performed instead:

```python
tf.cumsum([a, b, c], exclusive=True)  # => [0, a, a + b]
```

By setting the `reverse` kwarg to `True`, the cumsum is performed in the
opposite direction:

```python
tf.cumsum([a, b, c], reverse=True)  # => [a + b + c, b + c, c]
```

This is more efficient than using separate `tf.reverse` ops.

*   GPU
    *   GPU acceleration now supports quantized models by default
*   `DynamicBuffer::AddJoinedString()` will now add a separator if the first
    string to be joined is empty.
*   Adds support for cumulative sum (cumsum), both as builtin op and MLIR
    conversion.

### `TensorRT`

*   Issues a warning when the `session_config` parameter for the TF1 converter

    if (result.getType() != op.getType())
      result = rewriter.create<tensor::CastOp>(loc, op.getType(), result);
    rewriter.replaceOp(op, result);
    return success();
  }
};

// Converts the Cumsum or Cumprod TensorFlow op to the HLO ReduceWindow op by
// setting appropriate window dimensions, with the given aggregation op as the
// reduction function. The input tensor needs to have a static shape, and 'axis'