asagi4
ec8c40355c
Split documentation
2025-06-12 23:03:39 +03:00
asagi4
6e538e0abc
Fix markdown syntax
2025-06-12 22:42:45 +03:00
asagi4
25c44a1fbb
Documentation
2025-06-12 22:41:31 +03:00
asagi4
72d5490498
Add support for commenting out things with #
...
You can escape it with \#
Fixes #105
2025-06-12 22:26:34 +03:00
asagi4
3de4538326
Test cleanup
2025-06-12 21:38:56 +03:00
asagi4
f61af15d52
Update the description a bit
2025-06-10 19:05:52 +03:00
asagi4
8e59f140ff
v2.0.0-rc.8
2025-06-09 20:07:54 +03:00
asagi4
44044e962c
Very basic test for COUPLE
2025-06-09 20:06:25 +03:00
asagi4
04f36687c5
Fix skipping prompt segments by setting weight to 0
2025-06-09 20:05:50 +03:00
asagi4
3a8a360d03
Make testing less stupid
2025-06-09 19:53:05 +03:00
asagi4
d76331315a
Deduplicate tests
2025-06-09 19:15:55 +03:00
asagi4
e3a6050536
Don't call to() on every iteration
2025-06-09 18:37:53 +03:00
asagi4
7b001ace7b
Fix Attention Couple when combined with hooks on the CLIP (eg. LoRAs)
...
All clones of the AC hook must maintain the same state. This feels
a bit hacky though; there should be a better way
Fixes #119
2025-06-09 17:25:26 +03:00
asagi4
f1de65f257
Don't override existing hooks. Unfortunately, this doesn't make things quite work; hmm.
2025-06-09 16:22:25 +03:00
asagi4
11aaa0ac7b
Fix indexing error
2025-06-09 01:06:44 +03:00
asagi4
85de3ef0d3
Fix links
2025-06-08 23:09:16 +03:00
asagi4
a73260ff34
Split t5 tests
2025-06-08 23:05:11 +03:00
asagi4
50a2e0abbf
Change ATTN() to COUPLE() and remove need for AND
2025-06-08 23:01:59 +03:00
asagi4
5cf45ca264
Make functions generally callable without argument lists
2025-06-08 18:57:25 +03:00
asagi4
bd4a787400
Use a helper function to parse function splits
2025-06-08 18:56:30 +03:00
asagi4
c9e5bc25c3
Need to do imports after torch mock, otherwise running tests on CPU torch fails
2025-06-08 17:22:54 +03:00
asagi4
d11ffa6e25
Don't pass in a custom prefix to GraphBuilder
...
It breaks when PCLazyTextEncode etc. are called with list inputs.
Tests needed adjusting after the change.
Fixes #117
2025-06-08 16:49:37 +03:00
asagi4
8cc73a2e49
v2.0.0-rc.7
2025-06-08 01:17:08 +03:00
asagi4
27ae5f683e
Of course I forgot to test NegPiP
2025-06-08 01:15:16 +03:00
asagi4
68cda3663e
v2.0.0-rc.6
2025-06-08 01:02:16 +03:00
asagi4
3d46f705b6
Test downweighting too, and normalizations
2025-06-08 01:01:57 +03:00
asagi4
eec4bc4da9
Split old code into its own file for easy removal later
2025-06-08 00:55:43 +03:00
asagi4
3d3218e831
Tests for verifying refactor
2025-06-08 00:55:43 +03:00
asagi4
f4e57ec514
Switch on new implementation by default
2025-06-08 00:55:40 +03:00
asagi4
74f65c1b31
Use old from_masked batching for now
...
I can't figure out why from_masked works differently from down_weight
which also used that batching function but could be replaced with a simple
torch.cat
2025-06-08 00:50:26 +03:00
asagi4
4d94cca88f
Restore adv_encoding to original implementation to compare them
2025-06-08 00:44:56 +03:00
asagi4
4569ecccf9
Manual testing...
2025-06-08 00:44:31 +03:00
asagi4
192e6d30d4
refactor adv_encode
2025-06-07 21:41:13 +03:00
asagi4
67112f11e0
T5 makes STYLE(perp) return NaNs. Just replace them with 0
2025-06-07 21:41:13 +03:00
asagi4
f761dfac86
Fix comfy++ with more than one weight, see #115
...
I'm not sure if this is correct, but it at least doesn't fail.
2025-06-07 14:09:32 +03:00
asagi4
278e733835
Fix normalization validity check
2025-06-07 14:09:32 +03:00
asagi4
2bf65720eb
Fix STYLE(perp) exception, see #115
2025-06-07 14:09:32 +03:00
asagi4
99f3af92b7
Add tests for weighting and a way to run manual testing
2025-06-07 14:09:14 +03:00
asagi4
2437cd4daf
Merge pull request #116 from pamparamm/pooled_none_check
...
Partially resolve #115
2025-06-07 13:30:01 +03:00
asagi4
d262a7dc7a
Remove an extra clone.
2025-06-07 13:23:06 +03:00
Pam
6c319ad5b4
Partially resolve #115
2025-06-07 09:38:13 +05:00
asagi4
34056cac19
v2.0.0-rc.5
2025-06-06 19:52:02 +03:00
asagi4
8ae436abf1
Merge pull request #114 from pamparamm/negpip_option
...
Use ppm_negpip option to detect NegPiP
2025-06-06 16:49:20 +03:00
Pam
c2ce2ce023
Use ppm_negpip option to detect NegPiP
2025-06-06 16:24:15 +05:00
asagi4
7a9e69ec31
Fix errors found in testing
...
Who knew tests could be useful, too?
2025-06-05 23:49:02 +03:00
asagi4
aaff8dc7da
Encoding tests
2025-06-05 23:49:02 +03:00
asagi4
3e4722278a
v2.0.0-rc.4
2025-06-05 21:39:47 +03:00
asagi4
11cb430396
Support NegPiP without requiring a monkeypatch.
2025-06-05 21:38:31 +03:00
asagi4
9f1cbfd11c
Add a test suite for text encoding
...
Can only be run manually and imports ComfyUI main to configure search paths, but
at least it works...
2025-06-05 21:05:39 +03:00
asagi4
5b3a914f1d
Try not to fail in cases where masks have irregular shapes
2025-06-04 22:34:19 +03:00
asagi4