{d: 5 * 365 * 24 * time.Hour, want: "5y"},
		{d: 6 * 365 * 24 * time.Hour, want: "6y"},
		{d: 7 * 365 * 24 * time.Hour, want: "7y"},
		{d: 8*365*24*time.Hour - time.Millisecond, want: "7y364d"},
		{d: 8 * 365 * 24 * time.Hour, want: "8y"},
		{d: 8*365*24*time.Hour + 364*24*time.Hour, want: "8y"},
		{d: 9 * 365 * 24 * time.Hour, want: "9y"},
	}
	for _, tt := range tests {

  // For example, if new filter shape is [4, 4, 12, 64], old filter shape
  // is [7, 7, 3, 64] with block_size 2.
  // Below transforms will be applied to the filter:
  // 1. Reshape to [4, 4, 2, 2, 3, 64];
  // 2. Transpose to [4, 2, 4, 2, 3, 64];
  // 3. Reshape to [8, 8, 3, 64];
  // 4. Slice to [7, 7, 3, 64].
  SmallVector<int64_t, 6> first_reshape_shape = {
      new_filter_shape[0],
      new_filter_shape[1],

    based on the cost. As the size of the kernel is relatively small, this won't
    add additional cost to TPU device time. Below is the logic to transform the
    kernel of shape [7, 7, 3, 64] to [4, 4, 12, 84].

    ```python
    conv0 = tf.compat.v1.layers.Conv2D(
     filters=filters,
     kernel_size=kernel_size,
     strides=2,
     padding=('SAME' if strides == 1 else 'VALID'),

    %0 = "tf.Const"() {value = dense<1> : tensor<i64>} : () -> tensor<i64>
    %1 = "tf.Const"() {value = dense<0> : tensor<1x1xi32>} : () -> tensor<1x1xi32>
    %2 = "tf.Const"() {value = dense<[7, 7, 3, 64]> : tensor<4xi32>} : () -> tensor<4xi32>
    %3 = "tf.Const"() {value = dense<[[0, 0], [3, 3], [3, 3], [0, 0]]> : tensor<4x2xi32>} : () -> tensor<4x2xi32>
    %4 = "tf.Const"() {value = dense<0> : tensor<i32>} : () -> tensor<i32>

	0x3640: 0x000a, 0x3641: 0x000a, 0x3642: 0x000a, 0x3643: 0x000a, 0x3644: 0x000a, 0x3645: 0x000a,
	0x3646: 0x000a, 0x3647: 0x000a, 0x3648: 0x000a, 0x3649: 0x000a, 0x364a: 0x000a, 0x364b: 0x000a,
	0x364c: 0x000a, 0x364d: 0x000a, 0x364e: 0x000a, 0x364f: 0x000a, 0x3650: 0x000a, 0x3651: 0x000a,
	0x3652: 0x000a, 0x3653: 0x000a, 0x3654: 0x000a,
	// Block 0xda, offset 0x3680

	0x3640: 0x0001, 0x3641: 0x0001, 0x3642: 0x0001, 0x3643: 0x0001, 0x3644: 0x000c, 0x3645: 0x000c,
	0x3646: 0x000c, 0x3647: 0x000c, 0x3648: 0x000c, 0x3649: 0x000c, 0x364a: 0x000c, 0x364b: 0x0001,
	0x364c: 0x0001, 0x364d: 0x0001, 0x364e: 0x0001, 0x364f: 0x0001, 0x3650: 0x0001, 0x3651: 0x0001,
	0x3652: 0x0001, 0x3653: 0x0001, 0x3654: 0x0001, 0x3655: 0x0001, 0x3656: 0x0001, 0x3657: 0x0001,

	// Block 0xd9, offset 0x3640
	0x3641: 0x000a, 0x3642: 0x000a, 0x3643: 0x000a, 0x3644: 0x000a, 0x3645: 0x000a,
	0x3646: 0x000a, 0x3647: 0x000a, 0x3648: 0x000a, 0x3649: 0x000a, 0x364a: 0x000a, 0x364b: 0x000a,
	0x364c: 0x000a, 0x364d: 0x000a, 0x364e: 0x000a, 0x364f: 0x000a, 0x3651: 0x000a,
	0x3652: 0x000a, 0x3653: 0x000a, 0x3654: 0x000a, 0x3655: 0x000a, 0x3656: 0x000a, 0x3657: 0x000a,

	0x3640: 0x000c, 0x3641: 0x000c, 0x3642: 0x000c, 0x3643: 0x000c, 0x3644: 0x000c, 0x3645: 0x000c,
	0x3646: 0x000c, 0x3647: 0x000c, 0x3648: 0x000c, 0x3649: 0x000c, 0x364a: 0x000c, 0x364b: 0x000c,
	0x364c: 0x000c, 0x364d: 0x000c, 0x364e: 0x000c, 0x364f: 0x000c, 0x3650: 0x000c, 0x3651: 0x000c,
	0x3652: 0x000c, 0x3653: 0x000c, 0x3654: 0x000c, 0x3655: 0x000c, 0x3656: 0x000c, 0x3657: 0x000c,

* [Breaking change, client-go]: The WaitFor function returns, probably an ErrWaitTimeout, when the done channel is closed, even if the `WaitFunc` doesn't handle the done channel. ([#72364](https://github.com/kubernetes/kubernetes/pull/72364), [@kdada](https://github.com/kdada))

?0Y\030\332\301m\335?\324\203\257\327\220\343\315?H\362>E\035c\264?\315\241\r\"\300\036\357?\215\251\343\312\255G\356?\244\201pP\360b\337?}\346L\350J\336\352?49\010P\007\221\311?\325\021\267\364\222\253\353?\200\343\250{Io\227??\3106n\311\320\345?\364d\207\331\3314\327?\354\3045g\303\375\326?\264\353\034\242\306\010\334?\023\213z\230\251m\342?\374f3\264B\033\345?\354X\356\353\234\237\352?\200Ia\022(K\244?G\372)\240\010\367\355?(K\205o\305\200\336?\263\264\231\255\333\207\340?\236w\263A\310R\353?...