search

NVIDIA / apex

A PyTorch Extension: Tools for easy mixed precision and distributed training in Pytorch
BSD 3-Clause "New" or "Revised" License
8.39k stars 1.4k forks source link

Apex with windows, anaconda and jupyter noptebook #419

Open LiamDeth opened 5 years ago

LiamDeth commented 5 years ago

! pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ../nvidiaapex/repository/NVIDIA-apex-39e153a

import apex from apex.normalization.fused_layer_norm import FusedLayerNorm from pytorch_pretrained_bert import convert_tf_checkpoint_to_pytorch

this works fine but then, I still have:

ModuleNotFoundError: No module named 'fused_layer_norm_cuda'

When trying to do:

convert_tf_checkpoint_to_pytorch.convert_tf_checkpoint_to_pytorch( BERT_MODEL_PATH + 'bert_model.ckpt', BERT_MODEL_PATH + 'bert_config.json', WORK_DIR + 'pytorch_model.bin')

because it tries to do :

import importlib importlib.import_module("fused_layer_norm_cuda")

any idea ?

I have: nvcc --version giving release 10.1 and torch.version giving 1.0.1

ptrblck commented 5 years ago

Did the install log give you some information, if all CUDA extensions were built?