fjmscut comments

Repositories
Issues
Comments

Results 3 comments of


                                            fjmscut

NCCL + XLA fails for multi-GPU training.

same bug 'xla.gpu.all_reduce' failed, i found that "tensorflow.keras.applications.inception_resnet_v2.InceptionResNetV2" was called before all_reduce, the problem occur, but if i use dense layer as replacement, it work well.... sb. help me plz

NCCL + XLA fails for multi-GPU training.

i guess the former batchnormalize layer called cause the later all_reduce crash as the code above

[BUG][all_reduce] INVALID_ARGUMENT: You must feed a value for placeholder tensor

same bug when i implement gradient accumulate feature, sb. can help me ? plz .....