-
Notifications
You must be signed in to change notification settings - Fork 224
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[WIP] Initial PR for generating and loading state dict #1329
base: main
Are you sure you want to change the base?
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchchat/1329
Note: Links to docs will display an error until the docs builds have been completed. ❌ 9 New Failures, 3 Cancelled JobsAs of commit 8ade3c4 with merge base 70260eb (): NEW FAILURES - The following jobs have failed:
CANCELLED JOBS - The following jobs were cancelled. Please retry:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
@@ -148,6 +148,12 @@ def _add_model_config_args(parser, verb: str) -> None: | |||
help="Whether to compile the prefill. Improves prefill perf, but has higher compile times.", | |||
) | |||
|
|||
model_config_parser.add_argument( |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
after migration we shouldn't need anything special for quantized model right
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Not sure I follow? The things I'm testing out should work for tensor subclass right?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
oh I meant that with tensor subclass API, quantized checkpoint should be able to be loaded the same way as normal checkpoint.
the code path of loading a quantized model v.s. quantizing model on the fly might still make sense though, maybe just need to change the naming or something
No description provided.