Start with 16 MB and grow on demand. An experimental llama.cpp patch that reduced upfront KV allocation and avoided GPU OOM on Apple Silicon.

I tried to use macOS demand paging to save GPU memory in llama.cpp. It didn't work — and understanding why led to a better approach.

A technical analysis of llama.cpp's KV cache architecture, the llama_memory_i interface, and how PagedAttention can be implemented as a composable new memory strategy.

What happens when you hold GPU load constant and only increase memory pressure? I designed an experiment to isolate swap-induced slowdown from GPU contention on a 32GB Mac Mini.

Sweeping parallelism from 1 to 10 concurrent requests on a 30B model to find exactly where performance falls off a cliff — and why KV cache pre-allocation is the culprit.

Switching ollacode's system prompt from Korean to English and measuring real performance gains with ollama-bench. 60% faster TTFT, 55% fewer input tokens.

I built ollama-bench to measure local LLM performance degradation — token generation speed, prefill speed, TTFT, and memory usage over progressive requests.

Day 2 of ollacode development — optimizing token usage for local LLMs with English prompts, context compaction, and smart history management.

Day 1 of building ollacode — a lightweight CLI coding assistant powered by Ollama's qwen3-coder:30b model with Telegram bot integration.

Key findings from OpenZeppelin's audit of the EIP-4337 Account Abstraction implementation — deposit manipulation, silent token transfer failures, and the dangers of Solidity's transfer().