CUDA & AI Performance Engineer for LLM Inference

MWDN Ltd.; Madrid; On-site; EUR 90,000 - 150,000; Full time

Benefits offered by this job

PTO 29 days

10 paid recovery days

Free English classes

HR support

Independent contractor support

Job summary

MWDN Ltd. is seeking a highly technical engineer or researcher with deep expertise in CUDA, GPU programming, and LLM inference optimization to tackle challenging AI system problems and high-performance computing tasks.

Ideal candidates understand GPU architecture from low-level kernels to large-scale inference engines, and enjoy hands-on experimentation, performance profiling, and cross-framework work with vLLM, SGLang, TensorRT-LLM, and related tools.

Qualifications

• Deep hands-on experience with CUDA and GPU programming.
• Strong understanding of GPU architecture, memory hierarchy, parallel execution, and performance tuning.
• Experience with LLM inference, AI systems, GPU kernels, or high-performance computing.

Responsibilities

• Investigate and optimize LLM inference performance on modern GPUs.
• Analyze GPU execution, memory access patterns, kernel performance, and bottlenecks.
• Develop, optimize, debug, or reverse-engineer CUDA kernels and low-level GPU operations.
• Explore optimization techniques for matrix multiplication, attention mechanisms, and neural network layers.
• Benchmark performance and identify opportunities to improve latency, throughput, and GPU utilization.
• Tackle technically challenging problems with strong analytical thinking and hands-on experimentation.

Skills

CUDA programming

GPU architecture

LLM inference

Performance optimization

English communication

Tools

vLLM

SGLang

TensorRT-LLM

Triton

FlashAttention

Job description

MWDN Ltd. is seeking a highly technical engineer or researcher with deep expertise in CUDA, GPU programming, and LLM inference optimization to tackle challenging AI system problems and high-performance computing tasks.

Ideal candidates understand GPU architecture from low-level kernels to large-scale inference engines, and enjoy hands-on experimentation, performance profiling, and cross-framework work with vLLM, SGLang, TensorRT-LLM, and related tools.