```

## Verify the `username` and data shape

We verify that we get a `username`, and extract the scopes.

And then we validate that data with the Pydantic model (catching the `ValidationError` exception), and if we get an error reading the JWT token or validating the data with Pydantic, we raise the `HTTPException` we created before.

For that, we update the Pydantic model `TokenData` with a new property `scopes`.

//
// Below here we do the gradient algorithm. It works as follows:
//
// First we filter the tape to just the subset of operations we want to
// differentiate. In the process of doing so we count how many times each Tensor
// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output

We are going to include this `APIRouter` in the main `FastAPI` app, but first, let's check the dependencies and another `APIRouter`.

## Dependencies

We see that we are going to need some dependencies used in several places of the application.

So we put them in their own `dependencies` module (`app/dependencies.py`).

   * methods to accept null. We could avoid having unchecked casts in Converter.java itself if we
   * could perform a cast to LegacyConverter, but we can't because it's an internal-only class.
   *
   * TODO(cpovirk): So make it part of the open-source build, albeit package-private there?
   *
   * So we use uncheckedCastNullableTToT here. This is a weird usage of that method: The method is

      if (retryRoute != null) {
        // Lock in the route because retryRoute() is racy and we don't want to call it twice.
        nextRouteToTry = retryRoute
        return true
      }
    }

    // If we have a routes left, use 'em.
    if (routeSelection?.hasNext() == true) return true

    // If we haven't initialized the route selector yet, assume it'll have at least one route.

// scanners.
func (s *Scanner) Scan() bool {
	if s.done {
		return false
	}
	s.scanCalled = true
	// Loop until we have a token.
	for {
		// See if we can get a token with what we already have.
		// If we've run out of data but have an error, give the split function
		// a chance to recover any remaining, possibly empty token.
		if s.end > s.start || s.err != nil {

         */
        return 0;
    }


    /*
     * We overload this method from ServerMessageBlock because
     * we want writeAndXWireFormat to write the parameterWords
     * and bytes. This is so we can write batched smbs because
     * all but the first smb of the chaain do not have a header
     * and therefore we do not want to writeHeaderWireFormat. We
     * just recursivly call writeAndXWireFormat.
     */

```

There are some differences with the code for the SQLAlchemy tutorial.

We don't pass a `db` attribute around. Instead we use the models directly. This is because the `db` object is a global object, that includes all the connection logic. That's why we had to do all the `contextvars` updates above.

Aso, when returning several objects, like in `get_users`, we directly call `list`, like in:

```Python
list(models.User.select())
```

    using a job called "copybara".

**5. Copy to Google Internal codebase and run internal CI**

-   Once the PR is in the Google codebase, we make sure it integrates well with
    its dependencies and the rest of the system.
-   Rarely, If the tests fail at this stage, we cannot merge the code.
-   If needed, we may come to you to make some changes. At times, it may not be

  /** Ciphers in order we observed directly on the socket. */
  private lateinit var handshakeEnabledCipherSuites: List<String>

  /** Ciphers in order we observed on sslSocketFactory defaults. */
  private lateinit var defaultEnabledCipherSuites: List<String>

  /** Ciphers in order we observed on sslSocketFactory supported. */