Record: Tuning applied — 2026-08-26

Outcome

Applied the task-01 tuning recommendations from documents/14-service-tuning-recommendations.md to both hosts and verified.

What changed

Beast (/usr/local/etc/config.ini, baseline backed up as .bak-20260826):

• [*]: parallel = 4.
• Aliases on all 11 presets (coder32, coder14, qwen9, gpt-oss-20b, gemma-12b, devstral, r1-14b, gemma-26b, gpt-oss-120b, r1-70b, llama70).
• Four cache models promoted to GPU presets (ngl = 99): bartowski/Qwen2.5-Coder-14B-Instruct-GGUF:Q4_K_M, unsloth/Qwen3.5-9B-GGUF:Q4_K_M, unsloth/gpt-oss-20b-GGUF:UD-Q4_K_XL, unsloth/gemma-4-12B-it-qat-GGUF:Q4_K_XL.

Athena (/usr/local/etc/config.ini, baseline backed up as .bak-20260826):

• [*]: parallel = 4.
• Aliases on the presets (coder15, coder3, llama3, gemma-e4, coder32-fallback, r1-32b-fallback).
• New embeddings preset [ggml-org/Nomic-Embed-Text-V2-GGUF:Q8_0] with embeddings = on, alias nomic-embed.

Verification (all 2026-08-26)

• GET /v1/models on both routers shows the aliases; child args include --parallel 4 (beast checked).
• Chat completion routed through the coder32 alias answered correctly on beast.
• POST /v1/embeddings with model nomic-embed on athena returned a 768-dim vector — the RAG endpoint is live.
• Captures refreshed (dataflow.in/endpoints/, 13 beast / 15 athena); make test passes.

Commits

• (pending) tuning applied + applied-config archives.