// function. Otherwise return nullptr in result and error Status from the
// function.
//
// If `library_filename` has already been loaded, we return a cached handle.
// Device and Kernels/Ops are registered as globals when a library is loaded
// for the first time.
TF_Library* TF_LoadPluggableDeviceLibrary(const char* library_filename,
                                          TF_Status* status) {

        assertNotNull(hash);
        // For lmCompatibility 3,4,5 (NTLMv2), returns empty array
        assertEquals(0, hash.length);
    }

    // Test that changing lmCompatibility at runtime doesn't affect already loaded static value
    @ParameterizedTest
    @ValueSource(strings = { "0", "1", "2", "3", "4", "5" })
    void testLmCompatibilityStaticInitialization(String lmCompatibility) {

            extractorMap.put(key, newExtractors);
        } else {
            extractorMap.put(key, new Extractor[] { extractor });
        }

        if (logger.isDebugEnabled()) {
            logger.debug("Loaded {} : {}", key, extractor.getClass().getName());
        }
    }

    /**
     * Adds an extractor to the factory for all keys in the provided list.
     *

# TF_LoadLibrary/tf.load_op_library) register their ops and kernels with this
# shared object directly.
#
# For example, from Python tf.load_op_library loads a custom op library (via
# dlopen() on Linux), the library finds libtensorflow_framework.so (no
# filesystem search takes place, since libtensorflow_framework.so has already

    return statsCounterSupplier;
  }

  /**
   * Builds a cache, which either returns an already-loaded value for a given key or atomically
   * computes or retrieves it using the supplied {@code CacheLoader}. If another thread is currently
   * loading the value for this key, simply waits for that thread to finish and returns its loaded
   * value. Note that multiple threads can concurrently load values for distinct keys.
   *

          }
          // TODO(cpovirk): if reference has been cleared, remove it?
        }
        checkExceptionClassValidity(exceptionClass);

        /*
         * It's very unlikely that any loaded Futures class will see getChecked called with more
         * than a handful of exceptions. But it seems prudent to set a cap on how many we'll cache.

          }
          // TODO(cpovirk): if reference has been cleared, remove it?
        }
        checkExceptionClassValidity(exceptionClass);

        /*
         * It's very unlikely that any loaded Futures class will see getChecked called with more
         * than a handful of exceptions. But it seems prudent to set a cap on how many we'll cache.

             to quit<b>)</b>
     <span style="background-color:#007166"><font color="#D3D7CF"> INFO </font></span>  Started reloader process <b>[</b><font color="#34E2E2"><b>383138</b></font><b>]</b> using WatchFiles

                    if (!info.isExpired()) {
                        handles.put(info.getPath(), info);
                        guidToHandle.put(info.getCreateGuid(), info);
                        log.info("Loaded persisted handle: {}", info.getPath());
                    } else {
                        Files.delete(handleFile);
                        log.debug("Deleted expired persisted handle: {}", handleFile);

That's what we'll solve, let's load the model before the requests are handled, but only right before the application starts receiving requests, not while  the code is being loaded.

## Lifespan { #lifespan }

You can define this *startup* and *shutdown* logic using the `lifespan` parameter of the `FastAPI` app, and a "context manager" (I'll show you what that is in a second).