#### Debug builds

When [building Tensorflow](https://www.tensorflow.org/install/source), passing
`--config=dbg` to Bazel will build with debugging information and without
optimizations, allowing you to use GDB or other debuggers to debug C++ code. For
example, you can build the pip package with debugging information by running:

```bash

   * includes exceptional exits, so all control paths involving signalling must be protected by a
   * finally block.
   *
   * Further optimizations of this algorithm become increasingly subtle. A wait that terminates
   * without the guard being satisfied (due to timeout, but not interrupt) can then immediately exit

build:cuda --@local_config_cuda//:enable_cuda

# CUDA: This config refers to building CUDA op kernels with clang.
build:cuda_clang --config=cuda
# Enable TensorRT optimizations https://developer.nvidia.com/tensorrt
build:cuda_clang --config=tensorrt
build:cuda_clang --action_env=TF_CUDA_CLANG="1"
build:cuda_clang --@local_config_cuda//:cuda_compiler=clang

Anyway, in any of the cases above, FastAPI will still work asynchronously and be extremely fast.

But by following the steps above, it will be able to do some performance optimizations.

## Technical Details

Modern versions of Python have support for **"asynchronous code"** using something called **"coroutines"**, with **`async` and `await`** syntax.

     * system defaults will be used.
     *
     * Most applications should not call this method, and instead use the system defaults. Those
     * classes include special optimizations that can be lost if the implementations are decorated.
     *
     * If necessary, you can create and configure the defaults yourself with the following code:
     *
     * ```java

    }
    return (ListenableFuture<? extends T>[]) collection.toArray(new ListenableFuture<?>[0]);
  }

  // This can't be a TrustedFuture, because TrustedFuture has clever optimizations that
  // mean cancel won't be called if this Future is passed into setFuture, and then
  // cancelled.
  private static final class InCompletionOrderFuture<T extends @Nullable Object>
      extends AbstractFuture<T> {

-       | ((const unsigned char *) (const char *) (src))[idx + 2]) << 8	      \
-      | ((const unsigned char *) (const char *) (src))[idx + 3])
-# endif
 #else
 /* These are a few types we need for the optimizations if we cannot
    use unaligned memory accesses.  */
@@ -94,148 +72,11 @@ __STRING2_COPY_TYPE (8);
 
 /* Set N bytes of S to C.  */
 #if !defined _HAVE_STRING_ARCH_memset
-# if !__GNUC_PREREQ (3, 0)

// --------------------------------------------------------------------------
TF_SessionOptions* TF_NewSessionOptions() {
  TF_SessionOptions* out = new TF_SessionOptions;
  // Disable optimizations for static graph to allow calls to Session::Extend.
  out->options.config.mutable_experimental()
      ->set_disable_optimize_for_static_graph(true);
  return out;
}

<li>The global generator accessed by top-level functions is unconditionally randomly seeded.
Because the API guarantees no fixed sequence of results,
optimizations like per-thread random generator states are now possible.

<li>The <a href="/pkg/math/rand/v2/#Source"><code>Source</code></a>
interface now has a single <code>Uint64</code> method;

/// argument list with the out arguments. We only use the return type for an out
/// parameter if that is a plain C type, as this ensures ABI compatibility
/// (returning structures has issues in case compiler options affect
/// optimizations such as RVO). If a status needs to be returned from these
/// methods, the last argument is always a `TF_Status *` (or an array of such
/// pointers) owned by core TensorFlow and guaranteed to not be `nullptr`.
///