HiFloat4 Format for Language Model Pre-training on Ascend NPUs

HiFloat4 addresses a critical bottleneck in the non-NVIDIA AI ecosystem: efficient 4-bit training on Ascend hardware. The project shows 25 forks despite 0 stars, a signal often associated with institutional or corporate research teams (likely in the Chinese market) immediately adopting a paper's codebase for internal testing on NPU clusters. While technically sophisticated—requiring deep knowledge of Ascend's memory hierarchy and CANN operator development—the defensibility is limited by its platform-specific nature. The 'moat' here is purely the first-mover advantage in specialized hardware optimization. However, Huawei's own software teams (CANN/MindSpore) are the primary threat, as they are likely to integrate native FP4 support into their standard libraries, which would render this third-party implementation obsolete. Compared to NVIDIA-centric formats like NVFP4 or the OCP standard MXFP4, HiFloat4 is a niche but vital alternative for those operating outside the CUDA ecosystem. Its survival depends on whether it can become the community-standard library for Ascend-based LLM training before Huawei provides a first-party equivalent.