rlx-cli

Shared CLI plumbing for the per-model rlx-<family> binaries (and the optional rlx-run multiplexer). It centralizes the argument parsing, device selection, chat templating, output formatting, and weight-inspection helpers so each model crate's binary stays thin and consistent.

Modules

• args / lm_args — common flag parsing (--weights, --device, --prompt, sampling, …).
• device — parse_standard_device(...) maps cpu|metal|mlx|cuda|wgpu|vulkan|ane strings to rlx_runtime::Device.
• chat — chat-template application.
• format — token/latency/RTF output helpers.
• auto_dispatch — registry that routes rlx-run <family> … to the right model runner.
• inspect — GGUF/safetensors introspection (also exposed via the rlx-inspect binary).

Binary

# Inspect a checkpoint's tensors / metadata
cargo run -p rlx-cli --bin rlx-inspect -- <weights.gguf|model.safetensors>

How it fits

Consumed by every model runner crate (e.g. rlx-llama32, rlx-qwen3, rlx-whisper, rlx-phi) for their command-line front-ends.