Training Performance Engineer (AI)

TL;DR

Training Performance Engineer (AI): Optimizing large-scale foundation model training on Blackwell clusters with an accent on kernel-level performance, throughput, and cluster fabric efficiency. Focus on solving complex challenges in low-precision training, MoE parallelism, and custom attention-variant kernels to push MFU and uptime.

Location: Must be based in the Netherlands or Switzerland, with an expectation of at least 50% time in the office.

Company

hirify.global is a well-funded startup building a next-generation agentic clinical AI assistant designed to support complex diagnostic workflows and clinical decision-making.

What you will do

• Instrument and analyze training runs to identify and close utilization gaps.
• Benchmark NCCL collectives over InfiniBand and NVLink to optimize fabric performance.
• Drive low-precision training initiatives and validate performance gains.
• Tune MoE parallelism strategies (TP/PP/CP/EP/DP) to optimize communication costs.
• Implement and integrate custom attention-variant kernels into the training stack.

Requirements

• Must be based in the Netherlands or Switzerland.
• Deep experience with GPU systems, including kernel-level CUDA or Triton.
• Proficiency with CUTLASS, Flash Attention, PyTorch, and Nsight profiling.
• Production experience with NCCL on high-bandwidth interconnects like InfiniBand.
• Strong understanding of parallelism strategies under memory and MFU constraints.

Nice to have

• Experience with low-precision training (FP8, dynamic loss scaling).
• Knowledge of sparse, hybrid, or MLA attention at the kernel level.
• Proven track record of shipping large-scale MoE training in production.
• Experience with Megatron or NeMo frameworks.

Culture & Benefits

• Competitive salary and pension plan.
• 25 days of annual vacation.
• EUR 1000 annual learning and development budget.
• Regular offsites and team events.
• Flexible work environment with commuting subsidy.

Hiring process

• Screening call to align on motivation and fit.
• Technical take-home assessment.
• Technical assessment debrief and collaboration discussion.
• Final onsite interview to discuss long-term alignment.