Bringing Nunchaku 4-bit Diffusion Inference to Diffusers

Hugging Face has integrated Nunchaku 4-bit diffusion inference natively into Diffusers, enabling users to load quantized checkpoints with a simple from_pretrained() call and no local CUDA compilation. The SVDQuant-based Nunchaku engine runs transformer layers with 4-bit weights and activations (W4A4), reducing memory usage from about 24 GB to 12 GB and generating a 1024x1024 image in about 1.7 seconds on an RTX 5090. NVFP4 checkpoints require NVIDIA Blackwell GPUs, while INT4 variants support earlier generations.

Text-to-Image • 8B • Updated • 39 • 2 Bringing Nunchaku 4-bit Diffusion Inference to Diffusers Update on GitHub https://github.com/huggingface/blog/blob/main/nunchaku-diffusers.md Exploring Quantization Backends in Diffusers https://huggingface.co/blog/diffusers-quantization . Most of these backends are weight-only . This means that they store the weights in low precision and dequantize them back to high precision at compute time. This reduces memory usage significantly, but it usually does not make inference faster, and can even add a small latency overhead. SVDQuant https://arxiv.org/abs/2411.05007 , the quantization method behind the popular Nunchaku https://github.com/nunchaku-tech/nunchaku inference engine, takes a different approach. It runs the main transformer layers with 4-bit weights and activations W4A4 , reducing memory while also speeding up the denoising loop. The details are covered below, but until now, using these checkpoints required a separate inference library. With current Diffusers, loading a Nunchaku checkpoint is as simple as calling from pretrained , with no local CUDA compilation required thanks to the kernels https://github.com/huggingface/kernels package. In addition, the companion diffuse-compressor https://github.com/rootonchair/diffuse-compressor toolkit lets you quantize new architectures yourself and publish them as regular Diffusers repositories. Table of Contents Getting started with Nunchaku Lite getting-started-with-nunchaku-lite Background: SVDQuant and Nunchaku background-svdquant-and-nunchaku Introducing Nunchaku Lite introducing-nunchaku-lite Native loading in Diffusers native-loading-in-diffusers Getting more speed and lower memory getting-more-speed-and-lower-memory Benchmarks benchmarks Quantizing your own model quantizing-your-own-model Ready-to-use checkpoints ready-to-use-checkpoints Conclusion conclusion Acknowledgements acknowledgements Getting started with Nunchaku Lite First, install the requirements. You need a recent version of Diffusers and the Hugging Face kernels package: pip install -U diffusers transformers accelerate kernels bitsandbytes Then load a pre-quantized pipeline like any other Diffusers model: python import torch from diffusers import ErnieImagePipeline pipe = ErnieImagePipeline.from pretrained "lite-infer/ERNIE-Image-Turbo-nunchaku-lite-nvfp4 r32-bnb4-text-encoder", torch dtype=torch.bfloat16, .to "cuda" image = pipe prompt="A cinematic portrait of a red fox in a misty forest at sunrise, " "detailed fur, volumetric light", height=1024, width=1024, num inference steps=8, guidance scale=1.0, generator=torch.Generator "cuda" .manual seed 42 , .images 0 image.save "output.png" No custom pipeline class or separate inference engine is needed, and there is nothing to compile locally. The NVFP4 kernels are downloaded from the Hub through the Nunchaku Lite kernels page https://huggingface.co/kernels/rootonchair/nunchaku-lite-kernels the first time they are used. This checkpoint pairs a Nunchaku NVFP4 transformer with a bitsandbytes NF4 text encoder, and generates a 1024x1024 image in about 1.7 seconds on an RTX 5090 with a peak memory usage of about 12 GB, compared with about 24 GB for the BF16 pipeline. You can find more details about the Nunchaku Lite checkpoint format in the official Diffusers documentation https://huggingface.co/docs/diffusers/main/en/quantization/nunchaku . NVFP4 checkpoints require an NVIDIA Blackwell GPU RTX 50 series, RTX PRO 6000, B200 . For earlier generations, use the INT4 variants. See the hardware support table below for details. Background: SVDQuant and Nunchaku SVDQuant is the quantization method behind Nunchaku , its reference CUDA inference engine. Standard 4-bit quantization is difficult for diffusion transformers because both weights and activations contain large outliers. SVDQuant handles this by moving activation outliers into the weights, representing the hardest part of each weight matrix with a small 16-bit low-rank branch, and quantizing the remaining residual to 4 bits. Nunchaku makes this fast with fused kernels for the 4-bit path and the low-rank branch. Introducing Nunchaku Lite The original Nunchaku engine https://github.com/nunchaku-ai/nunchaku gets much of its speed from model-specific fused execution paths quantizing-models-with-structural-rewrites , such as fused QKV projections and fused GELU/MLP kernels. Those optimizations are tied to each architecture's module layout and checkpoint format, so supporting a new model family usually requires model-specific integration work. Nunchaku Lite is the new integration path in Diffusers. With it, Diffusers can load Nunchaku-style checkpoints without a custom pipeline or a separate inference engine. Under the hood, Nunchaku Lite patches the relevant nn.Linear modules of a stock Diffusers model with runtime SVDQ/AWQ linear layers before the checkpoint is loaded. The CUDA kernels come from the Hub through the kernels package. Two kernel families are used: : 4-bit weights and activations with the SVDQuant low-rank correction. This layer is used for the transformer's attention and MLP projections, where nearly all of the compute is spent, and is available in INT4 and NVFP4 variants. svdq w4a4 : 4-bit weights with 16-bit activations, used for adaptive normalization and modulation projections such as FLUX awq w4a16 adanorm single / adanorm zero or Qwen-Image modulation layers. These layers are memory-bound and precision-sensitive, making AWQ a good fit to preserve precision while still saving memory and space. The trade-off is that, without architecture-specific fused kernels and modules, Nunchaku Lite cannot match the speedup of the original Nunchaku engine. However, the bare-bones implementation still delivers around 30% speedup while retaining the same level of VRAM reduction . Native loading in Diffusers If you have used bitsandbytes or torchao in Diffusers, the mechanics will feel familiar. A Nunchaku Lite model repository is an ordinary Diffusers repository. The only special part is a quantization config block inside the transformer's config.json : "quantization config": { "quant method": "nunchaku lite", "compute dtype": "bfloat16", "svdq w4a4": { "precision": "nvfp4", "group size": 16, "rank": 32, "targets": "layers.0.self attention.to q", "layers.0.self attention.to k", "..." }, "awq w4a16": { "precision": "int4", "group size": 64, "targets": "adaLN modulation.1", "..." } } This config tells Diffusers which modules were quantized, which scheme they use, and which Nunchaku Lite runtime layer to instantiate SVDQW4A4Linear or AWQW4A16Linear . Because the quantized model keeps the exact module structure of the dense one, everything downstream schedulers, LoRA loading hooks, offloading, torch.compile sees a normal Diffusers model. Hardware support Nunchaku Lite uses different kernel variants depending on the GPU generation and checkpoint precision: | Scheme | Precision | Supported GPUs | |---|---|---| svdq w4a4 | nvfp4 | Blackwell RTX 50 series, RTX PRO 6000, B200 | svdq w4a4 | int4 | Turing / Ampere / Ada RTX 30 & 40 series, A100, L40S | awq w4a16 | int4 | Turing / Ampere / Ada RTX 30 & 40 series, A100, L40S | Volta and Hopper GPUs are currently not supported by the 4-bit kernels. The quantizer validates the GPU's CUDA capability at load time and raises a clear error instead of producing incorrect outputs. Getting more speed and lower memory Nunchaku Lite can be combined with other Diffusers memory and speed optimizations. torch.compile. Compiling the transformer improves the end-to-end speedup from 1.35x to 1.8x: pipe.transformer.compile fullgraph=True or compile repeated blocks for faster compilation pipe.transformer.compile repeated blocks fullgraph=True Quantized text encoders. The transformer is not the only component with a large memory footprint. Text encoders such as T5 or Qwen3 can occupy several gigabytes on their own. Further quantizing the text encoder with bitsandbytes NF4 reduces peak VRAM by about 22% in our benchmark. Offloading. Diffusers offloading helpers such as enable model cpu offload and enable sequential cpu offload work as usual if you need to fit the pipeline onto a smaller GPU. Benchmarks All numbers below were measured on an NVIDIA RTX PRO 6000 Blackwell at 1024x1024 using rootonchair/ERNIE-Image-Turbo-nunchaku-lite-int4-bnb4-text-encoder https://huggingface.co/rootonchair/ERNIE-Image-Turbo-nunchaku-lite-int4-bnb4-text-encoder . End-to-end latency and memory | Configuration | Full pipeline | Denoise loop | Peak VRAM | Speedup | |---|---|---|---|---| | BF16 baseline | 3.00 s | 2.86 s | 31.1 GB | 1.0x | | Nunchaku Lite NVFP4 | 2.27 s | 2.13 s | 20.6 GB | 1.35x | Nunchaku Lite NVFP4 + torch.compile | 1.68 s | 1.53 s | 20.6 GB | 1.8x | | Nunchaku Lite NVFP4 + NF4 text encoder | 2.29 s | 2.13 s | 16.0 GB | 1.35x | As shown above, Nunchaku reduces peak VRAM by up to 50% while still improving latency by roughly 30%. The remaining overhead comes largely from extra kernel launches, which torch.compile can mitigate, bringing the full pipeline down to 1.68 s, or 1.8x faster than the BF16 baseline. Image quality Quantizing your own model Nunchaku Lite support in Diffusers is architecture-agnostic, and the diffuse-compressor https://github.com/rootonchair/diffuse-compressor toolkit provides an end-to-end SVDQuant workflow for Diffusers models: calibrate, quantize, package, and publish. Below, we walk through quantizing FLUX.2 Klein 4B as an example. It covers the main steps: inspect the model, calibrate and quantize the transformer, package the result as a Diffusers pipeline, then verify and push it to the Hub. The full tutorial https://github.com/rootonchair/diffuse-compressor/blob/main/docs/quantize new hf model.md covers every flag in detail. 1. Inspect what will be quantized The generic scanner walks the model and decides what to target: compatible linears inside the repeated transformer-block stack become SVDQ W4A4 targets, recognized modulation linears become AWQ W4A16 targets, and everything else stays dense. python examples/text to image/quantize hf.py black-forest-labs/FLUX.2-klein-4B \ --precision int4 --rank 32 --inspect-config Always read this report before quantizing. For FLUX.2 Klein 4B, the expected result is 100 SVDQ targets, 3 AWQ targets, and 6 dense outer linears, with no missing patterns or duplicate names. 2. Run quantization The following command runs SVDQuant on the transformer and writes the quantized checkpoint to outputs/checkpoints/svdq-int4 r32-flux-2-klein-4b.safetensors : python examples/text to image/quantize hf.py black-forest-labs/FLUX.2-klein-4B \ --precision int4 \ --output outputs/checkpoints/svdq-int4 r32-flux-2-klein-4b.safetensors Replace --precision int4 with nvfp4 to build Blackwell-native weights. 3. Package a Diffusers pipeline The converter combines the quantized transformer with the base pipeline's other components, writes the compact nunchaku lite configuration into transformer/config.json , and can optionally convert text encoders to NF4: python examples/convert nunchaku lite diffusers.py \ --checkpoint outputs/checkpoints/svdq-int4 r32-flux-2-klein-4b.safetensors \ --model-id black-forest-labs/FLUX.2-klein-4B \ --bnb4-text-encoder text encoder \ --compute-dtype bfloat16 \ --output-dir outputs/diffusers/FLUX.2-klein-4B-nunchaku-lite-int4-bnb4-text-encoder 4. Load, verify, and push to the Hub python import torch from diffusers import DiffusionPipeline pipe = DiffusionPipeline.from pretrained "outputs/diffusers/FLUX.2-klein-4B-nunchaku-lite-int4-bnb4-text-encoder", device map="cuda", image = pipe "A glass robot in a greenhouse, cinematic lighting", num inference steps=4, guidance scale=1.0, generator=torch.Generator "cuda" .manual seed 12345 , .images 0 Once the outputs look good, run pipe.push to hub "your-name/your-model-nunchaku-lite-int4" . Other users can then load it with the same from pretrained pattern shown above. Quantizing models with structural rewrites Note that the generic path assumes the architecture can be quantized without structural rewrites. For additional speedup, the original Nunchaku engine rewrites groups of Diffusers layers as fused modules. The generic path cannot infer these changes on its own, such as combining separate Q, K, and V projections into one module or splitting a fused projection across several modules. FLUX.1-dev's QKV projection is a concrete example. Diffusers defines three separate modules https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer flux.py L313-L329 : self.to q = torch.nn.Linear query dim, self.inner dim, bias=bias self.to k = torch.nn.Linear query dim, self.inner dim, bias=bias self.to v = torch.nn.Linear query dim, self.inner dim, bias=bias The Nunchaku FLUX module combines those layers https://github.com/nunchaku-ai/nunchaku/blob/main/nunchaku/models/transformers/transformer flux v2.py L63-L79 into one quantized to qkv module: to qkv = fuse linears other.to q, other.to k, other.to v self.to qkv = SVDQW4A4Linear.from linear to qkv, kwargs This grouped module is required because Nunchaku's fused operator consumes the QKV projection, Q/K normalization, and rotary embeddings together. By comparison, the default Diffusers path https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer flux.py L45-L116 executes them separately: query = attn.to q hidden states key = attn.to k hidden states value = attn.to v hidden states query = query.unflatten -1, attn.heads, -1 key = key.unflatten -1, attn.heads, -1 value = value.unflatten -1, attn.heads, -1 query = attn.norm q query key = attn.norm k key if image rotary emb is not None: query = apply rotary emb query, image rotary emb, sequence dim=1 key = apply rotary emb key, image rotary emb, sequence dim=1 The Nunchaku path https://github.com/nunchaku-ai/nunchaku/blob/main/nunchaku/models/attention processors/flux.py L69-L93 supplies the grouped projection, normalization modules, and rotary embeddings to one fused operator: qkv = fused qkv norm rottary hidden states, attn.to qkv, attn.norm q, attn.norm k, image rotary emb This is the structural rewrite that the generic path cannot infer. Diffusers has three destination modules with to q , to k , and to v parameter prefixes, while Nunchaku has one grouped module under to qkv . A model-specific target config or adapter must state that the Q, K, and V parameters should be concatenated along the output dimension, in that order, and loaded into to qkv . Structural rewrites like these are described by a model-specific target config during quantization and handled by a small runtime adapter when the checkpoint is loaded. The FLUX.2 Klein 4B quantization script https://github.com/rootonchair/diffuse-compressor/blob/main/examples/text to image/quantize flux2 klein 4b.py provides a concrete target-config example for producing a structurally rewritten checkpoint, while rootonchair/nunchaku-lite https://github.com/rootonchair/nunchaku-lite provides the runtime adapters needed to load grouped QKV tensors, split fused projections, and other fused operations. For the complete workflow, you can check the Adding A New Model https://github.com/rootonchair/diffuse-compressor/blob/main/docs/adding new model.md guide. Ready-to-use checkpoints To get started right away, check out the following repositories: rootonchair/ERNIE-Image-Turbo-nunchaku-lite-int4-bnb4-text-encoder https://huggingface.co/rootonchair/ERNIE-Image-Turbo-nunchaku-lite-int4-bnb4-text-encoder : INT4 ERNIE-Image-Turbo with a bitsandbytes NF4 text encoder rootonchair/ERNIE-Image-Turbo-nunchaku-lite-nvfp4-bnb4-text-encoder https://huggingface.co/rootonchair/ERNIE-Image-Turbo-nunchaku-lite-nvfp4-bnb4-text-encoder : NVFP4 ERNIE-Image-Turbo with a bitsandbytes NF4 text encoder OzzyGT/Krea 2 Turbo nunchaku lite nvfp4 https://huggingface.co/OzzyGT/Krea 2 Turbo nunchaku lite nvfp4 : NVFP4 Krea 2 Turbo checkpoint lite-infer https://huggingface.co/lite-infer : more Nunchaku Lite checkpoints and collections Conclusion Nunchaku's SVDQuant kernels are one of the most effective ways to run diffusion transformers efficiently on consumer hardware, and they are now natively supported in Diffusers. Pre-quantized checkpoints load with from pretrained , and the diffuse-compressor toolkit makes it possible to quantize new architectures without waiting for engine support. By quantizing both weights and activations, the W4A4 path lowers memory use while improving denoising latency, keeping image quality close to the BF16 original. If you quantize and publish a new model, we would love to hear about it. Share it on the Hub and let us know If you have any questions about this feature, feel free to join our Discord https://discord.gg/G7tWnz98XR . To learn more, check out the following resources: Diffusers Nunchaku documentation https://huggingface.co/docs/diffusers/quantization/nunchaku The integration PR huggingface/diffusers 14100 https://github.com/huggingface/diffusers/pull/14100 SVDQuant paper https://arxiv.org/abs/2411.05007 and the Nunchaku engine https://github.com/nunchaku-tech/nunchaku diffuse-compressor https://github.com/rootonchair/diffuse-compressor - Previous posts: Exploring Quantization Backends in Diffusers https://huggingface.co/blog/diffusers-quantization and Memory-efficient Diffusion Transformers with Quanto and Diffusers https://huggingface.co/blog/quanto-diffusers Acknowledgements Thanks to the Diffusers maintainers for reviews and guidance throughout the integration, and to the MIT HAN Lab / Nunchaku team for the original SVDQuant work. Thanks to Marc Sun for providing feedback on the blog post. Thanks to Álvaro Somoza for trying out nunchaku-lite and for providing feedback. rootonchair is also grateful to SilverAI for supporting this work and providing the environment in which much of this development took place.