TF BERT not FP16 compatible?

[volker42maru]

🐛 Bug

Information

Model I am using (Bert, XLNet ...): TFBertForQuestionAnswering

Language I am using the model on (English, Chinese ...): English

The problem arises when using:

• my own modified scripts:

The tasks I am working on is:

• an official GLUE/SQUaD task: SQUaD

To reproduce

Simple example to reproduce error:

import tensorflow as tf
from transformers import TFBertForQuestionAnswering

# turn on mp (fp16 operations)
tf.keras.mixed_precision.experimental.set_policy('mixed_float16')

model = TFBertForQuestionAnswering.from_pretrained('bert-base-uncased')

The error occurs here:
transformers/modeling_tf_bert.py", line 174, in _embedding
embeddings = inputs_embeds + position_embeddings + token_type_embeddings

And this is the error:
tensorflow.python.framework.errors_impl.InvalidArgumentError: cannot compute AddV2 as input #1(zero-based) was expected to be a half tensor but is a float tensor [Op:AddV2] name: tf_bert_for_question_answering/bert/embeddings/add/

Expected behavior

I want to use TF BERT with mixed precision (for faster inference on tensor core GPUs). I know that full fp16 is not working out-of-the-box, because the model weights need to be in fp16 as well. Mixed precision, however, should work because only operations are performed in fp16.

I get some dtype issue. Seems the mode is not fp16 compatible yet? Will this be fixed in the future?

Environment info

• transformers version: 2.5.0
• Platform: ubuntu 16.04
• Python version: 3.6.9
• PyTorch version (GPU?): 1.4.0 (GPU)
• Tensorflow version (GPU?): 2.1.0 (GPU)
• Using GPU in script?: sort of
• Using distributed or parallel set-up in script?: nope

[bamps53]

I've aced same issue. Maybe it's hard coded the data type somewhere? Have you found solution?

[rzepinskip]

Tried this on Colab TPU, same error.

[ben74]

Same here, would be convenient as hell :)

[patrickvonplaten]

Having the same error also for transformers version 2.11.0.
Here some code to easily reproduce the error:

#!/usr/bin/env python3
from transformers import TFBertModel, BertTokenizer
from tensorflow.keras.mixed_precision import experimental as mixed_precision

policy = mixed_precision.Policy('mixed_float16')
mixed_precision.set_policy(policy)

tok = BertTokenizer.from_pretrained("bert-base-uncased")
model = TFBertModel.from_pretrained("bert-base-uncased")
input_ids = tok("The dog is cute", return_tensors="tf").input_ids
model(input_ids)  # throws error on GPU

[chrisabbott]

Encountering the same issue here:

import tensorflow as tf
from transformers.modeling_tf_distilbert import TFDistilBertModel

tf.keras.mixed_precision.experimental.set_policy('mixed_float16')
model = TFDistilBertModel.from_pretrained('distilbert-base-uncased')

[patrickvonplaten]

Put this issue on my TF ToDo-List :-)

[Hazarapet]

+1

[QixinLi]

Hi @patrickvonplaten, is this problem fixed？
I got the same error recently with version 3.0.2

[patrickvonplaten]

This is still an open problem...I didn't find the time yet to take a look! Will link this issue to the TF projects.

[xuxingya]

This is already solved in new version.
position_embeddings = tf.cast(self.position_embeddings(position_ids), inputs_embeds.dtype) token_type_embeddings = tf.cast(self.token_type_embeddings(token_type_ids), inputs_embeds.dtype) embeddings = inputs_embeds + position_embeddings + token_type_embeddings