AI Infrastructure Expert — KV Cache System

Brak informacji o wynagrodzeniu

SeniorOn-site·WarszawaFull-time

#472369·Dodano 2 dni temu·0

Źródło: nofluffjobs.com

Aplikuj teraz

Tech Stack / Keywords

LLMCacheStorageNetworkingDegree

Podsumowanie oferty

Design and implement high-performance KV Cache systems for large-model inference, addressing memory and communication bottlenecks in ultra-long-context and high-concurrency serving. The work includes core cache mechanisms, distributed inference optimization, and performance benchmarking in collaboration with inference engine, scheduler, and hardware teams.

Wymagania

Solid understanding of LLM inference fundamentals and KV Cache internals.
Hands-on experience building KV Cache systems, general cache or storage subsystems, or high-performance networking stacks is strongly preferred.
Proficiency in performance profiling and tuning, including diagnosing memory, bandwidth, and communication bottlenecks, with a proven track record of performance optimization on large-scale systems.
Strong algorithmic mindset. Candidates with competitive programming backgrounds (such as ICPC, IMC, IOI, IMO, or IOAI), first-author papers on algorithm or system performance optimization, or open-source contributions are prioritized.
Ability to decompose complex problems independently and drive solutions from theoretical modeling through engineering implementation.
Master's or Ph.D. degree in Computer Science, Computer Architecture, High Performance Computing, or a related field; exceptional bachelor's candidates with strong relevant backgrounds may also be considered.

Nice to have:

Familiarity with mainstream inference frameworks, including vLLM, SGLang, and TensorRT-LLM, and experience reading source code for KV Cache modules.
Hands-on experience with heterogeneous memory, remote memory (RDMA), distributed caching, page-based storage, or custom memory pool implementations.
Participation in systems, storage, or HPC competitions, or research experience with papers at top-tier systems conferences such as OSDI, SOSP, ATC, or NSDI.
Familiarity with multi-GPU or multi-NPU collective communication and optimization of NCCL-style collective primitives.

Obowiązki

Research, design, and implement core KV Cache mechanisms, including intelligent eviction policies, KV compression, page-based memory management, hierarchical caching across heterogeneous memory (VRAM, host RAM, and external storage), pre-scheduling, and prefetching.
Optimize KV sharding, migration, and sharing for distributed inference, and resolve communication bottlenecks for KV data transfer across multiple accelerators and nodes.
Profile system performance to identify memory-bandwidth and communication-overhead bottlenecks; improve end-to-end throughput through co-optimization of algorithmic policies and low-level system engineering.
Design benchmark suites and simulation frameworks to validate KV strategy effectiveness, generate performance reports, and drive production deployment of optimized solutions.
Collaborate with inference engine, scheduler, and hardware teams to deliver end-to-end optimizations for large language model inference systems.

Benefity

Private healthcare package.
Sport subscriptions and sport association options.
Benefit Platform options, including cinema and theater tickets, shopping cards, and discounts.
Special employee discounts through partnerships with local companies.
15-minute chair-based office massages.

Opieka zdrowotna

Karta sportowa

Dodatkowe benefity

HUAWEI

3 aktywne oferty

Zobacz wszystkie oferty

Aplikuj teraz