Age | Commit message (Collapse) | Author | Lines | |
---|---|---|---|---|
2023-06-07 | Merge pull request #11066 from aljungberg/patch-1 | AUTOMATIC1111 | -1/+1 | |
Fix upcast attention dtype error. | ||||
2023-06-06 | Fix upcast attention dtype error. | Alexander Ljungberg | -1/+1 | |
Without this fix, enabling the "Upcast cross attention layer to float32" option while also using `--opt-sdp-attention` breaks generation with an error: ``` File "/ext3/automatic1111/stable-diffusion-webui/modules/sd_hijack_optimizations.py", line 612, in sdp_attnblock_forward out = torch.nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=0.0, is_causal=False) RuntimeError: Expected query, key, and value to have the same dtype, but got query.dtype: float key.dtype: float and value.dtype: c10::Half instead. ``` The fix is to make sure to upcast the value tensor too. | ||||
2023-06-04 | Merge pull request #10990 from vkage/sd_hijack_optimizations_bugfix | AUTOMATIC1111 | -1/+1 | |
torch.cuda.is_available() check for SdOptimizationXformers | ||||
2023-06-04 | fix the broken line for #10990 | AUTOMATIC | -1/+1 | |
2023-06-03 | torch.cuda.is_available() check for SdOptimizationXformers | Vivek K. Vasishtha | -1/+1 | |
2023-06-01 | revert default cross attention optimization to Doggettx | AUTOMATIC | -3/+3 | |
make --disable-opt-split-attention command line option work again | ||||
2023-06-01 | revert default cross attention optimization to Doggettx | AUTOMATIC | -3/+3 | |
make --disable-opt-split-attention command line option work again | ||||
2023-05-31 | rename print_error to report, use it with together with package name | AUTOMATIC | -2/+1 | |
2023-05-29 | Add & use modules.errors.print_error where currently printing exception info ↵ | Aarni Koskela | -4/+2 | |
by hand | ||||
2023-05-21 | Add a couple `from __future__ import annotations`es for Py3.9 compat | Aarni Koskela | -0/+1 | |
2023-05-19 | Apply suggestions from code review | AUTOMATIC1111 | -38/+28 | |
Co-authored-by: Aarni Koskela <akx@iki.fi> | ||||
2023-05-19 | fix linter issues | AUTOMATIC | -1/+1 | |
2023-05-18 | make it possible for scripts to add cross attention optimizations | AUTOMATIC | -3/+132 | |
add UI selection for cross attention optimization | ||||
2023-05-11 | Autofix Ruff W (not W605) (mostly whitespace) | Aarni Koskela | -16/+16 | |
2023-05-10 | ruff auto fixes | AUTOMATIC | -7/+7 | |
2023-05-10 | autofixes from ruff | AUTOMATIC | -1/+0 | |
2023-05-08 | Fix for Unet NaNs | brkirch | -0/+3 | |
2023-03-24 | Update sd_hijack_optimizations.py | FNSpd | -1/+1 | |
2023-03-21 | Update sd_hijack_optimizations.py | FNSpd | -1/+1 | |
2023-03-10 | sdp_attnblock_forward hijack | Pam | -0/+24 | |
2023-03-10 | argument to disable memory efficient for sdp | Pam | -0/+4 | |
2023-03-07 | scaled dot product attention | Pam | -0/+42 | |
2023-01-25 | Add UI setting for upcasting attention to float32 | brkirch | -60/+99 | |
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. | ||||
2023-01-23 | better support for xformers flash attention on older versions of torch | AUTOMATIC | -24/+18 | |
2023-01-21 | add --xformers-flash-attention option & impl | Takuma Mori | -2/+24 | |
2023-01-21 | extra networks UI | AUTOMATIC | -5/+5 | |
rework of hypernets: rather than via settings, hypernets are added directly to prompt as <hypernet:name:weight> | ||||
2023-01-06 | Added license | brkirch | -0/+1 | |
2023-01-06 | Change sub-quad chunk threshold to use percentage | brkirch | -9/+9 | |
2023-01-06 | Add Birch-san's sub-quadratic attention implementation | brkirch | -25/+99 | |
2022-12-20 | Use other MPS optimization for large q.shape[0] * q.shape[1] | brkirch | -4/+6 | |
Check if q.shape[0] * q.shape[1] is 2**18 or larger and use the lower memory usage MPS optimization if it is. This should prevent most crashes that were occurring at certain resolutions (e.g. 1024x1024, 2048x512, 512x2048). Also included is a change to check slice_size and prevent it from being divisible by 4096 which also results in a crash. Otherwise a crash can occur at 1024x512 or 512x1024 resolution. | ||||
2022-12-10 | cleanup some unneeded imports for hijack files | AUTOMATIC | -3/+0 | |
2022-12-10 | do not replace entire unet for the resolution hack | AUTOMATIC | -28/+0 | |
2022-11-23 | Patch UNet Forward to support resolutions that are not multiples of 64 | Billy Cao | -0/+31 | |
Also modifed the UI to no longer step in 64 | ||||
2022-10-19 | Remove wrong self reference in CUDA support for invokeai | Cheka | -1/+1 | |
2022-10-18 | Update sd_hijack_optimizations.py | C43H66N12O12S2 | -0/+3 | |
2022-10-18 | readd xformers attnblock | C43H66N12O12S2 | -0/+15 | |
2022-10-18 | delete xformers attnblock | C43H66N12O12S2 | -12/+0 | |
2022-10-11 | Use apply_hypernetwork function | brkirch | -10/+4 | |
2022-10-11 | Add InvokeAI and lstein to credits, add back CUDA support | brkirch | -0/+13 | |
2022-10-11 | Add check for psutil | brkirch | -4/+15 | |
2022-10-11 | Add cross-attention optimization from InvokeAI | brkirch | -0/+79 | |
* Add cross-attention optimization from InvokeAI (~30% speed improvement on MPS) * Add command line option for it * Make it default when CUDA is unavailable | ||||
2022-10-11 | rename hypernetwork dir to hypernetworks to prevent clash with an old ↵ | AUTOMATIC | -1/+1 | |
filename that people who use zip instead of git clone will have | ||||
2022-10-11 | fixes related to merge | AUTOMATIC | -1/+2 | |
2022-10-11 | replace duplicate code with a function | AUTOMATIC | -29/+15 | |
2022-10-10 | remove functorch | C43H66N12O12S2 | -2/+0 | |
2022-10-09 | Fix VRAM Issue by only loading in hypernetwork when selected in settings | Fampai | -3/+3 | |
2022-10-08 | make --force-enable-xformers work without needing --xformers | AUTOMATIC | -1/+1 | |
2022-10-08 | add fallback for xformers_attnblock_forward | AUTOMATIC | -1/+4 | |
2022-10-08 | simplify xfrmers options: --xformers to enable and that's it | AUTOMATIC | -7/+13 | |
2022-10-08 | emergency fix for xformers (continue + shared) | AUTOMATIC | -8/+8 | |