A performance comparison of the new #PyTorch2 with the well established #PyTorch 1.
The benchmarks cover different areas of #deeplearning, such as image classification (#resnet) and language models (#BERT).
#pytorch2 #pytorch #deeplearning #resnet #BERT
A performance comparison of the new #PyTorch2 with the well established #PyTorch 1.
The benchmarks cover different areas of #deeplearning, such as image classification (#resnet) and language models (#BERT).
#pytorch2 #pytorch #deeplearning #resnet #BERT
Our lab will offer the two classes High Performance Computing and Efficient Machine Learning in the upcoming semester. New topics include BF16 support in SVE, recent features of PyTorch 2 and inference on mobile devices: https://scalable.uni-jena.de/teaching/2023/03/15/summer-semester.html
#fsujena #hpc #ml #bfloat16 #sve #arm #graviton3 #pytorch2 #quantization #snapdragon
#fsujena #hpc #ml #bfloat16 #SVE #arm #Graviton3 #pytorch2 #quantization #snapdragon
Got #Speechbrain on #pytorch2 within #python311 up and running. So far just recipe tested, but seems to run faster whilst CPU decoding.
Building it was not that easy though. First, I experienced troubles while linking against glibc, but it was due to switching back and forth to conda environment, because I needed gcc 11.1 (my Fedora 36 has gcc 12+, and it has issues preprocessing some templates among the pytorch dependencies). Moreover, I needed linkage to the system's #cuda12, which was present at a conda repository
I was too lazy to find the way through, exporting correct variables and/or setting correct paths here and there. So I decided to first build custom gcc 11.1 and eventually pytorch.
I visited https://ftp.gnu.org/gnu/gcc/gcc-11.1.0/, but then I struggled once again. Now it was compiling gcc itself. Even then I had very little patience finding the true reason, but was lucky enough to find a patch enabling succesful compilation of the gcc: https://reviews.llvm.org/D129471
#speechbrain #pytorch2 #python311 #cuda12