onsite
ML Model Serving Engineer
ML Model Serving Engineer
As an ML Model Serving Engineer, you will be responsible for optimizing and maintaining a high-performance serving layer for various machine learning models, including LLM, speech, and vision models. You will work closely with ML infrastructure and training teams to ensure fast, cost-effective, accurate, and reliable model serving, leveraging and extending frameworks like VLLM and SGLang.
About the role
About Sesame
Sesame believes in a future where computers are lifelike - with the ability to see, hear, and collaborate with us in ways that feel natural and human. With this vision, we're designing a new kind of computer, focused on making voice agents part of our daily lives. Our team brings together founders from Oculus and Ubiquity6, alongside proven leaders from Meta, Google, and Apple, with deep expertise spanning hardware and software. Join us in shaping a future where computers truly come alive.
Responsibilities
- Turbocharge our serving layer, consisting of a variety of LLM, speech, and vision models.
- Partner with ML infrastructure and training engineers to build a fast, cost-effective, accurate, and reliable serving layer to power a new consumer product category.
- Modify and extend LLM serving frameworks like VLLM and SGLang to take advantage of the latest techniques in high-performance model serving.
- Work with the training team to identify opportunities to produce faster models without sacrificing quality.
- Use techniques like in-flight batching, caching, and custom kernels to speed up inference.
- Find ways to reduce model initialization times without sacrificing quality.
Required Qualifications
- Expert in some differentiable array computing framework, preferably PyTorch.
- Expert in optimizing machine learning models for serving reliably at high throughput, with low latency.
- Significant systems programming experience; ex. Experience working on high-performance server systems—you’d be just as comfortable with the internals of VLLM as you would with a complex PyTorch codebase.
- Significant performance engineering experience; ex. Bottleneck analysis in high-scale server systems or profiling low-level systems code.
- Always up to date on the latest techniques for model serving optimization.
Preferred Qualifications
- Familiarity with high-performance LLM serving; ex. experience with VLLM, SGlang deployment, and internals.
- Experience with a public cloud platform such as GCP, AWS, or Azure.
- Experience deploying and scaling inference workloads in the cloud using Kubernetes, Ray, etc.
- You like to ship and have a track record of leading complex multi-month projects without assistance.
- You’re excited to learn new things and work in a multitude of roles.