kijai
24a7edfca6
update, works
2024-11-05 08:41:50 +02:00
kijai
3535a846a8
update from main
2024-11-05 07:08:14 +02:00
Yoshimasa Niwa
5ca4bbf319
Workaround pad problem on mps
...
When using `torch.nn.functional.pad` with tensor that size is
larger than 2^16 (65526), the output tensor would be broken.
This patch moves tensor to CPU to workaround the problem.
It doesn't too much impacts in terms of speed of vea on mps.
2024-11-05 12:46:13 +09:00
kijai
56b5dbbf82
Add different RMSNorm functions for testing
...
Initial testing for me shows that the RMSNorm from flash_attn.ops.triton.layer_norm is ~8-10% faster, apex is untested as I don't currently have it installed.
2024-11-04 14:11:58 +02:00
kijai
4a7458ffd6
restore MochiEdit compatiblity
...
temporary
2024-11-03 19:18:29 +02:00
kijai
0dc011d1b6
cleanup and align more to comfy code, switch to using cpu seed as well
2024-11-03 18:36:23 +02:00
kijai
3cf9289e08
cleanup code
2024-11-03 01:47:34 +02:00
kijai
a6e545531c
test
2024-11-03 00:32:51 +02:00
kijai
85c996d7b8
Add MochiSigmaSchedule node, better denoise formula
2024-11-01 19:36:40 +02:00
kijai
a5b06b02ad
tiled encoding
2024-11-01 18:03:47 +02:00
kijai
69ab797b8c
Add sampler preview
2024-11-01 06:18:41 +02:00
kijai
ac5de728ad
Add VAE encoder
2024-11-01 05:22:49 +02:00
kijai
d971a19410
spatial VAE decoder fixes
2024-10-30 22:06:55 +02:00
kijai
f0f939b20b
cleanup, fix untiled spatial vae decode
2024-10-30 21:12:34 +02:00
kijai
3395aa8ca0
cleanup, sampler output name fix
2024-10-27 20:02:44 +02:00
kijai
195da244df
make cfg 1.0 not do uncond, set steps by sigma schedule
2024-10-27 19:52:16 +02:00
kijai
3613700752
possible sdpa kernel fixes and add optional cfg scheduling
2024-10-27 12:23:01 +02:00
kijai
e20eb66f93
cleanup
2024-10-27 03:00:52 +02:00
kijai
c5c136cb11
fix
2024-10-26 17:57:02 +03:00
kijai
ddfb3a6bf2
backends
2024-10-26 17:49:15 +03:00
kijai
f29f739707
support cublas_ops with GGUF
...
pretty big speed boost on 4090 at least, needs this installed:
https://github.com/aredden/torch-cublas-hgemm
2024-10-26 16:42:25 +03:00
kijai
0d15c0bd69
torch compile for vae loader
2024-10-26 03:24:25 +03:00
kijai
b932036af3
Update asymm_models_joint.py
2024-10-25 22:37:08 +03:00
kijai
e66735527c
tweak
2024-10-25 20:04:01 +03:00
kijai
2e22529c99
clean
2024-10-25 19:51:15 +03:00
kijai
bd844331b2
Update t2v_synth_mochi.py
2024-10-25 19:41:50 +03:00
kijai
aa30132268
temporary monkey patch for torch compile Windows bug
2024-10-25 19:41:10 +03:00
kijai
25eeab3c4c
torch.compile support
...
works in Windows with torch 2.5.0 and Triton from https://github.com/woct0rdho/triton-windows
2024-10-25 18:15:30 +03:00
kijai
36a4275b3b
Add alternative VAE decoding node
...
This was actually unused code in the VAE model, only does spatial tiling though, but seams look better
2024-10-25 15:30:20 +03:00
kijai
813bbe8f4b
Add model and vae loader nodes
2024-10-24 21:38:06 +03:00
kijai
2c67025577
remove prints
2024-10-24 17:16:26 +03:00
kijai
f4c13b1ef4
Add first GGUF test version
2024-10-24 17:05:50 +03:00
kijai
d699fae213
cleanup, possibly support older GPUs
2024-10-24 14:27:11 +03:00
kijai
257c526125
Cleanup, fix seed gen, better warnings for decoder
2024-10-24 12:45:01 +03:00
kijai
c673508188
tweaks
2024-10-24 02:59:52 +03:00
Jukka Seppänen
00a550e81c
Should work without flash_attn (thanks @logtd), add sage_attn
...
tested to work in Linux at least
2024-10-24 02:25:57 +03:00
Jukka Seppänen
1ba3ac8e25
Revert "works without flash_attn (thanks @juxtapoz!)"
...
This reverts commit a1b1f86aa3b6780a4981157b8e2e37b0a1017568.
2024-10-24 02:23:46 +03:00
Jukka Seppänen
a1b1f86aa3
works without flash_attn (thanks @juxtapoz!)
...
at least on Linux, also sage_attn
2024-10-24 02:19:00 +03:00
kijai
83097a6b63
Update t2v_synth_mochi.py
2024-10-24 00:40:52 +03:00
kijai
bd954ec132
Update t2v_synth_mochi.py
2024-10-24 00:36:38 +03:00
kijai
508eaa22df
Update t2v_synth_mochi.py
2024-10-24 00:34:50 +03:00
kijai
1cd5409295
Add bf16 model
2024-10-24 00:00:38 +03:00
kijai
a32064eefb
fix non-accelerate model loading
2024-10-23 20:40:42 +03:00
kijai
57640ab0f8
Update t2v_synth_mochi.py
2024-10-23 19:38:01 +03:00
kijai
fb880273a0
update
2024-10-23 17:04:50 +03:00
kijai
db87f8e608
Add accelerate
2024-10-23 16:05:19 +03:00
kijai
34e029bacc
fix dtype selection
2024-10-23 15:49:37 +03:00
kijai
4efb7c85df
cleanup
2024-10-23 15:45:20 +03:00
kijai
b80cb4a691
initial commit
2024-10-23 15:34:22 +03:00
