Support ONNX conversion and pipeline for SD3

scripts/convert_stable_diffusion_3_checkpoint_to_onnx.py

@@ -0,0 +1,292 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import os
+import shutil
+from pathlib import Path
+
+import onnx
+import torch
+from packaging import version
+from torch.onnx import export
+
+from diffusers import OnnxRuntimeModel, OnnxStableDiffusion3Pipeline, StableDiffusion3Pipeline
+
+
+is_torch_less_than_1_11 = version.parse(version.parse(torch.__version__).base_version) < version.parse("1.11")
+
+
+def onnx_export(
+    model,
+    model_args: tuple,
+    output_path: Path,
+    ordered_input_names,
+    output_names,
+    dynamic_axes,
+    opset,
+    use_external_data_format=False,
+):
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    # PyTorch deprecated the `enable_onnx_checker` and `use_external_data_format` arguments in v1.11,
+    # so we check the torch version for backwards compatibility
+    if is_torch_less_than_1_11:
+        export(
+            model,
+            model_args,
+            f=output_path.as_posix(),
+            input_names=ordered_input_names,
+            output_names=output_names,
+            dynamic_axes=dynamic_axes,
+            do_constant_folding=True,
+            use_external_data_format=use_external_data_format,
+            enable_onnx_checker=True,
+            opset_version=opset,
+        )
+    else:
+        export(
+            model,
+            model_args,
+            f=output_path.as_posix(),
+            input_names=ordered_input_names,
+            output_names=output_names,
+            dynamic_axes=dynamic_axes,
+            do_constant_folding=True,
+            opset_version=opset,
+        )
+
+
+@torch.no_grad()
+def convert_models(model_path: str, output_path: str, opset: int, fp16: bool = False):
+    dtype = torch.float16 if fp16 else torch.float32
+    if fp16 and torch.cuda.is_available():
+        device = "cuda"
+    elif fp16 and not torch.cuda.is_available():
+        raise ValueError("`float16` model export is only supported on GPUs with CUDA")
+    else:
+        device = "cpu"
+    pipeline = StableDiffusion3Pipeline.from_pretrained(model_path, torch_dtype=dtype).to(device)
+    output_path = Path(output_path)
+
+    # TEXT ENCODER
+    num_tokens = pipeline.text_encoder.config.max_position_embeddings
+    text_hidden_size = pipeline.text_encoder.config.hidden_size
+    text_input = pipeline.tokenizer(
+        "A sample prompt",
+        padding="max_length",
+        max_length=pipeline.tokenizer.model_max_length,
+        truncation=True,
+        return_tensors="pt",
+    )
+    onnx_export(
+        pipeline.text_encoder,
+        # casting to torch.int32 until the CLIP fix is released: https://github.com/huggingface/transformers/pull/18515/files
+        model_args=(
+            text_input.input_ids.to(device=device, dtype=torch.int32),
+            None,
+            None,
+            None,
+            True,
+        ),
+        output_path=output_path / "text_encoder" / "model.onnx",
+        ordered_input_names=["input_ids"],
+        output_names=["last_hidden_state", "pooler_output", "hidden_states"],
+        dynamic_axes={
+            "input_ids": {0: "batch", 1: "sequence"},
+        },
+        opset=opset,
+    )
+    del pipeline.text_encoder
+
+    num_tokens = pipeline.text_encoder_2.config.max_position_embeddings
+    text_hidden_size = pipeline.text_encoder_2.config.hidden_size
+    text_input = pipeline.tokenizer_2(
+        "A sample prompt",
+        padding="max_length",
+        max_length=pipeline.tokenizer_2.model_max_length,
+        truncation=True,
+        return_tensors="pt",
+    )
+    onnx_export(
+        pipeline.text_encoder_2,
+        # casting to torch.int32 until the CLIP fix is released: https://github.com/huggingface/transformers/pull/18515/files
+        model_args=(
+            text_input.input_ids.to(device=device, dtype=torch.int32),
+            None,
+            None,
+            None,
+            True,
+        ),
+        output_path=output_path / "text_encoder_2" / "model.onnx",
+        ordered_input_names=["input_ids"],
+        output_names=["last_hidden_state", "pooler_output", "hidden_states"],
+        dynamic_axes={
+            "input_ids": {0: "batch", 1: "sequence"},
+        },
+        opset=opset,
+    )
+    del pipeline.text_encoder_2
+
+    text_input = pipeline.tokenizer_3(
+        "A sample prompt",
+        padding="max_length",
+        max_length=pipeline.tokenizer_3.model_max_length,
+        truncation=True,
+        return_tensors="pt",
+    )
+    onnx_export(
+        pipeline.text_encoder_3,
+        # casting to torch.int32 until the CLIP fix is released: https://github.com/huggingface/transformers/pull/18515/files
+        model_args=(text_input.input_ids.to(device=device, dtype=torch.int32)),
+        output_path=output_path / "text_encoder_3" / "model.onnx",
+        ordered_input_names=["input_ids"],
+        output_names=["last_hidden_state"],
+        dynamic_axes={
+            "input_ids": {0: "batch", 1: "sequence"},
+        },
+        opset=opset,
+    )
+    del pipeline.text_encoder_3
+
+    # TRANSFORMER
+    in_channels = pipeline.transformer.config.in_channels
+    sample_size = pipeline.transformer.config.sample_size
+    joint_attention_dim = pipeline.transformer.config.joint_attention_dim
+    pooled_projection_dim = pipeline.transformer.config.pooled_projection_dim
+    transformer_path = output_path / "transformer" / "model.onnx"
+    onnx_export(
+        pipeline.transformer,
+        model_args=(
+            torch.randn(2, in_channels, sample_size, sample_size).to(device=device, dtype=dtype),
+            torch.randn(2, num_tokens, joint_attention_dim).to(device=device, dtype=dtype),
+            torch.randn(2, pooled_projection_dim).to(device=device, dtype=dtype),
+            torch.randn(2).to(device=device, dtype=dtype),
+        ),
+        output_path=transformer_path,
+        ordered_input_names=["hidden_states", "encoder_hidden_states", "pooled_projections", "timestep"],
+        output_names=["out_sample"],  # has to be different from "sample" for correct tracing
+        dynamic_axes={
+            "hidden_states": {0: "batch", 1: "channels", 2: "height", 3: "width"},
+            "encoder_hidden_states": {0: "batch", 1: "sequence", 2: "embed_dims"},
+            "pooled_projections": {0: "batch", 1: "projection_dim"},
+            "timestep": {0: "batch"},
+        },
+        opset=opset,
+        use_external_data_format=True,  # UNet is > 2GB, so the weights need to be split
+    )
+    model_path = str(transformer_path.absolute().as_posix())
+    transformer_dir = os.path.dirname(model_path)
+    transformer = onnx.load(model_path)
+    # clean up existing tensor files
+    shutil.rmtree(transformer_dir)
+    os.mkdir(transformer_dir)
+    # collate external tensor files into one
+    onnx.save_model(
+        transformer,
+        model_path,
+        save_as_external_data=True,
+        all_tensors_to_one_file=True,
+        location="weights.pb",
+        convert_attribute=False,
+    )
+    del pipeline.transformer
+
+    # VAE ENCODER
+    vae_encoder = pipeline.vae
+    vae_in_channels = vae_encoder.config.in_channels
+    vae_sample_size = vae_encoder.config.sample_size
+    # need to get the raw tensor output (sample) from the encoder
+    vae_encoder.forward = lambda sample, return_dict: vae_encoder.encode(sample, return_dict)[0].sample()
+    onnx_export(
+        vae_encoder,
+        model_args=(
+            torch.randn(1, vae_in_channels, vae_sample_size, vae_sample_size).to(device=device, dtype=dtype),
+            False,
+        ),
+        output_path=output_path / "vae_encoder" / "model.onnx",
+        ordered_input_names=["sample", "return_dict"],
+        output_names=["latent_sample"],
+        dynamic_axes={
+            "sample": {0: "batch", 1: "channels", 2: "height", 3: "width"},
+        },
+        opset=opset,
+    )
+
+    # VAE DECODER
+    vae_decoder = pipeline.vae
+    vae_latent_channels = vae_decoder.config.latent_channels
+    vae_out_channels = vae_decoder.config.out_channels
+    # forward only through the decoder part
+    vae_decoder.forward = vae_encoder.decode
+    onnx_export(
+        vae_decoder,
+        model_args=(
+            torch.randn(1, vae_latent_channels, sample_size, sample_size).to(device=device, dtype=dtype),
+            False,
+        ),
+        output_path=output_path / "vae_decoder" / "model.onnx",
+        ordered_input_names=["latent_sample", "return_dict"],
+        output_names=["sample"],
+        dynamic_axes={
+            "latent_sample": {0: "batch", 1: "channels", 2: "height", 3: "width"},
+        },
+        opset=opset,
+    )
+    del pipeline.vae
+
+    onnx_pipeline = OnnxStableDiffusion3Pipeline(
+        vae_encoder=OnnxRuntimeModel.from_pretrained(output_path / "vae_encoder"),
+        vae_decoder=OnnxRuntimeModel.from_pretrained(output_path / "vae_decoder"),
+        text_encoder=OnnxRuntimeModel.from_pretrained(output_path / "text_encoder"),
+        tokenizer=pipeline.tokenizer,
+        text_encoder_2=OnnxRuntimeModel.from_pretrained(output_path / "text_encoder_2"),
+        tokenizer_2=pipeline.tokenizer_2,
+        text_encoder_3=OnnxRuntimeModel.from_pretrained(output_path / "text_encoder_3"),
+        tokenizer_3=pipeline.tokenizer_3,
+        transformer=OnnxRuntimeModel.from_pretrained(output_path / "transformer"),
+        scheduler=pipeline.scheduler,
+    )
+
+    onnx_pipeline.save_pretrained(output_path)
+    print("ONNX pipeline saved to", output_path)
+
+    del pipeline
+    del onnx_pipeline
+    _ = OnnxStableDiffusion3Pipeline.from_pretrained(output_path, provider="CPUExecutionProvider")
+    print("ONNX pipeline is loadable")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_path",
+        type=str,
+        required=True,
+        help="Path to the `diffusers` checkpoint to convert (either a local directory or on the Hub).",
+    )
+
+    parser.add_argument("--output_path", type=str, required=True, help="Path to the output model.")
+
+    parser.add_argument(
+        "--opset",
+        default=14,
+        type=int,
+        help="The version of the ONNX operator set to use.",
+    )
+    parser.add_argument("--fp16", action="store_true", default=False, help="Export the models in `float16` mode")
+
+    args = parser.parse_args()
+
+    convert_models(args.model_path, args.output_path, args.opset, args.fp16)

src/diffusers/__init__.py

@@ -439,6 +439,7 @@
             "OnnxStableDiffusionPipeline",
             "OnnxStableDiffusionUpscalePipeline",
             "StableDiffusionOnnxPipeline",
+            "OnnxStableDiffusion3Pipeline",
         ]
     )
 
@@ -878,6 +879,7 @@
             OnnxStableDiffusionPipeline,
             OnnxStableDiffusionUpscalePipeline,
             StableDiffusionOnnxPipeline,
+            OnnxStableDiffusion3Pipeline,
         )
 
     try:

src/diffusers/pipelines/__init__.py

@@ -352,6 +352,20 @@
         ]
     )
 
+try:
+    if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ..utils import dummy_torch_and_transformers_and_onnx_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_onnx_objects))
+else:
+    _import_structure["stable_diffusion_3"].extend(
+        [
+            "OnnxStableDiffusion3Pipeline",
+        ]
+    )
+
 try:
     if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
         raise OptionalDependencyNotAvailable()
@@ -675,6 +689,16 @@
                 StableDiffusionOnnxPipeline,
             )
 
+        try:
+            if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
+                raise OptionalDependencyNotAvailable()
+        except OptionalDependencyNotAvailable:
+            from ..utils.dummy_torch_and_transformers_and_onnx_objects import *
+        else:
+            from .stable_diffusion_3 import (
+                OnnxStableDiffusion3Pipeline,
+            )
+
         try:
             if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
                 raise OptionalDependencyNotAvailable()

src/diffusers/pipelines/stable_diffusion_3/__init__.py

@@ -8,6 +8,7 @@
     is_flax_available,
     is_torch_available,
     is_transformers_available,
+    is_onnx_available,
 )
 
 
@@ -27,6 +28,18 @@
     _import_structure["pipeline_stable_diffusion_3_img2img"] = ["StableDiffusion3Img2ImgPipeline"]
     _import_structure["pipeline_stable_diffusion_3_inpaint"] = ["StableDiffusion3InpaintPipeline"]
 
+try:
+    if not (is_transformers_available() and is_onnx_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_onnx_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_onnx_objects))
+else:
+    _import_structure["pipeline_onnx_stable_diffusion_3"] = [
+        "OnnxStableDiffusion3Pipeline",
+    ]
+
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
         if not (is_transformers_available() and is_torch_available()):
@@ -38,6 +51,14 @@
         from .pipeline_stable_diffusion_3_img2img import StableDiffusion3Img2ImgPipeline
         from .pipeline_stable_diffusion_3_inpaint import StableDiffusion3InpaintPipeline
 
+    try:
+        if not (is_transformers_available() and is_onnx_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_onnx_objects import *
+    else:
+        from .pipeline_onnx_stable_diffusion_3 import OnnxStableDiffusion3Pipeline
+
 else:
     import sys