical optimization framework encompassing model parallelism strategies, compute-communication co-scheduling, and customized low-level operators. By integrating solver-guided synthetic data to construct a domain-specific instruction tuning dataset, we achieve, for the first time, efficient full-parameter post-training of the DeepSeek-V4 model family on Ascend super-nodes. The system attains a Model FLOPs Utilization (MFU) of 34.22%, yielding a 2.93× speedup over open-source baselines. On operations research tasks, it achieves a zero-shot Pass@1 accuracy of 71.81%, significantly outperforming both GPT-5.4-Mini and the original DeepSeek-V4-Flash model.

Abstract

Full-parameter post-training of trillion-parameter-scale MoE models introduces substantial system-level challenges for large-scale distributed training, including severe memory pressure, non-overlapped communication overhead, and inefficient kernel execution. While most large-scale LLM training systems are built around GPU-based clusters, this report presents an end-to-end optimization practice on the Ascend NPU SuperPOD. Using the DeepSeek-V4 model family as the target workload, we develop a hierarchical optimization framework spanning model-level parallelism, computation-communication orchestration, and low-level kernel execution. The resulting system achieves 34.22% Model FLOPs Utilization (MFU) with a 2.93x improvement over the open-source baseline recipe while maintaining training stability. Building on this optimized infrastructure, we further establish a CPT and SFT workflow for complex Operations Research (OR) tasks. We refer to the integrated framework as SLAI T-Rex. Using DeepSeek-V4-Flash, we develop OR-oriented CPT and SFT data pipelines that combine collected domain resources with 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 the highest average zero-shot Pass@1 score among the evaluated models, reaching 71.81% and outperforming GPT-5.4-Mini and the base DeepSeek-V4-Flash model by 3.98 and 11.27 percentage points, respectively. Overall, this work demonstrates a full-stack pathway from efficient trillion-parameter model post-training on Ascend infra to domain-specialized Flash models for solver-grounded mathematical modeling, advancing frontier-model systems for complex reasoning.

Problem

• post-training
• trillion-parameter MoE models
• distributed training
• memory pressure
• communication overhead

Innovation

• full-parameter post-training
• MoE models
• Ascend SuperPOD
• computation-communication orchestration
• solver-grounded SFT