You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The used fbgemm @ c3a26e1 is buggy when compiling with "--cpu_only".
As specified in CMakeLists.txt, "src/permute_pooled_embedding_ops_gpu.cpp" will only compile when "NOT FBGEMM_CPU_ONLY", which means the method "permute_pooled_embs_auto_grad" won't be generated when "--cpu_only". However, this method is used by column_wise sharding.
To correctly support "--cpu_only", see the pr in FBGEMM#987
The text was updated successfully, but these errors were encountered:
Thanks @RabbitWhite1 for reporting this issue! We are working on fixing this issue with pytorch/FBGEMM#950 as you mentioned. Currently the major blocker on landing the PR is due to some internal backward compatibility issues. Will push on landing it soon. cc @geyyer
The used fbgemm @ c3a26e1 is buggy when compiling with "--cpu_only".
As specified in CMakeLists.txt, "src/permute_pooled_embedding_ops_gpu.cpp" will only compile when "NOT FBGEMM_CPU_ONLY", which means the method "permute_pooled_embs_auto_grad" won't be generated when "--cpu_only". However, this method is used by column_wise sharding.
To correctly support "--cpu_only", see the pr in FBGEMM#987
The text was updated successfully, but these errors were encountered: