[wip] attention refactor #2143
Closed
[wip] attention refactor #2143
Conversation
This file contains 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
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. |
williamberman
force-pushed
the
attention_refactor
branch
2 times, most recently
from
2658f5a to
c6709ed
Compare
williamberman
commented
Jan 27, 2023
src/diffusers/models/attention.py
Outdated
Comment on lines
119
to
137
| attn = CrossAttention( | ||
| self.channels, | ||
| heads=self.num_heads, | ||
| dim_head=dim_head, | ||
| bias=True, | ||
| upcast_softmax=True, | ||
| norm_num_groups=self.group_norm.num_groups, | ||
| processor=processor, | ||
| eps=self.group_norm.eps, | ||
| rescale_output_factor=self.rescale_output_factor, | ||
| ) | ||
|
|
||
| # compute next hidden_states | ||
| hidden_states = self.proj_attn(hidden_states) | ||
| attn.group_norm = self.group_norm | ||
| attn.to_q = self.query | ||
| attn.to_k = self.key | ||
| attn.to_v = self.value | ||
| attn.to_out[0] = self.proj_attn | ||
|
|
||
| hidden_states = hidden_states.transpose(-1, -2).reshape(batch, channel, height, width) | ||
| hidden_states = attn(hidden_states) |
There was a problem hiding this comment.
Creating CrossAttention on the fly like this causes some of the mps tests that rely on reproducibility to fail. Could this have something to do with the mps warm up passes? cc @pcuenca
williamberman
force-pushed
the
attention_refactor
branch
10 times, most recently
from
e11a51a to
52f16b8
Compare
williamberman
commented
Jan 30, 2023
|
|
||
| from ...configuration_utils import ConfigMixin, register_to_config | ||
| from ...schedulers.scheduling_utils import SchedulerMixin | ||
|
|
||
|
|
||
| warnings.filterwarnings("ignore") |
There was a problem hiding this comment.
This warnings filter silences all warnings. Need to remove to see the attention block deprectation warning
williamberman
force-pushed
the
attention_refactor
branch
from
cbec387 to
ae18d1d
Compare
williamberman
force-pushed
the
attention_refactor
branch
from
ae18d1d to
06f3e9b
Compare
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: #1880
AttentionBlock's forward is called, dynamically create aCrossAttentionmodule with the same parameters and call its forward method instead.