AIAny
Icon for item

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

Performs full-parameter post-training of trillion-parameter MoE DeepSeek-V4 models on an Ascend NPU SuperPOD, using a hierarchical optimization of model parallelism, communication orchestration, and kernel execution to increase Model FLOPs Utilization. Also builds CPT/SFT pipelines with solver-verified synthetic data for Operations Research, reporting strong zero-shot Pass@1 results.

Introduction

Full-parameter post-training at trillion-scale MoE strains conventional GPU-centric stacks: extreme memory pressure, poor overlap between compute and comms, and low kernel utilization. This paper shows an end-to-end engineering pathway on Ascend NPU SuperPOD that treats the infra challenge as the main research question rather than just a benchmarking footnote — and pairs system wins with a domain-specialized dataset+model workflow for Operations Research.

Key Findings
  • System-level gains: a hierarchical optimization across model-parallel layouts, compute-communication orchestration, and low-level kernel tuning yields 34.22% Model FLOPs Utilization (MFU), a 2.93x improvement over the open-source baseline while preserving training stability. So what: higher MFU means more effective use of available NPU cycles and lower wall-clock cost per step at trillion scale.
  • Full-parameter post-training feasible on NPUs: the work demonstrates practical techniques to map MoE and large transformer kernels to Ascend SuperPOD, addressing memory pressure and communication bottlenecks that typically force parameter freezing or heavy sharding.
  • Domain specialization pipeline: establishes a CPT (Continued Pre-Training) and SFT workflow for Operations Research that mixes collected domain resources with solver-verified synthetic documents, producing ~10K high-quality SFT samples. So what: solver-grounded synthetic data improves model ability to generate solver-ready formulations.
  • Downstream performance: the specialized DeepSeek-V4-Flash variant achieves 71.81% average zero-shot Pass@1 on evaluated OR tasks, outperforming GPT-5.4-Mini by ~3.98 percentage points and the base DeepSeek-V4-Flash by ~11.27 points. So what: targeted post-training plus solver-verified SFT yields measurable reasoning gains for optimization problems.
Who It's For & Tradeoffs

Great fit if you manage large MoE/LLM training at datacenter scale and are exploring non-GPU accelerators (Ascend NPUs), or if you need solver-grounded fine-tuning data for mathematical/OR reasoning. The paper is valuable for systems engineers, ML infra teams, and teams building domain-specialized LLMs. Look elsewhere if you need reproducible, self-serve recipes for GPU clusters (the optimizations target Ascend SuperPOD and require platform-specific engineering) or if your priority is lightweight fine-tuning on a single GPU.

Method highlights
  • Hierarchical optimization: adjusts model-parallel decomposition, overlaps communication with computation where possible, and applies kernel-level tuning to improve NPU utilization.
  • CPT+SFT data engineering: creates solver-verified synthetic optimization documents and assembles a 10K-sample SFT dataset spanning multiple task types and representations, aimed at making model outputs compatible with downstream solvers.
Where It Fits

This paper sits at the intersection of large-model systems engineering and domain-specialized model adaptation: less about novel model architectures and more about making full-parameter post-training practical at trillion scale on alternative accelerators, then leveraging that capacity to produce stronger OR-focused reasoning models.

Information

  • Websitearxiv.org
  • AuthorsDongfang Li, Xiaodong Luo, Ruoyu Sun, Xuhui Chen, Linyuan Qiu, Jian Meng, Zhengxuan Lu, Yiting Wang, Yucheng Xie, Tao Guo
  • Published date2026/07/22

More Items

Proposes Riemannian Isometric Policy Optimization (RIPO) to fix exploration collapse in PPO-style RL for LLMs by aligning policy updates with the policy manifold's Riemannian geometry, improving exploration–exploitation balance and optimization stability across competition benchmarks.

Studies train-time knowledge injection via hypernetworks that generate fixed LoRA adapters from large fact corpora, empirically characterizing power-law scaling across hypernetwork depth, width, and target model size and reporting improved OOD generalization.

Quantifies active visual observation in multimodal LLMs with ActiveVision, a 17-task benchmark that forces repeated perception rather than one-shot description. Finds frontier MLLMs fail badly (top model 10.6% vs humans 96.1%) and that model-generated vision code does not close the gap.