nunchaku-chroma 0.1.3

Chroma model support for Nunchaku - quantized inference for diffusion models

Nunchaku Chroma

Adds Chroma model support to Nunchaku for fast 4-bit quantized inference.

Installation

Requirements

• Python 3.10+
• PyTorch 2.0+
• CUDA-capable GPU

Install from PyPI

pip install nunchaku nunchaku-chroma

Install from source

pip install nunchaku
git clone https://github.com/spooknik/nunchaku-chroma.git
cd nunchaku-chroma
pip install -e .

Usage

Diffusers API

from diffusers import ChromaPipeline
from nunchaku_chroma import NunchakuChromaTransformer2DModel
import torch

# Load quantized transformer
transformer = NunchakuChromaTransformer2DModel.from_pretrained(
    "path/to/quantized-chroma.safetensors",
    device="cuda",
    torch_dtype=torch.bfloat16,
)

# Create pipeline
pipe = ChromaPipeline.from_pretrained(
    "lodestones/Chroma",
    transformer=transformer,
    torch_dtype=torch.bfloat16,
).to("cuda")

# Generate
image = pipe(
    "A beautiful sunset over mountains",
    num_inference_steps=20,
    guidance_scale=4.0,
).images[0]

image.save("output.png")

LoRA Support

from nunchaku_chroma import NunchakuChromaTransformer2DModel

# Load quantized model
transformer = NunchakuChromaTransformer2DModel.from_pretrained(
    "path/to/quantized-chroma.safetensors",
    device="cuda",
    torch_dtype=torch.bfloat16,
)

# Apply single LoRA with strength
transformer.update_lora_params("path/to/lora.safetensors", strength=0.8)

# Adjust strength dynamically (without reloading)
transformer.set_lora_strength(0.5)

# Reset to base model
transformer.reset_lora()

# Apply multiple LoRAs with different strengths
transformer.update_lora_params_multi([
    ("lora1.safetensors", 0.8),
    ("lora2.safetensors", 0.5),
])

ComfyUI

For ComfyUI integration, use the comfyui-nunchaku-chroma custom node:

# Install dependencies
pip install nunchaku nunchaku-chroma

# Install the ComfyUI node
cd ComfyUI/custom_nodes
git clone https://github.com/spooknik/comfyui-nunchaku-chroma.git

# Restart ComfyUI

The following nodes will be available:

• Nunchaku Chroma DiT Loader - Load quantized Chroma models
• Nunchaku Chroma LoRA Loader - Apply a single LoRA
• Nunchaku Chroma LoRA Stack - Apply multiple LoRAs at once

Supported LoRA Formats

• ComfyUI/Kohya format: lora_unet_double_blocks_*, lora_unet_single_blocks_*
• Diffusers format: transformer_blocks.*.lora_A.weight, etc.

License

Apache-2.0