К содержанию
myOffer Вакансии
Все вакансии

tensordyne

Senior Software Engineer, Inference Performance

Опубликовано

География
Munich
Уровень
Senior
PythonRustC++

Актуальность нужно проверить

Откройте публикацию в источнике и уточните, продолжается ли набор.

Источник проверяли 14 ноября

Проверить в источнике

Доступ проверим при нажатии.

О вакансии

Вакансия «Senior Software Engineer, Inference Performance» в компании tensordyne. Место работы: Munich.

Чем предстоит заниматься

  • You'll build the software stack that takes a model from code down to the hardware: the abstractions, kernels, libraries, and tools that engineers and agents use to reach peak performance. Help optimize the latest AI models on our silicon.
  • Our hardware is designed to be straightforward to program, so the goal is not to hide it. The goal is to expose the right level of control at each layer, from our compiler-driven Python SDK to our low-level Rust programming stack, and to make every performance gap measurable and every fix permanent.
  • We're looking for a builder who understands hardware deeply and develops with agents by default. You'll own substantial parts of one of these layers and work closely with engineers across models, compilers, runtime, and hardware. Your focus is performance: where cycles, bandwidth, and communication go, and how to win them back. Number formats and model accuracy are owned by our conversion and emulation team; you'll partner with them where precision and speed meet.
  • Build the stack. Design APIs, kernel abstractions, and libraries in our Python SDK or Rust stack so model engineers and agents can reach high performance at the layer that fits their problem.
  • Make models fast. Take priority workloads, from LLMs to multimodal models, to peak performance on real silicon. Profile, find the bottleneck across kernels, compiler, runtime, and parallelism, and fix it at the right layer.
  • Close the gap between model and measurement. Compare measured performance against analytical models and simulation. Explain every material gap, and turn the explanation into a fix or a better model.
  • Build the agentic performance loop. Give agents the profilers, traces, counters, and benchmark harnesses to form a hypothesis, change code, and measure the result in minutes. Make the results trustworthy enough to act on.
  • Set the bar for performance. Define how we measure, track regressions, and report results.
  • Build for the next chip. Design abstractions that carry forward to future hardware generations. Work with hardware architects to test how architectural choices affect real workloads.
  • What we need from you
  • You understand hardware performance. You've optimized AI workloads on GPUs or custom accelerators and can reason about compute utilization, memory hierarchy, latency hiding, synchronization, and multi-chip scaling.
  • You build excellent software. You've owned substantial systems from design through production. You're strong in Python and/or at least one systems language such as Rust, C++, or C, and rigorous about interfaces, abstractions, and testing. We care about engineering quality more than a specific language.
  • You understand LLM inference end to end. You can read a new model's implementation and reason about prefill, decode, attention variants, KV caches, MoE, and parallelism, and how they map to compute, memory, and communication.
  • You build with agents. Agentic development is how you work, not an experiment. You'd rather build the loop that finds the optimization than find it by hand, and you know when the loop still needs a human.
  • You build things others succeed with. You've designed APIs, libraries, or performance tooling that other engineers adopted.
  • You drive problems to resolution. You form clear hypotheses, measure, and follow the evidence across software and hardware boundaries. You take ownership of the result.
  • If you hit a few of those boxes, we'd like to speak to you.
  • We'd be thrilled if you had
  • Experience with embedded, bare-metal, or low-level runtime programming.
  • Experience with MLIR, other compiler frameworks, or kernel DSLs such as Triton, Pallas, NKI, or CuTe.
  • Contributions to LLM serving stacks such as vLLM or SGLang, or a track record of taking end-to-end inference workloads to competitive performance on a hardware target.

Источник: Arbeitnow. Условия и актуальность вакансии подтверждает работодатель.

Подходит вакансия?

Перейдите к исходному объявлению или контакту работодателя.

Проверить в источнике

Доступ проверим при нажатии.

Можно искать с поддержкой

Сопровождение поиска

Личный специалист поможет с резюме, подбором вакансий и откликами. Вы сосредоточитесь на интервью.

Делегировать поиск

Менторинг + поиск

Практика с профильным ментором: план подготовки, задачи и обратная связь. После подготовки — поиск работы.

О менторинге

Можно искать с поддержкой

Сопровождение поиска

Личный специалист поможет с резюме, подбором вакансий и откликами. Вы сосредоточитесь на интервью.

Делегировать поиск

Менторинг + поиск

Практика с профильным ментором: план подготовки, задачи и обратная связь. После подготовки — поиск работы.

О менторинге

Больше подходящих вакансий

Какой поддержки вам не хватает?

Можно передать поиск команде или сначала подготовиться с ментором.

Сопровождение поиска

Личный специалист поможет с резюме, подбором вакансий и откликами. Вы сосредоточитесь на интервью.

Делегировать поиск

Менторинг + поиск

Практика с профильным ментором: план подготовки, задачи и обратная связь. После подготовки — поиск работы.

О менторинге

Каталог остаётся бесплатным. Подробности программ откроются в новой вкладке.