Module auto_dispatch 

Auto-dispatch: pick a registered model runner from a weights path.

auto_runner_name(path) resolves the path (file or directory), sniffs the model family (GGUF general.architecture for .gguf, sidecar config.json model_type for safetensors), and maps it to the short runner name a callsite registered with register_cli (e.g. "qwen3", "gemma").

auto_dispatch(path, args) is a one-shot: sniff, look up, run.

Used by skill so callers don’t need to hardcode Qwen3Runner vs GemmaRunner per family.

Structs

SniffedRunner

Result of sniffing a weights path.

UnimplementedArch

A catalog arch that RLX recognizes but has not yet implemented a runner for. Returned by known_unimplemented_arch so error messages can point at the PLAN.md milestone that unblocks the family.

Enums

SniffedFrom

Source the sniffer used to identify the model family.

Functions

arch_runner_name

Map a GGUF general.architecture tag to the short runner name.

auto_dispatch

Sniff path, look up its runner in the registry, and run it with args.

auto_runner_name

Sniff path and return only the runner short name.

auto_sniff

Resolve path to a single weight file, then sniff the runner.

known_unimplemented_arch

Look up an arch / model_type in the unimplemented-families table.

known_unimplemented_keys

Snapshot of every (key, family) pair currently in the unimplemented table — useful for rlx-run check --list-unimplemented style tooling.

model_type_runner_name

Map an HF config.json model_type value to a short runner name.

run_auto

Entry point for an rlx-run auto WEIGHTS [args...] subcommand.