EOF
    find . | tr '[A-Z]' '[a-z]' | sort | uniq -d | tee $BATS_FILE_TMPDIR/repeats
    [[ ! -s $BATS_FILE_TMPDIR/repeats ]]
}

# It's unclear why, but running this on //tensorflow/... is faster than running
# only on affected targets, usually. There are targets in //tensorflow/lite that
# don't pass --nobuild, so they're on their own.
#
# Although buildifier checks for formatting as well, "bazel build nobuild"

    `tensorflow/python/saved_model/load_test.py`:

    (Requires `python>=3.7`)

    ```bash
    bazel test ${flags} //tensorflow/python/saved_model:load_test --test_filter=*LoadTest.test_capture_variables*
    ```

    **Note:** You can add `--test_sharding_strategy=disabled` to the `flags` to
    disable the sharding so that all the test outputs are in one file. However,

// creating a new op every time. If `raw_device_name` is `NULL` or empty, it
// does not set the device name. If it's not `NULL`, then it attempts to parse
// and set the device name. It's effectively `TFE_OpSetDevice`, but it is faster
// than separately calling it because if the existing op has the same
// `raw_device_name`, it skips parsing and just leave as it is.
TF_CAPI_EXPORT extern void TFE_OpReset(TFE_Op* op_to_reset,

FROM mcr.microsoft.com/dotnet/framework/sdk:4.8-windowsservercore-ltsc2019

# Set default powershell policy for this script (ProgressPreference='SilentlyContinue' makes
# downloads with Invoke-WebRequest not show the progress bar and is MUCH faster).
SHELL ["powershell.exe", "-ExecutionPolicy", "Bypass", "-Command", "$ErrorActionPreference='Stop'; $ProgressPreference='SilentlyContinue'; $VerbosePreference = 'Continue';"]

EOF
    find . | tr '[A-Z]' '[a-z]' | sort | uniq -d | tee $BATS_FILE_TMPDIR/repeats
    [[ ! -s $BATS_FILE_TMPDIR/repeats ]]
}

# It's unclear why, but running this on //tensorflow/... is faster than running
# only on affected targets, usually. There are targets in //tensorflow/lite that
# don't pass --nobuild, so they're on their own.
#
# Although buildifier checks for formatting as well, "bazel build nobuild"

//  - target: dy in dx/dy
//  - source: dx in dx/dy
//  - tensor: one of the many inputs or outputs of an operation
//
// Below here we do the gradient algorithm. It works as follows:
//
// First we filter the tape to just the subset of operations we want to
// differentiate. In the process of doing so we count how many times each Tensor
// is used as an input to an op (so we know when we're done computing gradients