-
Notifications
You must be signed in to change notification settings - Fork 24.7k
Closed
Labels
module: binariesAnything related to official binaries that we release to usersAnything related to official binaries that we release to usersmodule: regressionIt used to work, and now it doesn'tIt used to work, and now it doesn'ttopic: binariestriage review
Description
🐛 Describe the bug
I use torch +cpu
variants in images I run on VMs without GPUs because the CUDA libraries are huge, so if I don't need them then I definitely don't want them.
When I use poetry lock
on poetry 1 or 2 with torch 2.7.0+cpu
in my pyproject.toml, the cuda libraries and triton are added as dependencies (11GB image), while with 2.5.1+cpu
, they were not (3.7GB image).
I'm reporting this as a bug because I assume it was unintentional, given that the +cpu addition seems to imply the user won't be running on a GPU.
Versions
I already deleted the big image with torch 2.7.0+cpu
:(
cc @ezyang @gchanan @zou3519 @kadeng @msaroufim @seemethere @malfet @osalpekar @atalman
Metadata
Metadata
Assignees
Labels
module: binariesAnything related to official binaries that we release to usersAnything related to official binaries that we release to usersmodule: regressionIt used to work, and now it doesn'tIt used to work, and now it doesn'ttopic: binariestriage review