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.

   * 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

    }
    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> {

     * 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

// --------------------------------------------------------------------------
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;
}

-       | ((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)

      generateCancellationCauses = false;
    }
    GENERATE_CANCELLATION_CAUSES = generateCancellationCauses;
  }

  /**
   * Tag interface marking trusted subclasses. This enables some optimizations. The implementation
   * of this interface must also be an AbstractFuture and must not override or expose for overriding
   * any of the public methods of ListenableFuture.
   */

/// 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`.
///

#include "tensorflow/c/eager/tfe_tensorhandle_internal.h"
#include "tensorflow/core/common_runtime/eager/eager_operation.h"
#include "tensorflow/core/common_runtime/function_optimization_registry.h"
#include "tensorflow/core/common_runtime/optimization_registry.h"
#include "tensorflow/core/distributed_runtime/rpc/grpc_server_lib.h"
#include "tensorflow/core/framework/function.h"
#include "tensorflow/core/graph/graph.h"