Crate infraqueue_language_model 

Structs

CandleRunParams

Parameters for local Candle generation.

TrainExample

TrainParams

TrainResult

Functions

build_chat_messages

Build OpenAI-compatible chat messages from optional system and required user strings.

build_fallback_prompt

Build a combined prompt string as a fallback for backends that expect a single prompt. If sys is empty, returns user as-is; otherwise formats as system+user sections.

generate_local_candle

Run a minimal local Candle generation using HF Hub llama-family checkpoints. Returns the generated assistant text (without the prompt).

preload_local_candle

Optionally preload the model so the first request is fast.

train_logit_bias

Train a persistent global logit-bias adapter from prompt/assistant examples. This does not modify base model weights; instead it learns a bias vector (size=vocab) that is added to the logits during generation. The bias is saved to ADAPTER_DIR and automatically loaded on startup and applied for all subsequent generations.