echarlaix's picture
echarlaix HF staff
update loading instructions
57906db
metadata
license: apache-2.0
tags:
  - int8
  - Intel® Neural Compressor
  - neural-compressor
  - PostTrainingDynamic
datasets:
  - cnn_dailymail
metrics:
  - rougeLsum

INT8 DistilBart finetuned on CNN DailyMail

Post-training dynamic quantization

This is an INT8 PyTorch model quantized with huggingface/optimum-intel through the usage of Intel® Neural Compressor.

The original fp32 model comes from the fine-tuned model sshleifer/distilbart-cnn-12-6.

Below linear modules (21/133) are fallbacked to fp32 for less than 1% relative accuracy loss:

'model.decoder.layers.2.fc2', 'model.encoder.layers.11.fc2', 'model.decoder.layers.1.fc2', 'model.decoder.layers.0.fc2', 'model.decoder.layers.4.fc1', 'model.decoder.layers.3.fc2', 'model.encoder.layers.8.fc2', 'model.decoder.layers.3.fc1', 'model.encoder.layers.11.fc1', 'model.encoder.layers.0.fc2', 'model.encoder.layers.3.fc1', 'model.encoder.layers.10.fc2', 'model.decoder.layers.5.fc1', 'model.encoder.layers.1.fc2', 'model.encoder.layers.3.fc2', 'lm_head', 'model.encoder.layers.7.fc2', 'model.decoder.layers.0.fc1', 'model.encoder.layers.4.fc1', 'model.encoder.layers.10.fc1', 'model.encoder.layers.6.fc1'

Evaluation result

INT8 FP32
Accuracy (eval-rougeLsum) 41.4707 41.8117
Model size 722M 1249M

Load with optimum:

# transformers <= 4.23.0
from optimum.intel import INCModelForSeq2SeqLM

model_id = "Intel/distilbart-cnn-12-6-int8-dynamic"
int8_model = INCModelForSeq2SeqLM.from_pretrained(model_id)