RL Inference Engineer: Low-Precision ML at Scale

SpaceX Super Intelligence; Palo Alto (CA); On-site; USD 180,000 - 440,000; Full time

Benefits offered by this job

Equity

Health insurance

401(k) retirement plan

Disability insurance

Life insurance

Job summary

SpaceXSI’s RL infrastructure team seeks an engineer to advance low precision RL training and inference across diverse workloads. You will design, implement, and optimize the inference stack, and collaborate with the modelling team to deploy novel RL techniques at scale.

You will tackle performance bottlenecks on large-scale RL systems, profiling across teams and stacks, and contribute to building robust, efficient, and extensible solutions for experimental and production workloads.

Qualifications

• Experience in building, debugging, and optimizing efficiency of large-scale distributed systems.
• Experience in LLM inference.
• Proficiency in programming languages such as Python, C++ and/or Rust; frameworks such as PyTorch, Jax, CUDA.
• Willingness to dive deep and solve hardcore problems at all levels of the stack.

Responsibilities

• Design and optimize our inference stack for all shapes of RL workloads at SpaceXSI, from small scale ablations to production training runs.
• Analyze, profile and address performance bottlenecks in large scale RL systems
• Work closely with the modelling team to efficiently implement novel RL techniques and algorithms

Skills

Distributed systems

LLM inference

Python

C++

Rust

Deep learning

Problem solving

Tools

PyTorch

Jax

CUDA

Job description

SpaceXSI’s RL infrastructure team seeks an engineer to advance low precision RL training and inference across diverse workloads. You will design, implement, and optimize the inference stack, and collaborate with the modelling team to deploy novel RL techniques at scale.

You will tackle performance bottlenecks on large-scale RL systems, profiling across teams and stacks, and contribute to building robust, efficient, and extensible solutions for experimental and production workloads.