jbilcke-hf HF Staff commited on
Commit
7e21e7f
·
1 Parent(s): c550ede

testing torch 2.7.1

Browse files
Files changed (1) hide show
  1. requirements.txt +19 -5
requirements.txt CHANGED
@@ -1,11 +1,25 @@
1
  --find-links https://download.pytorch.org/whl/torch_stable.html
2
- torch==2.8.0
3
- torchvision==0.23.0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  torchdata==0.11.0
5
  torchao==0.12.0
6
- torchcodec
7
-
8
- flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.8cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
9
 
10
  # something broke in Transformers > 4.55.4
11
  transformers==4.55.4
 
1
  --find-links https://download.pytorch.org/whl/torch_stable.html
2
+ # we seem to have an issue with Torch 2.8
3
+ # I believe it works but it is incompatible with older weights formats?
4
+ # it looks like they changed the checkpoint format or something
5
+ # python3.10/site-packages/torch/distributed/checkpoint/default_planner.py", line 471, in create_default_local_load_plan
6
+ # RuntimeError: Missing key in checkpoint state_dict: lr_scheduler._is_initial.
7
+ #
8
+ #torch==2.8.0
9
+ #torchvision==0.23.0
10
+ #torchdata==0.11.0
11
+ #torchao==0.12.0
12
+ #torchcodec
13
+ # flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.8cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
14
+ #
15
+ # our solution for now is to use torch 2.7
16
+ #
17
+ torch==2.7.1
18
+ torchvision==0.22.1
19
  torchdata==0.11.0
20
  torchao==0.12.0
21
+ torchcodec=0.5.0
22
+ flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.7cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
 
23
 
24
  # something broke in Transformers > 4.55.4
25
  transformers==4.55.4