device map legacy attention block weight conversion #3804
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
williamberman
commented
Jun 15, 2023
| @@ -78,6 +78,7 @@ def __init__( | |||
| self.upcast_softmax = upcast_softmax | |||
| self.rescale_output_factor = rescale_output_factor | |||
| self.residual_connection = residual_connection | |||
| self.dropout = dropout | |||
There was a problem hiding this comment.
Added so can re-create the dropout module when converting back to new weight format
|
The documentation is not available anymore as the PR was closed or merged. |
sayakpaul
reviewed
Jun 16, 2023
| # (which look like they should be private variables?), so we can't use the standard hooks | ||
| # to rename parameters on load. We need to mimic the original weight names so the correct | ||
| # attributes are available. After we have loaded the weights, we convert the deprecated | ||
| # names to the new non-deprecated names. Then we _greatly encourage_ the user to convert |
There was a problem hiding this comment.
Do we have guidance available for the users on how they should perform the conversion?
There was a problem hiding this comment.
Ah nevermind. I guess you meant once we load the old attention block weight names and run conversion internally, and suggest users save the pipeline. Right?
There was a problem hiding this comment.
Yeah just this blurb here
f"Taking `{str(e)}` while using `accelerate.load_checkpoint_and_dispatch` to mean {pretrained_model_name_or_path}"
" was saved with deprecated attention block weight names. We will load it with the deprecated attention block"
" names and convert them on the fly to the new attention block format. Please re-save the model after this conversion,"
" so we don't have to do the on the fly renaming in the future. If the model is from a hub checkpoint,"
" please also re-upload it or open a PR on the original repository."
sayakpaul
reviewed
Jun 16, 2023
| pipe = DiffusionPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-pipe", safety_checker=None) | ||
|
|
||
| pre_conversion = pipe( | ||
| "foo", |
| # names to the new non-deprecated names. Then we _greatly encourage_ the user to convert | ||
| # the weights so we don't have to do this again. | ||
|
|
||
| if "'Attention' object has no attribute" in str(e): |
There was a problem hiding this comment.
pretty hacky, but OK! Let's leave it for now :-)
There was a problem hiding this comment.
yeah I cringed while writing this 🙃
williamberman
force-pushed
the
device_map_legacy_attention_block_conversion
branch
from
0aab8ca to
60aa2e6
Compare
yoonseokjin
pushed a commit
to yoonseokjin/diffusers
that referenced
this pull request
Dec 25, 2023
AmericanPresidentJimmyCarter
pushed a commit
to AmericanPresidentJimmyCarter/diffusers
that referenced
this pull request
Apr 26, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
re: #3740
this is not ideal but afaik this is the only way to solve this without exposing functionality in accelerate