13 days ago
China (city not specified)Senior
Responsibilities
- Optimize TensorFlow, PyTorch, vLLM, and SGLang for AMD GPUs and contribute improvements upstream.
- Develop and tune GPU kernels and performance-critical operators to improve throughput and latency.
- Optimize LLM architectures including Llama, Qwen, and DeepSeek using techniques such as FlashAttention, PagedAttention, speculative decoding, and quantization.
- Profile and optimize compute, memory, communication, system, and inference bottlenecks across multi-GPU and multi-node systems.
- Use compiler technologies and graph compilers to accelerate deep learning and inference pipelines.
- Integrate emerging optimization methods such as weight-only quantization into production systems.
- Collaborate with internal GPU library teams and open-source maintainers on upstream integration.
- Deliver maintainable, reliable, production-quality performance-critical software.
Requirements
- Deep practical experience with vLLM or SGLang and modern LLMs such as DeepSeek and Qwen.
- Strong knowledge of Transformer, Attention, Mixture-of-Experts, and KV-cache concepts.
- Hands-on experience with FlashAttention, PagedAttention, continuous batching, and INT8, INT4, GPTQ, or AWQ quantization.
- Experience profiling and optimizing compute, memory, and communication bottlenecks across multi-GPU and multi-node environments.
- Experience running and optimizing large-scale workloads on heterogeneous clusters.
- Experience integrating optimized GPU kernels into TensorFlow or PyTorch for large-scale training and inference.
- Strong Python and C++ programming skills, debugging and testing experience, and a record of delivering maintainable performance-critical software.
- Open-source contribution experience and strong self-motivation.
- Bachelor’s and/or Master’s degree in Computer Science, Computer Engineering, Electrical Engineering, or a related field.
- Experience with CUDA C++ GPU-kernel development or practical experience using Triton to develop and optimize deep learning operators.
- GPU-kernel experience with HIP, CUDA, assembly language, CK, CUTLASS, or Triton and knowledge of GCN or RDNA architectures is a plus.
- Knowledge of LLVM, ROCm, TVM, or MLIR and distributed inference techniques such as tensor parallelism and pipeline parallelism is preferred.
Benefits
- AMD benefits are referenced through the company’s benefits-at-a-glance information.
About AMD
AMD designs and sells CPUs, GPUs, and adaptive/embedded computing products for PCs, data centers, gaming, and edge devices. Its portfolio includes Ryzen and EPYC processors, Radeon and Instinct graphics, and adaptive SoCs from its Xilinx acquisition, sold to OEMs, cloud providers, and device makers. Founded in 1969 and headquartered in Santa Clara, it is a public company on NASDAQ and supplies semi-custom chips for major game consoles.