addd479729
Fix Generate Text ignoring thinking=false on Gemma4 E2B/E4B ( #15278 )
2026-08-12 22:44:14 -04:00
2220d111c8
Query pytorch for aotriton support instead of listing its lib directory ( #15412 )
2026-08-12 21:40:11 -04:00
725e6ec606
Support anima tunes with extra blocks. ( #15555 )
2026-08-12 16:22:50 -04:00
bd34f338ac
Fix float64 device in ltx diffusion decoder. ( #15516 )
2026-08-12 03:55:08 -04:00
27bca654eb
Fix KSamplerAdvanced with add_noise disabled on nested latents ( #15447 )
2026-08-11 17:58:00 -04:00
024cbc5fc1
Remove potentially problematic process_tokens method. ( #15507 )
2026-08-11 16:45:52 -04:00
bbb4b04caa
Don't depend on transformers for mistral and llama tokenizers. ( #15503 )
2026-08-11 15:54:44 -04:00
d9f9d2ba12
Fix some clip vision regression. ( #15506 )
2026-08-11 15:54:15 -04:00
2a19bbf014
Fix for broken tiled audio decode. ( #15502 )
2026-08-11 14:07:32 -04:00
57ce8e1a27
Add support for LTX 2.5 ( #15499 )
...
---------
Co-authored-by: kijai <40791699+kijai@users.noreply.github.com>
2026-08-11 13:47:39 -04:00
62b3c94bd4
Fix peak memory issue with H3. ( #15486 )
2026-08-11 02:09:29 -07:00
bf4c9a08fc
Implement comfy kitchen attention. ( #15479 )
...
Add a ModelAttentionBackend node to manually select the attention for models in the workflows. Currently supports pytorch attention or comfy kitchen attention.
Add --use-ck-attention to enable comfy kitchen attention as the default attention backend for all models (might break some).
2026-08-10 22:03:08 -07:00
4f3544d131
Make cu130 warning more visible. ( #15463 )
2026-08-10 23:25:01 -04:00
2a68ce33b4
Optimize MiniMax-H3 VAE ( #15446 )
2026-08-09 14:24:48 -04:00
cbbc9dab1f
Make a context manager for cast_bias_weight and use it. ( #14750 )
2026-08-08 22:38:34 -04:00
cd84f47efe
Make it easier to debug nested tensors. ( #15383 )
2026-08-07 22:14:04 -04:00
344b43989e
Support asym w4a8_int ( #15308 )
...
* support asym w4a8_int
* Simplify
* Fixes
2026-08-07 08:32:57 -07:00
93cb5edb98
Fix audio carry to wrappers ( #15390 )
2026-08-07 08:30:21 -07:00
2340099d93
Fix full offload on minimax audio vae. ( #15377 )
2026-08-07 00:26:16 -04:00
a464ac3358
feat: Support Wan-Animate2 (CORE-358) ( #15362 )
2026-08-06 21:08:16 -04:00
88fec4b605
Skip creating useless noise in res_multistep sampler. ( #15339 )
2026-08-06 18:17:44 -04:00
bdcb886a47
Fix sampler issues for audio with minimax, support more samplers. ( #15243 )
2026-08-06 13:36:34 -07:00
563b98eefb
Fix MiniMax H3 latent noise mask sampling ( #15322 )
2026-08-05 23:47:13 -04:00
15989f87ca
Speedup LTX and Wan ( #15138 )
2026-08-05 21:15:48 -04:00
bbda83647d
Support int8_convrot VAE ( #15334 )
2026-08-05 21:12:23 -04:00
40dbdc1bef
restore SDPA non-cudnn small attention bypass ( #15296 )
...
This is performance critical for ACE step.
2026-08-04 15:25:25 -04:00
16e3f3034f
fix(minimax): cast raw parameters to input device in H3 VAEs ( #15268 )
2026-08-03 19:06:12 -04:00
e377e26304
Don't pin too much memory on Linux systems with no swap partition. ( #15266 )
2026-08-03 16:29:47 -04:00
57500fc5bc
feat: Support MiniMax-H3 (CORE-375) ( #15224 )
2026-08-02 22:28:29 -04:00
b53e247c94
rename comfy/logging.py to comfy/internal_logging.py ( #15231 )
...
This avoids name collision (circular imports) for external custom nodes,
for which the comfy path is pushed into sys.path so Python's own logging module
is shadowed otherwise.
fixes : #15229
2026-08-02 16:20:27 -04:00
611f2a4e0f
fix pin registration priority ( #15226 )
...
This priority scheme was broken in the case where you have pin
registration exhaustion while loading a VBAR that gets a big evicition.
The weight would stay in the loaded set but inherit the MRU priority
against other workflow models WRT pin registration which leads to async
offload without pinning.
Fix by universally promiting active pin registration above workflow
pins without concern for the weights/weights-loaded split. This diverges
from the actual budgeting where the split still makes sense.
2026-08-02 08:16:06 -07:00
8084083d4b
comfy-aimdo 0.4.11 ( #15215 )
...
Changes:
Remove sequential scan hint
Prefer NVML pressure on windows
Add async malloc clamp option (unused by comfy so far)
Workaround AMD windows GPU virtual address space leak
The largest change is the NVML pressure, which works around a cuMemGetInfo
drift from actual VRAM in some circumstances.
2026-08-02 06:55:37 -04:00
f06a187f50
Handle case where swap memory query fails on windows. ( #15219 )
2026-08-02 01:12:12 -04:00
532a16f3b9
Disable gradients on diffusion models. ( #15218 )
2026-08-01 20:15:51 -07:00
49a7422892
Support latent previews for nested latents ( #15196 )
2026-08-01 19:57:15 -04:00
a1c421994c
Expand k, v when attention backend would fall back to math because gqa. ( #15190 )
2026-07-31 15:17:56 -07:00
7c806288d5
ops: apply the custom placeholder logic to Linux too ( #15181 )
...
Windows has proven this logic works for a long time and there are
corner cases where this materialization actual consumes real RAM
on linux.
Its not as bad as the original windows commit charge surge, but
its still a detectable transient leak. So simplify and unify.
2026-07-31 10:44:09 -07:00
de5625a64a
Delay dynamic pin cleanup until model destruction ( #15183 )
2026-07-31 10:29:47 -07:00
831710d257
Don't assume sampler_function has a __name__ attribute in detail logging ( #15179 )
2026-07-31 19:12:17 +03:00
f73e8cde88
Fallback to cudnn attention on linux if flash attention doesn't work. ( #15146 )
2026-07-29 18:10:28 -07:00
e651b7bef5
Fix LTXAV crash when sampling without an audio latent ( #15132 )
2026-07-28 20:24:08 -07:00
c38171ddb9
Support Pruna LTX VAE ( #15129 )
2026-07-28 18:25:55 -04:00
fbe6d3ca8f
Add configurable DETAIL logging side channel ( #15064 )
2026-07-28 17:31:45 -04:00
c01175530e
Load weights to process RAM with MRU policy using pinning infrastructure ( #15027 )
2026-07-28 17:05:57 -04:00
3d41e3ea4e
Support int8 convrot embedding lookup ( #15035 )
2026-07-28 17:02:57 -04:00
a3572c4832
Allow using float fps for LTXVEmptyLatentAudio ( #15106 )
2026-07-27 13:04:14 -07:00
806e092ed4
Fix MageFlow on cards that don't support bf16 ( #15081 )
2026-07-25 21:01:51 -04:00
f966a2b38c
Optimize ideogram model using comfy kitchen rms rope. ( #15080 )
2026-07-25 15:09:26 -04:00
45ffd5430b
feat: Support MageFlow (CORE-372) ( #15026 )
2026-07-24 23:14:01 -04:00
36aec0d086
cli_args: bump clamp to 128BGB ( #15068 )
...
Some long running chaos testing on a 512GB RAM RTX6000 pro showed that this
is a little bit too low for common template workflows switching around. The
original number was just a guess from me, so go with the scientific result
instead.
2026-07-24 19:48:52 -04:00
Alexander Brown
Alex Harper
comfyanonymous
Jukka Seppänen
Alexis Rolland
blepping
endman100
rattus
rivadart
Oliver Freyermuth
Barish Ozbay