Citing MetaEra, Cambricon completes DeepSeek-V4 model adaptation, code open-sourced

ME News reports that on April 24 (UTC+8), according to monitoring by Beating, Cambricon announced it has completed adaptation of the 285B DeepSeek-V4-Flash and 1.6T DeepSeek-V4-Pro models on the day of V4’s release. Leveraging the vLLM inference framework, the adaptation code has been open-sourced on GitHub. This rapid adaptation was enabled by two key factors: first, Cambricon’s proprietary NeuWare software stack natively supports mainstream frameworks such as PyTorch and vLLM, enabling swift model migration; second, Cambricon’s chips natively support mainstream low-precision data formats, allowing precision validation without additional format conversion. To optimize for V4’s new architecture, Cambricon developed the proprietary fused operator library Torch-MLU-Ops to accelerate modules such as Compressor and mHC, and implemented high-performance kernel functions for sparse/compressed Attention and GroupGemm using BangC. At the inference framework level, Cambricon has integrated five-dimensional hybrid parallelism (TP/PP/SP/DP/EP), communication-computation overlap, low-precision quantization, and PD separation deployment into vLLM. While the V4 technical report only mentioned validation on NVIDIA GPUs and Huawei Ascend NPUs, this adaptation was independently accomplished by Cambricon. Stimulated by the V4 announcement, China’s domestic chip sector strengthened, with Cambricon’s stock surging intraday. (Source: BlockBeats)