Adapting AutoRound for a New Diffusion Model Architecture

Overview

AutoRound's new diffusion path uses auto_round/compressors/diffusion_mixin.py, auto_round/calibration/diffusion.py, and the quantizer implementations under auto_round/algorithms/quantization/. This skill covers what code changes are needed when a new diffusion model doesn't work out-of-the-box. Common reasons for adaptation:

• Transformer block type not registered in DIFFUSION_OUTPUT_CONFIGS
• Non-standard pipeline API (not compatible with pipe(prompts, ...))
• Hybrid architecture with both AR and diffusion components
• Model not detected as a diffusion model

Step 0: Diagnose the Problem

from auto_round import AutoRound

ar = AutoRound(
    "your-org/your-diffusion-model",
    scheme="W4A16",
    iters=2,
    nsamples=2,
    num_inference_steps=5,
)
ar.quantize_and_save(output_dir="./test_output", format="fake")

Step 1: Ensure Model Detection

AutoRound detects diffusion models by checking for model_index.json in the model directory:

# auto_round/utils/model.py
def is_diffusion_model(model_or_path):
    # Checks for model_index.json presence

If your model doesn't have model_index.json, either create one in the model directory or pass diffusion-specific options through new-architecture AutoRound kwargs:

ar = AutoRound(
    model,
    num_inference_steps=5,
)

Pipeline Loading

diffusion_load_model() uses AutoPipelineForText2Image.from_pretrained() and extracts pipe.transformer as the quantizable model. If your model uses a different attribute (e.g., pipe.unet), this needs adjustment in auto_round/utils/model.py.

Step 2: Register Transformer Block Output Config

This is the most common adaptation needed. DIFFUSION_OUTPUT_CONFIGS maps transformer block class names to their output tensor names. Without this, calibration crashes because AutoRound doesn't know how to collect activations.

Find your block class name

import diffusers

pipe = diffusers.AutoPipelineForText2Image.from_pretrained("your-model")
for name, module in pipe.transformer.named_modules():
    if hasattr(module, "forward") and "block" in name.lower():
        print(f"{name}: {type(module).__name__}")

Register in DIFFUSION_OUTPUT_CONFIGS

Edit auto_round/algorithms/quantization/base.py:

class BaseQuantizers:
    DIFFUSION_OUTPUT_CONFIGS = {
        "FluxTransformerBlock": ["encoder_hidden_states", "hidden_states"],
        "FluxSingleTransformerBlock": ["encoder_hidden_states", "hidden_states"],
        # Add your block type:
        "YourTransformerBlock": ["hidden_states"],  # output tensor names in order
    }

The list must match the exact order of tensors returned by the block's forward() method.

How to determine output tensor names

1. Read the block's forward() method in diffusers source code
2. Identify what tensors it returns (usually hidden_states, sometimes also encoder_hidden_states)
3. List them in the order they're returned

Example: If forward() returns (hidden_states, encoder_hidden_states):

BaseQuantizers.DIFFUSION_OUTPUT_CONFIGS["YourBlock"] = ["hidden_states", "encoder_hidden_states"]

Example: If forward() returns just hidden_states:

BaseQuantizers.DIFFUSION_OUTPUT_CONFIGS["YourBlock"] = ["hidden_states"]

Step 3: Handle Non-Standard Pipeline API

If your model's inference API differs from the standard pipe(prompts, guidance_scale=..., num_inference_steps=...), provide a custom pipeline function.

Option A: Add a custom pipeline dispatch in DiffusionCalibrator

Update auto_round/calibration/diffusion.py so DiffusionCalibrator.calib() dispatches through a small helper instead of calling pipe(...) directly:

class DiffusionCalibrator(LLMCalibrator):
    ...

    def _run_pipeline(self, pipe, prompts, generator):
        if getattr(pipe, "_autoround_pipeline_fn", None) is not None:
            pipe._autoround_pipeline_fn(
                pipe,
                prompts,
                guidance_scale=self.compressor.guidance_scale,
                num_inference_steps=self.compressor.num_inference_steps,
                generator=generator,
            )
            return
        pipe(
            prompts,
            guidance_scale=self.compressor.guidance_scale,
            num_inference_steps=self.compressor.num_inference_steps,
            generator=generator,
        )

Option B: Attach a model-specific function during model loading

For a known model family, attach _autoround_pipeline_fn in auto_round/utils/model.py or auto_round/special_model_handler.py:

pipe._autoround_pipeline_fn = your_model_pipeline_fn

Option C: Add a dedicated branch in DiffusionCalibrator

For full control, update auto_round/calibration/diffusion.py so DiffusionCalibrator.calib() dispatches through your custom pipeline function:

class DiffusionCalibrator(LLMCalibrator):
    ...

    def _run_pipeline(self, pipe, prompts):
        c = self.compressor
        generator = (
            None if c.generator_seed is None else torch.Generator(device=pipe.device).manual_seed(c.generator_seed)
        )
        pipe.your_custom_generate(
            prompts,
            steps=c.num_inference_steps,
            cfg=c.guidance_scale,
            generator=generator,
        )

Step 4: Add Hybrid AR+DiT Support

For models with both autoregressive and diffusion components (e.g., GLM-Image).

4a. Register AR component

Add hybrid routing through the new architecture. Start with auto_round/autoround.py, auto_round/compressors/entry.py, and auto_round/compressors/diffusion_mixin.py. If a reusable AR-component registry is needed, place it near the new routing code:

HYBRID_AR_COMPONENTS = [
    "vision_language_encoder",  # GLM-Image
    "your_ar_component",  # Your model's AR attribute name
]

The attribute name must match what exists on the diffusers pipeline object (i.e., pipe.your_ar_component).

4b. Register DiT block output config

Add the DiT-specific output config in BaseQuantizers.DIFFUSION_OUTPUT_CONFIGS:

BaseQuantizers.DIFFUSION_OUTPUT_CONFIGS["YourDiTBlock"] = ["hidden_states", "encoder_hidden_states"]

4c. Register AR block handler

In auto_round/special_model_handler.py, add a block handler for the AR component so AutoRound knows which layers to quantize:

def _get_your_hybrid_multimodal_block(model, quant_vision=False):
    block_names = []
    if quant_vision and hasattr(model, "vision_encoder"):
        block_names.append([f"vision_encoder.blocks.{i}" for i in range(len(model.vision_encoder.blocks))])
    block_names.append([f"language_model.layers.{i}" for i in range(len(model.language_model.layers))])
    return block_names


SPECIAL_MULTIMODAL_BLOCK["your_model_type"] = _get_your_hybrid_multimodal_block

Hybrid quantization flow

The new hybrid flow should run two phases:

1. Phase 1 (AR): Quantizes the AR component using text calibration data (MLLM-style)
2. Phase 2 (DiT): Quantizes the DiT component using diffusion pipeline calibration

ar = AutoRound(
    "your-hybrid-model",
    dataset="coco2014",  # DiT calibration
    ar_dataset="NeelNanda/pile-10k",  # AR calibration
    quant_ar=True,
    quant_dit=True,
)

Step 5: Add Custom Calibration Dataset (Optional)

If your model needs a specific dataset format:

Edit the diffusion calibration path used by the new architecture:

• auto_round/calibration/diffusion.py for how diffusion prompts are loaded and consumed
• auto_round/calib_dataset.py for reusable dataset registration helpers

def get_diffusion_dataloader(dataset_name, nsamples, ...):
    # Add handling for your dataset format
    if dataset_name == "your_custom_dataset":
        return _load_your_dataset(dataset_name, nsamples)
    ...

The default coco2014 dataset works for most text-to-image models. Custom datasets need a TSV file with id and caption columns.

Step 6: Test

def test_your_diffusion_model():
    ar = AutoRound(
        "your-org/your-diffusion-model",
        scheme="W4A16",
        iters=2,
        nsamples=4,
        num_inference_steps=5,
        guidance_scale=7.5,
    )
    compressed_model, layer_config = ar.quantize()
    assert len(layer_config) > 0, "No layers quantized"
    ar.save_quantized(output_dir="./test_output", format="fake")

For hybrid models, test both phases:

ar = AutoRound(
    "your-hybrid-model",
    quant_ar=True,
    quant_dit=True,
    iters=2,
    nsamples=4,
)

Checklist

• is_diffusion_model() detects model
• Transformer block class name identified
• DIFFUSION_OUTPUT_CONFIGS entry added with correct output tensor names and order
• Pipeline runs without errors during calibration
• Custom pipeline dispatch added in DiffusionCalibrator if non-standard API
• For hybrid: AR component registered in HYBRID_AR_COMPONENTS
• For hybrid: AR block handler in SPECIAL_MULTIMODAL_BLOCK
• For hybrid: DiT output config in BaseQuantizers.DIFFUSION_OUTPUT_CONFIGS
• Quantization produces valid layers (check "Quantized X/Y layers" log)
• Export to fake format works
• README.md + README_CN.md updated

Key Files

Reference: Existing Adaptations