Suzihao focused on optimizing the gemma3 model’s decoding performance on Ascend hardware within the rjg-lyh/vllm-ascend repository. They developed an Ascend-specific GemmaRMSNorm class, leveraging torch_npu and PyTorch to accelerate RMS normalization during model inference. By modularizing the optimization, Suzihao enabled future hardware-specific improvements while directly reducing normalization time in the decoding process. Their work addressed a key performance bottleneck in deep learning model deployment on NPUs, demonstrating a strong grasp of model optimization and hardware acceleration. Over the month, Suzihao delivered a targeted feature that improved throughput and maintainability for deep learning workflows on specialized hardware.