Title: Boundary-Protection W8A8 HiFloat8 Quantization for Large-Scale Text-to-Video Diffusion Transformers

Abstract:

This paper introduces a post-training quantization (PTQ) methodology designed for Wan2.1-T2V-14B, a 14-billion-parameter text-to-video diffusion transformer. The approach specifically targets the W8A8 HiFloat8 (HiF8) format for deployment on Ascend 910B NPUs. A primary obstacle in quantizing video DiT models lies in the heterogeneous activation distributions found across different transformer blocks. Specifically, boundary blocks—comprising the initial and final segments—possess distinct statistical characteristics compared to the intermediate blocks, which renders uniform quantization strategies ineffective.

To address this, we performed a comprehensive per-block activation analysis across all 40 WanAttentionBlocks. The insights gained informed the development of a boundary-protection strategy, which preserves the first two and last three blocks in BF16 precision while applying W8A8 HiF8 quantization to the remaining 35 blocks. Our evaluation demonstrates that this PTQ method achieves performance that matches or slightly surpasses the BF16 baseline across all five VBench dimensions. Within the context of a five-prompt evaluation set, this indicates an absence of measurable accuracy degradation. Furthermore, an ablation study examining four distinct protection configurations revealed that full boundary protection results in the highest average VBench score, thereby validating the data-driven nature of the block selection process. Finally, we explore quantization-aware training (QAT) as a supplementary fine-tuning phase, analyzing the specific scenarios where QAT fails to deliver superior results over standard PTQ on single-card hardware.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC