Single message in a chat conversation. ``content`` is a string or list of multimodal parts. Assistant messages with only ``tool_calls`` may s

Manages a llama-server subprocess for GGUF model inference. Lifecycle: 1. load_model(): start llama-server with the GGUF file 2.

OpenAI Responses API request.

Stand-in for the load route: records calls and simulates a load.

OpenAI-compatible chat completion request. Non-OpenAI extension fields are marked with 'x-unsloth'.

Async proxy for OpenAI-compatible external LLM APIs.

Raised when a native path grant is missing, invalid, or unsafe.

Recoverable failure -- caller should fall back (exit code 2).

Training orchestration backend — subprocess-based. Launches a fresh subprocess per job, communicates via mp.Queue.

Small model stub that satisfies MLXTrainer constructor probes.

Manages a llama.cpp embedding subprocess and talks to it over HTTP.

Request to load a model for inference

Request schema for starting training

List with `.put` so worker._send_status can send into it in tests.

Converts llama-server's OpenAI-format streaming chunks into Anthropic SSE. Used for the client-side tool-use pass-through path: the client (e.g.

Converts generate_chat_completion_with_tools() events into Anthropic Messages SSE strings.

Stand-in for ``core.inference.tools.execute_tool``.

Delta content for a streaming chunk.

Wall-clock + memory tracker for a named phase; records into a metrics dict.

Low-rank gradient projector with optional INT4/INT8 quantized projection matrices and layer-adaptive subspace update scheduling. SVD of the g

Response after loading a model

Discovered local model candidate.

ASGI request logger that avoids BaseHTTPMiddleware streaming wrappers.

Per-response ledger for local agentic tool loops.

Minimal model stand-in; routing reads nothing meaningful off it before dispatch.

A subprocess handle whose liveness the test drives directly.

Behaves like an HF Dataset for iteration, len(), and index access.

StreamingResponse without Starlette's legacy AnyIO task-group wrapper.

Run ``request_tool_decision`` in a thread and capture its result.

A unit of extracted text. ``page_number`` is 1-based (None if N/A).

The repo's executable code could not be fully fetched/listed to scan. Raised (not returned empty) so the gate distinguishes "code PRESENT but unr

Small VLM model stub for MLX vision trainer constructor probes.

Per-layer attention metadata. NOTE(djsaunde): Constructed on every forward pass (not once per layer) since it can be invalid across

Per-call info required to run attention.

Generic response for export operations.

Response for listing GGUF quantization variants in a HuggingFace repo.

Register cancel_event in _CANCEL_REGISTRY for the block's duration.

A single SSE chunk in OpenAI streaming format.

A single choice in a non-streaming response.

The assistant's complete response message.

A single GGUF quantization variant in a HuggingFace repo.

Result of a provider connectivity test.

Top-level Responses API response object.

Drop-in zoo collator that validates local video paths on every batch (deduped across batches), applying formatting_func first so formatter-ma