Backend Inference Runtime Engineer Graduate

About the Role

ByteDance's Data AML is a Machine Learning mid-platform providing training and inference systems for recommendation/advertising for businesses like Douyin, Jinri Toutiao, and Xigua Video. It offers Machine Learning computing power and conducts research on general and innovative algorithms. As a graduate, you will have opportunities to pursue bold ideas, tackle complex challenges, and unlock limitless growth. Successful candidates must commit to an onboarding date by the end of the year.

Key Responsibilities

• Responsible for the iteration of the underlying architecture of the large model inference engine and end-to-end GPU performance optimization, through means such as operator fusion and compilation optimization.
• Deeply optimize GPU memory access, computing pipeline, and Stream asynchronous scheduling to eliminate inference computing bottlenecks, improve single-card inference throughput, and reduce inference latency.
• Adapt to all series of GPU/NPU hardware architectures, refine the universality of the inference engine and hardware adaptability, and build a high-performance, low-loss underlying base for large model inference.
• Lead the design, development, and optimization of distributed parallel solutions for large model inference scenarios.
• Focus on implementing multi-dimensional parallel strategies such as tensor parallelism (TP), pipeline parallelism (PP), sequence parallelism, and MoE expert parallelism.
• Address core issues such as multi-card splitting and deployment of ultra-large models, high cross-card communication overhead, load imbalance, and low parallel efficiency.
• Follow up on cutting-edge technologies such as global large model inference, GPU high-performance computing, distributed parallelism, and cache optimization.
• Benchmark against mainstream inference frameworks like vLLM and TensorRT-LLM.
• Complete the implementation of solutions and technological innovation, continuously iterate and optimize performance and cost advantages of the inference system.
• Build the core technological barriers of the team.

Requirements & Tech Stack

Minimum Qualifications:

• Individuals completing or recently completed a Bachelor's/ Master's degree in Software Development, Computer Science, Computer Engineering, or a related technical discipline.
• Solid foundation in computer low-level knowledge.
• Proficient in C/C++ and Python programming.
• Skilled in CUDA programming and familiar with GPU hardware architecture principles.
• Well-versed in GPU memory models, computing scheduling, and communication mechanisms.
• Proficiently master the underlying development and implementation of various basic operators in Deep learning.
• Well-versed in GPU adaptation and optimization of core operators such as matrix operations, normalization, and activation functions.
• Able to independently complete operator handwritten reconstruction, memory access optimization, vectorization acceleration, and precision alignment.
• Familiar with the end-to-end process of deep learning inference compilation.
• Understand core compilation technologies such as computational graph optimization, operator fusion, constant folding, memory reuse, scheduling optimization, and quantization compilation.
• Proficient in using GPU performance analysis tools such as Nsight and Profiler.
• Possess the thinking of software-hardware collaborative optimization.
• Capable of outputting systematic optimization solutions and completing implementation iterations.
• Adaptable to the requirements of industrial-level high-concurrency, low-latency inference business.
• Possess good cross-team collaboration skills, communication and presentation skills, and document writing skills.
• Strong sense of responsibility and stress tolerance.

Preferred Qualifications:

• Thoroughly understand the core principles of large model inference.
• Proficiently master the core technologies of model parallelism.
• Experience in implementing distributed inference solutions such as tensor parallelism, pipeline parallelism, and sequence parallelism.
• Familiar with multi-card communication, load balance, and parallel efficiency optimization methods.
• Experience in secondary development and performance optimization of mainstream large model inference frameworks such as vLLM, SGLang, TensorRT-LLM, etc. is preferred.