Add example deployment scripts (#553)

.pre-commit-config.yaml

@@ -28,7 +28,7 @@ repos:
       types_or: [python, jupyter]
 
   - repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.7.1
+    rev: v1.8.0
     hooks:
     - id: mypy
       entry: python3 -m mypy --config-file pyproject.toml

deploy/README.md

@@ -0,0 +1,53 @@
+# Model Deployment with BentoML and Triton Inference Server
+
+1. Install the required dependencies with poetry.
+```bash
+poetry install --with deploy
+```
+2. Serialize trained model and move it to the `model_repo` directory. Then create
+   a `config.pbtxt` file for the model.
+
+   **Example - torchxrayvision model**
+   ```python
+   import torch
+   import torchxrayvision as xrv
+
+
+   model = xrv.models.ResNet(weights="resnet50-res512-all").eval().cuda()
+
+   dummy_input = (-1024 - 1024) * torch.rand(1, 1, 512, 512) + 1024
+   dummy_input = dummy_input.cuda()
+
+   torch.jit.trace(model, dummy_input).save("model_repo/resnet50_res512_all/1/model.pt")
+   ```
+   See `model_repo/resnet50_res512_all/config.pbtxt` for an example of a pytorch model configuration file.
+
+   **Example - sklearn model**
+   ```python
+   from skl2onnx import to_onnx
+
+
+   onnx_model = to_onnx(
+      <sklearn_model>,
+      <input_data>,
+      options={"zipmap": False},
+   )
+   with open("model_repo/<model_name>/1/model.onnx", "wb") as f:
+      f.write(onnx_model.SerializeToString())
+   ```
+   See `model_repo/heart_failure_prediction/config.pbtxt` for an example of an ONNX model configuration file.
+3. Create a service with BentoML with a triton runner. See `service.py` for an example.
+4. Define a bentofile to specify which files to include in the bento. See `bentofile.yaml` for an example.
+5. Build a bento.
+```bash
+bentoml build --do-not-track
+```
+6. Containerize the bento.
+```bash
+bentoml containerize -t model-service:alpha --enable-features=triton --do-not-track model-service:latest
+```
+
+7. Run the container with docker.
+```bash
+docker run -d --gpus=1 --rm -p 3000:3000 model-service:alpha
+```

deploy/bentofile.yaml

@@ -0,0 +1,12 @@
+service: service:svc
+include:
+  - /model_repo
+  - /*.py
+exclude:
+  - /__pycache__
+python:
+  packages:
+  - bentoml[triton]
+  - torchxrayvision==1.2.1
+docker:
+  base_image: nvcr.io/nvidia/tritonserver:24.01-py3

deploy/deployment.yaml

@@ -0,0 +1,5 @@
+monitoring:
+  enabled: true
+  type: default
+  options:
+    output_dir: ./monitoring

deploy/model_repo/densenet121_res224_all/config.pbtxt

@@ -0,0 +1,41 @@
+backend: "pytorch"
+name: "densenet121_res224_all"
+max_batch_size: 64
+dynamic_batching {
+    max_queue_delay_microseconds: 100
+}
+input {
+ name: "INPUT__0"
+ data_type: TYPE_FP32
+ dims: 1
+ dims: 224
+ dims: 224
+}
+output {
+ name: "OUTPUT__0"
+ data_type: TYPE_FP32
+ dims: -1
+ dims: 18
+}
+instance_group [
+  {
+    count: 1
+    kind: KIND_GPU
+    gpus: [0]
+  }
+]
+model_warmup [
+  {
+    name : "random sample"
+    count: 1
+    batch_size: 1
+    inputs {
+      key: "INPUT__0"
+      value: {
+        data_type: TYPE_FP32
+        dims: [1, 224, 224]
+        random_data: true
+      }
+    }
+  }
+]

deploy/model_repo/heart_failure_prediction/config.pbtxt

@@ -0,0 +1,24 @@
+backend: "onnxruntime"
+name: "heart_failure_prediction"
+max_batch_size: 0
+input {
+  name: "X"
+  data_type: TYPE_FP32
+  dims: [-1, 21]
+}
+output {
+  name: "label"
+  data_type: TYPE_INT64
+  dims: -1
+}
+instance_group [
+  {
+    count: 1
+    kind: KIND_CPU
+  }
+]
+optimization { execution_accelerators {
+  cpu_execution_accelerator : [ {
+    name : "openvino"
+  }]
+}}

deploy/model_repo/resnet50_res512_all/config.pbtxt

@@ -0,0 +1,38 @@
+backend: "pytorch"
+name: "resnet50_res512_all"
+max_batch_size: 32
+dynamic_batching {
+    max_queue_delay_microseconds: 100
+}
+input {
+ name: "INPUT__0"
+ data_type: TYPE_FP32
+ dims: 1
+ dims: 512
+ dims: 512
+}
+output {
+ name: "OUTPUT__0"
+ data_type: TYPE_FP32
+ dims: -1
+ dims: 18
+}
+instance_group [
+  {
+    count: 1
+    kind: KIND_GPU
+    gpus: [0]
+  }
+]
+model_warmup [{
+  name : "random sample"
+  batch_size: 1
+  inputs {
+    key: "INPUT__0"
+    value: {
+      data_type: TYPE_FP32
+      dims: [1, 512, 512]
+      random_data: true
+    }
+  }
+}]

deploy/service.py

@@ -0,0 +1,115 @@
+"""Model serving service with Triton Inference Server as backend."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any, Literal
+
+import bentoml
+import numpy as np
+import torchxrayvision as xrv
+from torchvision import transforms
+
+
+if TYPE_CHECKING:
+    from PIL.Image import Image
+
+
+def get_transform(image_size: int) -> transforms.Compose:
+    """Get image transformation for model inference."""
+    return transforms.Compose(
+        [
+            xrv.datasets.XRayCenterCrop(),
+            xrv.datasets.XRayResizer(image_size),
+        ],
+    )
+
+
+triton_runner = bentoml.triton.Runner(
+    "triton_runner",
+    "src/model_repo",
+    tritonserver_type="http",
+    cli_args=[
+        "--exit-on-error=true",  # exits if any error occurs during initialization
+        "--http-restricted-api=model-repository:access-key=admin",  # restrict access to load/unload APIs
+        "--model-control-mode=explicit",  # enable explicit model loading/unloading
+        "--load-model=resnet50_res512_all",
+    ],
+)
+svc = bentoml.Service("model-service", runners=[triton_runner])
+
+
+@svc.api(  # type: ignore
+    input=bentoml.io.Multipart(im=bentoml.io.Image(), model_name=bentoml.io.Text()),
+    output=bentoml.io.JSON(),
+)
+async def classify_xray(im: Image, model_name: str) -> dict[str, float]:
+    """Classify X-ray image using specified model."""
+    img = np.asarray(im)
+    img = xrv.datasets.normalize(
+        img,
+        img.max(),
+        reshape=True,  # normalize image to [-1024, 1024]
+    )
+
+    model_repo_index = await triton_runner.get_model_repository_index()
+    available_models = [model["name"] for model in model_repo_index]
+    if model_name not in available_models:
+        raise bentoml.exceptions.InvalidArgument(
+            f"Expected model name to be one of {available_models}, but got {model_name}",
+        )
+
+    img_size = 224
+    if "resnet" in model_name:
+        img_size = 512
+
+    img = get_transform(img_size)(img)
+
+    if len(img.shape) == 3:
+        img = img[None]  # add batch dimension
+
+    InferResult = await getattr(triton_runner, model_name).async_run(img)  # noqa: N806
+    return dict(
+        zip(xrv.datasets.default_pathologies, InferResult.as_numpy("OUTPUT__0")[0]),
+    )
+
+
+@svc.api(  # type: ignore
+    input=bentoml.io.NumpyNdarray(dtype="float32", shape=(-1, 21)),
+    output=bentoml.io.NumpyNdarray(dtype="int64", shape=(-1,)),
+)
+async def predict_heart_failure(X: np.ndarray) -> np.ndarray:  # type: ignore
+    """Run inference on heart failure prediction model."""
+    InferResult = await triton_runner.heart_failure_prediction.async_run(  # noqa: N806
+        X,
+    )
+    return InferResult.as_numpy("label")  # type: ignore[no-any-return]
+
+
+# Triton Model management API
+@svc.api(input=bentoml.io.JSON(), output=bentoml.io.JSON())  # type: ignore
+async def model_config(input_model: dict[Literal["model_name"], str]) -> dict[str, Any]:
+    """Retrieve model configuration from Triton Inference Server."""
+    return await triton_runner.get_model_config(input_model["model_name"])  # type: ignore
+
+
+@svc.api(input=bentoml.io.Text(), output=bentoml.io.JSON())  # type: ignore
+async def unload_model(input_model: str, ctx: bentoml.Context) -> dict[str, str]:
+    """Unload a model from memory."""
+    await triton_runner.unload_model(
+        input_model,
+        headers=ctx.request.headers,
+    )  # noqa: E501
+    return {"unloaded": input_model}
+
+
+@svc.api(input=bentoml.io.Text(), output=bentoml.io.JSON())  # type: ignore
+async def load_model(input_model: str, ctx: bentoml.Context) -> dict[str, str]:
+    """Load a model into memory."""
+    await triton_runner.load_model(input_model, headers=ctx.request.headers)
+    return {"loaded": input_model}
+
+
+@svc.api(input=bentoml.io.Text(), output=bentoml.io.JSON())  # type: ignore
+async def list_models(_: str, ctx: bentoml.Context) -> list[str]:
+    """Return a list of models available in the model repository."""
+    return await triton_runner.get_model_repository_index(headers=ctx.request.headers)  # type: ignore