a7eb750594
Update the version to 0.2.34 ( #2793 )
Python package / build (3.10) (push) Has been cancelled
2023-12-09 08:36:29 -08:00
d61d43e4e4
Update README.md (vicuna-v1.3 -> vicuna-1.5) ( #2592 )
2023-10-21 14:06:08 -07:00
c478bbf367
Update train code to support the new tokenizer ( #2498 )
2023-09-29 21:56:19 -07:00
7f633245f4
To add a python script to shutdown serve ( #2196 )
2023-08-12 17:05:07 -07:00
6ea823e0e4
1. add shell scripts for shutdowning serve; 2. add a feature to launch all serve related to openai-api-server in one cmd; ( #2141 )
2023-08-08 06:43:05 -07:00
8976f570b2
Fix rope scaling in the training code ( #2124 )
2023-07-31 19:29:24 -07:00
bb027012fe
Show tqdm progress bar ( #2065 )
2023-07-24 08:14:56 -10:00
b5e3ed7869
Do not auto split train/test in the training script ( #2018 )
2023-07-19 12:06:29 -07:00
ef6dbe9662
Make flash_attn an option for train_lora ( #1927 )
...
Co-authored-by: minhvn <minhvn@rikkeisoft.com >
2023-07-15 04:36:26 -07:00
3285dfb8b2
Improve test process ( #1785 )
2023-06-26 11:11:15 -07:00
4960ca702c
Fix the random seed for train/test split ( #1522 )
2023-05-29 11:58:15 +04:00
ccc1b51aee
Improve data cleaning and training scripts ( #597 )
...
Co-authored-by: Ying Sheng <sqy1415@gmail.com >
2023-04-25 02:53:58 -07:00
00d9e6675b
Release new weights v1.1 ( #392 )
2023-04-12 12:22:52 -07:00
dc69abce16
Refactor code in train.py (_add_speaker_and_signal function ) and download_checkpoint.py ( create new download_checkpoint function) ( #207 )
2023-04-05 22:34:42 -07:00
e9ce8b53e7
Fix tokenizer issue ( #167 )
2023-04-04 15:40:49 -07:00
f9ce55d447
Update commit hash for transformers ( #149 )
2023-04-03 12:50:44 -07:00
9e786447ba
Fix the training script for syncing lastest ckpt ( #120 )
...
* Add announcement to readme
* typo
* fix-ckpt-sync
2023-04-01 13:36:24 -07:00
73f49a9899
[Training] Make the checkpointing more efficient ( #108 )
...
* make the checkpointing more efficient
* Add elimination of the old checkpoints
* revert train.py
* add comment
2023-03-31 23:03:58 -07:00
826ce631e1
Minor fix ( #84 )
...
* fix train yaml
* update
2023-03-30 12:55:03 -07:00
672fd1cc2a
fix train yaml ( #82 )
2023-03-30 09:43:06 -07:00
020f6ae756
First pass for cleaning up fine-tuning part ( #62 )
...
* First pass for cleaning up finetuning
* Fix data upload
* fix
* fix data
* update gitignore
* Fix data related description
* minor fix
* fix readme
* move data
2023-03-29 23:53:41 -07:00
31a1b9000b
Rename ChatServer -> FastChat ( #65 )
2023-03-30 02:53:54 +00:00
8489a56099
Clean up the training yaml ( #54 )
...
* Fix for installation
* minor fix
* cleanup training yaml
* rename to vicuna
* cleanup train vicuna
* fix
* Fix the training script for alpaca
2023-03-26 15:19:50 -07:00
1abed3669c
Simplify flash attn and add USE_SPLIT ( #42 )
...
* try xformers
Add support for flash attn training
Add comments
Update training script
remove the pre-download as that can cause out of disk space
Update readme
* simplify flash attn
* Add use_split
* remove accident upload
* fix use_slit suffix
* add grad accum
* Use the latest data
* correct the path for data
2023-03-24 10:00:09 -07:00
cec435320b
Add Flash Attention support for the model ( #37 )
...
* try xformers
* Add support for flash attn training
* Add comments
* Update training script
* remove the pre-download as that can cause out of disk space
* Update readme
2023-03-23 00:52:53 -07:00
c72089534b
Train 7b model on ShareGPT ( #31 )
...
* Reorganize training scripts (#23 )
* bug fix and add checkpoint
* train-7b scripts
* Add default value
* update lazy supervised dataset
* minor fix
* Add sys prompt and replace roles
* fix
* Cache preprocess
* bug fix & remove duplicated grad checkpoint option
* final fix
* add arg to control lazy preprocess or not
* update training script
* update
* Longer save step and larger save limit
* Add date
* readme
* Remove file
* remove requirements.txt
* Fix skypilot installation
* fix
* Fix GC_SCALE
---------
Co-authored-by: Lianmin Zheng <lianminzheng@gmail.com >
Co-authored-by: ZYHowell <yhzhuang@cmu.ed >
2023-03-22 15:16:50 -07:00
8ae118c71c
Correct gradient checkpointing ( #25 )
...
* Add skypilot alpaca training script
* Fix readme
* Change pytorch version and able to launch now
* update comment
* training script for 13b
* fix resume
* longer save steps
* instruction
* make it faster
* train script for 1024 length
* Add seq len
* Fix seq len
* remove train.yaml
* multi-node training
* fix multiple nodes
* gradient checkpointing now working
* add readme
* update readme
2023-03-21 22:16:12 -07:00
db4a5e0631
Reorganize training scripts ( #23 )
2023-03-21 15:19:04 +00:00
4954e06df5
Add yaml for 13B training and seq length support ( #19 )
2023-03-21 07:31:53 -07:00
c0690b64c6
Rename chatserver.server -> chatserver.serve ( #22 )
2023-03-21 14:04:04 +00:00
bd38d42ee3
Add serving with SkyPilot ( #20 )
2023-03-21 03:35:23 -07:00
ed644bcb81
add multi-node config ( #9 )
2023-03-20 23:17:12 -07:00
279227d26f
Automatically refresh models & Download model checkpoints ( #17 )
2023-03-20 20:06:25 -07:00
9aa518530d
Add a simple chatbot ( #2 )
2023-03-19 18:24:00 -07:00
b64c8936c2
Alpaca training script with SkyPilot ( #1 )
2023-03-19 01:09:59 -07:00
43fa440c02
move code
2023-03-19 00:37:04 +00:00
Lianmin Zheng