Add CUDA EP in StableDiffusion demo (#17788)

Add CUDA EP to the demo of stable diffusion.

### A100 Performance
Test | Engine Property | Batch Size | TRT Latency (ms) | ORT_TRT Latency
(ms) | ORT_CUDA Latency (ms) | TORCH Latency (ms)
-- | -- | -- | -- | -- | -- | --
SD 1.5, 50 steps, 512x512 | Static Input Shape | 1 | 861 | 851 | 861 |
N/A
SD 1.5, 50 steps, 512x512 | Dynamic Input Shape, Optimized for batch
size 1 and image size 512x512 | 1 | 974 | 1079 | 928 | 1222
SD 1.5, 50 steps, 768x768 | Dynamic Input Shape, Optimized for batch
size 1 and image size 512x512 | 1 | 2492 | OOM | 1901 | 1971
SD 1.5, 50 steps, 768x768 | Dynamic Input Shape, Optimized for batch
size 1 and image size 512x512 | 4 |9091 | OOM | 6785 | 6700

We can see that ORT_CUDA is the most robust one for handling dynamic
input shape. PyTorch could be a good choice if you run large batch size.

The above result is from one A100-SXM4-80GB GPU (in
Standard_ND96amsr_A100_v4 Azure VM) with 50 steps to generate 512x512 or
768x768 images using StableDiffusion 1.5. Onnxruntime-gpu is built from
source, and the following packages or libraries are used in this test:
* tensorrt==8.6.1.post1
* torch==2.2.0.dev20230920+cu121
* transformers==4.31.0
* diffusers==0.19.3
* onnx==1.14.1
* onnx-graphsurgeon==0.3.27
* polygraphy==0.47.1
* protobuf==3.20.2
* onnxruntime-gpu==1.17.0 (built from source of main branch)
* CUDA 12.2.2
* cuDNN 8.9.5.29
* python 3.10.13

For static input shape, the engine is built with static batch size and
static image shape, and cuda graph is enabled.

For dynamic input shape, the engine is built to support dynamic batch
size and dynamic image shape, and cuda graph is disabled. The TensorRT
engine is built for batch size 1~4, image size 256x256 ~ 1024x1024, and
the optimized image size is 512x512.

The script to test static and dynamic input shape are like the
following:
```
prompt="a cute magical flying dog, fantasy art drawn by disney concept artists, highly detailed, digital paintining"
for e in TRT ORT_TRT ORT_CUDA
do
  python demo_txt2img.py --engine $e "$prompt"
  python demo_txt2img.py --engine $e --disable-cuda-graph --build-dynamic-batch --build-dynamic-shape "$prompt"
  python demo_txt2img.py --engine $e --disable-cuda-graph --build-dynamic-batch --build-dynamic-shape --height 768 --width 768 "$prompt"
done
```

Performance of PyTorch is from commands like the following:
```
python benchmark.py -e torch -v 1.5 --enable_torch_compile -b 1 --height 512 --width 512
python benchmark.py -e torch -v 1.5 --enable_torch_compile -b 1 --height 768 --width 768
python benchmark.py -e torch -v 1.5 --enable_torch_compile -b 4 --height 768 --width 768
```

onnxruntime/python/tools/transformers/models/stable_diffusion/README.md

@@ -72,38 +72,52 @@ cd onnxruntime/onnxruntime/python/tools/transformers/models/stable_diffusion
 
 Below is an example to optimize Stable Diffusion 1.5 in Linux. For Windows OS, please change the format of path to be like `.\sd` instead of `./sd`.
 
+It is recommended to create a Conda environment with Python 3.10 for the following setup:
+```
+conda create -n py310 python=3.10
+conda activate py310
+```
+
 ### Setup Environment (CUDA)
 
-It is recommended to create a Conda environment with Python 3.8, 3.9 or 3.10, and run the model with CUDA 11.8.
-If you use CUDA 12.*, you will need build onnxruntime-gpu from source.
+First, we need install CUDA 11.8 or 12.1, [cuDNN](https://docs.nvidia.com/deeplearning/cudnn/install-guide/index.html) 8.5 or above, and [TensorRT 8.6.1](https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html) in the machine.
+
+#### CUDA 11.8:
+
+In the Conda environment, install PyTorch 2.1 or above, and other required packages like the following:
 ```
-conda create -n py38 python=3.8
-conda activate py38
-pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu118
+pip install torch --index-url https://download.pytorch.org/whl/nightly/cu118
 pip install --upgrade polygraphy onnx-graphsurgeon --extra-index-url https://pypi.ngc.nvidia.com
-pip install -r requirements-cuda.txt
+pip install -r requirements-cuda11.txt
 ```
-ONNX Runtime requires CUDA and [cuDNN](https://developer.nvidia.com/rdp/cudnn-download) for GPU inference. CUDA 11.8 and cuDNN 8.5 or above are recommended.
 
-#### Install Nightly (Optional)
+We cannot directly `pip install tensorrt` for CUDA 11. Follow https://github.com/NVIDIA/TensorRT/issues/2773 to install TensorRT for CUDA 11 in Linux. For Windows, pip install the tensorrt wheel in the downloaded TensorRT zip file instead.
 
-Skip this step if you use onnxruntime-gpu package from official releases.
+#### CUDA 12.*:
+The official package of onnxruntime-gpu 1.16.* is built for CUDA 11.8. To use CUDA 12.*, you will need [build onnxruntime from source](https://onnxruntime.ai/docs/build/inferencing.html).
 
-To try latest optimizations, you can install [ort-nightly-gpu](https://aiinfra.visualstudio.com/PublicPackages/_artifacts/feed/ORT-Nightly/PyPI/ort-nightly-gpu/) package like the following:
+```
+git clone --recursive https://github.com/Microsoft/onnxruntime.git
+cd onnxruntime
+pip install -r requirements-dev.txt
+```
+Follow [example script for A100 in Ubuntu](https://github.com/microsoft/onnxruntime/blob/26a7b63716e3125bfe35fe3663ba10d2d7322628/build_release.sh)
+or [example script for RTX 4090 in Windows](https://github.com/microsoft/onnxruntime/blob/8df5f4e0df1f3b9ceeb0f1f2561b09727ace9b37/build_trt.cmd) to build and install onnxruntime-gpu wheel.
 
+Then install other python packages like the following:
 ```
-pip uninstall onnxruntime-gpu
-pip install ort-nightly-gpu -i https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT-Nightly/pypi/simple/
+pip install torch --index-url https://download.pytorch.org/whl/nightly/cu121
+pip install --upgrade polygraphy onnx-graphsurgeon --extra-index-url https://pypi.ngc.nvidia.com
+pip install -r requirements-cuda12.txt
 ```
+Finally, `pip install tensorrt` for Linux. For Windows, pip install the tensorrt wheel in the downloaded TensorRT zip file instead.
 
 ### Setup Environment (ROCm)
 
-It is recommended that the users run the model with ROCm 5.4 or newer and Python 3.8, 3.9 or 3.10.
+It is recommended that the users run the model with ROCm 5.4 or newer and Python 3.10.
 Note that Windows is not supported for ROCm at the moment.
 
 ```
-conda create -n py38 python=3.8
-conda activate py38
 wget https://repo.radeon.com/rocm/manylinux/rocm-rel-5.4/torch-1.12.1%2Brocm5.4-cp38-cp38-linux_x86_64.whl
 pip install torch-1.12.1+rocm5.4-cp38-cp38-linux_x86_64.whl
 pip install -r requirements-rocm.txt

onnxruntime/python/tools/transformers/models/stable_diffusion/demo_utils.py

@@ -34,12 +34,15 @@ class RawTextArgumentDefaultsHelpFormatter(argparse.ArgumentDefaultsHelpFormatte
 def parse_arguments(is_xl: bool, description: str):
     parser = argparse.ArgumentParser(description=description, formatter_class=RawTextArgumentDefaultsHelpFormatter)
 
+    engines = ["ORT_TRT", "TRT"] if is_xl else ["ORT_CUDA", "ORT_TRT", "TRT"]
+
     parser.add_argument(
         "--engine",
         type=str,
-        default="ORT_TRT",
-        choices=["ORT_TRT", "TRT"],
-        help="Backend engine. Default is OnnxRuntime CUDA execution provider.",
+        default=engines[0],
+        choices=engines,
+        help="Backend engine in {engines}. "
+        "ORT_CUDA is CUDA execution provider; ORT_TRT is Tensorrt execution provider; TRT is TensorRT",
     )
 
     supported_versions = PipelineInfo.supported_versions(is_xl)
@@ -106,7 +109,7 @@ def parse_arguments(is_xl: bool, description: str):
     parser.add_argument(
         "--onnx-opset",
         type=int,
-        default=17,
+        default=None,
         choices=range(14, 18),
         help="Select ONNX opset version to target for exported models.",
     )
@@ -163,6 +166,16 @@ def parse_arguments(is_xl: bool, description: str):
 
     args = parser.parse_args()
 
+    if (
+        args.engine in ["ORT_CUDA", "ORT_TRT"]
+        and (args.force_onnx_export or args.force_onnx_optimize)
+        and not args.force_engine_build
+    ):
+        raise ValueError(
+            "For ORT_CUDA or ORT_TRT, --force_onnx_export and --force_onnx_optimize are not supported. "
+            "Please use --force_engine_build instead."
+        )
+
     # Validate image dimensions
     if args.height % 8 != 0 or args.width % 8 != 0:
         raise ValueError(
@@ -173,6 +186,9 @@ def parse_arguments(is_xl: bool, description: str):
         print("[I] CUDA Graph is disabled since dynamic input shape is configured.")
         args.disable_cuda_graph = True
 
+    if args.onnx_opset is None:
+        args.onnx_opset = 14 if args.engine == "ORT_CUDA" else 17
+
     print(args)
 
     return args
@@ -197,7 +213,7 @@ def repeat_prompt(args):
 
 def init_pipeline(pipeline_class, pipeline_info, engine_type, args, max_batch_size, batch_size):
     onnx_dir, engine_dir, output_dir, framework_model_dir, timing_cache = get_engine_paths(
-        args.work_dir, pipeline_info, engine_type
+        work_dir=args.work_dir, pipeline_info=pipeline_info, engine_type=engine_type
     )
 
     # Initialize demo
@@ -214,7 +230,24 @@ def init_pipeline(pipeline_class, pipeline_info, engine_type, args, max_batch_si
         engine_type=engine_type,
     )
 
-    if engine_type == EngineType.ORT_TRT:
+    if engine_type == EngineType.ORT_CUDA:
+        # Build CUDA EP engines and load pytorch modules
+        pipeline.backend.build_engines(
+            engine_dir=engine_dir,
+            framework_model_dir=framework_model_dir,
+            onnx_dir=onnx_dir,
+            onnx_opset=args.onnx_opset,
+            opt_image_height=args.height,
+            opt_image_width=args.height,
+            opt_batch_size=batch_size,
+            force_engine_rebuild=args.force_engine_build,
+            device_id=torch.cuda.current_device(),
+            disable_cuda_graph_models=[
+                "clip2",  # TODO: Add ArgMax cuda kernel to enable cuda graph for clip2.
+                "unetxl",
+            ],
+        )
+    elif engine_type == EngineType.ORT_TRT:
         # Build TensorRT EP engines and load pytorch modules
         pipeline.backend.build_engines(
             engine_dir,

onnxruntime/python/tools/transformers/models/stable_diffusion/diffusion_models.py

@@ -298,9 +298,16 @@ def get_input_profile(self, batch_size, image_height, image_width, static_batch,
     def get_shape_dict(self, batch_size, image_height, image_width):
         return None
 
+    def fp32_input_output_names(self) -> List[str]:
+        """For CUDA EP, we export ONNX model with FP32 first, then convert it to mixed precision model.
+        This is a list of input or output names that are kept as float32 during converting.
+        For the first version, we will use same data type as TensorRT.
+        """
+        return []
+
     def optimize_ort(self, input_onnx_path, optimized_onnx_path, to_fp16=True):
         optimizer = self.get_ort_optimizer()
-        optimizer.optimize(input_onnx_path, optimized_onnx_path, to_fp16)
+        optimizer.optimize(input_onnx_path, optimized_onnx_path, to_fp16, keep_io_types=self.fp32_input_output_names())
 
     def optimize_trt(self, input_onnx_path, optimized_onnx_path):
         onnx_graph = onnx.load(input_onnx_path)
@@ -416,7 +423,7 @@ def get_sample_input(self, batch_size, image_height, image_width):
         self.check_dims(batch_size, image_height, image_width)
         return (torch.zeros(batch_size, self.text_maxlen, dtype=torch.int32, device=self.device),)
 
-    def add_hidden_states_graph_output(self, model: ModelProto, optimized_onnx_path):
+    def add_hidden_states_graph_output(self, model: ModelProto, optimized_onnx_path, use_external_data_format=False):
         graph: GraphProto = model.graph
         hidden_layers = -1
         for i in range(len(graph.node)):
@@ -457,7 +464,29 @@ def add_hidden_states_graph_output(self, model: ModelProto, optimized_onnx_path)
 
         onnx_model = OnnxModel(model)
         onnx_model.add_node(cast_node)
-        onnx_model.save_model_to_file(optimized_onnx_path)
+        onnx_model.save_model_to_file(optimized_onnx_path, use_external_data_format=use_external_data_format)
+
+    def optimize_ort(self, input_onnx_path, optimized_onnx_path, to_fp16=True):
+        optimizer = self.get_ort_optimizer()
+        if not self.output_hidden_state:
+            optimizer.optimize(
+                input_onnx_path, optimized_onnx_path, to_fp16, keep_io_types=[], keep_outputs=["text_embeddings"]
+            )
+        else:
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                # Save to a temporary file so that we can load it with Onnx Runtime.
+                logger.info("Saving a temporary model to add hidden_states to graph output ...")
+                tmp_model_path = os.path.join(tmp_dir, "model.onnx")
+
+                model = onnx.load(input_onnx_path)
+                self.add_hidden_states_graph_output(model, tmp_model_path, use_external_data_format=True)
+                optimizer.optimize(
+                    tmp_model_path,
+                    optimized_onnx_path,
+                    to_fp16,
+                    keep_io_types=[],
+                    keep_outputs=["text_embeddings", "hidden_states"],
+                )
 
     def optimize_trt(self, input_onnx_path, optimized_onnx_path):
         onnx_graph = onnx.load(input_onnx_path)
@@ -598,6 +627,9 @@ def get_sample_input(self, batch_size, image_height, image_width):
             torch.randn(2 * batch_size, self.text_maxlen, self.embedding_dim, dtype=dtype, device=self.device),
         )
 
+    def fp32_input_output_names(self) -> List[str]:
+        return ["sample", "timestep"]
+
 
 class UNetXL(BaseModel):
     def __init__(
@@ -703,6 +735,9 @@ def get_sample_input(self, batch_size, image_height, image_width):
             },
         )
 
+    def fp32_input_output_names(self) -> List[str]:
+        return ["sample", "timestep"]
+
 
 # VAE Decoder
 class VAE(BaseModel):
@@ -773,6 +808,9 @@ def get_sample_input(self, batch_size, image_height, image_width):
         latent_height, latent_width = self.check_dims(batch_size, image_height, image_width)
         return (torch.randn(batch_size, 4, latent_height, latent_width, dtype=torch.float32, device=self.device),)
 
+    def fp32_input_output_names(self) -> List[str]:
+        return ["latent", "images"]
+
 
 def get_tokenizer(pipeline_info: PipelineInfo, framework_model_dir, hf_token, subfolder="tokenizer"):
     tokenizer_dir = os.path.join(framework_model_dir, pipeline_info.name(), subfolder)

onnxruntime/python/tools/transformers/models/stable_diffusion/engine_builder.py

@@ -77,9 +77,8 @@ def get_cached_model_name(self, model_name):
 
     def get_onnx_path(self, model_name, onnx_dir, opt=True):
         engine_name = self.engine_type.name.lower()
-        onnx_model_dir = os.path.join(
-            onnx_dir, self.get_cached_model_name(model_name) + (f".{engine_name}" if opt else "")
-        )
+        directory_name = self.get_cached_model_name(model_name) + (f".{engine_name}" if opt else "")
+        onnx_model_dir = os.path.join(onnx_dir, directory_name)
         os.makedirs(onnx_model_dir, exist_ok=True)
         return os.path.join(onnx_model_dir, "model.onnx")