Huawei Ascend AI Chips Achieve Zero Logdiff in Training-Inference Consistency, Boosting Performance by 20%-60%

On August 6, Huawei Computing officially announced that Ascend now provides a complete set of training-inference consistent operators based on the AscendC programming language, achieving numerical consistency (Logdiff = 0) between inference and training in tests with the Qwen3-30B-MoE model. Through FA operator optimization, Ascend has achieved a 20%–60% performance improvement under Eager mode. To address numerical inconsistencies between the Rollout engine and the training engine, Ascend implemented four key optimizations: unifying attention semantics, locking the reduce accumulation order, aligning online Softmax tiling strategies, and aligning precision along critical computation paths—ensuring numerical consistency between training and inference. The related infrastructure has been open-sourced and will soon include a "Training-Inference Consistency Issue Identification Skill" to detect residual logdiffs and pinpoint their root causes.