LoRA and X-LoRA adapters

LoRA (Low-Rank Adaptation) adapters add task-specific fine-tuning on top of a base model without modifying the base weights. X-LoRA loads several adapters at once and lets the model select among them per request.

Loading a LoRA

mistral.rs reads adapter_config.json from the LoRA repo for the targeted modules and rank. Pass multiple adapters as a semicolon-separated list.

CLI

mistralrs run -m <base-model> --lora <lora-repo>

Multiple adapters:

mistralrs run -m <base-model> --lora "<lora-repo-1>;<lora-repo-2>"

Python

from mistralrs import Runner, Which

runner = Runner(
    which=Which.Lora(
        model_id="<base-model>",
        adapter_model_ids=["<lora-repo-1>", "<lora-repo-2>"],
    )
)

Rust

use mistralrs::{LoraModelBuilder, TextModelBuilder};

let model = LoraModelBuilder::from_text_model_builder(
    TextModelBuilder::new("<base-model>"),
    vec!["<lora-repo-1>", "<lora-repo-2>"],
)
.build()
.await?;

Full examples: lora-zephyr (Python), lora (Rust).

X-LoRA

X-LoRA loads multiple adapters with a learned scaling head that selects per-token weighting. The ordering file maps adapters to the scaler's output positions.

CLI

mistralrs run \
  -m <base-model> \
  --xlora <xlora-repo> \
  --xlora-order <ordering-file.json>

Flag rules:

• --xlora conflicts with --lora.
• --xlora-order and --tgt-non-granular-index are only valid alongside --xlora.
• --tgt-non-granular-index <n> controls how often the X-LoRA scaler recomputes. Without it, the scaler recomputes every token.

Python

from mistralrs import Runner, Which

runner = Runner(
    which=Which.XLora(
        model_id="<base-model>",
        xlora_model_id="<xlora-repo>",
        order="<ordering-file.json>",
        # tgt_non_granular_index=...,
    )
)

Rust

use std::fs::File;
use mistralrs::{XLoraModelBuilder, TextModelBuilder};

let model = XLoraModelBuilder::from_text_model_builder(
    TextModelBuilder::new("<base-model>"),
    "<xlora-repo>",
    serde_json::from_reader(File::open("<ordering-file.json>")?)?,
)
.build()
.await?;

Full examples: xlora-zephyr (Python), xlora (Rust).

AnyMoE

AnyMoE goes a step further than adapters: it composes several fine-tunes of the same base model into a MoE (Mixture of Experts) configuration at inference time, training only a small per-layer router.

• It is exposed through the Rust SDK (AnyMoeModelBuilder) and the Python SDK (AnyMoeConfig, AnyMoeExpertType); it is not configurable via the CLI.
• Expert checkpoints must share the base model architecture, and a small JSON calibration dataset is required to train the router.
• The AnyMoeConfig docstrings in the AnyMoE Python reference cover finding the prefix/mlp values from a model's model.safetensors.index.json.

Full examples: anymoe (Python), anymoe and anymoe-lora (Rust).