kijai
a5ae70624d
Support Stand-In LoRA
...
https://github.com/WeChatCV/Stand-In
2025-08-12 21:51:38 +03:00
kijai
057ee642f6
Add async prefetch as option for block swapping
...
Can speed up block swapping on some systems, added debug option as well so you can see if enabling the option gives you any benefit.
2025-08-11 16:17:16 +03:00
kijai
7c93aea182
rename fp8 optimization in code for clarity as it does more than that with loras
2025-08-11 14:32:57 +03:00
kijai
341932b58f
Fix unmerged lora application on unquantized models
2025-08-11 01:34:25 +03:00
kijai
8185e9b3cb
Better fp8 linear layer patching with torch.compile
2025-08-09 16:18:33 +03:00
kijai
cf8e403f88
Fix gguf
2025-08-09 13:28:20 +03:00
kijai
0da602f8e9
silence the img key error spam when using 2.1 I2V loras on 2.2 or T2V
2025-08-09 12:17:55 +03:00
kijai
ceedc9e493
Update nodes_model_loading.py
2025-08-09 10:49:42 +03:00
kijai
48fa904ad8
fp8 matmul for scaled models
...
Fp8 matmul (fp8_fast) doesn't seem feasible with unmerged LoRAs as you'd need to first upcast, then apply LoRA, then downcast back to fp8 and that is too slow. Direct adding in fp8 is also not possible since that's just not something fp8 dtypes support.
2025-08-09 10:17:11 +03:00
kijai
9be0eb133c
Fix lora scheduling in multilora loader
2025-08-08 13:47:08 +03:00
kijai
e1acc2c4e7
Support Fun 2.2
2025-08-08 11:49:04 +03:00
kijai
92eac2a9e1
typo
2025-08-05 18:25:42 +03:00
kijai
bacb2a2053
Update nodes_model_loading.py
2025-08-05 01:44:42 +03:00
kijai
8e71286b6d
Allow torch.compile VAE decoder
...
Slight speedup especially for 5B...
2025-08-05 01:31:31 +03:00
kijai
78a7e20838
Fix lora schedule with multilora node
2025-08-04 16:51:53 +03:00
kijai
3552478d02
Update nodes_model_loading.py
2025-08-04 00:04:21 +03:00
kijai
801f543a7a
Allow LoRA timestep scheduling
...
LoRA strength can now be a list of floats that represents the strength at any given step
2025-08-04 00:02:19 +03:00
kijai
309a93c221
Update nodes_model_loading.py
2025-08-02 21:13:29 +03:00
kijai
c93e69fb9a
Merge branch 'main' of https://github.com/kijai/ComfyUI-WanVideoWrapper
2025-08-02 19:53:26 +03:00
kijai
b5199ec7e0
Fix control lora
2025-08-02 19:53:20 +03:00
rkfg
ed247d78b6
Reduce tdqm spam
2025-08-02 16:38:27 +03:00
kijai
83bb0e7f35
fp8 text encoder fixes
2025-08-01 17:42:32 +03:00
kijai
9edab74562
Default non_blocking to False
...
too many RAM issues
2025-08-01 02:29:07 +03:00
kijai
7eebde487f
typo
2025-08-01 00:54:54 +03:00
kijai
4eabdcce78
wrong way around
2025-07-30 18:45:19 +03:00
kijai
138b377d1d
Fix loading fp8 models where everything is fp8...
2025-07-30 17:34:56 +03:00
kijai
ec066008a8
Make start/end step work with vid2vid
2025-07-30 15:59:22 +03:00
kijai
1570287db1
Update nodes_model_loading.py
2025-07-29 18:10:02 +03:00
kijai
15a57d22f7
fix slightly increased memory use
2025-07-29 18:08:32 +03:00
kijai
234cffbf24
Fix comfy org scaled models giving black output
2025-07-29 08:05:48 +03:00
kijai
fb7073259f
detect 2.2 I2V
...
untested
2025-07-28 22:43:37 +03:00
kijai
d6e9af7090
2.2 stuff
2025-07-28 21:38:40 +03:00
kijai
9a1ab1c656
only show lora metadata if it actually exists
2025-07-26 14:32:53 +03:00
kijai
34c2cc23a9
Update nodes_model_loading.py
2025-07-25 20:49:41 +03:00
kijai
a35eb7dfa4
Basic sageattn3 support
...
Probably going to need further timestep scheduling as the quality of sage3 on it's own is pretty bad.
2025-07-25 20:48:34 +03:00
kijai
76750eec05
Allow loading aitoolkit/lycoris format
2025-07-25 17:24:43 +03:00
kijai
b3820055d0
Fix GGUF with LoRAs and support GGUF witgh SetLoras -node
2025-07-24 14:59:59 +03:00
kijai
35e637cbfd
Allow merging LoRA to fp8 scaled models
2025-07-24 11:51:36 +03:00
kijai
d6425cab02
Update nodes_model_loading.py
2025-07-24 00:08:39 +03:00
kijai
838803d053
Update nodes_model_loading.py
2025-07-23 17:58:28 +03:00
kijai
bdce322e65
remove loras if SetLora bypassed/disconnected
2025-07-23 17:51:59 +03:00
kijai
5e56e3649f
fix control lora
2025-07-23 16:20:30 +03:00
kijai
fa93f5b3c1
Add optional prompt disk caching
...
Caches text embeds to disk with unique hash based on T5 name, dtype and prompts. If found on disk skips encoding, persists after restarts.
2025-07-23 00:52:56 +03:00
kijai
9c45f3a97d
Update nodes_model_loading.py
2025-07-23 00:10:14 +03:00
kijai
eea896dace
Update nodes_model_loading.py
2025-07-22 18:25:03 +03:00
kijai
70fcdff3c5
Update nodes_model_loading.py
2025-07-22 17:36:41 +03:00
kijai
bfd6af8141
Update nodes_model_loading.py
2025-07-22 15:41:32 +03:00
kijai
98fa84bb17
Update nodes_model_loading.py
2025-07-22 13:58:43 +03:00
kijai
3988accdf3
Update nodes_model_loading.py
2025-07-21 23:44:24 +03:00
kijai
296baa30ce
Add WanVideoSetLoRAs
...
Node to set the LoRA weights to use with the unmerged LoRA mode, not able to merge LoRAs but allows instant LoRA switching without any loading times. The effect of unmerged LoRAs is stronger and differs from merged LoRAs.
2025-07-21 22:39:17 +03:00
kijai