Endpoints Map — verified 2026-08-26

Live-state facts captured from GET /v1/models on both routers (last refreshed 2026-09-03). Captured responses: dataflow.in/endpoints/beast-models.json, dataflow.in/endpoints/athena-models.json.

Inference Hosts

Host Role Hardware RAM Backend
beast.lan (beast.lan) heavy server AMD Instinct MI25, 16 GB VRAM (Vulkan, GGML_VK_VISIBLE_DEVICES=1) 256 GB (dual Xeon, NUMA) llama.cpp router
athena.lan (athena.lan) desktop (this machine) AMD RX 5500 XT, 8 GB VRAM 128 GB (Ryzen 9 5900X, 12 threads) llama.cpp router
minerva.lan (minerva.lan) travel laptop unverified unverified none yet (planned)

Endpoints

Endpoint Base URL Default context Auth
beast GPU router http://beast.lan:2001/v1 32768 none configured (--api-key not set)
beast CPU router http://beast.lan:2002/v1 65536 (gpt-oss-120b) none configured
athena router http://athena.lan:2001/v1 8192 none configured
DSH web GUI (athena) http://127.0.0.1:3080 — —

Both routers answer OpenAI-compatible chat completions with --tools all. Model ids are preset section names or hf-repo cache names; the routers list 6 (beast GPU :2001), 1 (beast CPU :2002) and 17 (athena) models as of the 2026-09-03 capture.) with the preset file at /usr/local/etc/config.ini on each host (to confirm; the imported copies in sources/models/config-*.ini are older snapshots). Each model entry in the capture carries its full child-server command line (status.args) and the generated preset text, including ctx-size, ngl, and --alias.

Request Flow

  1. A client (DSH, aider, llama.vscode, curl) sends an OpenAI-compatible request to a router base URL with a model id.
  2. The router loads the matching child server (or uses the cached one) and forwards the request.
  3. GPU models (no explicit ngl; fit trims layers to VRAM, excess on CPU) run on the host GPU; CPU models (ngl 0) run in system RAM with numa = distribute on beast.
  4. The router returns the OpenAI-formatted streamed response.

Concurrency

Router mode runs one child server per active model. A CPU-heavy model (for example gpt-oss-120b) and a GPU model can be active concurrently, which is the basis of the routing policy in documents/10-routing-policy.md. Each child serves one request at a time unless parallel is set in the preset.