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

Senior Diffusion Distillation & GPU Inference Engineer

Berlin, Germany · Full-time · Hybrid

AI Specialism

AI Infrastructure

Workplace

Hybrid

Location

Berlin, Germany

Compensation

€105k – €150k

annual package

Role Dossier & Specification

REF: SENIOR-DIFFUSION-D

THE CLIENT & MANDATE

Our client is one of the most celebrated frontier generative media laboratories in the world, founded in Germany by the core scientific pioneers behind modern latent diffusion and rectified-flow visual generation. Operating out of Berlin and Freiburg, their open-weight and API visual foundation models power global creative suites, design platforms, e-commerce studios, and consumer AI products at tens of millions of daily generations.

They have retained Talent247 to hire a Senior Diffusion Distillation & GPU Inference Engineer in Berlin. Working shoulder-to-shoulder with the researchers inventing next-generation image, editing, and video Diffusion Transformers (DiTs), you will compress 50-step multi-billion-parameter flow-matching models into sub-second, interactive 1-to-4-step production engines running at world-record GPU efficiency.

KEY RESPONSIBILITIES

  • Research, implement, and scale step-distillation and architectural compression recipes—including adversarial diffusion distillation, consistency/rectified-flow trajectory distillation, and width/depth pruning—for 10B+ parameter image and video DiTs.
  • Author and tune custom CUDA, CUTLASS, and Triton kernels for quantized attention (FP8, NVFP4, INT4), fused AdaLN modulation, 3D spatiotemporal RoPE, and high-resolution VAE decoding.
  • Profile end-to-end serving stacks with Nsight Systems and Nsight Compute across H100, H200, and Blackwell B200 GPUs to eliminate host-overhead, memory-bandwidth stalls, and multi-GPU communication bottlenecks.
  • Implement sequence/context parallelism (USP/Ring Attention) and dynamic resolution batching to serve megapixel image editing and high-frame-rate video generation concurrently.
  • Build perceptual and human-preference evaluation pipelines ensuring distilled and quantized checkpoints preserve typography accuracy, photorealistic detail, and prompt adherence.
  • Release industry-leading open-source reference inference implementations alongside commercial API rollouts.

TECHNICAL ENVIRONMENT & OFFERING

Their stack runs on PyTorch, C++20, CUDA/CUTLASS, Triton, torch.compile, TensorRT, and dedicated multi-hundred-H100/B200 clusters. Working hybrid from Berlin (or Freiburg), our client offers €105,000–€150,000 EUR base salary plus high-upside frontier-lab equity, 30 days vacation, and full German visa/relocation support.

Core Requirements

ESSENTIAL QUALIFICATIONS

  • 5+ years of experience in ML performance engineering, GPU kernel optimization, or generative computer vision model compression.
  • Deep proficiency in PyTorch, C++, and CUDA or Triton kernel programming, with proven speedups on large transformer or diffusion architectures.
  • Hands-on experience with diffusion/flow-matching distillation techniques (consistency models, adversarial/distribution-matching distillation, CFG distillation) or low-bit quantization (FP8/INT4/NVFP4).
  • Strong hardware-level intuition for Tensor Cores, shared memory tiling, CUDA graphs, torch.compile/Inductor, and multi-GPU NCCL collectives.
  • Track record shipping production-grade inference code under high-throughput commercial API traffic.

PREFERRED / BONUS EXPERIENCE

  • Publications at CVPR, ICCV, ECCV, NeurIPS, or MLSys, or recognized open-source contributions to diffusers, sglang, TensorRT, FlashAttention, or Triton.
  • Experience optimizing 3D causal video VAEs and real-time streaming video generation.

Position Telemetry

Location Berlin, Germany (Hybrid)
Department AI Infrastructure
Employment type Full-time
Workplace type Hybrid
Salary range
€105k – €150k annual package