Skip to content
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.

Sign up for GitHub

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

Jump to bottom

sd : taesd faster decoding implementation #87

Closed
wants to merge 69 commits into from
Closed

sd : taesd faster decoding implementation #87

wants to merge 69 commits into from

Conversation

FSSRepo
Copy link
Contributor

@FSSRepo FSSRepo commented Nov 26, 2023 •
edited
Loading

Unused

FSSRepo and others added 30 commits November 20, 2023 22:36
@FSSRepo @leejet
set ggml url to FSSRepo/ggml
75f6183
@FSSRepo @leejet
ggml-alloc integration
b89ca2c
@leejet
update ggml to FSSRepo/sd-cpp-fixes latest commit
c149384
@FSSRepo @leejet
some fixes
c778941
@FSSRepo @leejet
convert.py some ignore tensor model_ema.diffusion_model
6f8bba4
@FSSRepo @leejet
update ggml submodule + remove duplicated code
d406517
@FSSRepo @leejet
fix timings
78b87a5
@FSSRepo @leejet
update submodule + clip model gpu backend
d348a1d
@FSSRepo @leejet
update submodule + unet cuda backend (broken)
f1556fe
@FSSRepo @leejet
update submodule + offload all functions to gpu
3dc28a7
@FSSRepo @leejet
delete unused executable
665cca5
@FSSRepo @leejet
fix submodule
7bff159
@FSSRepo @leejet
fix submodule
34e9a43
@FSSRepo @leejet
change branch submodule
cb122aa
@FSSRepo @leejet
fix cmake error
a575655
@FSSRepo @leejet
fix sd 2.x alloc error
899a4a8
@FSSRepo @leejet
gguf format + native converter
1e7f055
@FSSRepo @leejet
add converter docs
8d8bc76
@FSSRepo @leejet
warning unsopported models
0330249
@FSSRepo @leejet
correct some print verbose
c250948
@FSSRepo @leejet
update ggml submodule + cuda speedup
88defcf
@FSSRepo @leejet
clip text model use cuda backend
c0bd14c
@FSSRepo @leejet
prepare the converter to support lora
0472945
@FSSRepo @leejet
converter: bf16 unsupported error
cf442f5
@FSSRepo @leejet
lora info + verbose converter
35fd078
@FSSRepo @leejet
linux fix segmentation fault
ed737bf
@FSSRepo @leejet
avoid upload dev output image
c877027
@FSSRepo @leejet
fix issue with lora alphas f16
429e4bf
@FSSRepo @leejet
convert support xl models + some info for next PR
95c3459
@FSSRepo @leejet
converter support all sd-1.5 models
9189cce
FSSRepo and others added 28 commits November 21, 2023 00:24
@FSSRepo @leejet
full offload to gpu
ea2c82d
@FSSRepo @leejet
vae encoder offload to gpu
2a1f820
@FSSRepo @leejet
adapt LoRA support to ggml_backend
84bc2bd
@FSSRepo
remove some unused code + converter low memory usage
db8ffd3
@FSSRepo
convert support bfloat16
4fcbf9b
@FSSRepo
update README.md
f8e5eaa
@leejet
fix extract_and_remove_lora
8242631
@FSSRepo
avoid unnecessary memory usage
ec268ff
@FSSRepo
show lora warning
66a02d4
@FSSRepo
flush stdout - pretty progress
d02d4b2
@FSSRepo
added batch-count option to CLI
75d0a4f
@FSSRepo
useful comments
fa749b1
@FSSRepo
improve LoRA detection
33d9149
@FSSRepo
improve pretty progress
8a3256f
@FSSRepo
fix and improve timings
3b688b5
@FSSRepo
delete leaked images
750366e
@FSSRepo
fix img2img
1bd91ca
@FSSRepo
simplify saving images
0938476
@leejet
ignore models
eca9455
@leejet
refactor command-line parameter formats for consistency
3dc5855 
short options should be only a single character
@leejet
a few modifications
7f1a59a
@leejet
clear the compilation warnings
27dfe3c
@leejet
standardize naming conventions
60d9f8a 
employ PascalCase for class/struct/enumnames and underscore_case for variables/functions/methods.
@leejet
attempt to fix a build failure in some compilation environments
b9a02c1
@leejet
attempt to fix a build failure in macOS
40d3f7e
@leejet
format code using .clang-format
9affd38
@FSSRepo
Merge branch 'leejet:master' into master
d14c8c9
@FSSRepo
dummy implementation TAESD
78d1e3b
@FSSRepo FSSRepo closed this Nov 26, 2023
@leejet
Copy link
Owner

leejet commented Nov 26, 2023

I don't know why the quality degrades a bit, probably because the weights of the convolutions are in f16, there must be a slight drop in quality, or it must be inherent to the decoder model.

Because this is a distilled model and the structure is different, it is normal that there are differences.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@FSSRepo @leejet