[Huawei Official: Ascend Supernode Series Fully Supports DeepSeek V4] Golden Finance reports that on April 24, DeepSeekV4-Pro and DeepSeekV4-Flash were officially released and open-sourced. Ascend has consistently supported the DeepSeek series models, and this time, through close collaboration in chip and model technology, the entire Ascend Supernode product line now supports the DeepSeekV4 series models. The Ascend 950 reduces Attention computation and memory access overhead by integrating kernel and multi-stream parallel technologies, significantly improving inference performance. Combined with various quantization algorithms, it enables high-throughput, low-latency deployment of DeepSeekV4 model inference. The Ascend A3 Supernode series products are also fully adapted, and to facilitate quick fine-tuning for users, a training reference implementation based on the Ascend A3 Supernode is provided. (Huawei Computing)