rlx-models

The umbrella facade crate for the RLX model zoo. It re-exports the shared config/weight/graph plumbing (rlx_core, rlx_flow) and, behind cargo features, the per-family model modules (qwen3, gemma, whisper, bert, nomic, vision, …). Depend on this crate to reach several model families through one dependency — or depend on a specific rlx-<family> member directly when you only need one.

Features

Model modules are feature-gated and off by default — the default features (qwen35-tokenizer, flux2-tokenizer, flux2-image, embed) do not pull in the model modules. Enable what you need:

cargo build -p rlx-models --features "qwen3,qwen35,gemma"

Usage

// with --features embed
use rlx_models::embed::{Pooling, RlxBertModel, BertTokenizer, embed_with_rlx};

let tok = BertTokenizer::from_dir(model_dir, 512)?;
let mut model = RlxBertModel::load(&config, &weights)?;
let vecs = embed_with_rlx(&mut model, &tok, &["hello", "world"], Pooling::Mean)?;
# anyhow::Ok(())

Each gated module (rlx_models::qwen3, ::gemma, ::whisper, …) simply forwards to the corresponding member crate.

Binary

# rlx-run: multiplexer dispatching to per-family runners
cargo run -p rlx-models --bin rlx-run -- <family> [args…]

How it fits

Sits on top of every model member — e.g. rlx-qwen3, rlx-gemma, rlx-whisper, rlx-embed, rlx-sam — plus rlx-cli for the command-line front-ends.