Support Flux quantization

.azure-pipelines/scripts/ut/run_ut_cuda.sh

@@ -46,6 +46,7 @@ function run_unit_test() {
     CMAKE_ARGS="-DGGML_CUDA=on -DLLAVA_BUILD=off" uv pip install llama-cpp-python
     uv pip install 'git+https://github.com/ggml-org/llama.cpp.git#subdirectory=gguf-py'
     uv pip install -r requirements.txt
+    uv pip install -r requirements_diffusion.txt
 
     uv pip list
     export COVERAGE_RCFILE=${REPO_PATH}/.azure-pipelines/scripts/ut/.coverage

auto_round/__init__.py

@@ -14,7 +14,7 @@
 from auto_round.autoround import AutoRound
 
 # support for old api
-from auto_round.autoround import AutoRoundLLM, AutoRoundMLLM, AutoRoundAdam
+from auto_round.autoround import AutoRoundLLM, AutoRoundMLLM, AutoRoundAdam, AutoRoundDiffusion
 from auto_round.utils import LazyImport
 
 

auto_round/__main__.py

@@ -230,6 +230,25 @@ def __init__(self, *args, **kwargs):
             help="the template for building training dataset. It can be a custom one.",
         )
 
+        ## ===================== diffusion model ==================
+        self.add_argument(
+            "--guidance_scale",
+            default=7.5,
+            type=float,
+        )
+
+        self.add_argument(
+            "--num_inference_steps",
+            default=50,
+            type=int,
+        )
+
+        self.add_argument(
+            "--generator_seed",
+            default=None,
+            type=int,
+        )
+
         ## ======================= eval =======================
         self.add_argument(
             "--tasks",
@@ -258,6 +277,22 @@ def __init__(self, *args, **kwargs):
             "--eval_model_dtype", default=None, type=str, help="the torch_dytpe to load the model for evaluation."
         )
 
+        ## ======================= diffusion model eval =======================
+        self.add_argument("--prompt_file", default=None, type=str, help="the prompt file to load prmpt.")
+
+        self.add_argument("--prompt", default=None, type=str, help="the prompt for test.")
+
+        self.add_argument(
+            "--metrics",
+            "--metric",
+            default="clip",
+            help="support clip, clip-iqa, imagereward",
+        )
+
+        self.add_argument(
+            "--image_save_dir", default="./tmp_image_save", type=str, help="path to save generated images"
+        )
+
 
 def setup_parser():
     parser = BasicArgumentParser()
@@ -427,6 +462,7 @@ def tune(args):
         )
 
     from auto_round.compressors import (
+        DiffusionExtraConfig,
         ExtraConfig,
         MLLMExtraConfig,
         SchemeExtraConfig,
@@ -466,9 +502,15 @@ def tune(args):
     mllm_config = MLLMExtraConfig(
         quant_nontext_module=args.quant_nontext_module, extra_data_dir=args.extra_data_dir, template=args.template
     )
+    diffusion_config = DiffusionExtraConfig(
+        guidance_scale=args.guidance_scale,
+        num_inference_steps=args.num_inference_steps,
+        generator_seed=args.generator_seed,
+    )
     extra_config.tuning_config = tuning_config
     extra_config.scheme_config = scheme_config
     extra_config.mllm_config = mllm_config
+    extra_config.diffusion_config = diffusion_config
 
     autoround: BaseCompressor = AutoRound(
         model=model_name,
@@ -524,6 +566,45 @@ def tune(args):
     model.eval()
     clear_memory()
 
+    eval_model_dtype = get_model_dtype(args.eval_model_dtype, "auto")
+
+    # diffusion model has different evaluation path
+    if getattr(autoround, "diffusion", False):
+        pipe = autoround.pipe
+        pipe.to(model.dtype)
+        pipe.transformer = model
+        device_str = detect_device(device_str)
+        pipe = pipe.to(device_str)
+        if pipe.dtype != eval_model_dtype and eval_model_dtype != "auto":
+            pipe.to(getattr(torch, eval_model_dtype))
+
+        gen_kwargs = {
+            "guidance_scale": args.guidance_scale,
+            "output_type": "pil",
+            "num_inference_steps": args.num_inference_steps,
+            "generator": (
+                None
+                if args.generator_seed is None
+                else torch.Generator(device=pipe.device).manual_seed(args.generator_seed)
+            ),
+        }
+        if not os.path.exists(args.image_save_dir):
+            os.makedirs(args.image_save_dir)
+
+        if args.prompt is not None:
+            outputs = pipe(prompt=args.prompt, **gen_kwargs)
+            outputs.images[0].save(os.path.join(args.image_save_dir, "img.png"))
+            logger.info(
+                f"Image generated with prompt {args.prompt} is saved as {os.path.join(args.image_save_dir, 'img.png')}"
+            )
+
+        if args.prompt_file is not None:
+            from auto_round.compressors.diffusion import diffusion_eval
+
+            metrics = args.metrics.split(",")
+            diffusion_eval(pipe, args.prompt_file, metrics, args.image_save_dir, 1, gen_kwargs)
+        return
+
     lm_eval_version = get_library_version("lm-eval")
 
     eval_folder = folders[-1]
@@ -545,8 +626,6 @@ def tune(args):
 
     import time
 
-    eval_model_dtype = get_model_dtype(args.eval_model_dtype, "auto")
-
     if (autoround.act_bits <= 8 and formats[-1] == "fake") or eval_gguf_model:
         if eval_gguf_model:
             # for file in os.listdir(eval_folder):

auto_round/autoround.py

@@ -20,13 +20,14 @@
 from auto_round.compressors import (
     AdamCompressor,
     BaseCompressor,
+    DiffusionCompressor,
     ExtraConfig,
     LLMCompressor,
     MLLMCompressor,
 )
 from auto_round.logger import deprecated, logger
 from auto_round.schemes import QuantizationScheme
-from auto_round.utils import is_mllm_model
+from auto_round.utils import is_diffusion_model, is_mllm_model
 
 
 class AutoRound:
@@ -77,7 +78,7 @@ def __new__(
         seed: int = 42,
         # for adam
         enable_adam: bool = False,
-        # for MLLM
+        # for MLLM and Diffusion
         extra_config: ExtraConfig = None,
         **kwargs,
     ) -> BaseCompressor:
@@ -145,9 +146,17 @@ def __new__(
         if (extra_config and not extra_config.mllm_config.is_default()) or is_mllm_model(model):
             logger.info("using MLLM mode for multimodal model.")
             model_cls.append(MLLMCompressor)
+            if extra_config:
+                extra_config.diffusion_config = None
+        elif (extra_config and not extra_config.diffusion_config.is_default()) or is_diffusion_model(model):
+            logger.info("using Diffusion mode for diffusion model.")
+            model_cls.append(DiffusionCompressor)
+            if extra_config:
+                extra_config.mllm_config = None
         else:
             if extra_config:
                 extra_config.mllm_config = None
+                extra_config.diffusion_config = None
             model_cls.append(LLMCompressor)
 
         if enable_adam:
@@ -540,3 +549,83 @@ def __init__(
             seed=seed,
             **kwargs,
         )
+
+
+@deprecated("AutoRound")
+class AutoRoundDiffusion(DiffusionCompressor):
+    """Class for automatic rounding-based quantization with Diffusion models.
+
+    Args:
+        model: The PyTorch model to be quantized.
+        tokenizer: An optional tokenizer for processing input data, is not used for diffusion models.
+        guidance_scale (float): Control how much the image generation process follows the text prompt.
+                                The more it is, the more closely it follows the prompt (default is 7.5).
+        num_inference_steps (int): The reference number of denoising steps (default is 50).
+        generator_seed (int): A sees that controls the initial noise from which an image is generated (default is None).
+        scheme: (str| dict | QuantizationScheme ): A preset scheme that defines the quantization configurations.
+        layer_config (dict): Configuration for weight quantization (default is None).
+        dataset: The path or name of the calib dataset.
+        iters (int): Number of iterations (default is 200).
+        seqlen (int): Length of the sequence.
+        nsamples (int): Number of samples (default is 128).
+        batch_size (int): Batch size for training (default is 8).
+        gradient_accumulate_steps (int): Number of gradient accumulation steps (default is 1).
+        low_gpu_mem_usage (bool): Whether to use low GPU memory (default is False).
+        device_map (str | dict | int | torch.device, optional): Device placement map. Defaults to 0.
+        enable_torch_compile (bool): Whether to enable torch compile to optimize quant_block/layer
+        **kwargs: Additional keyword arguments.
+    """
+
+    bits: int | None
+    group_size: int | None
+    sym: bool | None
+    data_type: str | None
+    act_bits: int | None
+    act_group_size: int | None
+    act_sym: bool | None
+    act_data_type: str | None
+    act_dynamic: bool | None
+    super_bits: int | None
+    super_group_size: int | None
+
+    def __init__(
+        self,
+        model: Union[object, str],
+        tokenizer=None,
+        guidance_scale: float = 7.5,
+        num_inference_steps: int = 50,
+        generator_seed: int = None,
+        scheme: Union[str, dict, QuantizationScheme] = "W4A16",
+        layer_config: dict[str, Union[str, dict, QuantizationScheme]] = None,
+        dataset: Union[str, list, tuple, torch.utils.data.DataLoader] = "coco2014",
+        iters: int = 200,
+        seqlen: int = 2048,
+        nsamples: int = 128,
+        batch_size: int = 8,
+        gradient_accumulate_steps: int = 1,
+        low_gpu_mem_usage: bool = False,
+        device_map: Union[str, torch.device, int, dict] = 0,
+        enable_torch_compile: bool = False,
+        seed: int = 42,
+        **kwargs,
+    ):
+        super().__init__(
+            model=model,
+            tokenizer=None,
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            generator_seed=generator_seed,
+            scheme=scheme,
+            layer_config=layer_config,
+            dataset=dataset,
+            iters=iters,
+            seqlen=seqlen,
+            nsamples=nsamples,
+            batch_size=batch_size,
+            gradient_accumulate_steps=gradient_accumulate_steps,
+            low_gpu_mem_usage=low_gpu_mem_usage,
+            device_map=device_map,
+            enable_torch_compile=enable_torch_compile,
+            seed=seed,
+            **kwargs,
+        )

auto_round/compressors/__init__.py

@@ -14,7 +14,9 @@
 
 from auto_round.compressors.base import *
 from auto_round.compressors.mllm.compressor import MLLMCompressor
+from auto_round.compressors.diffusion.compressor import DiffusionCompressor
 from auto_round.compressors.config import (
+    DiffusionExtraConfig,
     ExtraConfig,
     MLLMExtraConfig,
     SchemeExtraConfig,