Inference Engineering, Co-op

Inferact•San Francisco, CA

3d•Onsite

About The Position

Inferact's mission is to grow vLLM as the world's AI inference engine and accelerate AI progress by making inference cheaper and faster. Founded by the creators and core maintainers of vLLM, we sit at the intersection of models and hardware, a position that took years to build. We're looking for exceptional University of Waterloo co-op students who want to work on the systems that determine how fast, efficiently, and reliably frontier AI models run on frontier workloads at scale. This is not a sandboxed internship project. We will match your strengths and interests to a real engineering problem across the vLLM stack, from model execution and low-level accelerator code to distributed serving and the cloud platform that makes it all usable. You'll work alongside the creators and core maintainers of vLLM on work intended to ship into open source, production systems, or the tooling that supports both. You will have a primary technical track, meaningful ownership, and mentorship from a small, senior team, with opportunities to collaborate across models, compilers, accelerators, networking, and distributed systems. The goal is to let exceptional students learn at the frontier while making contributions used by developers and AI teams around the world.

Requirements

Currently pursuing a bachelor's, master's, or PhD degree in computer science, engineering, mathematics, or a related technical field, and eligible for a University of Waterloo co-op work term.
Strong programming ability in Python, C++, Rust, Go, or another systems-oriented language.
Strong computer science fundamentals and the ability to learn from research papers, technical documentation, and complex systems code.
Evidence that you have built and debugged nontrivial software through coursework, research, a prior internship, open-source work, or an ambitious side project.
A builder mindset: you define what success means, measure results, validate correctness, communicate clearly, and keep iterating until the system works.

Nice To Haves

Depth in at least one relevant area such as ML or inference systems, distributed systems, GPU or accelerator programming, compilers, high-performance computing, operating systems, networking, Kubernetes, or cloud infrastructure.
Hands-on experience with one or more relevant technologies such as PyTorch, vLLM, SGLang, TensorRT-LLM, CUDA, Triton, TileLang, ROCm/HIP, JAX/XLA, Pallas, Kubernetes, Helm, Terraform, Ray, or SLURM.
Experience profiling or benchmarking software, validating numerical or systems correctness, or building tests that protect against performance regressions.
Experience with systems beyond a single local process or device, including multi-GPU, multi-node, research-cluster, high-throughput, or production-like workloads.
Contributed to open-source ML, systems, compiler, or infrastructure projects, especially vLLM or adjacent projects.
Built a research system, benchmark suite, compiler or kernel project, distributed service, or infrastructure tool that demonstrates unusual technical depth and initiative.
Worked across more than one layer of the stack and enjoy moving between algorithms, runtimes, systems, and hardware rather than staying inside a single abstraction.
Created a technical artifact such as a paper, design document, blog post, demo, or talk that clearly explains what you built and what you learned.

Responsibilities

Bring new model architectures and inference techniques to life in vLLM.
Implement ideas from research papers; support mixture-of-experts, multimodal, diffusion, and agentic workloads; and improve scheduling, continuous batching, KV-cache memory management, prefix caching, and hybrid model serving.
Build the distributed serving data plane that lets vLLM run across many GPUs and nodes.
Work on tensor, expert, or context parallelism; prefill/decode separation and KV-cache transport; fault tolerance and multi-tenancy; and high-performance communication using NCCL, DeepEP, NVSHMEM, RDMA, or InfiniBand.
Raise the hardware performance ceiling by writing and optimizing attention, GEMM, sampling, KV-cache, fused, and quantization kernels.
Use CUDA, Triton, TileLang, CUTLASS/CUTE, or related tools; reason about memory hierarchy, occupancy, and tensor cores; and prove speedups through profiling, correctness tests, and reproducible benchmarks.
Help make vLLM first-class on AMD accelerators.
Work across ROCm, HIP, Triton, CK, AITER, kernels, runtime paths, quantization, compiler integration, and performance-regression infrastructure while learning how AMD-specific execution, memory, and toolchain constraints shape inference.
Help make vLLM fast and correct on Google TPUs.
Build backend, runtime, and compiler integrations with JAX, XLA, Pallas, MLIR, and related tooling; inspect compiler artifacts; work on lowering, fusion, and code generation; and benchmark production-relevant serving across correctness, latency, and throughput.
Build the operational platform that makes large-scale inference deployable and reliable.
Work on Kubernetes and custom operators, topology-aware GPU scheduling, zero-downtime vLLM rollouts, token-aware routing, observability, infrastructure-as-code, automated recovery, and bring-your-own-cloud or multi-cloud fleet management.