diffusers/pr_13098/en/quantization/autoround.md

# AutoRound

[AutoRound](https://github.com/intel/auto-round) is an advanced quantization toolkit. It achieves high accuracy at ultra-low bit widths (2-4 bits) with minimal tuning by leveraging sign-gradient descent and providing broad hardware compatibility. See our papers [SignRoundV1](https://arxiv.org/pdf/2309.05516) and [SignRoundV2](https://arxiv.org/abs/2512.04746) for more details.

Install `auto-round`(version ≥ 0.13.0):

```bash
pip install "auto-round>=0.13.0"
```

To use the Marlin kernel for faster CUDA inference, install `gptqmodel`:

```bash
pip install "gptqmodel>=5.8.0"
```

## Load a quantized model

Load a pre-quantized AutoRound model by passing `AutoRoundConfig` to [from_pretrained()](/docs/diffusers/pr_13098/en/api/models/overview#diffusers.ModelMixin.from_pretrained). The method works with any model that loads via [Accelerate](https://hf.co/docs/accelerate/index) and has `torch.nn.Linear` layers.

You can use [PipelineQuantizationConfig](/docs/diffusers/pr_13098/en/api/quantization#diffusers.PipelineQuantizationConfig) to quantize specific components of a pipeline:

```python
import torch
from diffusers import DiffusionPipeline, PipelineQuantizationConfig, AutoRoundConfig

pipeline_quant_config = PipelineQuantizationConfig(
    quant_mapping={"transformer": AutoRoundConfig(backend="auto")}
)
pipe = DiffusionPipeline.from_pretrained(
    "INCModel/Z-Image-W4A16-AutoRound",
    quantization_config=pipeline_quant_config,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)

image = pipe("a cat holding a sign that says hello").images[0]
image.save("output.png")
```

Or load a quantized model component directly:

```python
import torch
from diffusers import ZImageTransformer2DModel, ZImagePipeline, AutoRoundConfig

model_id = "INCModel/Z-Image-W4A16-AutoRound"

quantization_config = AutoRoundConfig(backend="auto")
transformer = ZImageTransformer2DModel.from_pretrained(
    model_id,
    subfolder="transformer",
    quantization_config=quantization_config,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)

pipe = ZImagePipeline.from_pretrained(
    model_id,
    transformer=transformer,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)

image = pipe("a cat holding a sign that says hello").images[0]
image.save("output.png")
```

> [!NOTE]
> AutoRound in Diffusers only supports loading *pre-quantized* models. To quantize a model from scratch, use the [AutoRound CLI or Python API](https://github.com/intel/auto-round) directly, then load the result with Diffusers.

## torch.compile

AutoRound is compatible with [`torch.compile`](../optimization/fp16#torchcompile) for faster inference. You can compile the quantized transformer (DiT) for better performance:

```python
import torch
from diffusers import DiffusionPipeline, PipelineQuantizationConfig, AutoRoundConfig

pipeline_quant_config = PipelineQuantizationConfig(
    quant_mapping={"transformer": AutoRoundConfig(backend="auto")}
)
pipe = DiffusionPipeline.from_pretrained(
    "INCModel/Z-Image-W4A16-AutoRound",
    quantization_config=pipeline_quant_config,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)

pipe.transformer = torch.compile(pipe.transformer, mode="default", fullgraph=False)
```

## Backends

AutoRound supports multiple inference backends for Weight-only quantized model. The backend controls which kernel handles dequantization during the forward pass. Set the `backend` parameter in `AutoRoundConfig` to choose one:

| Backend | Value | Device | Requirements | Notes |
|---------|-------|--------|--------------|-------|
| **Auto** | `"auto"` | Any | — | Default. Automatically selects the best available backend. |
| **PyTorch** | `"torch"` | CPU / CUDA | — | Pure PyTorch implementation. Broadest compatibility. |
| **Triton** | `"tritonv2"` | CUDA | `triton` | Triton-based kernel for GPU inference. |
| **ExllamaV2** | `"exllamav2"` | CUDA | `gptqmodel>=5.8.0` | Good CUDA performance via the ExllamaV2 kernel. |
| **Marlin** | `"marlin"` | CUDA | `gptqmodel>=5.8.0` | Best CUDA performance via the Marlin kernel. |

```python
from diffusers import AutoRoundConfig

# Auto-select (default)
config = AutoRoundConfig()

# Explicit Triton backend for CUDA
config = AutoRoundConfig(backend="tritonv2")

# Marlin backend for best CUDA performance (requires gptqmodel>=5.8.0)
config = AutoRoundConfig(backend="marlin")

# ExllamaV2 backend for good CUDA performance (requires gptqmodel>=5.8.0)
config = AutoRoundConfig(backend="exllamav2")

# PyTorch backend for CPU/CUDA inference
config = AutoRoundConfig(backend="torch")
```

## Save and load

AutoRound requires data calibration to quantize a model. This is done outside of Diffusers using the [AutoRound library](https://github.com/intel/auto-round) directly:

```python
from auto_round import AutoRound

autoround = AutoRound(
    "Tongyi-MAI/Z-Image",
    scheme="W4A16",  # W4G128 symmetric
    enable_torch_compile=True,
    num_inference_steps=3,
    guidance_scale=7.5,
    dataset="coco2014",
)
autoround.quantize_and_save("Z-Image-W4A16-AutoRound")
```

For more details on calibration options, see the [AutoRound documentation](https://github.com/intel/auto-round).

```python
import torch
from diffusers import ZImageTransformer2DModel, ZImagePipeline

model_id = "INCModel/Z-Image-W4A16-AutoRound"

# The inference backend will be automatically selected.
pipe = ZImagePipeline.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)

image = pipe("a cat holding a sign that says hello").images[0]
image.save("output.png")
```

### Supported Quantization Schemes

AutoRound supports several Schemes:

- **W4A16**(bits:4,group_size:128,sym:True,act_bits:16)
- **W8A16**(bits:8,group_size:128,sym:True,act_bits:16)
- **W3A16**(bits:3,group_size:128,sym:True,act_bits:16)
- **W2A16**(bits:2,group_size:128,sym:True,act_bits:16)
- **GGUF:Q4_K_M**(all Q*_K,Q*_0,Q*_1 provided by llamacpp are supported)
- **NVFP4**(Experimental feature, recommend exporting to `llm_compressor` format.data_type nvfp4,act_data_type nvfp4,static_global_scale,group_size 16)
- **MXFP4**(**Research feature, no real kernel**, Standard MXFP4, data_type mxfp,act_data_type mxfp,bits 4, act_bits 4, group_size 32)
- **MXINT4**(**Research feature, no real kernel**, Standard MXINT4, data_type mxint,act_data_type mxint,bits 4, act_bits 4, group_size 32)
- **MXFP4_RCEIL**(**Research feature,no real kernel**, NVIDIA's variant, data_type mxfp,act_data_type mxfp_rceil,bits 4, act_bits 4, group_size 32)
- **MXFP8**(**Research feature, no real kernel**, data_type mxfp,act_data_type mxfp_rceil,group_size 32)
- **FPW8A16**(**Research feature, no real kernel**, data_type fp8,group_size 0->per tensor )
- **FP8_STATIC**(**Research feature, no real kernel**, data_type:fp8,act_data_type:fp8,group_size -1 ->per channel, act_group_size=0->per tensor)

Besides, you could modify the `group_size`, `bits`, `sym` and many other configs you want, though there are maybe no real kernels.

## Resources

- [Pre-quantized AutoRound models on the Hub](https://huggingface.co/models?search=autoround)