资深AI框架工程师

该职位来源于猎聘 Job Title: Senior Engineer / Specialist, LLM Inference Performance Optimization Department: AI Infrastructure / Performance Optimization Team Employment Type: Full-time Position Overview We are seeking a highly experienced engineer specializing in large language model (LLM) inference performance optimization. You will be a core member of our team, responsible for building and optimizing the LLM inference performance with high-throughput, low-latency on AMD Instinct GPUs. If you are passionate about pushing performance boundaries and have deep, hands-on expertise with cutting-edge technologies like vLLM or SGLang, we invite you to join us.

Key Responsibilities

• Core System Optimization: Lead the development, tuning, and customization of LLM performance optimization on AMD GPUs, leveraging and extending frameworks like vLLM or SGLang to address performance bottlenecks in production environments.
• Performance Analysis & Tuning: Conduct end-to-end performance profiling using specialized tools. Perform deep optimization of compute-bound operators (e.g., Attention), memory I/O, and communication to significantly increase throughput and reduce latency.
• Model Architecture Adaptation: Demonstrate expertise in mainstream LLM architectures (e.g., DeepSeek, Qwen, Llama, ChatGLM) and optimize inference for their specific characteristics (e.g., RoPE, SWA, MoE, GQA).
• Algorithm & Principle Application: Leverage your deep understanding of core algorithms (Transformer, Attention, MoE) to implement advanced optimization techniques such as PagedAttention, FlashAttention, continuous batching, quantization, and model compression.
• Technology Foresight & Implementation: Research and prototype state-of-the-art optimization techniques (e.g., Speculative Decoding, Weight-Only Quantization) and drive their adoption into production systems.

  Qualifications Mandatory Requirements:

• Expertise in Inference Frameworks: Proven, hands-on experience with vLLM or SGLang, including deep understanding of their source code, deployment, configuration, and performance tuning. (Please describe relevant projects in your resume).
• Mastery of Model Architectures: In-depth understanding and practical experience with inference workflows of mainstream LLMs (e.g., DeepSeek, Qwen), including their tokenizers, model configurations, and architecture definitions.
• Strong Theoretical Foundation: Solid grasp of the principles behind Transformer, Self-Attention, MoE, KV Cache, and their impact on inference performance.
• Proven Optimization Experience: Familiarity with end-to-end LLM inference optimization techniques such as PagedAttention, FlashAttention, continuous/dynamic batching, and quantization (INT8/INT4/GPTQ/AWQ), demonstrated with successful case studies.
• Programming Skills: Proficiency in Python and strong software engineering best practices.

  Preferred Qualifications (Plus):

• Low-Level Development Skills: Experience with CUDA C programming for writing and debugging high-performance GPU kernels; or practical experience using Triton to develop and optimize deep learning operators.
• Compiler Knowledge: Understanding or practical experience with compiler technologies like TVM or MLIR is a significant advantage.
• Distributed Systems Experience: Hands-on experience with distributed inference for large-scale models (e.g., Tensor Parallel, Pipeline Parallel).
• Education: Master's or Ph.D. in Computer Science, Artificial Intelligence, Electrical Engineering, or a related field.