PyTree manipulation is now widely used in PyTorch. It becomes a fairly significant perf bottleneck during torch.fx tracing using the pure-Python approach of PyTree. We are pushing to use C++-based PyTree implementation. The lack of conda package in the Anaconda default channel feels like a serious hindrance for this.
I have asked the Anaconda team to add optree in the default channel (ContinuumIO/anaconda-issues#13131). But the average response time on the anaconda issue tracker is roughly weeks. I wonder if it is possible to add optree in the pytorch channel.
XuehaiPan
commented
1 year ago
PyTree manipulation is now widely used in PyTorch. It becomes a fairly significant perf bottleneck during
torch.fxtracing using the pure-Python approach of PyTree. We are pushing to use C++-based PyTree implementation. The lack of conda package in the Anaconda default channel feels like a serious hindrance for this.I have asked the Anaconda team to add
optreein the default channel (ContinuumIO/anaconda-issues#13131). But the average response time on the anaconda issue tracker is roughly weeks. I wonder if it is possible to addoptreein the pytorch channel.Ref: Conda-Forge feedstock: https://github.com/conda-forge/optree-feedstock