* and from it, import the function `get_token_header`.

That would refer to some package above `app/`, with its own file `__init__.py`, etc. But we don't have that. So, that would throw an error in our example. 🚨

But now you know how it works, so you can use relative imports in your own apps no matter how complex they are. 🤓

### Add some custom `tags`, `responses`, and `dependencies`

                    /*expected_shape*/ make_shape({kUnknownDim, kUnknownDim}));

  TFE_DeleteOp(matmul_op);
  // TODO(bgogul): Add some death tests where status is not OK.
}

TEST_F(ShapeInferenceTest, InfersShapesFromInputTensors) {
  // Prepare some tensors for shape.
  TF_Tensor* tensor_1X6 = Int32Tensor({1, 6});
  CHECK_EQ(TF_OK, TF_GetCode(status_)) << TF_Message(status_);

Then your API will (let's imagine):

* Send the invoice to some customer of the external developer.
* Collect the money.
* Send a notification back to the API user (the external developer).
    * This will be done by sending a POST request (from *your API*) to some *external API* provided by that external developer (this is the "callback").

## The normal **FastAPI** app

## Bodies of arbitrary `dict`s

You can also declare a body as a `dict` with keys of some type and values of some other type.

This way, you don't have to know beforehand what the valid field/attribute names are (as would be the case with Pydantic models).

	// If x-amz-content-sha256 is set and the value is not
	// 'UNSIGNED-PAYLOAD' we should validate the content sha256.
	switch v[0] {
	case unsignedPayload, unsignedPayloadTrailer:
		return true
	case emptySHA256:
		// some broken clients set empty-sha256
		// with > 0 content-length in the body,
		// we should skip such clients and allow
		// blindly such insecure clients only if
		// S3 strict compatibility is disabled.

# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
# This script dumps some information about the environment. It's most useful
# for verifying changes to the TFCI scripts system, and most users won't need
# to interact with it at all.
source "${BASH_SOURCE%/*}/utilities/setup.sh"

 * limitations under the License.
 */

plugins {
    id("io.gitlab.arturbosch.detekt")
}

detekt {
    // enable all default rules
    buildUponDefaultConfig = true

    // customize some of the rules, until we can fix the offending cases
    config.convention(project.isolated.rootProject.projectDirectory.file("gradle/detekt.yml"))

    // also check the project build file
    source.from(project.buildFile)
}

            bytes,
            Doc(
                """
                The bytes to write to the file.
                """
            ),
        ],
    ) -> None:
        """
        Write some bytes to the file.

        You normally wouldn't use this from a file you read in a request.

        To be awaitable, compatible with async, this is run in threadpool.
        """
        return await super().write(data)

```

In this example, we didn't declare any Pydantic model. In fact, the request body is not even <abbr title="converted from some plain format, like bytes, into Python objects">parsed</abbr> as JSON, it is read directly as `bytes`, and the function `magic_data_reader()` would be in charge of parsing it in some way.

Nevertheless, we can declare the expected schema for the request body.

### Custom OpenAPI content type

#ifndef TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_
#define TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_

// Run a function containing a MatMul op and check its output.
// If heavy_load_on_streaming_rpc is true, send some rpc requests before the one
// which creates a remote input, to simulate a scenario that the remote input
// is not ready when we start running an op or a function.