Skip to content
Quantoros

Mentoring · GPU & ML Performance

GPU programming and ML performance mentoring

Write and profile your own GPU kernels and make models faster: from the first vector-add kernel to a custom PyTorch operator in C++/CUDA with autograd and tests.

Who this is for

  • ML engineers who want to understand and remove performance bottlenecks.
  • C++ and systems developers moving into ML infrastructure and inference.
  • Quant developers who need fast numerical code on GPUs.

What you will build

The GPU & ML Performance track has 8 sprints. It opens after Deep Learning on the ML path or after Quant Development.

  • The first GPU kernel: vector addition.
  • Operator fusion: a softmax kernel.
  • Softmax occupancy and throughput.
  • Tiled matrix multiplication.
  • Block ordering and matmul autotuning.
  • Layer normalisation forward pass.
  • A PyTorch operator in C++ and CUDA.
  • Testing and autograd for your own operator.

How mentoring works

  • A free intro call to check your C++, Python and deep learning background.
  • Weekly 1:1 calls and review of kernels, benchmarks and profiles.
  • Portfolio projects you can show in ML infrastructure interviews.

Questions

Do I need my own GPU?

No. Free GPUs in Colab and Kaggle are enough for the track.

CUDA or Triton?

Both: Triton for productive kernels, C++/CUDA for custom PyTorch operators.

What background do I need?

Python, basic C++ and an understanding of how neural networks train.

Read next

Mentoring directions