- 
                Notifications
    You must be signed in to change notification settings 
- Fork 248
Update packaging in AOTI path #896
Conversation
| 🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchchat/896
 Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 32c5807 with merge base f20f5e7 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. | 
b23d414    to
    24a1760      
    Compare
  
    24a1760    to
    db26ea0      
    Compare
  
    db26ea0    to
    0cf4e99      
    Compare
  
    0cf4e99    to
    6ec5b35      
    Compare
  
    94c0079    to
    50c6617      
    Compare
  
    adb33fb    to
    0146b38      
    Compare
  
    | Weird that the pin bump in #1144 doesn't have the same errors as we see here.... Side question: Is it possible for us to push -l 2 into the dso as metadata as well or does that require looping through pytorch/pytorch? It's an arg when running the binary, but we know AOT what the arg should be | 
4311d16    to
    937ab9a      
    Compare
  
    f285434    to
    2c8b7b5      
    Compare
  
    925febe    to
    271f1d1      
    Compare
  
    b9d6ef5    to
    eae63fe      
    Compare
  
    9ff2e7d    to
    93e6727      
    Compare
  
    93e6727    to
    5cd2862      
    Compare
  
    | This command fails out of the box. I had to create the directory  | 
| command to build the runner fails  | 
| 
 That should be fixed with pytorch/pytorch#138919 | 
| 
 Seems like we need to bump the pin to pytorch as well? | 
| https://github.com/pytorch/torchchat/pull/1319/files There's a bump here that's not actively being worked on, seems like it's not a free bump (+ the infra issue we're seeing here) | 
| currently running into this issue when trying to run using the runner.  | 
| 
 path issue. Just need to update the instructions in the first comment so it's 
 | 
Added a
aoti_packagepath, dependent on pytorch/pytorch#129895. Follow up will be to delete the--output-dso-path.To export, use the
--output-aoti-package-pathto specify a file with a.pt2extension. This will generate an artifact containing all the AOTI generated filespython3 torchchat.py export stories15M --output-aoti-package-path exportedModels/stories15M_artifacts_cpu.pt2 --device cpuIf we look into the contents of the package, we can see the following:
To run with Python:
python3 torchchat.py generate stories15M --aoti-package-path exportedModels/stories15M_artifacts_cpu.pt2 --prompt "Hello my name is"To run with the C++ runner, first build the runner binary:
To run:
AOTI changes in PyTorch also now allow users to not need to specify --device when generating, as we will save this information during export time as metadata and use it during runtime to determine what device to use. A followup can be to extend the metadata to save model information so that users do not need to specify tokenizer version to use during runtime.