Steven Liu Email & Phone Number

Beijing, China

Beijing, China

Project Background:1. Developing an end-to-end deep learning compiler for a proprietary AI Engine based on the open-source AI compiler project TVM. The AI Engine is a new innovative processing technology developed by AMD as part of the Versal™ Adaptive Compute Acceleration Platform (ACAP) architecture. The solution based on TVM supports models in formats such as ONNX and PyTorch, performs graph optimization, fusion, and partitioning, and ultimately generates models that can run efficiently on CPU and AIE.Job Responsibilities:1. Build a front-end optimization framework based on TVM to achieve high-quality graph optimization and partitioning for cross-platform models.Extend TVM to support front-end optimizations that comply with AIE, including expanding specific AIE front-end operators, supporting platform-dependent and platform-independent passes, supporting operator partitioning and specific operator fusion, and building a universal Graph Partition framework tailored to AIE's hardware characteristics.2. Extend the TVM backend framework to support new operators and optimize instruction generation, including batch matmul and element-wise operators. This involves supporting symmetric and asymmetric quantization logic and parallelization acceleration for operators, supporting element-wise padding scenarios, and adding new backend passes to optimize the final instruction generation, thereby improving operator optimization efficiency.3. Provide model analysis and profiling tools to offer a unified analysis report for new models.

Shanghai, China

Project background:1. Responsible for developing the deep learning compilation framework HALO. This framework is mainly aimed at Alibaba Cloud's server-side and end-side deep learning algorithm deployment services, while improving system throughput. Job content:1. Base on Freshippo's business scenario, participate in the overall image process design, and focus on optimizin the online model inference implement based on Qualcomm Cloud hardware AIC100 to effectively improve the service throughput and efficiency.2. Adapt to the hardware platform of the cloud deep learning chip Qualcomm AIC100 for HALO, and implement the interface layer based on Freshippo's business scenario.

Shanghai, China

Project background:1. Responsible for the model deployment service of OpenMMLab. According to the more than 130 deep learning algorithm for image recognition and processing field, try to provide unified demployment plan, and realize the efficient algorithm of deep learning model on the server side.Job content: 1. Responsible for the development of OpenMMLab deep learning model deployment tools: for the deep learning model to provide a unified, standardized ONNX file export service, contained the export model accuracy repair, CUDA operator support, automated testing and other requirements.2. Responsible for OpenMMLab the maintenance of more than ten open source github repository, including reply to issue,review pull request, etc.