search

NVIDIA / FasterTransformer

Transformer related optimization, including BERT, GPT
Apache License 2.0
5.85k stars 891 forks source link

ft-tensorflow-quantization cannot work with tensorflow==1.14.0 #150

Closed zhuango closed 3 years ago

zhuango commented 3 years ago

I recently run PTQ with bert-tf-quantization tool. Following FasterTransformer README, I have tensorflow==1.14.0 in my environment, which should be sopported. But I got an error saying "Dense object has no attribute _compute_dtype". Then I installed tensorflow==1.15.0, and problem solved. Is the README out of date or something wrong with my environment ?

byshiue commented 3 years ago

Thanks for your feedback. In the document, we mean that the FT TensorFlow op should support on TensorFlow 1.13, 1.14 and 1.15.

But we don't verify all workflows on different environment. We will modify the document to distinguish the "verified environment" and "possible environment" in the future.

zhuango commented 3 years ago

Thanks ! @byshiue