mirror of
https://github.com/hanzoai/engine.git
synced 2026-11-20 14:17:52 +00:00
Wire the documented-but-unwired engine training path: for llama-family targets, hanzo train now fine-tunes LoRA against the engine's own loaded (dequantized) weights instead of hanzo-train's standalone loader. No new flag — the arch decides. - hanzo-engine: load_llama_base_for_training(id) builds an engine Llama, extracts base_weights_for_training + training_config, and returns a hanzo_train::BaseModel; Ok(None) for non-llama arch (caller falls back). Drop the module's dead_code allow now that the seam is driven. - hanzo-train: split run_sft into run_sft + run_sft_with_client so both the standalone and engine-built TrainingClient share one loop. - hanzo-cli: train branches on the engine base, else standalone.