janpauldahlke/dsh-slot-health
A live health pane for local LLM endpoints in the dsh web rightbar: llama.cpp slots and Ollama models, sampled once a second and named by fingerprint.
A busy llama.cpp slot or a resident Ollama model reads as a dashboard instead of a raw metrics dump: one rightbar tab, one sample per second, engines named by fingerprint. For llama.cpp that means per-slot busy state, ctx fill, decoded tokens, TTFT and free context plus a server window of prompt and decode tok/s from /metrics; for Ollama it means honest up-states and a MODELS card that lists what the endpoint reports as loaded.
Install
dsh plugin --profile web add github:janpauldahlke/dsh-slot-healthThe GitHub line from the README; the author states it was built and verified against live backends on the same box — llama.cpp on :8080 and Ollama on :11434, including a dual-GPU OLLAMA_SCHED_SPREAD setup with a 35B MoE. The endpoint origin is configured in cordis.patch.yml and defaults to http://127.0.0.1:8080, so watching Ollama means pointing it at :11434.
Compatibility
Supported engines: llama.cpp llama-server (full support), Ollama (honest up-states plus a MODELS card), and anything else reachable, fingerprinted as unknown/stub. For llama.cpp the pane shows per-slot ctx, busy time, decoded tokens, prompt fill, TTFT and free context, plus a server window with prompt and decode tok/s, acceptance and peaks taken from /metrics. Busy state is derived from is_processing only, which is the kind of detail that keeps a loaded slot from reading as idle.
Details
- Repo: janpauldahlke/dsh-slot-health
- Category: Models & Providers
- Stars: 0
- Version: installed from the repo's default branch
Alternatives
WNJXYK/dsh-codex-oauth · tianxia--/dsh-llm-local-token · pumbleblock/dsh-subagent-model-router