Senior Scientist: Efficient LLM Inference & Model Optimization

Userc Dehradun; Greater London; Hybrid; GBP 90,000 - 130,000; Full time

Job summary

Nebius in London is seeking a Senior Applied Scientist to drive efficient LLM inference and model optimization. You will own focused research, run rigorous experiments, and deliver production-ready prototypes and artifacts.

Your work spans quantization, QAT, distillation, speculative decoding, KV-cache reuse/compression, long-context inference, MoE routing, and model/runtime co-optimization. Build prototypes with PyTorch, Triton, CUDA, and evaluate latency, throughput, and cost for production

Qualifications

• PhD in CS/ML or closely related field.
• Strong Python and PyTorch implementation skills.
• Deep knowledge of LLMs, VLMs, transformer inference, decoding and production-serving trade-offs.

Responsibilities

• Lead research on efficient LLM and VLM inference, from hypotheses to production handoff.
• Develop methods across quantization, QAT, distillation, speculative decoding, KV-cache reuse/compression, long-context inference, MoE routing, and co-optimization.
• Build prototypes using PyTorch, Triton, CUDA tooling, and collaborate with engineers to productionise.
• Study LLM request routing, scheduling, cache-aware load balancing, and prefill–decode disaggregation for latency, throughput and cost.
• Investigate multi-node inference, WideEP, expert placement and load balance for scaling models.
• Design rigorous evaluations covering quality, latency, throughput, stability, memory and cost per token, under consistent workloads.
• Collaborate across teams to prioritise research with measurable production impact and share results publicly.

Skills

Python

PyTorch

LLMs

Transformer inference

Quantization

Model compression

Experimental design

Education

PhD in CS/ML or related field

Tools

PyTorch internals

Triton

CUDA

TensorRT-LLM

FlashAttention

Job description

Nebius in London is seeking a Senior Applied Scientist to drive efficient LLM inference and model optimization. You will own focused research, run rigorous experiments, and deliver production-ready prototypes and artifacts.

Your work spans quantization, QAT, distillation, speculative decoding, KV-cache reuse/compression, long-context inference, MoE routing, and model/runtime co-optimization. Build prototypes with PyTorch, Triton, CUDA, and evaluate latency, throughput, and cost for production