# EAI Final Project Proposal
- [11/8 Proposal Report](https://1drv.ms/p/s!AitIPL4ojI9ohxy560VZenV6bUkZ?e=nyJevx)
---
- [whisper-medium](https://huggingface.co/openai/whisper-medium)
- [mT5-multilingual-XLSum](https://huggingface.co/csebuetnlp/mT5_multilingual_XLSum)
- [finetuned-bart-for-conversation-summary](https://huggingface.co/kabita-choudhary/finetuned-bart-for-conversation-summary)

## Scenario - Task:
- LLM
- Model:
- [mT5-multilingual-XLSum](https://huggingface.co/csebuetnlp/mT5_multilingual_XLSum)
- Task :
- (Conversation to Summary)
- TODO : 確認chinese summary inference效果?

## Compression:
- Profile tools:
- NSYS: NVIDIA Nsight Systems
- NCU: NVIDIA Nsight Compute
- Methods:
- Knowledge distilation (finetune)
- Pruning
- [Token pruning](https://arxiv.org/abs/2012.09852)
- Quantization
- PTQ (Post-training quantization)
- QAT (Quantization aware training)
- Runtime [early exit](https://arxiv.org/abs/2006.11979)

## Scenario - Deployment:
- Inference:
- server
- A100
- web
- text/image
- linebot
- text(Conversation)

## Reference:
- [huggingface Model](https://huggingface.co/models)
- [T5 model paper: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683)
- [XL-Sum: Large-Scale Multilingual Abstractive Summarization for 44 Languages](https://arxiv.org/abs/2106.13822)