GPU Tooling · Inference
An Easier Way to Diagnose flash-attn Installation Problems
A practical guide to separating FlashAttention wheel, PyTorch, CUDA, platform, and GPU compatibility problems.
Archive
Writing
Personal and practical notes from my day-to-day work.
Transformers · Inference
A Simple Utility for Context Caching with Hugging Face
Build a reusable prefix cache for independent Transformers requests, verify token boundaries, and check cached output against an uncached baseline.
African NLP · Tokenization
Encoding, Tokenization, and African Languages: Why UTF-8 Matters
How Unicode encoding, normalization, and byte-level tokenization interact across Swahili, Yorùbá, and Amharic text.