2051c4100b
Apply emphasis mode changes immediately ( #2423 )
...
Previously emphasis mode changes applied only on model load.
Also added handling for mode 'None' (literal interpretation), which previously gave results identical to mode 'Ignore'.
2024-12-08 19:25:39 +00:00
19a9a78c9b
fix/workaround for potential memory leak ( #2315 )
...
unload old models, based on reference count <= 2
(in practise only noticed extra copies of JointTextEncoder, never KModel or IntegratedAutoencoderKL)
#2281 #2308 and others
2024-11-14 22:05:54 +00:00
6e1a7908b4
Fallback to estimated prediction if .yaml prediction not suitable ( #2273 )
2024-11-06 23:49:18 -05:00
f4afbaff45
Only use .yaml config prediction if actually set ( #2272 )
2024-11-06 23:44:12 -05:00
05b01da01f
Fix ldm .yaml loading typo ( #2248 )
2024-11-02 10:29:13 -04:00
90a6970fb7
Compatibility for ldm .yaml configs ( #2247 )
2024-11-02 10:16:51 -04:00
d50f390c7e
rewrite mu shift
2024-10-31 22:57:26 -07:00
6f4350d65f
Fix typo in .yaml config load ( #2228 )
2024-10-31 07:09:27 -04:00
b691b1e755
Fix .yaml config loading ( #2224 )
2024-10-30 16:18:44 -04:00
41a21f66fd
Restore embedding filepath. ( #2030 )
2024-10-26 23:53:48 +01:00
edeb2b883f
Fix loading diffusers format VAEs ( #2171 )
2024-10-24 14:27:57 -04:00
edc46380cc
Automatically enable ztSNR when applicable ( #2122 )
2024-10-19 20:33:34 -04:00
aba35cde5f
Fix Zero Terminal SNR option
2024-10-19 12:28:14 -04:00
f620f55e56
Fallback to already detected prediction type if none applicable found
2024-10-19 07:47:56 -04:00
5ec47a6b93
Fix model prediction detection
...
Closes #1109
2024-10-19 06:18:02 -04:00
cc378589a4
(T5) pad chunks to length of largest chunk ( #1990 )
...
When the prompt is chunked using the BREAK keyword, chunks will be padded to the minimum size of 256 tokens - but chunks can be longer. torch.stack then fails if all chunks are not the same size, so find the largest and pad all to match.
#1988 (doesn't quite ID the real issue, prompts longer than 255 tokens work fine)
2024-10-07 11:37:06 +01:00
2467c88c50
fix for XPU ( #1997 )
...
use float32 for XPU, same as previous fix for MPS
2024-10-06 14:33:47 +01:00
8bd7e0568f
Get Flux working on Apple Silicon ( #1264 )
...
Co-authored-by: Conor Nash <conor@nbs.consulting >
2024-09-13 15:40:11 +01:00
c13b26ba27
Rephrase low GPU warning ( #1761 )
...
Make emphasis that it's performance degradation, and for this diffusion process.
2024-09-09 14:30:22 -03:00
efe6fed499
add a way to exchange variables between modules
2024-09-08 20:22:04 -07:00
f40930c55b
fix
2024-09-08 17:24:53 -07:00
44eb4ea837
Support T5&Clip Text Encoder LoRA from OneTrainer
...
requested by #1727
and some cleanups/licenses
PS: LoRA request must give download URL to at least one LoRA
2024-09-08 01:39:29 -07:00
a8a81d3d77
fix offline quant lora precision
2024-08-31 13:12:23 -07:00
79b25a8235
move codes
2024-08-31 11:31:02 -07:00
33963f2d19
always compute on-the-fly lora weights when offload
2024-08-31 11:24:23 -07:00
70a555906a
use safer codes
2024-08-31 10:55:19 -07:00
1f91b35a43
add signal_empty_cache
2024-08-31 10:20:22 -07:00
ec7917bd16
fix
2024-08-30 15:37:15 -07:00
d1d0ec46aa
Maintain patching related
...
1. fix several problems related to layerdiffuse not unloaded
2. fix several problems related to Fooocus inpaint
3. Slightly speed up on-the-fly LoRAs by precomputing them to computation dtype
2024-08-30 15:18:21 -07:00
f04666b19b
Attempt #1575
2024-08-30 09:41:36 -07:00
4c9380c46a
Speed up quant model loading and inference ...
...
... based on 3 evidences:
1. torch.Tensor.view on one big tensor is slightly faster than calling torch.Tensor.to on multiple small tensors.
2. but torch.Tensor.to with dtype change is significantly slower than torch.Tensor.view
3. “baking” model on GPU is significantly faster than computing on CPU when model load.
mainly influence inference of Q8_0, Q4_0/1/K and loading of all quants
2024-08-30 00:49:05 -07:00
3d62fa9598
reduce prints
2024-08-29 20:17:32 -07:00
95e16f7204
maintain loading related
...
1. revise model moving orders
2. less verbose printing
3. some misc minor speedups
4. some bnb related maintain
2024-08-29 19:05:48 -07:00
d339600181
fix
2024-08-28 09:56:18 -07:00
81d8f55bca
support pytorch 2.4 new normalization features
2024-08-28 09:08:26 -07:00
0abb6c4686
Second Attempt for #1502
2024-08-28 08:08:40 -07:00
f22b80ef94
restrict baking to 16bits
2024-08-26 06:16:13 -07:00
388b70134b
fix offline loras
2024-08-25 20:28:40 -07:00
b25b62da96
fix T5 not baked
2024-08-25 17:31:50 -07:00
cae37a2725
fix dequant of unbaked parameters
2024-08-25 17:24:31 -07:00
13d6f8ed90
revise GGUF by precomputing some parameters
...
rather than computing them in each diffusion iteration
2024-08-25 14:30:09 -07:00
f82029c5cf
support more t5 quants ( #1482 )
...
lets hope this is the last time that people randomly invent new state dict key formats
2024-08-24 12:47:49 -07:00
f23ee63cb3
always set empty cache signal as long as any patch happens
2024-08-23 08:56:57 -07:00
2ab19f7f1c
revise lora patching
2024-08-22 11:59:43 -07:00
68bf7f85aa
speed up nf4 lora in offline patching mode
2024-08-22 10:35:11 -07:00
95d04e5c8f
fix
2024-08-22 10:08:21 -07:00
14eac6f2cf
add a way to empty cuda cache on the fly
2024-08-22 10:06:39 -07:00
909ad6c734
fix prints
2024-08-21 22:24:54 -07:00
0d8eb4c5ba
fix #1375
2024-08-21 11:01:59 -07:00
4e3c78178a
[revised] change some dtype behaviors based on community feedbacks
...
only influence old devices like 1080/70/60/50.
please remove cmd flags if you are on 1080/70/60/50 and previously used many cmd flags to tune performance
2024-08-21 10:23:38 -07:00
DenOfEquity