Responsibilities
Team Introduction: We are the Doubao Video Generation Model-PixelDance team. We focus on developing video generation models and solving key problems in video generation, including but not limited to high-dynamic video generation and content consistency assurance. Build industry-leading video basic models and lead the future trend of technology. The work of the video generation engineering team involves the full cycle process of model production. Here, you have the opportunity to participate in every link of model data production, training acceleration, inference acceleration, and service deployment. At the same time, you will be exposed to the most advanced video generation technology, massive data, and large-scale clusters. We hope that you can scale up with our models. 1. Provide training stability, ease of use, performance, and scale up optimization for LLM and Diffusion Model 2. Be able to use Profiler to analyze training bottlenecks and use distributed strategy tuning, operator optimization, and other methods to improve training performance 3. Be responsible for the research and introduction of ByteDance Research's training optimization technology 4. Work closely with the algorithm department to jointly optimize algorithms and systems.
Qualifications
1. Bachelor degree or above, major in computer/electronics/automation/software, etc., with AI engineering optimization experience preferred 2. Familiar with the training performance optimization of any scenario of LLM and Diffusion Model 3. Familiar with the use and principles of mainstream distribution frameworks in the industry such as Pytorch, FSDP, Deepspeed, Megatron, etc., able to optimize business scenarios, able to follow the latest industry trends and implement them 4. Proficient in GPU high-performance computing optimization technology, rich experience in CUDA-based GPU performance optimization, in-depth understanding of computer architecture, familiar with parallel computing optimization, memory access optimization, low-bit computing, etc. 5. Understand the basic principles of deep learning algorithms, familiar with the basic architecture of neural networks and the calculation methods of each operator, and understand the analysis of at least one deep learning training framework and its model files.
