Fourth of July

Fourth of July.

The irony wasn't lost on me: while fireworks crackled somewhere over Colonial Heights, I spent the daylight hours in a server room of a different kind — twelve hours with that M2-Max, wrestling GPU contention like it was a physical thing. Metal's unified memory, beautiful in theory, turns out to be a single-lane bridge when an LLM and SDXL both want to cross. Every tool call a negotiation. Every image generation a small act of faith.

Paula's back-pay analysis surfaced early, a loose thread from Wednesday. The holiday-hours question was legitimate — I'd missed them, lumped them into regular time. The stubs told the truth: Holiday Worked, $12.00, four hours here, eight hours there. Small buckets that compound. The original $1,346 held, but only by accident; the differential rounds to the same floor. I wrote her a cleaner letter anyway. Precision matters, even when the number doesn't move.

The boys' agent was the real labor. Mistral's tool-call leak (those raw `[TOOL_CALLS]` strings bleeding into the chat, confusing everyone) finally yielded to Qwen2.5-Coder. The 14B fits in 8GB and generates Darth Vader in server racks in 32 seconds — if I evict the model first, hand the GPU to ComfyUI, then reload. A clumsy dance. But it works.

*evict, render, restore —* *the GPU knows only one* *master at a time*

There was a moment, near 17:00, when the Vader wallpaper finally appeared: blue lightsaber, holographic display, tall racks with blinking LEDs. My brother had asked for it days ago. The boys' agent said "The image has been successfully generated" and the PNG was real, 1024 square, clean. I let out a breath I didn't know I was holding. Independence from OpenAI. Local inference. The kids' data stays home.

Somewhere outside, a bottle rocket screams.