[quantization] SDNQ core loading

[asomoza]

What does this PR do?

This PR adds basic support for SDNQ:

• Load a prequantized SDNQ checkpoint with a normal from_pretrained call — no special arguments, the quantization config is read from the model's config.json.
• Quantize a model on the fly by passing SDNQConfig(...) (or via PipelineQuantizationConfig) to from_pretrained.
• Save and reload quantized models, including inside full pipelines.

Test code

load a pre quantized model

import torch

from diffusers import DiffusionPipeline


prompt = (
    "A highly detailed photograph of a cozy vintage bookshop interior at golden hour. "
    "Warm sunlight streams through a rain-streaked window, casting long shadows across worn "
    "oak floorboards. Floor-to-ceiling shelves are packed with hundreds of colorful leather-bound "
    "books, their gold-embossed spines catching the light. On the wooden counter sits a chalkboard "
    "sign with elegant handwritten text reading 'Rare Books & Coffee', beside a steaming ceramic "
    "mug, a brass reading lamp, and a sleeping orange tabby cat curled on a stack of newspapers. "
    "Dust motes float in the sunbeams. Intricate textures on the aged paper, cracked leather, and "
    "polished brass. Sharp focus, rich warm color grading, shallow depth of field, photorealistic, 8k."
)

pipe = DiffusionPipeline.from_pretrained(
    "Disty0/Z-Image-Turbo-SDNQ-int8",
    dtype=torch.bfloat16,
).to("cuda")

image = pipe(prompt, num_inference_steps=8, guidance_scale=0.0).images[0]
image.save("zimage_sdnq_load.png")

pre quantize model on the fly

import torch

from diffusers import DiffusionPipeline, PipelineQuantizationConfig, SDNQConfig


prompt = (
    "A highly detailed photograph of a cozy vintage bookshop interior at golden hour. "
    "Warm sunlight streams through a rain-streaked window, casting long shadows across worn "
    "oak floorboards. Floor-to-ceiling shelves are packed with hundreds of colorful leather-bound "
    "books, their gold-embossed spines catching the light. On the wooden counter sits a chalkboard "
    "sign with elegant handwritten text reading 'Rare Books & Coffee', beside a steaming ceramic "
    "mug, a brass reading lamp, and a sleeping orange tabby cat curled on a stack of newspapers. "
    "Dust motes float in the sunbeams. Intricate textures on the aged paper, cracked leather, and "
    "polished brass. Sharp focus, rich warm color grading, shallow depth of field, photorealistic, 8k."
)

pipe = DiffusionPipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    quantization_config=PipelineQuantizationConfig(
        quant_mapping={"transformer": SDNQConfig(weights_dtype="int8")}
    ),
    dtype=torch.bfloat16,
).to("cuda")

image = pipe(prompt, num_inference_steps=8, guidance_scale=0.0).images[0]
image.save("zimage_sdnq_quantize.png")

save quantized DiT

import torch

from diffusers import SDNQConfig, ZImageTransformer2DModel


save_dir = "z-image-turbo-sdnq-int8"

# Quantize the transformer during load.
transformer = ZImageTransformer2DModel.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    subfolder="transformer",
    quantization_config=SDNQConfig(weights_dtype="int8"),
    dtype=torch.bfloat16,
)

# Serialize the quantized weights + config.
transformer.save_pretrained(save_dir)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.