For our client - who develops global information and communications technology (ICT) solutions, focusing heavily on software engineering for mobile devices, telecommunications infrastructure, and advanced digital services, we are looking for a world-class compiler and performance optimization expert,
What can we offer:
- On side job in Warsaw, Poland
- Felxible working hours
- Employment based on contract of employment (first probation period for 3 months)
- Long term project
- Work in one of the most advanced research development centers
About the Role:
You would join deep learning infrastructure team and take ownership of building a high-performance Triton compiler and kernel optimization framework, driving the next generation of AI workloads on NPUs. This is a highly technical role that sits at the intersection of AI compilation, NPU programming, and system performance engineering.
Key Responsibilities:
- Lead the design and development of the Triton compiler and performance optimization framework, enabling high-performance operator implementations on NPUs.
- Implement state-of-the-art Triton kernels (e.g., Attention, MatMul, LayerNorm, Conv, Softmax) with best-in-class efficiency.
- Optimize memory access patterns and parallel scheduling, deeply understanding cache behavior, register allocation, and SM occupancy limits.
- Drive end-to-end performance optimization by integrating Triton with framework backends (e.g., PyTorch, XLA) and runtime stacks..
- Research and apply auto-tuning, kernel fusion, and operator scheduling technologies to maximize performance scalability.
- Mentor team members in Triton kernel development and establish standard processes for performance analysis and optimization.
- Stay on top of cutting-edge compiler technologies (MLIR, TVM, Hidet, Cutlass) and introduce innovative ideas to push performance boundaries.
- Conduct performance modeling and workload fingerprinting for large models (LLM, Diffusion, etc.) to guide system-level optimization.
What We’re Looking For:
- Master’s degree or above in Computer Architecture, Compiler Theory, High Performance Computing, or related field; PhD preferred.
- 5+ years of experience in NPU/GPU programming, operator optimization, or compiler development.
- Deep understanding of accelerator architectures and performance bottleneck analysis (compute units, vector lanes, memory hierarchy, caches, etc.)..
- Proficiency in Triton, PTX, or LLVM IR for low-level programming and optimization.
- Familiarity with PyTorch, TensorFlow, or JAX, and their graph execution and operator scheduling mechanisms.
- Proven ability to independently develop, benchmark, and optimize complex kernels.
- Skilled with performance profiling tools (e.g., perf, torch.profiler, and other vendor-neutral or runtime profilers) for quantitative analysis and performance modeling.
- Strong system design and software engineering skills, balancing performance, maintainability, and generality.
- Good english communication skills.
Nice to have:
- Open-source contributions to Triton, LLVM, TVM, MLIR, or PyTorch.
- Experience with AI training or inference systems such as TensorRT, vLLM, DeepSpeed, OneFlow, or OpenXLA.
- Publications or patents in kernel fusion, memory tiling, or async pipeline optimization.
- Experience with distributed inference optimization (tensor/pipeline parallelism, ZeRO, PagedAttention).
- Proven cross-platform optimization experience across different accelerator vendors and architectures.