Commit Graph

1871 Commits

Author SHA1 Message Date
Wing Lian
4a0ab11fcf chore: lint 2025-01-13 14:05:56 -05:00
Wing Lian
73b6b0a580 chore: lint 2025-01-13 13:56:16 -05:00
Wing Lian
9db5072407 make sure to use tensorboard to capture loss for checks 2025-01-13 13:56:16 -05:00
Wing Lian
42d3e36a6f fix adapter model check 2025-01-13 13:56:15 -05:00
Wing Lian
b12d93bedf make sure to use the correct tokenizer 2025-01-13 13:56:15 -05:00
Wing Lian
08ec9c0e5b make sure to set tokenizer from l3 70b and save safetensors 2025-01-13 13:56:15 -05:00
Wing Lian
9abac55f92 lower lr 2025-01-13 13:56:15 -05:00
Wing Lian
800e7fa41e set lora_dropout explicitly 2025-01-13 13:56:15 -05:00
Wing Lian
5a1c1b82d4 make the kd e2e fit in vram for ci and add lora version 2025-01-13 13:56:15 -05:00
Wing Lian
efb3f70d38 rename test files so it gets picked up 2025-01-13 13:56:15 -05:00
Wing Lian
58d9896777 linting 2025-01-13 13:56:15 -05:00
Wing Lian
f7963083b8 add kd trainer e2e test 2025-01-13 13:56:15 -05:00
Wing Lian
f0b6581f8c reward model doesn't work well with batched 2025-01-13 13:56:15 -05:00
Wing Lian
27bb21c459 improve check for batched 2025-01-13 13:56:15 -05:00
Wing Lian
74d98ca6d8 fix reward trainer calls for tokenization 2025-01-13 13:56:14 -05:00
Wing Lian
ec4dfb02c8 reward can use same batch check 2025-01-13 13:56:14 -05:00
Wing Lian
28ef5e8d5a tweak check for batched prompt data 2025-01-13 13:56:14 -05:00
Wing Lian
5ed2823855 ensure that batch vs single is done properly 2025-01-13 13:56:14 -05:00
Wing Lian
fb0775d264 improve iterable support 2025-01-13 13:56:12 -05:00
Wing Lian
7cd0a317cb support streaming for processing sft datasts? 2025-01-13 13:41:36 -05:00
Wing Lian
1cc3a2d16c make loss torch script compat 2025-01-13 13:41:36 -05:00
Wing Lian
287d2ca8d5 kd sample packing 2025-01-13 13:41:36 -05:00
Wing Lian
03b86df506 be a bit pickier about loading dynamic prompt strategies 2025-01-13 13:41:36 -05:00
Wing Lian
2ed4246949 more info on preprocess for kd and fix import 2025-01-13 13:41:35 -05:00
Wing Lian
35bc2e2d3f remove duplicate code 2025-01-13 13:41:35 -05:00
Wing Lian
94f1094805 add copyrights 2025-01-13 13:41:35 -05:00
Wing Lian
a0070bf94e increase logging around loading plugins 2025-01-13 13:41:35 -05:00
Wing Lian
2ee2ffd834 make plugin setup concise 2025-01-13 13:41:35 -05:00
Wing Lian
723b0a2dee remove moved class from import 2025-01-13 13:41:35 -05:00
Wing Lian
327739c9e3 move more things to kd plugin 2025-01-13 13:41:35 -05:00
Wing Lian
8aafe142f2 refactor kd chat template loader 2025-01-13 13:41:35 -05:00
Wing Lian
a0d6d8895e support for custom trainer classes from plugins 2025-01-13 13:41:34 -05:00
Wing Lian
55b33cc44d handle token/logprob shifting 2025-01-13 13:41:34 -05:00
Wing Lian
69ed25e82c remove references to triton kd for now 2025-01-13 13:41:34 -05:00
Wing Lian
2ea8b7e518 add license block 2025-01-13 13:41:34 -05:00
Wing Lian
aa081e0e76 refactor so we can easily add new loss functions 2025-01-13 13:41:34 -05:00
Wing Lian
3f97ec45fb chore: lint 2025-01-13 13:41:34 -05:00
Wing Lian
7b5a24b0d2 var naming and add todo 2025-01-13 13:41:34 -05:00
Wing Lian
4ddd089d0a fix kd loss so it's causal (fixes repeating tokens) 2025-01-13 13:41:34 -05:00
Wing Lian
b88128d067 use kd_alpha in the correct loss method 2025-01-13 13:41:32 -05:00
Wing Lian
2e6422a711 hash for temperature too 2025-01-13 13:40:19 -05:00
Wing Lian
6ad809287b better rescaling for temperatures 2025-01-13 13:40:19 -05:00
Wing Lian
e376e00386 don't use triton for now 2025-01-13 13:40:19 -05:00
Wing Lian
23d7ae6caa fix kwarg 2025-01-13 13:40:19 -05:00
Wing Lian
19638590d5 v3 2025-01-13 13:40:18 -05:00
Wing Lian
73f5b83431 no torch.tensor 2025-01-13 13:40:18 -05:00
Wing Lian
9b1164b841 no log etc 2025-01-13 13:40:18 -05:00
Wing Lian
5a7d6f6175 no torch.exp inside triton kernel 2025-01-13 13:40:18 -05:00
Wing Lian
a803c3d3ee v2 trial 2025-01-13 13:40:18 -05:00
Wing Lian
48ccf55752 no where support 2025-01-13 13:40:18 -05:00