 kijai
|
ec0fdf1fc4
|
consistently use safe_load
|
2024-12-19 01:53:00 +02:00 |
|
|
kijai
|
fb34e91f0e
|
Support official scaled fp8 model
|
2024-12-18 10:29:56 +02:00 |
|
|
kijai
|
46f569f571
|
clearer error on frame count mismatch
|
2024-12-18 00:34:12 +02:00 |
|
|
kijai
|
ecae5f270e
|
small fixes
|
2024-12-14 22:28:02 +02:00 |
|
|
kijai
|
4a81c5ec94
|
Add CFG as optional node
experimental, using CFG halves the inference speed but allows using negative prompt, possibly more useful when using LoRAs
|
2024-12-13 16:59:49 +02:00 |
|
|
kijai
|
706c3a07ef
|
Add nodes for text embed saving/loading
This way you can encode prompts in separate workflow to avoid ever having to load the text encoders for sampling if you're low on RAM
|
2024-12-13 13:07:34 +02:00 |
|
|
kijai
|
9a4abbc535
|
Allow using comfy clip_l, fix fp8 fastmode mem use
|
2024-12-08 20:34:31 +02:00 |
|
|
kijai
|
d49a711cd8
|
Update models.py
|
2024-12-08 02:00:41 +02:00 |
|
|
kijai
|
f5d80350c0
|
remove prints
|
2024-12-08 01:47:06 +02:00 |
|
|
kijai
|
e9e058436e
|
block_swap tweaks
|
2024-12-08 01:44:16 +02:00 |
|
|
kijai
|
929781949e
|
Add STG start/end percent
|
2024-12-07 21:51:37 +02:00 |
|
|
kijai
|
0cbcc20a89
|
initial STG testing
|
2024-12-06 17:09:56 +02:00 |
|
|
kijai
|
2e8f1f2fbd
|
fix flash_attn, make sdpa default
|
2024-12-06 11:12:10 +02:00 |
|
|
kijai
|
31950d0b55
|
possibly fix up sdpa
|
2024-12-06 03:15:02 +02:00 |
|
|
kijai
|
3197107b5e
|
fixup block_swap start
|
2024-12-04 03:53:51 +02:00 |
|
|
kijai
|
1073ef9156
|
custom block_swap to reduce VRAM use
thanks to 2kpr for the initial block swap code!
|
2024-12-04 02:13:48 +02:00 |
|
|
kijai
|
fa45b717fe
|
Update models.py
|
2024-12-03 19:14:00 +02:00 |
|
|
kijai
|
73b49a30a6
|
initial commit
|
2024-12-03 19:09:50 +02:00 |
|