Performance Engineer, Inference Engine

Anthropic•San Francisco, CA

12d•$350,000 - $850,000•Hybrid

About The Position

Anthropic's inference engine is the software between the accelerator kernels and the routing layer. It manages the entire token path in between: batching requests, laying the model out across chips, managing memory for weights and activations, coordinating every forward pass, and managing model state across requests. Built in-house, it runs on all of our accelerator platforms, serving Claude to millions of users and running our research workloads. You will work on building and optimizing this system at Anthropic scale: improving throughput, cost, reliability, and latency across all accelerator and cloud platforms. You are intimately familiar with the hardware and bandwidth numbers (FLOPs, HBM, PCIe, RDMA, network links, etc.) and can model a problem quickly: where the time and bytes go, and what sets the bound. The role is deeply technical and high-impact, and suits engineers who enjoy working across accelerator programming, high-performance systems that seamlessly coordinate between host and device, and large-scale distributed systems. Familiarity with the transformer architecture is a plus.

Requirements

A working mental model of LLM inference: how prefill and decode land on an accelerator's compute, memory, and interconnect, and what the host is doing meanwhile
Proven quick learner: ramped fast in deep, unfamiliar systems and shipped consequential changes quickly
Strong systems programming (Rust, C++, or similar), with care for code quality and tests
Analytical about performance: observe and profile first, form a hypothesis, test it, then change the code and measure again
Low ego: ask the naive question, take feedback well, pick up slack outside your job description
Enjoy pair programming (we love to pair!) and care about the societal impacts of your work

Nice To Haves

Experience inside an LLM serving engine and a sense of where its abstractions strain
GPU/Accelerator programming
OS internals
Language modeling with transformers
Experience building an allocator, cache, scheduler, or high-bandwidth transport
Fluency in Rust
Experience making systems reproducible: determinism, replay, property-based tests

Responsibilities

Keep device utilization high. Accelerators should never be waiting due to other overheads.
Reuse instead of recompute. Keep model state cached and reuse it whenever that is cheaper than computing it again.
Measure, model, then change. We build the observability to see where the gaps are, model the impact of potential improvements, deploy them, and go around again, with Claude speeding up every turn of that loop.
Tokens you can trust. Ensuring model quality matters more than efficiency. We build the infrastructure to ensure Claude maintains its intelligence across platforms and over time.
Safety on every token. We work closely with our safeguards and safety teams. The inference engine is the backbone behind our production safety systems, ensuring efficiency without compromising robustness.