Headline
Cutting LLM Memory by 84%: A Deep Dive into Fused Kernels
Fusing the linear projection and cross-entropy loss into a single tiled Triton kernel cuts peak training memory by 84%, by never materializing the full logits tensor.
May 2026
Learning Triton One Kernel at a Time: Softmax
A numerically-stable, online softmax kernel fuses the max and sum reduction passes into one, cutting global memory reads from three to two while exceeding PyTorch's throughput.
Towards Data Science · Dec 2025
Learning Triton One Kernel at a Time: Matrix Multiplication
An illustrated introduction to tiled matrix multiplication, the memory hierarchy and memory coalescing.
Towards Data Science · Oct 2025
Learning Triton One Kernel at a Time: Vector Addition
A first-principles vector-add kernel in Triton demonstrates the GPU execution hierarchy — threads, warps, blocks, streaming multiprocessors — before any optimization is possible.
Towards Data Science · Sep 2025
Other articles
Reinforcement Learning
| Rainbow: The Colorful Evolution of Deep Q-Networks | Towards Data Science | Jul 2024 |
| Breaking Down State-of-the-Art PPO Implementations in JAX | Towards Data Science | May 2024 |
| A Gentle Introduction to Deep Reinforcement Learning in JAX | Towards Data Science | Nov 2023 |
| Vectorize and Parallelize RL Environments with JAX: Q-learning at the Speed of Light | Towards Data Science | Oct 2023 |
| Temporal-Difference Learning and the Importance of Exploration: An Illustrated Guide | Towards Data Science | Sep 2023 |
LLMs
| Implementing a Transformer Encoder from Scratch with JAX and Haiku | Towards Data Science | Nov 2023 |
Autonomous Driving
| LatentVLA: Latent Reasoning Models for Autonomous Driving | Towards Data Science | May 2026 |
| AlpamayoR1: Large Causal Reasoning Models for Autonomous Driving | Towards Data Science | Feb 2026 |