Skip to main content
TALENT247 // SPECIALIST AI EXECUTIVE SEARCH & TECHNICAL RECRUITMENT
INF · AI Infrastructure

Senior CUDA & Distributed LLM Inference Engineer

Paris, France · Full-time · Hybrid

AI Specialism

AI Infrastructure

Workplace

Hybrid

Location

Paris, France

Compensation

€110k – €155k

annual package

Role Dossier & Specification

REF: SENIOR-CUDA-DISTRI

THE CLIENT & MANDATE

Our client is Europe's premier frontier foundation model laboratory headquartered in Paris, renowned globally for pioneering high-efficiency dense and sparse Mixture-of-Experts (MoE) language, code, and vision models. Because their models run everywhere—from massive multi-tenant cloud APIs to sovereign air-gapped defence clusters and single-node enterprise appliances—squeezing every last drop of memory bandwidth and compute utilization out of modern GPUs is existential to their platform advantage.

They have retained Talent247 to appoint a Senior CUDA & Distributed LLM Inference Engineer in Paris. Joining their core Inference Runtime & Kernels group, you will co-design low-level GPU kernels and distributed serving architectures alongside model researchers so newly trained architectures launch on day zero at world-leading throughput and minimal time-to-first-token.

KEY RESPONSIBILITIES

  • Author, profile, and maintain custom C++20, CUDA/PTX, CUTLASS, and Triton kernels for grouped-query/sliding-window attention, sparse MoE routing/grouped GEMMs, and fused RoPE/SwiGLU/RMSNorm operators on Hopper and Blackwell GPUs.
  • Implement ultra-low-overhead quantization formats (FP8, NVFP4, INT4 weight-only and W4A8) paired with calibration pipelines that preserve reasoning and coding benchmark scores.
  • Scale multi-node disaggregated prefill/decode serving, expert parallelism over NVLink and InfiniBand RDMA, continuous batching, and radix prefix KV-cache sharing.
  • Collaborate upstream and internally across vLLM, TensorRT-LLM, and custom Rust/C++ serving runtimes to ensure sovereign on-premise and cloud API parity.
  • Partner with pre-training architects before major training runs lock in, modeling KV-cache footprint, expert-routing balance, and hardware FLOP utilization to ensure hardware-friendly architectures.
  • Build deterministic kernel correctness suites and hardware-level roofline profilers catching regressions on every commit.

TECHNICAL ENVIRONMENT & OFFERING

Their runtime stack spans C++20, CUDA/CUTLASS/PTX, Triton, Rust, Python/PyTorch, vLLM/TensorRT-LLM, NCCL/NVSHMEM, and dedicated H100/H200/B200 clusters. Working hybrid in central Paris, our client offers €110,000–€155,000 EUR base salary plus premier frontier-lab BSPCE equity, comprehensive benefits, and full visa/relocation support.

Core Requirements

ESSENTIAL QUALIFICATIONS

  • 5+ years of low-level GPU performance engineering, HPC systems, or ML inference runtime development.
  • Expert proficiency in C++17/20 and CUDA (including PTX/SASS inspection, CUTLASS/CuTe, Tensor Core MMA instructions, CUDA Graphs, and Nsight Compute/Systems profiling) and/or OpenAI Triton.
  • Deep production understanding of modern LLM serving mechanics: PagedAttention, continuous batching, speculative decoding, FP8/INT4 quantization, and tensor/pipeline/expert parallelism.
  • Strong grasp of multi-GPU interconnect behavior across NVLink, NVSwitch, PCIe, and GPUDirect RDMA / NCCL collectives.
  • Pragmatic, benchmark-driven engineering approach focused on end-to-end p99 latency and tokens-per-watt/dollar.

PREFERRED / BONUS EXPERIENCE

  • Core contributions to vLLM, SGLang, TensorRT-LLM, FlashAttention, llama.cpp, or PyTorch CUDA internals.
  • Experience bringing up inference kernels on next-generation Blackwell (B200/GB200) or AMD MI300X architectures.

Position Telemetry

Location Paris, France (Hybrid)
Department AI Infrastructure
Employment type Full-time
Workplace type Hybrid
Salary range
€110k – €155k annual package