 celestinoalan
|
2a94bfff26
Append epoch rather than best val. loss to val_loss
|
hai 11 meses |
|
celestinoalan
|
d6ae2031c3
Fix fine-tuning training loss accumulation (#725)
|
hai 1 ano |
 Sanyam Bhutani
|
5fdeb55058
chore: update train_utils.py (#690)
|
hai 1 ano |
 Lucas Ventura
|
2774065891
Improve model checkpoint saving logic (#691)
|
hai 1 ano |
 Ikko Eltociear Ashimine
|
139824aafa
chore: update train_utils.py
|
hai 1 ano |
 Kai Wu
|
c18a0d277f
changed dataset to ocrvqa
|
hai 1 ano |
|
Kai Wu
|
bd22f407d5
changed to aid2 dataset
|
hai 1 ano |
|
Kai Wu
|
79dbe05a94
batch fine-tuning lmm working
|
hai 1 ano |
|
Kai Wu
|
12da109823
Merge branch 'main' into lmm_finetune
|
hai 1 ano |
|
Kai Wu
|
bb990be967
not working, need create dataloader function
|
hai 1 ano |
 Matthias Reso
|
778e31e35c
Fix checkpoint saving (#650)
|
hai 1 ano |
|
Kai Wu
|
ee204ccb98
working now
|
hai 1 ano |
|
Kai Wu
|
b566582a86
finetune not working with fsdp
|
hai 1 ano |
|
Matthias Reso
|
eca526526c
Use new get_model_state_dict api for save_pretrained peft model (#629)
|
hai 1 ano |
|
Matthias Reso
|
7a8c52cb38
Remove pkg_resources.packaging
|
hai 1 ano |
 simwiki
|
66e1867120
Fix save metric FileNotFoundError when finetuning
|
hai 1 ano |
|
Kai Wu
|
26e877fd42
changed readme, unified the context interface and added get_flops_per_sec()
|
hai 1 ano |
|
Kai Wu
|
d9558c11ca
changed context name and add more docs
|
hai 1 ano |
|
Kai Wu
|
03f1ca7817
fixed some typo to pass spellcheck
|
hai 1 ano |
|
Kai Wu
|
7b1a9413d2
fixed a typo
|
hai 1 ano |
|
Kai Wu
|
41434dc825
formatted and removed duplicated or unused function get_total_flops() and byte2mb()
|
hai 1 ano |
|
Kai Wu
|
f2e80bae22
created a FlopMeasure class on top of FlopCounterMode instead of keep of copy of our own tflop_counter.py
|
hai 1 ano |
|
Kai Wu
|
69e46887b4
handling incorrect profiling early stop caused by max_train_steps and add profiler.step() for each train step
|
hai 1 ano |
|
Kai Wu
|
34e0bf4c6e
second draft of this feature, seems to be working now
|
hai 1 ano |
|
Kai Wu
|
a35519ee90
fixed typo and handling unexpected exit
|
hai 1 ano |
|
Kai Wu
|
2a5de9b448
first draft of flop counter feature
|
hai 1 ano |
|
Kai Wu
|
e6f69f84ad
add max_steps_reached to reduce redundancy
|
hai 1 ano |
|
Kai Wu
|
fa0a389f74
add max_step feature for training and eval
|
hai 1 ano |
 jpgard
|
6954b16b3b
only save training params on rank 0
|
hai 1 ano |
 Hamid Shojanazeri
|
761b7e6e51
adding wandb_run ro eval
|
hai 1 ano |
