Home

Diffusion Language Model Serving Engine

Diffulex is a Diffusion Language Model Serving Engine built on PagedAttention-style runtime primitives. It provides a unified engine for KV cache management, block scheduling, prefix reuse, MoE execution, CUDA graph replay, and model-specific diffusion samplers.

Diffulex is also the runtime engine behind the Multi-Block Diffusion Language Models (MBD-LMs) line of work. Native Block Diffusion LMs perform Single-Block Diffusion (SingleBD): each forward pass refines one noisy block conditioned on a clean cached prefix. This preserves KV caching but leaves blocks sequential, creating a store bubble where the GPU runs a forward that produces no new output. Multi-Block Diffusion (MultiBD) removes this bottleneck by maintaining a bounded running-set of consecutive blocks, enabling decode-store overlap and inter-block parallelism. MBD-LMs are BD-LMs post-trained with Multi-block Teacher Forcing (MultiTF) so the model can handle practical MultiBD running-set states — and Diffulex executes them with an optimized Block Buffer runtime that preserves static input shapes for CUDA Graph replay. In the engine, MultiBD is exposed as decoding_strategy=multi_bd.

For reproducing the MBD-LMs experiments, use the Diffulex mbd-lms branch (CUDA 12). For engine development, open-source contributions, or exploring new decoding algorithms and turning them into runnable systems, use the main branch. main contains ongoing runtime and model-specific optimizations, so its behavior and performance profile may differ from the experiment reproduction branch. The main branch requires CUDA 13.

Where to Start

Goal	Start here
Understand MultiBD in the engine	Multi-Block Diffusion
Install Diffulex and run one command	Quickstart
Set up Python, CUDA, and vLLM dependencies	Installation
Run GSM8K or other lm-eval benchmarks	Benchmark
Start the HTTP server	Server
Tune engine or YAML parameters	Configuration
Use Diffulex as a research backend	Research Engine
Add a model, strategy, or kernel	Developer Guide

Current Scope

Diffulex focuses on cache-aware block-wise dLLM decoding. The main supported runtime pieces are:

• PagedAttention-style KV cache management for diffusion decoding.

• Strategy-specific schedulers and request state.

• Prefix caching for block-causal Multi-Block Diffusion.

• Tensor and data parallel inference paths.

• Optional vLLM-backed common layers and MoE kernels.

• Benchmark and HTTP serving entry points.

For new algorithms, Diffulex main is intended to be a research backend rather than only a benchmark runner. Its Block Buffer, paged KV cache, scheduler, sampler, and Triton kernel boundaries are designed so block-level generation ideas can be implemented as strategy components. See Research Engine for the implementation map.

Model Families

Model family	model_name	Typical strategy	Status
Dream / D2F-Dream	dream	d2f	Supported
DiffuCoder / D2F-DiffuCoder	diffucoder	d2f	Supported
Dream reasoner	dream_reasoner	multi_bd	Supported
Stable-DiffCoder	stable_diffcoder	multi_bd	Supported
LLaDA / D2F-LLaDA	llada	d2f	Supported
Fast-dLLM-v2	fast_dllm_v2	multi_bd or fast_dllm_v2	Supported
SDAR	sdar	multi_bd	Supported
SDAR-MoE	sdar_moe	multi_bd	Supported
LLaDA2 family	llada2, llada2_mini, llada2_moe, llada2dot1_mini, llada2_mini_dmax	multi_bd, dmax, or fast_dllm_v2	Supported
DiffusionGemma	diffusion_gemma	diffusion_gemma	Supported

Use Models for compatibility details before mixing model names, strategies, and sampling modes.