hoelee/stable-diffusion-webui
1
0
Fork 0
mirror of https://github.com/AUTOMATIC1111/stable-diffusion-webui.git synced 2025-08-08 13:19:54 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Add UI setting for upcasting attention to float32

Browse Source 
Adds "Upcast cross attention layer to float32" option in Stable Diffusion settings. This allows for generating images using SD 2.1 models without --no-half or xFormers.

In order to make upcasting cross attention layer optimizations possible it is necessary to indent several sections of code in sd_hijack_optimizations.py so that a context manager can be used to disable autocast. Also, even though Stable Diffusion (and Diffusers) only upcast q and k, unfortunately my findings were that most of the cross attention layer optimizations could not function unless v is upcast also.
This commit is contained in:
brkirch
2023-01-25 00:23:10 -05:00
parent 84d9ce30cb
commit e3b53fd295
5 changed files with 108 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
modules/sub_quadratic_attention.py
View File

@@ -67,7 +67,7 @@ def _summarize_chunk(
max_score, _ = torch.max(attn_weights, -1, keepdim=True)
max_score = max_score.detach()
exp_weights = torch.exp(attn_weights - max_score)
exp_values = torch.bmm(exp_weights, value)
exp_values = torch.bmm(exp_weights, value) if query.device.type == 'mps' else torch.bmm(exp_weights, value.to(exp_weights.dtype)).to(value.dtype)
max_score = max_score.squeeze(-1)
return AttnChunk(exp_values, exp_weights.sum(dim=-1), max_score)
@@ -129,7 +129,7 @@ def _get_attention_scores_no_kv_chunking(
)
attn_probs = attn_scores.softmax(dim=-1)
del attn_scores
hidden_states_slice = torch.bmm(attn_probs, value)
hidden_states_slice = torch.bmm(attn_probs, value) if query.device.type == 'mps' else torch.bmm(attn_probs, value.to(attn_probs.dtype)).to(value.dtype)
return hidden_states_slice