Convolutional VAE for MNIST using Reactant failed to produce right results

[Yangyang-Tan]

The Reactant example in the document failed to produce the right results for the latest version.
https://lux.csail.mit.edu/dev/tutorials/intermediate/5_ConvolutionalVAE

[Yangyang-Tan]

I also get lots of warnings like the following:

E0000 00:00:1742930830.945941 2247005 buffer_comparator.cc:156] Difference at 4096: nan, expected 35.3139 E0000 00:00:1742930830.945957 2247005 buffer_comparator.cc:156] Difference at 4097: nan, expected 33.723 E0000 00:00:1742930830.945961 2247005 buffer_comparator.cc:156] Difference at 4098: nan, expected 30.5046 E0000 00:00:1742930830.945964 2247005 buffer_comparator.cc:156] Difference at 4099: nan, expected 31.9991 E0000 00:00:1742930830.945967 2247005 buffer_comparator.cc:156] Difference at 4100: nan, expected 32.3158 E0000 00:00:1742930830.945970 2247005 buffer_comparator.cc:156] Difference at 4101: nan, expected 31.243 E0000 00:00:1742930830.945989 2247005 buffer_comparator.cc:156] Difference at 4102: nan, expected 32.8456 E0000 00:00:1742930830.945992 2247005 buffer_comparator.cc:156] Difference at 4103: nan, expected 34.6195 E0000 00:00:1742930830.945995 2247005 buffer_comparator.cc:156] Difference at 4104: nan, expected 35.1943 E0000 00:00:1742930830.945999 2247005 buffer_comparator.cc:156] Difference at 4105: nan, expected 31.1287 2025-03-25 19:27:10.946004: E external/xla/xla/service/gpu/autotuning/gemm_fusion_autotuner.cc:1137] Results do not match the reference. This is likely a bug/unexpected loss of precision.

[avik-pal]

probably related to EnzymeAD/Reactant.jl#444

[avik-pal]

Can confirm that this is a bug. I will fix this upstream first EnzymeAD/Reactant.jl#1042 and then port over the changes

[Yangyang-Tan]

Hi Avik, this still failed for the latest build of Reactant.

[avik-pal]

yeah we need to use the config, and see if that gets rid of the buffer comparator failures. I am a bit swamped because of a deadline on 15th April, I will take a closer look after that