zeekay 6f13ef421a pkg/miner: pluggable inference-engine backend interface ...

Close three `// TODO: Integrate with actual ...` stubs in
pkg/miner/miner.go (runInference, runChat, runEmbedding) by introducing
a small pluggable backend interface. The miner now dispatches through
backend.InferenceBackend rather than inlining placeholder output.

Two reference backends ship in-tree:

- pkg/miner/backend/noop  — deterministic mock, byte-for-byte identical
  to the pre-refactor inline stubs. Default. Zero config, zero deps.
- pkg/miner/backend/openai — OpenAI-compatible HTTP adapter built on
  net/http. Covers llama.cpp, vllm, ollama, LocalAI, and api.openai.com
  via OPENAI_API_BASE. No new go.mod dependencies.

Config gains Backend / OpenAIBase / OpenAIAPIKey / OpenAIModel /
OpenAIEmbeddingModel fields. Unknown backend values fall back to noop so
operator typos don't crash-loop the miner. Existing miner_test.go passes
unchanged; DefaultConfig() behaviour is preserved.

Also wire a GPUStatsProvider hook on Miner so GetStats can surface
best-effort GPU telemetry (GPUUtilization, MemoryUsed) without pulling
NVML/MLX bindings into the main binary. Default is nil = no-op, matching
current behaviour; a follow-up PR will wire a real provider.

Closes luxfi/ai#1.

Contributed by kcolbchain (https://kcolbchain.com) · https://abhishekkrishna.com

2026-04-21 06:55:21 +05:30

..

openai_test.go

pkg/miner: pluggable inference-engine backend interface

2026-04-21 06:55:21 +05:30

openai.go

pkg/miner: pluggable inference-engine backend interface

2026-04-21 06:55:21 +05:30