Diffusers Integrates Nunchaku Lite for 4-Bit Diffusion Transformer Inference

Diffusers Integrates Nunchaku Lite for 4-Bit Diffusion Transformer Inference Hugging Face Diffusers has added native support for Nunchaku Lite, enabling 4-bit weights-and-activations (W4A4) diffusion transformer inference via SVDQuant. The integration provides significant VRAM reduction and latency improvements directly through standard from_pretrained() calls without requiring manual local CUDA compilation. SVDQuant enables 4-bit weights and activations (W4A4) by shifting activation outliers into weights, using a 16-bit low-rank branch alongside 4-bit residual weights. Nunchaku Lite brings native loading to Diffusers by patching nn.Linear layers with svdq_w4a4 and awq_w4a16 runtime layers, delivering roughly 30% speedups and up to 50% VRAM reductions. NVFP4 is supported on NVIDIA Blackwell GPUs, while INT4 variants support Turing, Ampere, and Ada architectures; Volta and Hopper GPUs are currently unsupported. Combining Nunchaku Lite with torch.compile increases pipeline speedups to 1.8x, while quantizing text encoders with bitsandbytes NF4 reduces peak VRAM by about 22%. The diffuse-compressor toolkit provides an architecture-agnostic pipeline to scan, quantize, convert, and upload custom diffusion models. SVDQuant and Nunchaku Lite Architecture Unlike weight-only quantization backends that dequantize to higher precision during computation, SVDQuant quantizes both weights and activations to 4-bit (W4A4). It addresses activation outliers in diffusion transformers by shifting them into the weights, isolating the difficult portions into a 16-bit low-rank branch, and quantizing residual weights to 4-bit. Nunchaku Lite integrates this into Diffusers without custom pipeline classes by patching nn.Linear modules with svdq_w4a4 for compute-heavy attention and MLP layers, and awq_w4a16 for precision-sensitive adaptive normalization and modulation layers. Kernels are fetched automatically from the Hub via the kernels package. Hardware Support and Compatibility Nunchaku Lite kernel support varies by GPU generation and target format. NVFP4 svdq_w4a4 kernels require NVIDIA Blackwell GPUs, including the RTX 50 series, RTX PRO 6000, and B200. Older architectures—specifically Turing, Ampere, and Ada (RTX 30 and 40 series, A100, L40S)—are supported through INT4 variants for svdq_w4a4 and awq_w4a16. Volta and Hopper GPUs are currently not supported by these 4-bit kernels, and the quantizer checks CUDA capability at load time to prevent incorrect executions. Performance and Memory Optimizations Tested on an NVIDIA RTX PRO 6000 running ERNIE-Image-Turbo at 1024x1024, Nunchaku Lite NVFP4 reduced peak VRAM from 31.1 GB in BF16 down to 20.6 GB while improving latency from 3.00 seconds to 2.27 seconds (a 1.35x speedup). Applying torch.compile lowers overall runtime to 1.68 seconds (a 1.8x speedup) by reducing kernel launch overhead. Further pairing the model with a bitsandbytes NF4 text encoder lowers peak VRAM to 16.0 GB, representing an overall memory reduction of roughly 50% compared to the BF16 baseline. Quantization Workflow and Structural Rewrites Users can quantize custom models using diffuse-compressor through a multi-step workflow: inspecting target modules, running quantization, packaging into a standard Diffusers pipeline, and uploading to the Hub. The generic scanner automatically targets transformer linears for SVDQ and modulation layers for AWQ. However, models requiring fused execution—such as FLUX.1-dev's grouped QKV projections combined with normalization and rotary embeddings—require model-specific target configurations and runtime adapters to execute structural module rewrites.

Diffusers 原生支持 Nunchaku 4位扩散模型推理与 SVDQuant 量化 Hugging Face Diffusers 现已原生集成 Nunchaku Lite，支持使用 SVDQuant 方法的 4-bit（W4A4）扩散 Transformer 模型推理。开发者无需单独编译 CUDA 内核或依赖专有推理引擎，即可通过标准的 from_pretrained() 流程加载轻量化模型，在大幅节省显存的同时提升推理速度。 SVDQuant 采用 W4A4（4 位权重与激活）量化方案，将激活异常值转移至权重，并结合 16 位低秩分支保持模型精度。 Diffusers 原生集成采用 Nunchaku Lite 方案，通过 kernels 运行时修补线性层，支持 svdq_w4a4 与 awq_w4a16 两种内核机制。 硬件方面，NVFP4 支持 Blackwell 架构，INT4 支持 Turing、Ampere 与 Ada 架构，暂不支持 Volta 和 Hopper 显卡。 基准测试显示，NVFP4 配合 torch.compile 与 NF4 文本编码器可将峰值显存降低高达近 50%，推理加速达 1.8 倍。 配套工具 diffuse-compressor 支持开发者针对自定义架构进行扫描、校准、量化、打包并发布模型至 Hugging Face Hub。 SVDQuant 与 Nunchaku 技术原理 传统扩散 Transformer 在 BF16 精度下往往需要 20-30 GB 显存，且多数仅量化权重的方案无法带来推理加速甚至存在延迟开销。SVDQuant 针对扩散模型中权重与激活均存在离群值的难题，将激活异常值移入权重矩阵中，通过 16 位低秩分支表示最复杂的权重部分，其余残差进行 4 位量化。Nunchaku 引擎通过融合低秩投影与量化/计算内核，消除了 16 位分支的显存访问瓶颈。 Nunchaku Lite 原生集成与硬件兼容 Nunchaku Lite 在 Diffusers 中对核心 Linear 模块进行动态修补，无需自定义 pipeline 类。计算密集的主干投影采用 svdq_w4a4，显存受限且对精度敏感的自适应归一化等层则采用 awq_w4a16。硬件支持上，NVFP4 内核适用于 Blackwell 架构（如 RTX 50 系列、RTX PRO 6000、B200），而 INT4 适用于 Turing、Ampere、Ada 架构（RTX 30/40 系列、A100、L40S）；Volta 与 Hopper 架构当前暂不支持。 性能表现与组合优化 在 NVIDIA RTX PRO 6000 测试 ERNIE-Image-Turbo 时，原生 Nunchaku Lite NVFP4 将生成时间从 BF16 的 3.00 秒缩短至 2.27 秒（加速 1.35x），显存从 31.1 GB 降至 20.6 GB。叠加 torch.compile 后生成时间进一步压至 1.68 秒（加速 1.8x）。若文本编码器结合 bitsandbytes NF4 量化，峰值显存可降至 16.0 GB（减少近一半），同时生成图像质量与 BF16 基线基本一致。 模型量化流程与结构重写支持 利用 diffuse-compressor 工具包，用户可通过检查配置、执行 SVDQuant 量化、打包 Diffusers 格式并推送到 Hub 的四步流程完成适配。对于标准架构可通用转换；若模型存在 QKV 投影融合等架构级改动（如 FLUX.1-dev 将分离的 Q/K/V 投影合并输入融合算子），则需提供特定目标的配置与运行时适配器完成张量维度的结构重写。

Large diffusion transformers can create stunning images (or even videos, audio snippets, and now text), but loading a modern text-to-image model in BF16 precision often requires 20-30 GB of VRAM, which puts these models out of reach of most consumer GPUs. Quantization is a powerful solution to this problem, and Diffusers already integrates several quantization backends such as bitsandbytes, GGUF, torchao, and Quanto, which we covered in Exploring Quantization Backends in Diffusers . 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 , the quantization method behind the popular 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 package. In addition, the companion diffuse-compressor toolkit lets you quantize new architectures yourself and publish them as regular Diffusers repositories. Table of Contents Getting started with Nunchaku Lite Background: SVDQuant and Nunchaku Introducing Nunchaku Lite Native loading in Diffusers Getting more speed and lower memory Benchmarks Quantizing your own model Ready-to-use checkpoints Conclusion 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: 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 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 . 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. Nunchaku fuses the low-rank down projection with the quantization kernel and the low-rank up projection with the 4-bit compute kernel, eliminating the memory access overhead of the 16-bit branch. Figure from the SVDQuant paper . Introducing Nunchaku Lite The original Nunchaku engine gets much of its speed from model-specific fused execution paths , 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: svdq_w4a4 : 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. awq_w4a16 : 4-bit weights with 16-bit activations, used for adaptive normalization and modulation projections such as FLUX 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 . 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 BF16 vs 4-bit outputs with identical seeds and settings. Quantizing your own model Nunchaku Lite support in Diffusers is architecture-agnostic, and the 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 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 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 : 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 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 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 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 provides a concrete target-config example for producing a structurally rewritten checkpoint, while 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 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 : INT4 ERNIE-Image-Turbo with a bitsandbytes NF4 text encoder 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 : NVFP4 Krea 2 Turbo checkpoint 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 . To learn more, check out the following resources: Diffusers Nunchaku documentation The integration PR (huggingface/diffusers#14100) SVDQuant paper and the Nunchaku engine diffuse-compressor Previous posts: Exploring Quantization Backends in Diffusers and Memory-efficient Diffusion Transformers with Quanto and 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.