Closed manupak closed 2 years ago
cc : @leandron @areusch @Lunderberg @mbrookhart
@Lunderberg @mbrookhart any idea on whether this is something trivial to fix? Otherwise I would suggest we revert #8306 while this is investigated.
:( I would probably hit this today as I get closer to finally being able to build the images.
I think the issue is that we updated docker to cuda 11, and the older version of TF isn't compatible with cuda11. I'll poke and see if I can figure out what's going on.
We might be able to still use the older version of TF, since I think it falls back to using the CPU implementation, but would add to the already long time to runtime of the CI. It also wouldn't be as direct of a comparison, since it would be comparing TVM+cuda to tflite+CPU.
I'd lean toward improving compatibility with TF 2.4.2, but haven't looked at that portion of TVM to know how difficult that would be.
Does anyone know where an API reference for TFLite is? I'm not super familiar with this frontend, but this does seem to be tflite returning the wrong number of inputs. I'm kind of wondering if they merged "ADD" and "ADD_N"
Can we merge this reasonably sooner ?
https://github.com/apache/tvm/pull/8375
@mbrookhart @Lunderberg @leandron @u99127
i think we should wait til #8177 is done, just to make sure the fix will stick.
I thinks this can now be closed ?
I think this is solved now.
With recent docker environment update to #8306, the tflite version is upgraded to 2.4.2.
I'm raising this, as when we re-generate Docker images, this will certainly show up. The error messages I see are as follows :
A specific error message from the above list will be as follows :
Therefore, I think we should stick with tflite=2.3.1 unless we include the changes to the tflite frontend / tests to make the CI green once the docker images are regenerated.