Intel/bart-large-cnn-int8-dynamic-inc

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 facebook/bart-large-cnn.

Below linear modules (40/193) are fallbacked to fp32 for less than 1% relative accuracy loss:

'model.decoder.layers.10.fc1', 'model.decoder.layers.0.fc2', 'model.decoder.layers.4.fc2', 'model.decoder.layers.1.fc2', 'model.decoder.layers.6.fc2', 'model.decoder.layers.2.fc2', 'model.decoder.layers.3.fc2', 'model.encoder.layers.11.fc2', 'model.decoder.layers.9.fc1', 'model.decoder.layers.5.fc2', 'model.decoder.layers.7.fc1', 'model.decoder.layers.8.fc1', 'model.encoder.layers.0.fc2', 'model.decoder.layers.11.fc1', 'model.encoder.layers.8.fc2', 'model.encoder.layers.11.fc1', 'model.decoder.layers.8.fc2', 'model.decoder.layers.2.fc1', 'model.decoder.layers.11.self_attn.v_proj', 'model.encoder.layers.9.fc1', 'model.decoder.layers.9.fc2', 'model.decoder.layers.7.fc2', 'model.decoder.layers.6.fc1', 'model.decoder.layers.0.fc1', 'model.decoder.layers.1.self_attn.v_proj', 'model.encoder.layers.3.fc1', 'model.encoder.layers.2.fc2', 'model.encoder.layers.7.fc2', 'model.decoder.layers.3.fc1', 'model.encoder.layers.1.fc2', 'model.encoder.layers.10.fc2', 'model.encoder.layers.8.fc1', 'lm_head', 'model.decoder.layers.6.self_attn.v_proj', 'model.decoder.layers.11.self_attn.out_proj', 'model.decoder.layers.11.encoder_attn.v_proj', 'model.encoder.layers.10.fc1', 'model.encoder.layers.6.fc1', 'model.decoder.layers.4.fc1', 'model.decoder.layers.1.fc1'

Evaluation result

	INT8	FP32
Accuracy (eval-rougeLsum)	41.2224	41.5274
Model size	625M	1669M

Load with optimum:

from optimum.intel import INCModelForSeq2SeqLM

model_id = "Intel/bart-large-cnn-int8-dynamic"
int8_model = INCModelForSeq2SeqLM.from_pretrained(model_id)