From October 11 to 13, 2023, during the China Mobile Global Partner Conference, ZTE's "Nebula R&D Model" was unveiled, aiming to assist developers in demand analysis, product design, programming, testing, version deployment, etc.
According to reports, the "Nebula R&D Big Model" supports a whitelist mechanism to effectively control the scope of use , as well as code characteristic value recognition to effectively identify sensitive code fragments, a sensitive word recognition mechanism to monitor and intercept sensitive content in real time, and a background audit mechanism to fully trace back security events etc.
ZTE said that in April 2023, the "Nebula R&D Large Model" was launched. So far, the number of daily active users has reached 12,000, the code adoption rate has reached 40%~45%, coding efficiency has improved by 30%, and overall R&D efficiency has improved by 10%. .
ZTE has injected domain data, Know-How knowledge accumulation, hundreds of thousands of technical documents in the communications field, and wireless/core network/cloud code corpus of 100 billion tokens into the large model to conduct incremental pre-training and use a parallel training framework.
ZTE said: “The self-developed deployment solution uses dynamic batch strategy, PagedAttention technology, combined with lossless model quantification, and the throughput is greatly improved. A single GPU (A800) can reach 1500 tokens/s, and only 4 GPU cards (A800) can meet the needs of more than 1,000 tokens/s. Human usage needs . Compared with conventional deployment solutions in the industry, the single-GPU throughput is increased by 10+ times and 20+ times respectively; combined with int4 quantization technology, the model size and video memory usage are reduced by half without reducing model accuracy."