Repository navigation
Releases
Releases · tpnam0901/GPT2
v0.3.0
v0.2.0
0.2.0 (2024-08-30)
Features
- add change vocab size for model (bf1ff05)
- add custom dataset preparation (580e809)
- add hosting server (78f5ffb)
- add inference script (a5f177d)
- server_api: add optional checkpoint (4099e95)
Bug Fixes
- duplicate cfg.log file (960f530)
- duplicate log in mlflow (bd83877)
- eos_token, response message, wrong key json (f75d810)
- infer token of server api, add eot token to input (20d13fe)
- miss best ckpt, load config (7616f7c)
- not passing parameters in the config (59ed96a)
- server_api: bot repeat the same sentence (4f7c717)
- server_api: special token (02f9e4d)
- variable type in args (c164c7e)
Documentation
v0.1.0
0.1.0 (2024-08-28)
Features
- add automatic mix precision (5c4831d)
- add crop block size of GPT model (3c66a57)
- add data distributed parallel training (f0bb8b0)
- add dataloader for .bin file - openwebtext dataset (2966136)
- add GPT2 config, weight converter, missing model file (c2ee7d7)
- add GPT2 model (bb53955)
- add log interval (6640c38)
- add loss function, learning rate schedular (4a5f625)
- add openwebtext preprocess (5d7daa9)
- add optimizers (634108f)
- add train config for openwebtext dataset, gpt2 model (92d20bd)
- add train script for GPT2, update config (13a3fb3)
Bug Fixes
- automatic mix precision for cpu (ffe3058)
- change cel to functional (66c25ef)
- duplicate log (a903e81)
- feed forward loss function, change variable name (6d799dd)
- gpt.py: remove model ff return loss (f12aa99)
- infinity validating step (a07d7b6)
- losses.py: missing parent init function (2b6a2df)
- openwebtext dataset link (4fe8add)
- same GPU in training multi GPU (b5ddc45)
- step calling in optimizer (4f1d309)
- x, y device in dataloader (9db7aad)
Documentation
- add training instruction (bcfc7b1)