TASK-DESCRIPTION

• Create: documents/12-model-retention.md
• Create: prompts/tasks/01-tune-llama-service-configs.md
• Inspect: dataflow.in/endpoints/beast-models.json
• Inspect: sources/models/_combine-amd-rx-5500.md

Sort the llama.cpp service configuration on beast and athena before any client-side (DSH) work.

The owner’s ordering decision: llama service config first, then the DSH settings. This task produces the tuning recommendations; it does not apply them to the live hosts.

Inspect the live presets on both hosts (/usr/local/etc/config.ini on beast and athena) and the captured router model lists in dataflow.in/endpoints/.

Create documents/12-model-retention.md with a model-by-model retention analysis:

• Mark each router model as keep, keep-with-tuning, or discard-candidate.
• Discard candidates must be flagged but never deleted (the owner decides); record the exact GGUF location in the HuggingFace cache (HF_HUB_CACHE).
• Base the analysis on execution role and the benchmark numbers in sources/models/_combine-amd-rx-5500.md (imported reference) and sources/models/logs/.
• Cover: triple 1.5B coder entries on beast, 3B coder on beast, Llama-3.2-1B, Mistral-Nemo-12B, the 27B family (Qwen3.5 Q4_K_M/Q4_K_S, Qwen3.8), duplicate quants of the same model, and the 70B/120B cache entries on athena.

Create prompts/tasks/01-tune-llama-service-configs.md (this file) as the executable task and record in it, as concrete tuning recommendations:

• Add alias = <name> lines to each preset section so clients can use short ids (coder32, devstral, r1-14b, gemma-26b, gpt-oss-120b, r1-70b, llama70; athena equivalents).
• Add parallel = 4 under [*] on both hosts so agent clients (DSH subagents) do not serialize.
• Promote useful cache models to presets where they fit the host GPU: on beast, Qwen/Qwen2.5-Coder-14B-Instruct-GGUF:Q4_K_M, unsloth/Qwen3.5-9B-GGUF:Q4_K_M, unsloth/gpt-oss-20b-GGUF:Q4_K_XL, unsloth/gemma-4-12B-it-qat-GGUF:Q4_K_XL; decide the fate of the 27B family.
• Configure an embeddings preset on athena for ggml-org/Nomic-Embed-Text-V2-GGUF:Q8_0 (verify the preset key with llama-server --help | grep -i embed) so /v1/embeddings works for RAG.
• MTP: benchmark unsloth/gpt-oss-120b-GGUF:Q4_K_XL on beast with --spec-type none versus --spec-type draft-mtp (llama-bench), and enable MTP in its preset when it helps. Record the result.
• Do not change the live hosts; the human reviews and applies.

TASK-OUTPUT

Provide a summary: files created, per-host tuning recommendation counts, and the MTP benchmark outcome if run. No commentary beyond the summary.

TASK-CONTEXT

The owner downloaded the first model set by interest, then benchmarked; some of that set is likely not useful. Nothing is discarded without the owner’s decision. Beast runs build 10628 (dev) with --spec-type supporting draft-mtp; athena runs build 10129. The owner’s benchmark shows gpt-oss-120b at 8.42 t/s generation on CPU versus 0.91 t/s for the 70B models — consistent with MTP speculation already active; verify explicitly. Feature requirements: prompts/features/01-endpoints-and-models.md.

TASK-FILES

TASK-VERIFY

• Run: make test from the repository root.
• Expected: exit status 0.

TASK-FEATURES

• prompts/features/01-endpoints-and-models.md

TASK-ACCEPTANCE

• ENDPOINTS-R001
• ENDPOINTS-R002
• ENDPOINTS-R003

OUTPUT: the content of documents/12-model-retention.md, as TASK-OUTPUT describes, and nothing else.