Boris Fomitchev

To find which libcudnn your particular Torch installation uses, look at cudnn/ffi.lua, it should have a line like : local libnames = {'libcudnn.so.4', 'libcudnn.4.dylib'}

Yes I can repro it with latest Nemo - though, with changed export script, in different place, during actual export() call. That error usually means that some training-only code gets...

This was my quick workaround - to replace instances of tensor_parallel.ColumnParallelLinear with my wrapper class below. Something like that should be implemented inside tensor_parallel.ColumnParallelLinear.forward instead: ``` class ColumnLinear(tensor_parallel.ColumnParallelLinear): # redefine...

TRT will most likely be faster that jitted PyTorch. @ariel415el : did you use FP16 ?

Still not published? UI model can't be run from test.py - different script is required.

Thanks! I actually found out a bit more resolution needed for scenes like conference cloud (which I used SFD+FER on). But that already sounds good - I will be using...

@kevinch-nv @rajeevsrao : Basically, most of the loops would require a sequence.

Same issue with radius() call. This is for the package built from the trunk, on A6000 box.

Any plans to include bfloat16 support on GPU soon ?

> @borisfom The changes seem to breaks three CI tests. Any idea? The failing tests seem to be those I did not touch - like test_packed_sequence. I can fix them...