Verified Run: OpenAI-Compatible Qwen on Linux Localhost
Date: 2026-06-25
This artifact pack records an inferctl validation against the generic openai_compat adapter path using a real local OpenAI-compatible /v1/models server. The server was llama.cpp serving a small Qwen GGUF model, but inferctl was configured only as kind = "openai_compat" to exercise the generic adapter rather than the dedicated llama.cpp backend.
Result
Status: pass with expected v0.1 warnings
inferctl discovered a localhost OpenAI-compatible provider, patched a clean config, validated the config, inspected backend and model state, selected the configured model for the code task, and produced deterministic triage output.
Environment
- inferctl commit:
d8f812a8d0087fe6cb382ab67d3f352c4fa6547b - inferctl version:
0.2.2-0.20260625174225-d8f812a8d008 - target: Linux amd64
- provider adapter under test:
openai_compat - provider server: llama.cpp
9789 (b3ce5cedf)exposing OpenAI-compatible/v1/models - endpoint:
http://127.0.0.1:8090 - selected route backend:
openai_compat_local - selected route model:
qwen2.5-0.5b-instruct-q4_k_m-openai-compat - source model:
Qwen/Qwen2.5-0.5B-Instruct-GGUF - quantization:
Q4_K_M
Commands Captured
inferctl version --jsoninferctl capabilities --jsoninferctl discover --jsoninferctl discover --kind openai_compat --jsoninferctl discover --kind openai_compat --format tomlinferctl config init --path init.scaffold.toml --jsoninferctl config patch --from-stdin --path inferctl.toml --json < patch.openai_compat.tomlINFERCTL_CONFIG=inferctl.toml inferctl config validate --jsonINFERCTL_CONFIG=inferctl.toml inferctl doctor --jsonINFERCTL_CONFIG=inferctl.toml inferctl backends --jsonINFERCTL_CONFIG=inferctl.toml inferctl models --jsonINFERCTL_CONFIG=inferctl.toml inferctl model qwen2.5-0.5b-instruct-q4_k_m-openai-compat --jsonINFERCTL_CONFIG=inferctl.toml inferctl route code --jsonINFERCTL_CONFIG=inferctl.toml inferctl triage --json
Caveats
- This validates provider workflow behavior and route selection, not model quality.
openai_compatin v0.1 does not support loaded-model inventory.doctorandtriagetherefore reportW_BACKEND_DEGRADED, androute codereportsW_MODEL_NOT_LOADEDeven though the provider endpoint is reachable and the model is listed by/v1/models.- The provider server software path overlaps with the separately validated llama.cpp run; the inferctl adapter path under test here is the generic
openai_compatbackend. config initis captured as the stock scaffold artifact. The final run config starts from a minimal meta/profile config and appliespatch.openai_compat.tomlbecause v0.1 config patching adds and updates keys but does not delete scaffold backends.
Redaction Review
No raw inferctl JSON fields required redaction. The artifacts use localhost provider URLs and a neutral model alias. They do not include private hostnames, home paths, Tailscale addresses, tokens, API key values, or local GGUF paths.