jobapplier

Member of Technical Staff - GPU Inference Engineer

Liquid AI

Skills

Required

  • GPU
  • vLLM / SGLang
  • serving framework
  • transformer
  • CUDA
  • kernels
  • LLM
  • GPUs
  • autoregressive decoding
  • Python
  • PyTorch
  • open-source
  • FlashInfer

Preferred

  • generation
  • GPU optimization
  • Triton kernels
  • synthetic data generation
  • Liquid Foundation Models
  • Hugging Face

Member of Technical Staff - GPU Inference Engineer

GPU inference sits underneath much of how we build and ship models at Liquid. Evaluation runs, synthetic data generation, and our API service all depend on how fast and how cheaply we can serve Liquid Foundation Models on GPUs, and demand for that work is growing. You will join the small team that owns our GPU serving stack, built on vLLM and SGLang, and make it faster and cheaper to run.

About Liquid AI

About Liquid AI

Spun out of MIT CSAIL, we build general-purpose AI systems that run efficiently across deployment targets, from data center accelerators to on-device hardware, ensuring low latency, minimal memory usage, privacy, and reliability. We partner with enterprises across consumer electronics, automotive, life sciences, and financial services. We are scaling rapidly and need exceptional people to help us get there.

Moves from evaluation throughput to data generation to API serving as priorities shift, and takes each one through to a measured result.

Works directly with the researchers and engineers who depend on inference, and treats their throughput as the result that counts.

The inference engine and configuration behind our API service are chosen on benchmark evidence, and have been profiled and tuned for latency and throughput

Responsibilities

San Francisco; Boston; Remote

  • This is a high-ownership role: you will be pointed at whichever inference bottleneck matters most and carry it from profiling to a benchmarked result. The work sits at the serving-framework level rather than in custom CUDA kernels, which is a separate role on the team.
  • Own and improve our GPU serving stack on vLLM and SGLang, from configuration and scheduling choices to changes in the frameworks themselves
  • Speed up large-scale evaluation and synthetic data generation, and design the inference pipelines those workloads run on
  • Select, profile, and benchmark the inference engine behind our API service, and adapt the API layer around it
  • Integrate Liquid Foundation Models into vLLM, SGLang, and Hugging Face Transformers, and maintain those public integrations as models and frameworks change
  • Diagnose throughput and latency problems alongside the researchers and serving engineers who hit them

Requirements

  • Explains why a workload is slow before changing it, working from what the model and the GPU are actually doing rather than from framework defaults.

Picks up the next problem and owns it

  • Hands-on experience running LLM inference on GPUs with a serving framework such as vLLM or SGLang, beyond default configuration
  • First-principles understanding of LLM inference: the transformer forward pass, autoregressive decoding on a GPU, and the latency, throughput, and memory trade-offs that follow
  • Solid grasp of GPU fundamentals as they bear on inference. You do not need to have written CUDA kernels
  • Experience profiling and benchmarking inference workloads and acting on what the numbers show
  • Strong Python and PyTorch, and comfort reading and changing large open-source codebases
  • Helpful, not required: Merged contributions to vLLM, SGLang, or FlashInfer

Preferred

  • Lower-level GPU optimization experience, such as CUDA or Triton kernels
  • Experience building high-throughput batch inference for evaluation or synthetic data generation
  • Evaluation and synthetic data generation run measurably faster and at lower cost per token on the same hardware, freeing GPU capacity for other work
  • Liquid models run correctly and efficiently in the public serving frameworks and libraries our users rely on, including vLLM, SGLang, and Transformers

Benefits

  • Problems standard stacks don't solve: Our hybrid architectures do not drop into off-the-shelf serving frameworks unchanged, and the inference you speed up feeds how every Liquid model is evaluated and trained. High ownership from day one. [INFERRED -- review this line]
  • Compensation & Equity: Competitive base salary plus equity, so you share in what we build. Final offers reflect skills, experience, and location.
  • Time Off: Unlimited PTO, plus company-wide Refill Days, long weekends when the whole company rests together.
  • Health & Wellbeing: Comprehensive health coverage, with details varying by location.
  • Paid Leave: Paid leave for pregnancy, parenting, medical needs, and caring for family.

Location Type

Hybrid

Easy Apply