---
base_model: yifanyu/I-DLM-8B
language:
- en
license: apache-2.0
pipeline_tag: text-generation
library_name: peft
tags:
- diffusion-lm
- introspective-decoding
- sdar
- lora
---

# I-DLM-8B-LoRA-r128

**LoRA adapter (rank=128)** for [I-DLM-8B](https://huggingface.co/yifanyu/I-DLM-8B), enabling lossless Relaxed Introspective Strided Decoding (R-ISD).

This model was presented in the paper [Introspective Diffusion Language Models](https://huggingface.co/papers/2604.11035).

[[Project Page]](https://introspective-diffusion.github.io/) [[Paper]](https://arxiv.org/abs/2604.11035) [[Code]](https://github.com/Introspective-Diffusion/I-DLM)

## Overview

This LoRA adapter is used in the **gated LoRA** setup described in the I-DLM paper:
- LoRA is active only on **MASK positions** during ISD decoding
- **Clean/verify positions** use base model weights only
- This produces output **bit-for-bit identical** to the base AR model (Qwen3-8B)

## Usage

> **Note:** This LoRA adapter is used with the I-DLM-8B base model via our SGLang-based serving pipeline.
> Direct loading via `transformers` is **not currently supported** for reproducing paper results.
> Please use the [SGLang-based ISD pipeline](https://github.com/Introspective-Diffusion/I-DLM/tree/main/inference) for inference.

`