Machine learning engineer in LLM inference optimization

FDE Times

Berlin

On-site

EUR 120,000 - 180,000

Full time

8 days ago

A complete application in a minute

Tailored resume and cover letter, ready to send.

Get past ATS filters

Benefits offered by this job

Конкурентная оплата

Возможности карьерного роста

Гибкость и автономия

Инновационная культура

Проекты с глобальным воздействием

Международная команда

Job summary

Nebius ищет инженера по инфраструктуре ИИ для оптимизации и развёртывания трансформерных моделей на уровне продакшн. Вам предстоит работать над снижением задержки, увеличением пропускной способности и эффективного использования GPU в рамках сложных рабочих нагрузок.

Кандидат должен обладать опытом в Python и PyTorch, с практикой разработки и развёртывания LLM/VLM. Ожидается умение работать в международной среде и взаимодействовать с различными командами.

Qualifications

Сильные навыки Python и PyTorch.
Практический опыт развёртывания и оптимизации LLM/VLM или высокоп throughput трансформер-инференса.
Знание современных инференс-стеков (vLLM, TensorRT-LLM, Triton, Ray Serve и пр.).
Понимание узких мест инференса трансформеров: KV-кэш, батчинг, память, пропускная способность.
Умение оценивать trade-offs по задержке, пропускной способности, качеству, загрузке и цене.

Responsibilities

Владение проектами по оптимизации для семейств моделей и точек доступа.
Сравнение инференс-движков и настройка конфигураций под нагрузку.
Диагностика регрессий качества моделей в продакшн.
Оптимизация задержки, пропускной способности и затрат у LLM/VLM.
Развертывание и настройка движков и пайплайнов для продакшн.

Tools

vLLM

SGLang

TensorRT-LLM

Triton Inference Server

NVIDIA Dynamo

Ray Serve

KServe

Job description

Описание

Nebius is leading a new era in cloud infrastructure for the global AI economy. We are building a full-stack AI cloud platform that supports developers and enterprises from data and model training through to production deployment, without the cost and complexity of building large in-house AI/ML infrastructure.

Задачи

Own optimization projects for specific model families, customer endpoints, or serving backends.
Compare inference engines and recommend serving configurations suited to individual workloads.
Diagnose model-quality and performance regressions during production rollouts.
Improve LLM and VLM endpoint latency, throughput, memory efficiency, GPU utilization, quality, and cost per token.
Deploy, configure, benchmark, and extend engines such as vLLM, SGLang, TensorRT-LLM, Triton Inference Server, and NVIDIA Dynamo.
Develop production model-compression workflows covering quantization, quantization-aware training, distillation, low-bit serving, and accuracy recovery.
Implement or integrate speculative decoding, draft-model approaches, KV-cache optimization, prefix caching, chunked prefill, continuous batching, and disaggregated prefill/decode serving.
For prefill–decode disaggregation (PDD), evaluate KV-cache transfer, worker placement, and capacity balancing to determine which workloads benefit from the architecture.
Design and improve LLM request routers and scheduling policies that balance worker utilization, queueing, request characteristics, and KV-cache locality while meeting latency and reliability targets.
Scale dense and mixture-of-experts inference across multiple GPU nodes.
Select and tune tensor, pipeline, data, and expert parallelism—including wide expert parallelism (WideEP)—with attention to hardware topology, expert load balance, and communication overhead.
Build reproducible benchmark harnesses measuring TTFT, TPOT, tokens per second per GPU, p95/p99 latency, GPU memory, reliability, and cost per token.
Validate architecture choices under representative traffic and consistent GPU budgets.
Collaborate with GPU kernel and platform engineers to trace bottlenecks across model code, kernels, runtime, scheduler, gateway, and cluster layers.
Produce design documents, performance reports, rollout plans, and customer-facing technical explanations.

Требования

Strong engineering skills in Python and PyTorch.
Hands‑on experience deploying or optimizing LLM, VLM, or high‑throughput transformer inference.
Practical knowledge of at least one modern inference stack, such as vLLM, SGLang, TensorRT-LLM, Triton Inference Server, NVIDIA Dynamo, Ray Serve, KServe, or an equivalent internal system.
A strong understanding of transformer inference bottlenecks involving KV cache, attention, memory bandwidth, batching, parallelism, and long‑context serving.
Ability to quantify trade‑offs among latency, throughput, quality, utilization, and cost.
Experience designing or optimizing distributed inference system architecture, with hands‑on work in one or more areas such as request routing, distributed scheduling, PDD, multi‑node inference, or expert parallelism.
Strong communication and collaboration skills across research, kernel, infrastructure, product, and customer teams.
Будет плюсом: Experience with quantization‑aware training, post‑training quantization, FP8, INT8, INT4, NVFP4, MXFP4, AWQ, GPTQ, or SmoothQuant;
Experience with distillation, speculative decoding, EAGLE, Medusa, multi‑token prediction, or related inference acceleration methods;
Experience supporting agentic workloads involving tool calling, structured outputs, streaming APIs, high concurrency, and multi‑step orchestration;
Familiarity with CUDA or Triton;
The role does not require kernel engineering to be the candidate’s primary specialization;
Contributions to open‑source projects such as vLLM, SGLang, TensorRT-LLM, FlashInfer, LMCache, PyTorch, Triton, Ray, or KServe;
Implementation experience with cache‑aware request routing, PDD, or WideEP, including diagnosing communication bottlenecks and load imbalance;
Familiarity with GPU communication libraries and interconnects, such as NCCL, NVLink, InfiniBand, or RoCE.

Условия