absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    /* Given upstream grad U and a matmul op A*B, the gradients are:
     *
     *    dA = U * B.T
     *    dB = A.T * U
     *
     *    where A.T means `transpose(A)`
     */
    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    // Get transpose attrs

  switch (dtype.code) {
    case DLDataTypeCode::kDLBool:
      if (dtype.bits != 8) {
        return tensorflow::errors::InvalidArgument(
            "Only DLPack bools of bitwidth 8 are supported, got: ", dtype.bits);
      }
      *tf_dtype = TF_DataType::TF_BOOL;
      return absl::OkStatus();

    case DLDataTypeCode::kDLUInt:
      switch (dtype.bits) {
        case 8:

function resultstore_extract_fallback {
  # In case the main script fails somehow.
  cat <<EOF
IMPORTANT: For bazel invocations that uploaded to ResultStore (e.g. RBE), you
can view more detailed results that are probably easier to read than this log.
Try the links below:
EOF
  # Find any "Streaming build results to" line, then print the last word in it,
  # and don't print duplicates

//
// This has three benefits:
//   * allows having newer plugins than the current core TensorFlow: the
//     additional entries in the plugin's table are just discarded;
//   * allows having older plugins than the current core TensorFlow (though
//     we are still warning users): the entries that core TensorFlow expects
//     but plugins didn't provide will be set to `nullptr` values and core

 different computation approaches and orders.
    *   To verify if oneDNN optimizations are on, look for a message with *"oneDNN custom operations are on"* in the log. If the exact phrase is not there, it means they are off.

* Making the `tf.function` type system fully available:

# Experimental filesystem C APIs for TensorFlow.
# Will be moved in proper place once all filesystems are converted to the
# modular framework.
load("//tensorflow:tensorflow.bzl", "tf_cc_test")
load("//tensorflow/core/platform:rules_cc.bzl", "cc_library")

package(
    # copybara:uncomment default_applicable_licenses = ["//tensorflow:license"],
    licenses = ["notice"],
)

# This is only for plugins
cc_library(
    name = "filesystem_interface",

/// This file builds classes needed to hold a filesystem implementation in the
/// modular world. Once all TensorFlow filesystems are converted to use the
/// plugin based approach, this file will replace the one in core/platform and
/// the names will lose the `Modular` part. Until that point, the `Modular*`
/// classes here are experimental and subject to breaking changes.
/// For documentation on these methods, consult `core/platform/filesystem.h`.

      "keras.api._v1.keras.__internal__.legacy.rnn_cell")
_current_module.nn.__path__ = [_module_dir] + _current_module.nn.__path__

del importlib

# Load all plugin libraries from site-packages/tensorflow-plugins if we are
# running under pip.
# TODO(gunan): Find a better location for this code snippet.
from tensorflow.python.framework import load_library as _ll
from tensorflow.python.lib.io import file_io as _fi

gast == 0.4.0
termcolor == 2.3.0
wrapt == 1.16.0
tblib == 2.0.0

# Install tensorboard, and keras
# Note that here we want the latest version that matches TF major.minor version
# Note that we must use nightly here as these are used in nightly jobs
# For release jobs, we will pin these on the release branch
keras-nightly ~= 3.0.0.dev
tb-nightly ~= 2.17.0.a

# Test dependencies
grpcio >= 1.24.3, < 2.0
portpicker == 1.6.0
scipy == 1.11.3

      const = 0;
  virtual Status SetInput(size_t index,
                          ImmediateExecutionTensorHandle* input) = 0;

  virtual ImmediateExecutionContext* GetContext() const = 0;

  // Following two methods are used to support custom device.
  // Return true if the inputs contain custom device tensor handle. It means
  // that the argument need to be handled by a custom device.
  virtual bool HasCustomDeviceInput() const = 0;