 Lengyue
|
18af8a9406
Add whisper with flash attn & use whisper to batch transcribe
|
2 rokov pred |
|
Lengyue
|
9920873cf2
Optimize schedule
|
2 rokov pred |
|
Lengyue
|
7031d18b3d
Let model fail on non-inf loss
|
2 rokov pred |
|
Lengyue
|
5f01798ab4
Fix non-inf gradient
|
2 rokov pred |
|
Lengyue
|
2a19adb045
Fix grad acc loss
|
2 rokov pred |
|
Lengyue
|
8f75f764ad
Fix incorrect batch time
|
2 rokov pred |
|
Lengyue
|
5f07060c6a
Optimize training loop
|
2 rokov pred |
|
Lengyue
|
5d65ec0bd5
fix typo
|
2 rokov pred |
|
Lengyue
|
b6b478701a
Rename dataloader to add support for valid loader
|
2 rokov pred |
|
Lengyue
|
fb23d5df2d
Better train loop tracking
|
2 rokov pred |
|
Lengyue
|
3fc0a688ca
Update Readme
|
2 rokov pred |
|
Lengyue
|
9beb5981ec
Remove unused ds config
|
2 rokov pred |
|
Lengyue
|
d211a7eac7
Better ETA
|
2 rokov pred |
|
Lengyue
|
2733ed600f
Accelerate training with fsdp
|
2 rokov pred |
|
Lengyue
|
01cf9bbafd
Fix dataset and training loop
|
2 rokov pred |
|
Lengyue
|
712908e673
Support stream reading
|
2 rokov pred |
|
Lengyue
|
e01dd330a1
optimize InterleaveDataset
|
2 rokov pred |
|
Lengyue
|
0282451dff
fix scheduler
|
2 rokov pred |
|
Lengyue
|
7650b2ac43
Update much robust dataloader
|
2 rokov pred |
|
Lengyue
|
60ac11d7fe
Update deps
|
2 rokov pred |
|
Lengyue
|
bce1f85209
Faster schedule
|
2 rokov pred |
|
Lengyue
|
d28ba743d3
Fix quantizer norm
|
2 rokov pred |
|
Lengyue
|
0dae48d2dc
Fix model training mode
|
2 rokov pred |
|
Lengyue
|
0b5216ef0e
Fix hubert config
|
2 rokov pred |
|
Lengyue
|
3297b2dc2c
Add hubert vq training
|
2 rokov pred |
|
Lengyue
|
311a8b64ea
Add hubert dataset and hubert model
|
2 rokov pred |
|
Lengyue
|
c88bf969b1
Update gitignore
|
2 rokov pred |
|
Lengyue
|
70797ef7d5
Add dockerignore
|
2 rokov pred |
|
Lengyue
|
900fc9dec9
Optimize data loading & gradient accumulate
|
2 rokov pred |
|
Lengyue
|
026d419e9a
Fix logger in config
|
2 rokov pred |
