Marcos Fernandez marcosfernandez.ai Book a call

I measure inference systems and write down what held.

Hands-on investigations of GPU serving, cache reuse across workers, and autonomous development under engineering control, run on hardware I own. Every write-up names the run it came from.

Book a free 30-minute call Free, thirty minutes, for questions about local inference, cache tiers and autonomous development. You pick the slot; the invite arrives with a Google Meet link and I read your question before we talk.
Latest measured write-upBuilding a 2-million-token KV cache across eight AMD R9700s. 25 September 2026. Restoring a roughly 100K-token prefix from the shared tier took about 12.6 seconds, against about 89 seconds to compute it cold; 64 of 64 cross-GPU restores succeeded. Newest commentary: Anthropic reinvents source control, 2 November.
Schematic of the restoration test, September 2026 run

Writing, newest first.

  1. Anthropic reinvents source control. Software engineering would like a word.Thirty thousand changes a day and a faster merge; what modular engineering was for.
    commentary, 2 Nov 2026
  2. The cloud is becoming a fleetIf compute goes to orbit, storage and networking follow it.
    commentary, 24 Oct 2026
  3. Building a 2-million-token KV cache across eight AMD R9700sThree weeks from four GPUs that could not see each other to a cache tier that outlives the worker.
    run, 25 Sep 2026
  4. August 2026: local AI reached a new milestoneThe gap between hosted and open-weight models is now a deployment decision, not a hobby.
    benchmarks, 1 Sep 2026
  5. Autonomous development patterns: from code generation to controlled engineeringWhat has to be true before an agent's output is allowed to count.
    method, 1 Jun 2026
How entries are taggedA run carries a configuration and a result. Benchmarks compare published numbers. Commentary argues a position. Method describes a way of working.

The current investigation: inference state that survives the worker.

Eight independent GPU lanes, a shared cache tier, and the question of whether long-context attention state can be restored by a worker other than the one that produced it. The write-up covers PCIe BAR mapping, peer-to-peer paths, attention backends and the restoration test, with the exact configuration beside every result.

The runEight AMD Radeon AI PRO R9700s, one host, vLLM 0.29.0 serving RedHatAI/Qwen3.8-27B-INT4 with FP8 KV, a shared object store for the cache. September 2026. Restoring a roughly 100K-token prefix from the shared tier took about 12.6 seconds, against about 89 seconds to compute it cold; 64 of 64 cross-GPU restores succeeded. Read the full run.

If you are weighing local inference, a cache tier, or how far to trust autonomous development, bring the question.

Thirty minutes, free. No deck and nothing to buy. We look at your constraint and you leave with a direction and, where I have one, the run that supports it.

Book a free 30-minute call
What happens after you bookYou pick a slot, the invite arrives with a Google Meet link, and I read your question before we talk.