Research
Research notes
Deeper technical explorations. Notes are marked 🌱 seedling 🌿 growing 🌳 evergreen to signal maturity.
- 🌳 Validating Numerical Precision for Low-Precision Ops Methods and tools to detect and validate numerical precision issues when implementing math ops in low-precision types like bfloat16.
- 🌳 Numerical Precision Challenges in ATen Operations Precision-loss pitfalls in low-precision (bf16/fp16) PyTorch ATen ops — catastrophic cancellation, absorption, ULP, and how to reason about them.
- 🌿 LLM Inference Optimization Where the time and memory actually go when serving large language models.
- 🌱 Barrett Reduction Replacing division by a constant modulus with a multiply and a shift.
- 🌱 Knuth Algorithm D Schoolbook long division on machine-word limbs, done carefully.
- 🌿 SIMD Integer Arithmetic Why integer division resists vectorization and how to work around it.