Conv3D crash when the data_format is 'channels_first' and using Tensorflow backend

[AGFACBNNR]

According to the document of Conv3D in keras website, Conv3D should accept inputs with data format 'channels_first' or 'channels_last'.
While in this colab, I got the following results.

[fchollet]

According to the error message, the lack of support is only on CPU -- GPU should work fine. There's no CPU kernel for channels_first Conv3D. We can't fix that on the Keras side except by doing a transpose/counter-transpose in that case, which would be very inefficient.

[AGFACBNNR]

Got it. I'll try it on GPU.

[AGFACBNNR]

@fchollet
Sorry for bothering again.
Surprisingly, I found that sometimes Conv3D can get an output when data_format is 'channels_first'.
In this colab, l1 and l2 have the same parameters, except for 'groups'. However, l1 can generate an output while l2 meets an error, as shown in the following. This is very strange. I thought 'groups' would not influence the data format of inputs.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No