1. Creating a PR and observing the presubmit test results
2. Running the CI scripts locally, as explained below
3. **Google employees only**: Google employees can use an internal-only tool
called "MLCI" that makes testing more convenient: it can execute any full CI job
against a pending change. Search for "MLCI" internally to find it.

You may invoke a CI script of your choice by following these instructions:

```bash

```
--repo_env=TF_PYTHON_VERSION=3.11
```

## Requirements updater

Requirements updater is a standalone tool, intended to simplify process of
updating requirements for multiple minor versions of Python.

It takes in a file with a set of dependencies, and produces a more detailed
requirements file for each version, with hashes specified for each
dependency required, as well as their sub-dependencies.

### How to update/add requirements

                                                              TF_Status*);

// Creates a new TensorHandle from memory residing in the physical device
// device_name. Takes ownership of the memory, and will call deleter to release
// it after TF no longer needs it or in case of error.
//
// Custom devices must use TFE_NewCustomDeviceTensorHandle instead.

  owner->reference.Unref();
  delete owner;
}

// Converts TF_DATAType to DLPack data type.
DLDataType GetDlDataType(TF_DataType data_type, TF_Status* status) {
  DLDataType dtype;
  dtype.lanes = 1;
  dtype.bits = TF_DataTypeSize(data_type) * 8;
  switch (data_type) {
    case TF_DataType::TF_BOOL:
      dtype.code = DLDataTypeCode::kDLBool;
      break;
    case TF_DataType::TF_HALF:

which is matrix inversion * tf.round now uses banker's rounding (round to even)
semantics to match NumPy. * `tf.split` now takes arguments in a reversed order
and with different keywords. In particular, we now match NumPy order as
`tf.split(value, num_or_size_splits, axis)`. * `tf.sparse_split` now takes
arguments in reversed order and with different keywords. In particular we now

    visibility = ["//visibility:public"],
)

# This condition takes precedence over :linux_x86_64
config_setting(
    name = "linux_x86_64_no_sse",
    values = {
        "cpu": "k8",
        "copt": "-mno-sse4.2",
    },
    visibility = ["//visibility:public"],
)

# This condition takes precedence over :linux_x86_64
# TODO(b/290533709): Remove this with PJRT build rule cleanup.

  ///
  /// The returned entries are paths relative to `path`.
  ///
  /// Plugins must allocate `entries` to hold all names that need to be returned
  /// and return the size of `entries`. Caller takes ownership of `entries`
  /// after the call.
  ///
  /// In case of error, plugins must set `status` to a value different than
  /// `TF_OK`, free memory allocated for `entries` and return -1.
  ///

namespace tensorflow {

// Set the shapes and types of the output's handle.
//
// The lengths of the arrays pointed to by `shapes`, `ranks`, and `types` must
// all be equal to `num_shapes_and_types`. If `ranks[i] != -1`, (i.e., if the
// rank is known), then it must be equal to the length of `shapes[i]`; if
// `ranks[i] == 1`, then `shapes[i]` may be nullptr.
//
// TODO(akshayka): Implement a corresponding getter method.

Status CreateParamsForInputs(AbstractContext* ctx,
                             absl::Span<AbstractTensorHandle* const> inputs,
                             std::vector<AbstractTensorHandle*>* params);

// A callable that takes tensor inputs and returns zero or more tensor outputs.
using Model = std::function<Status(AbstractContext*,
                                   absl::Span<AbstractTensorHandle* const>,

      block_size, block_count * block_size, 0, fetcher);
  std::vector<char> out;
  // Read blocks from the cache, and verify the LRU behavior based on the
  // fetcher calls that the cache makes.
  calls.push_back(0);
  // Cache miss - drains an element from `calls`.
  TF_EXPECT_OK(ReadCache(&cache, "", 0, 1, &out));
  // Cache hit - does not drain an element from `calls`.