Fix NCCL rewrite bug when rerunning sessions (assigned device id is not stable).
Fix collocate_gradients for initial losses. Remove NcclBroadcast gradient test for now. The generated AddN to accumulate the broadcast outputs before passing it to the gradient function is CPU only and cannot be collocated with NcclBroadcast on the GPU. PiperOrigin-RevId: 173306409
Loading
Please sign in to comment