Senior AI Performance Engineer - LLM Inference (vLLM)

WHAT YOU DO AT AMD CHANGES EVERYTHING At AMD, our mission is to build great products that accelerate next-generation computing experiences—from AI and data centers, to PCs, gaming and embedded systems. Grounded in a culture of innovation and collaboration, we believe real progress comes from bold ideas, human ingenuity and a shared passion to create something extraordinary. When you join AMD, you’ll discover the real differentiator is our culture. We push the limits of innovation to solve the world’s most important challenges—striving for execution excellence, while being direct, humble, collaborative, and inclusive of diverse perspectives. Join us as we shape the future of AI and beyond. Together, we advance your career. THE ROLE: AMD is looking for a performance-obsessed senior engineer to push AI inference performance to the limit on AMD GPUs, with vLLM as the primary serving framework. You will work end-to-end across the stack: profiling, diagnosing, and optimizing leading models running on vLLM across customer-relevant serving configurations (e.g. agentic coding, long-context, high-throughput serving). You will own hard performance problems on our most strategic customer engagements and leave behind measurable uplifts and reusable methodology. This is not a sustaining role: every engagement is different, every optimization leaves a lasting impact. THE PERSON: You can take an AI workload, understand it top to bottom, and make it faster on vLLM. You know the framework's internals well: PagedAttention and the KV cache manager, continuous batching and the scheduler, the V1 engine architecture, chunked prefill, and the paths that connect a user request down to the GPU kernel on the AMD (ROCm) backend. You are comfortable profiling a distributed vLLM deployment, diagnosing a kernel-level bottleneck, and clearly communicating your findings to both engineers and customer stakeholders. You understand GPU kernel performance well: not just how to use profiling tools, but how to reason about occupancy, cache behavior, memory coalescing, and instruction-level bottlenecks from first principles. You raise the bar through technical depth—you take on hard problems and help teammates grow along the way. You are AI-fluent, not just in the models you optimize, but in how you work: you leverage AI agents and tools daily to accelerate your workflows. You thrive under pressure, move fast, and measure everything. KEY RESPONSIBILITIES: • Drive performance optimization end-to-end on vLLM across leading models and customer-relevant serving configurations, closing competitive gaps through kernel and systems-level optimizations • Profile, diagnose, and resolve cross-stack performance bottlenecks in vLLM deployments, from GPU kernels and operator dispatch to the vLLM scheduler, PagedAttention/KV cache management, and multi-node communication • Diagnose kernel-level performance issues using profiling tools: identify occupancy limitations, L2 cache thrashing, register pressure, memory coalescing issues, etc, and translate findings into actionable optimizations • Contribute to customer-facing technical engagements: present findings, recommend optimizations, and deliver measurable performance uplifts on vLLM Integrate and optimize custom kernels (Triton, Gluon, CK, PyDSL, ASM, AITER) within vLLM, understanding dispatch paths, shape extraction, and backend selection • Optimize multi-node distributed inference on vLLM: communication-compute overlap, parallelism strategies (TP/PP/EP/DP), and scale-out performance • Contribute to shared performance optimization methodology that raises the bar across the team • Leverage AI agents to accelerate daily work and help define best practices for AI-assisted performance engineering • Upstream optimizations into vLLM and adjacent open-source frameworks such as SGLang and PyTorch PREFERRED EXPERIENCE: • 5+ years of software development experience in GPU computing, AI systems, or high-performance computing • Hands-on experience with vLLM internals (V1 engine, scheduler, PagedAttention/KV cache manager, chunked prefill, ROCm backend integration); familiarity with SGLang, TensorRT-LLM, or similar is a plus • Strong background in end-to-end workload profiling and bottleneck diagnosis: you can trace from user request through the vLLM engine to the GPU kernel and back • Understanding of GPU kernel performance characteristics: occupancy, register and LDS pressure, memory coalescing, cache utilization, wavefront scheduling, and instruction-level bottlenecks • Ability to read and reason about kernel-level profiling data and translate it into concrete optimization actions. You may not write kernels from scratch daily, but you can tell why one is slow and what needs to change • Understanding of model architectures (transformers, MoE, diffusion), inference paradigms (speculative decoding, prefill-decode disaggregation, continuous batching), and how they map to hardware and to vLLM's execution model Experience with custom kernel development or integration (HIP, CUDA, Triton, CK, or similar) • Understanding of multi-GPU and multi-node distributed systems: scale-up and scale-out topologies, RCCL/NCCL, RDMA, and communication-compute overlap • Strong proficiency in Python and C++ • Ability to engage with customers, present findings, and support technical decisions • Fluent in AI-assisted development: daily user of AI agents and tools • Strong Linux systems knowledge • Excellent written and verbal English communication skills ACADEMIC CREDENTIALS: Master's, or PhD in Computer Science, Computer Engineering, Electrical Engineering, or equivalent. Advanced degree preferred but exceptional industry experience valued equally. LOCATION: Helsinki, Finland or Stockholm, Sweden. #LI-MH3 #LI-HYBRID Benefits offered are described: AMD benefits at a glance. AMD does not accept unsolicited resumes from headhunters, recruitment agencies, or fee-based recruitment services. AMD and its subsidiaries are equal opportunity, inclusive employers and will consider all applicants without regard to age, ancestry, color, marital status, medical condition, mental or physical disability, national origin, race, religion, political and/or third-party affiliation, sex, pregnancy, sexual orientation, gender identity, military or veteran status, or any other characteristic protected by law. We encourage applications from all qualified candidates and will accommodate applicants’ needs under the respective laws throughout all stages of the recruitment and selection process. AMD may use Artificial Intelligence to help screen, assess or select applicants for this position. AMD’s “Responsible AI Policy” is available here. This posting is for an existing vacancy. THE ROLE: AMD is looking for a performance-obsessed senior engineer to push AI inference performance to the limit on AMD GPUs, with vLLM as the primary serving framework. You will work end-to-end across the stack: profiling, diagnosing, and optimizing leading models running on vLLM across customer-relevant serving configurations (e.g. agentic coding, long-context, high-throughput serving). You will own hard performance problems on our most strategic customer engagements and leave behind measurable uplifts and reusable methodology. This is not a sustaining role: every engagement is different, every optimization leaves a lasting impact. THE PERSON: You can take an AI workload, understand it top to bottom, and make it faster on vLLM. You know the framework's internals well: PagedAttention and the KV cache manager, continuous batching and the scheduler, the V1 engine architecture, chunked prefill, and the paths that connect a user request down to the GPU kernel on the AMD (ROCm) backend. You are comfortable profiling a distributed vLLM deployment, diagnosing a kernel-level bottleneck, and clearly communicating your findings to both engineers and customer stakeholders. You understand GPU kernel performance well: not just how to use profiling tools, but how to reason about occupancy, cache behavior, memory coalescing, and instruction-level bottlenecks from first principles. You raise the bar through technical depth—you take on hard problems and help teammates grow along the way. You are AI-fluent, not just in the models you optimize, but in how you work: you leverage AI agents and tools daily to accelerate your workflows. You thrive under pressure, move fast, and measure everything. KEY RESPONSIBILITIES: • Drive performance optimization end-to-end on vLLM across leading models and customer-relevant serving configurations, closing competitive gaps through kernel and systems-level optimizations • Profile, diagnose, and resolve cross-stack performance bottlenecks in vLLM deployments, from GPU kernels and operator dispatch to the vLLM scheduler, PagedAttention/KV cache management, and multi-node communication • Diagnose kernel-level performance issues using profiling tools: identify occupancy limitations, L2 cache thrashing, register pressure, memory coalescing issues, etc, and translate findings into actionable optimizations • Contribute to customer-facing technical engagements: present findings, recommend optimizations, and deliver measurable performance uplifts on vLLM Integrate and optimize custom kernels (Triton, Gluon, CK, PyDSL, ASM, AITER) within vLLM, understanding dispatch paths, shape extraction, and backend selection • Optimize multi-node distributed inference on vLLM: communication-compute overlap, parallelism strategies (TP/PP/EP/DP), and scale-out performance • Contribute to shared performance optimization methodology that raises the bar across the team • Leverage AI agents to accelerate daily work and help define best practices for AI-assisted performance engineering • Upstream optimizations into vLLM and adjacent open-source frameworks such as SGLang and PyTorch PREFERRED EXPERIENCE: • 5+ years of software development experience in GPU computing, AI systems, or high-performance computing • Hands-on experience with vLLM internals (V1 engine, scheduler, PagedAttention/KV cache manager, chunked prefill, ROCm backend integration); familiarity with SGLang, TensorRT-LLM, or similar is a plus • Strong background in end-to-end workload profiling and bottleneck diagnosis: you can trace from user request through the vLLM engine to the GPU kernel and back • Understanding of GPU kernel performance characteristics: occupancy, register and LDS pressure, memory coalescing, cache utilization, wavefront scheduling, and instruction-level bottlenecks • Ability to read and reason about kernel-level profiling data and translate it into concrete optimization actions. You may not write kernels from scratch daily, but you can tell why one is slow and what needs to change • Understanding of model architectures (transformers, MoE, diffusion), inference paradigms (speculative decoding, prefill-decode disaggregation, continuous batching), and how they map to hardware and to vLLM's execution model Experience with custom kernel development or integration (HIP, CUDA, Triton, CK, or similar) • Understanding of multi-GPU and multi-node distributed systems: scale-up and scale-out topologies, RCCL/NCCL, RDMA, and communication-compute overlap • Strong proficiency in Python and C++ • Ability to engage with customers, present findings, and support technical decisions • Fluent in AI-assisted development: daily user of AI agents and tools • Strong Linux systems knowledge • Excellent written and verbal English communication skills ACADEMIC CREDENTIALS: Master's, or PhD in Computer Science, Computer Engineering, Electrical Engineering, or equivalent. Advanced degree preferred but exceptional industry experience valued equally. LOCATION: Helsinki, Finland or Stockholm, Sweden. #LI-MH3 #LI-HYBRID Benefits offered are described: AMD benefits at a glance. AMD does not accept unsolicited resumes from headhunters, recruitment agencies, or fee-based recruitment services. AMD and its subsidiaries are equal opportunity, inclusive employers and will consider all applicants without regard to age, ancestry, color, marital status, medical condition, mental or physical disability, national origin, race, religion, political and/or third-party affiliation, sex, pregnancy, sexual orientation, gender identity, military or veteran status, or any other characteristic protected by law. We encourage applications from all qualified candidates and will accommodate applicants’ needs under the respective laws throughout all stages of the recruitment and selection process. AMD may use Artificial Intelligence to help screen, assess or select applicants for this position. AMD’s “Responsible AI Policy” is available here. This posting is for an existing vacancy.