Get a bound structured logger for a module (name is usually __name__).

Pull matching lines as evidence snippets (``max_matches=0`` means all). Records every matching line in full, not a truncated sample, so an extra

Drive a single stream and return the captured outbound request body.

Fetch a file from GitHub raw. None on 404; skips on transient network errors.

Return a per-turn USD cost breakdown (per-bucket + total). Unknown model -> ``priced`` False and USD fields 0.0 (token counts still report).

Receive new token IDs from model.generate().

Update training progress and notify callbacks

Open studio.db with WAL mode, create tables once per process, enable foreign keys.

Return the transformers tier required for *model_name*. Returns ``"510"`` for models needing transformers 5.10.x (Gemma 4 Unified), ``"550"``

Detect VLMs via the config architecture (works for fine-tunes); transformers-5.x models are checked in a .venv_t5/ subprocess. Cached per (model_n

Grep for `class Name`, `def name`, or `Name = ...` at any indent (no ast.parse).

No-op — the underlying client is shared across requests.

Embedding width, probed via a 1-text encode and cached per model (_resolve_model_path clears it when the effective repo changes). Unlo

Extract a user-friendly message from known llama-server errors.

Import unsloth and skip when the current platform is not using MLX.

Log ``error`` in full server-side and return an ``HTTPException`` whose ``detail`` is only ``public_message`` -- never the raw exception text.