THE ROLE:As a senior member of the LLM inference framework team, you will be responsible for building and optimizing production-grade single-node and distributed inference runtimes for large language models on AMD GPUs. You will work at the framework and runtime layer, driving performance, scalability, and reliability, enabling tensor parallelism, pipeline parallelism, expert parallelism (MoE), and single-node or multi-node inference at scale. Your work will directly power customer-facing deployments and benchmarking platforms (e.g., InferenceMax, MLPerf, strategic partners, and cloud providers) and will be upstreamed into open-source inference frameworks such as vLLM and SGLang to make AMD a first-class platform for LLM serving.
This role sits at the intersection of inference engines, distributed systems, and GPU runtime and kernel backends.
THE PERSON:
You are a systems-minded ML engineer who thinks in terms of throughput, latency, memory movement, and scheduling, not just model code.
You are comfortable reading and modifying large-scale inference frameworks, debugging performance across GPUs and nodes, and collaborating with kernel, compiler, and networking teams to close end-to-end performance gaps.
You enjoy working in open source and driving architecture-level improvements in inference platforms.
KEY RESPONSIBILITIES:Inference Framework & Runtime
- Architect and optimize distributed LLM inference runtimes based on in-house LLM engines or open-source stacks such as vLLM, SGLang, and llm-d
- Design and improve TP / PP / EP (MoE) hybrid execution, including KV-cache management, attention dispatch, and token scheduling
- Implement and optimize multi-node inference pipelines using RCCL, RDMA, and collective-based execution
Performance & Scalability
- Drive throughput, latency, and memory efficiency across single-GPU and multi-GPU clusters
- Optimize continuous batching, speculative decoding, KV-cache paging, prefix caching, and multi-turn serving
GPU & Backend Integration
- Work with AMD GPU libraries (AITER, HIPBLAS-LT, RCCL, ROCm runtime) to ensure inference frameworks efficiently use FP8 / FP4 GEMM and FlashAttention / MLA
- Collaborate with compiler teams (Triton, LLVM, ROCm) to unblock framework-level performance
Open Source & Customer Enablement
- Upstream features and performance fixes into vLLM, SGLang, and llm-d
- Enable customer PoCs and production deployments on AMD platforms
- Build and maintain benchmark-grade inference pipelines
PREFERRED EXPERIENCE: Inference Stack Knowledge
- Hands-on understanding of vLLM, SGLang, or similar inference stacks
- Experience with distributed inference scaling and a proven track record of contributing to upstream open-source projects
Deep Learning Integration
- Strong experience integrating optimized GPU performance into machine-learning frameworks (e.g., PyTorch, TensorFlow) for high-throughput and scalable inference
Kernel & Inference Frameworks
- Strong background in NVIDIA, AMD, or similar GPU architectures and kernel development
Software Engineering
- Expertise in Python and preferably experience in C/C++, including debugging, performance tuning, and test design for large-scale systems
High-Performance Computing
- Experience running large-scale workloads on heterogeneous GPU clusters, optimizing for efficiency and scalability
Compiler & Runtime Optimization
- Understanding of compiler and runtime systems, including LLVM, ROCm, and GPU code generation
ACADEMIC CREDENTIALS: - Master's or PhD in Computer Science, Computer Engineering, Electrical Engineering, or a related field.
#LI-JG1
Benefits offered are described: AMD benefits at a glance.
