# NVlabs/Fast-dLLM

Official implementation of "Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding"

Repository: https://github.com/NVlabs/Fast-dLLM
Canonical: https://ross.abutalabs.com/products/fast-dllm
Homepage: https://nvlabs.github.io/Fast-dLLM/
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-05-30T10:38:26+00:00

## Health v2 (maintenance only)
Score: 57/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 85, release rhythm 35, longevity 33
- inputs: {"age_days": 463, "days_push": 95, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1082, forks 143 (observed 2026-08-28T04:03:31.022434+00:00)

## What it is
NVIDIA's official implementation of Fast-dLLM, a family of training-free and fine-tuning-based acceleration techniques for diffusion-based large language models, vision-language models, and vision-language-action models. It enables KV cache reuse and confidence-aware parallel decoding to achieve up to 27.6x throughput improvement over standard diffusion LLM inference with minimal accuracy loss.

## Use cases
- accelerate diffusion llm inference
- speed up llada and dream text generation
- enable kv cache for bidirectional diffusion models
- parallel decode multiple tokens in diffusion language models
- convert autoregressive vlms to diffusion vlms
- efficient end-to-end autonomous driving with diffusion models

## When to choose
- you are running diffusion-based LLMs like LLaDA or Dream and need faster inference
- you want to reproduce ICLR 2026 research on diffusion LLM acceleration
- you need block-diffusion VLMs for multimodal or autonomous driving workloads
- you have NVIDIA GPUs and want training-free throughput gains

## When to avoid
- you use standard autoregressive LLMs where conventional inference is already fast
- you need a production serving stack rather than research code
- you lack GPU hardware or work outside the supported model backbones

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, deep-learning, gpu-computing
- domain: large-language-models, artificial-intelligence, deep-learning, gpu-computing, autonomous-vehicles, computer-vision
- platform: python
- tags: diffusion-llm, kv-cache, parallel-decoding, block-diffusion, speculative-decoding, inference-acceleration, vision-language-model, research-code, nvidia, iclr-2026, gpu, linux

## Member repositories
- NVlabs/Fast-dLLM (main) score 57

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:31.022434+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:51:13.550174+00:00, confidence not recorded.
  - readme: https://github.com/NVlabs/Fast-dLLM (fetched 2026-08-28T04:03:31.022434+00:00, sha bb764faf6ff6)
  - homepage: https://nvlabs.github.io/Fast-dLLM/ (fetched 2026-08-29T12:53:33.731054+00:00, sha 78c746cb7c42)
- Data as of 2026-08-30T08:39:29.467469+00:00.