File size: 1,765 Bytes
686d1e5 25121a8 686d1e5 25121a8 0033b45 25121a8 0033b45 25121a8 0033b45 afc3f72 25121a8 afc3f72 25121a8 0033b45 25121a8 0033b45 25121a8 9418e53 57906db |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 |
---
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](https://github.com/huggingface/optimum-intel) through the usage of [Intel® Neural Compressor](https://github.com/intel/neural-compressor).
The original fp32 model comes from the fine-tuned model [sshleifer/distilbart-cnn-12-6](https://huggingface.co/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:
```python
# 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)
```
|