Inference Optimization Engineer

About BentoML

BentoML is a leading inference platform provider that helps AI teams run large language models and other generative AI workloads at scale. With support from investors such as DCM, enterprises around the world rely on us for consistent scalability and performance in production. Our portfolio includes both open source and commercial products, and our goal is to help each team build its own competitive advantage through AI.

Role

As an Inference Optimization Engineer, you will improve the speed and efficiency of large language models at the GPU kernel level, through the inference engine, and across distributed architectures. You will profile real workloads, remove bottlenecks, and lift each layer of the stack to new performance ceilings. Every gain you unlock will flow straight into open source code and power fleets of production models, cutting GPU costs for teams around the world. By publishing blog posts and giving conference talks you will become a trusted voice on efficient LLM inference at large scale.

Example projects:

• https://bentoml.com/blog/structured-decoding-in-vllm-a-gentle-introduction

• https://bentoml.com/blog/benchmarking-llm-inference-backends

• https://bentoml.com/blog/25x-faster-cold-starts-for-llms-on-kubernetes

Responsibilities

• Latency & throughput - Identify bottlenecks and optimize inference efficiency in single-GPU, multi-GPU, and multi-node serving setups.

• Benchmarking - Build repeatable tests that model production traffic; track and report vLLM, SGLang, TRT-LLM, and future runtimes.

• Resource efficiency - Reduce memory use and compute cost with mixed precision, better KV-cache handling, quantization, and speculative decoding.

• Serving features - Improve batching, caching, load balancing, and model-parallel execution.

• Knowledge sharing - Write technical posts, contribute code, and present findings to the open-source community.

Qualifications

• Deep understanding of transformer architecture and inference engine internals.

• Hands-on experience speeding up model serving through batching, caching, load balancing.

• Experienced with inference engines such as vLLM, SGLang, or TRT-LLM (upstream contributions are a plus).

• Experienced with inference optimization techniques: quantization, distillation, speculative decoding, or similar.

• Proficiency in CUDA and use of profiling tools like Nsight, nvprof, or CUPTI. Proficiency in Triton and ROCm is a bonus.

• Track record of blog posts, conference talks, or open-source projects in ML systems is a bonus.

Why join us

• Direct impact – optimize distributed LLM inference and large GPU clusters worldwide and cut real GPU costs.

• Technical scope – operate distributed LLM inference and large GPU clusters worldwide.

• Customer reach – support organizations around the globe that rely on BentoML.

• Influence – mentor teammates, guide open-source contributors, and become a go-to voice on efficient inference in the community.

• Remote work – work from where you are most productive and collaborate with teammates in North America and Asia.

• Compensation – competitive salary, equity, learning budget, and paid conference travel.