return classpath.removeIf(new Spec<File>() {
                @Override
                public boolean isSatisfiedBy(File file) {
                    return file.getName().startsWith("jaxb-impl-");
                }
            });
        }
        return classpath;
    }

    private Set<Module> allRequiredModulesOf(String... names) {
        Set<Module> modules = new HashSet<>();

     * Base64 encode user info for HTTP Basic Authentication.
     *
     * Try to use {@literal java.util.Base64} encoder which is available starting with Java 8.
     * Fallback to {@literal javax.xml.bind.DatatypeConverter} from JAXB which is available starting with Java 6 but is not anymore in Java 9.
     * Fortunately, both of these two Base64 encoders implement the right Base64 flavor, the one that does not split the output in multiple lines.
     *

      </component>
      <component group="javax.xml.bind" name="jaxb-api" version="2.3.0">
         <artifact name="jaxb-api-2.3.0.jar">
            <pgp value="B38E195D438D44A1A8D2D2203575E1C767076CA8"/>
         </artifact>
      </component>
      <component group="javax.xml.bind" name="jaxb-api" version="2.3.1">
         <artifact name="jaxb-api-2.3.1.jar">
            <pgp value="70CD19BFD9F6C330027D6F260315BFB7970A144F"/>

				<exclusion>
					<groupId>javax.activation</groupId>
					<artifactId>javax.activation-api</artifactId>
				</exclusion>
				<exclusion>
					<groupId>javax.xml.bind</groupId>
					<artifactId>jaxb-api</artifactId>
				</exclusion>
				<exclusion>
					<groupId>jakarta.annotation</groupId>
					<artifactId>jakarta.annotation-api</artifactId>
				</exclusion>
				<exclusion>

AR_IMAGE_PATH="us-central1-docker.pkg.dev/tensorflow-sigs/tensorflow/build-arm64"

# Build for both JAX and TF usage.  We do these in one place because they share
# almost all of the same cache layers
export DOCKER_BUILDKIT=1
for target in jax tf; do
  IMAGE="gcr.io/tensorflow-sigs/build-arm64:$target-$TAG"
  AR_IMAGE="$AR_IMAGE_PATH:$target-$TAG"
  docker pull "$IMAGE" || true

limitations under the License.
==============================================================================*/

// The full pipeline of converting jax random include 2 steps.
// 1. Rename the jax random functions to tflite wrapped functions with the aid
//    of "jax.named_call". For example, in the dumped hlo, the
//    jax.random.uniform will have name "tfl_wrapped_jax_random_uniform".
// 2. Replace the body of "tfl_wrapped_jax_random_uniform" and

// CHECK:           return %[[VAL_2]] : tuple<tensor<3x4xf32>>
// CHECK:         }
func.func @tfl_wrapped_jax_random_normal(%arg0: tensor<2xui32>) -> tuple<tensor<3x4xf32>> {
  // This is a fake jax random normal body.
  %0 = stablehlo.constant dense<0.0> : tensor<12xf32>
  %1 = "stablehlo.reshape"(%0) : (tensor<12xf32>) -> tensor<3x4xf32>
  %2 = "stablehlo.tuple"(%1) : (tensor<3x4xf32>) -> tuple<tensor<3x4xf32>>

        ConstBoolAttrTrue,
        (returnType (GetNhwcReturnTypeFromNchw $old_val))),
      (Arith_ConstantOp
        ConstantAttr<RankedI32ElementsAttr<[4]>,"{0, 3, 1, 2}">))>;

// pattern to lower a stablehlo.composite of `jax` and `pytorch` image resize fuctions
// in `nearest` mode and with NHWC inputs to a tflite.resize_nearest_neighbor op.
def LegalizeNhwcResizeNearestNeighbor2dComposite: Pat<
    (MHLO_CompositeOp:$old_val
    (variadic $input),

keras-nightly ~= 3.0.0.dev
tb-nightly ~= 2.18.0.a

# Test dependencies
grpcio >= 1.24.3, < 2.0
portpicker == 1.6.0
scipy == 1.11.3
requests >= 2.31.0
packaging==23.2
setuptools==68.2.2

    #   h5py
    #   jax
    #   keras-nightly
    #   ml-dtypes
    #   opt-einsum
    #   scipy
    #   tb-nightly
opt-einsum==3.3.0 \
    --hash=sha256:2455e59e3947d3c275477df7f5205b30635e266fe6dc300e3d9f9646bfcea147 \
    --hash=sha256:59f6475f77bbc37dcf7cd748519c0ec60722e91e63ca114e68821c0c54a46549
    # via
    #   -r requirements.in
    #   jax
packaging==23.2 \