Created by: luotao1

I found TeamCity Always Fails due to Insufficient GPU memory to allocation. A PR should be re-run several times to pass the TeamCity check. How should we solve this problem fundamentally?

Following is Conv2dTransposeOp error:

[14:32:02] :	 [Step 1/1] ======================================================================
[14:32:02] :	 [Step 1/1] ERROR: test_check_grad_no_filter (__main__.TestConv2dTransposeOp)
[14:32:02] :	 [Step 1/1] ----------------------------------------------------------------------
[14:32:02] :	 [Step 1/1] Traceback (most recent call last):
[14:32:02] :	 [Step 1/1]   File "test_conv2d_transpose_op.py", line 69, in test_check_grad_no_filter
[14:32:02] :	 [Step 1/1]     no_grad_set=set(['Filter']))
[14:32:02] :	 [Step 1/1]   File "/paddle/python/paddle/v2/fluid/tests/op_test.py", line 386, in check_grad
[14:32:02] :	 [Step 1/1]     output_names, no_grad_set)
[14:32:02] :	 [Step 1/1]   File "/paddle/python/paddle/v2/fluid/tests/op_test.py", line 501, in _get_gradient
[14:32:02] :	 [Step 1/1]     for p_name in inputs_with_np for item in inputs_with_np[p_name]
[14:32:02] :	 [Step 1/1]   File "/paddle/python/paddle/v2/fluid/tests/op_test.py", line 501, in <dictcomp>
[14:32:02] :	 [Step 1/1]     for p_name in inputs_with_np for item in inputs_with_np[p_name]
[14:32:02] :	 [Step 1/1]   File "/paddle/python/paddle/v2/fluid/tests/op_test.py", line 424, in _numpy_to_lod_tensor
[14:32:02] :	 [Step 1/1]     tensor.set(np_value, place)
[14:32:02] :	 [Step 1/1] EnforceNotMet: ptr_ should not be null
[14:32:02] :	 [Step 1/1] Insufficient GPU memory to allocation. at [/paddle/paddle/framework/tensor.h:143]