Commit Graph

1762 Commits

Author SHA1 Message Date
strint
aab0e244f7 fix MMAP_MEM_THRESHOLD_GB default 2025-10-23 14:44:51 +08:00
strint
f3c673d086 Merge branch 'master' of https://github.com/siliconflow/ComfyUI into refine_offload 2025-10-22 21:15:28 +08:00
comfyanonymous
9cdc64998f
Only disable cudnn on newer AMD GPUs. (#10437) 2025-10-21 19:15:23 -04:00
strint
98ba311511 add env 2025-10-21 19:06:34 +08:00
strint
80383932ec lazy rm file 2025-10-21 18:00:31 +08:00
strint
08e094ed81 use native mmap 2025-10-21 17:00:56 +08:00
strint
fff56de63c fix format 2025-10-21 11:59:59 +08:00
strint
2d010f545c refine code 2025-10-21 11:54:56 +08:00
strint
2f0d56656e refine code 2025-10-21 11:38:17 +08:00
comfyanonymous
2c2aa409b0
Log message for cudnn disable on AMD. (#10418) 2025-10-20 15:43:24 -04:00
strint
05c2518c6d refact mmap 2025-10-21 02:59:51 +08:00
strint
8aeebbf7ef fix to 2025-10-21 02:27:40 +08:00
strint
49561788cf fix log 2025-10-21 02:03:38 +08:00
strint
e9e1d2f0e8 add mmap tensor 2025-10-21 00:40:14 +08:00
strint
4ac827d564 unload partial 2025-10-20 18:27:38 +08:00
strint
21ebcada1d debug free mem 2025-10-20 16:22:50 +08:00
comfyanonymous
b4f30bd408
Pytorch is stupid. (#10398) 2025-10-19 01:25:35 -04:00
comfyanonymous
dad076aee6
Speed up chroma radiance. (#10395) 2025-10-18 23:19:52 -04:00
comfyanonymous
0cf33953a7
Fix batch size above 1 giving bad output in chroma radiance. (#10394) 2025-10-18 23:15:34 -04:00
comfyanonymous
5b80addafd
Turn off cuda malloc by default when --fast autotune is turned on. (#10393) 2025-10-18 22:35:46 -04:00
comfyanonymous
9da397ea2f
Disable torch compiler for cast_bias_weight function (#10384)
* Disable torch compiler for cast_bias_weight function

* Fix torch compile.
2025-10-17 20:03:28 -04:00
strint
49597bfa3e load remains mmap 2025-10-17 21:43:49 +08:00
strint
6583cc0142 debug load mem 2025-10-17 18:28:25 +08:00
strint
5c3c6c02b2 add debug log of cpu load 2025-10-17 16:33:14 +08:00
comfyanonymous
b1293d50ef
workaround also works on cudnn 91200 (#10375) 2025-10-16 19:59:56 -04:00
comfyanonymous
19b466160c
Workaround for nvidia issue where VAE uses 3x more memory on torch 2.9 (#10373) 2025-10-16 18:16:03 -04:00
strint
e5ff6a1b53 refine log 2025-10-16 22:47:03 +08:00
strint
9352987e9b add log 2025-10-16 22:25:17 +08:00
strint
c1eac555c0 add debug log 2025-10-16 21:42:48 +08:00
strint
2b222962c3 add debug log 2025-10-16 21:42:02 +08:00
strint
fa19dd4620 debug offload 2025-10-16 17:00:47 +08:00
strint
6e33ee391a debug error 2025-10-16 16:45:08 +08:00
Faych
afa8a24fe1
refactor: Replace manual patches merging with merge_nested_dicts (#10360) 2025-10-15 17:16:09 -07:00
Jedrzej Kosinski
493b81e48f
Fix order of inputs nested merge_nested_dicts (#10362) 2025-10-15 16:47:26 -07:00
comfyanonymous
1c10b33f9b
gfx942 doesn't support fp8 operations. (#10348) 2025-10-15 00:21:11 -04:00
comfyanonymous
3374e900d0
Faster workflow cancelling. (#10301) 2025-10-13 23:43:53 -04:00
comfyanonymous
dfff7e5332
Better memory estimation for the SD/Flux VAE on AMD. (#10334) 2025-10-13 22:37:19 -04:00
comfyanonymous
e4ea393666
Fix loading old stable diffusion ckpt files on newer numpy. (#10333) 2025-10-13 22:18:58 -04:00
comfyanonymous
c8674bc6e9
Enable RDNA4 pytorch attention on ROCm 7.0 and up. (#10332) 2025-10-13 21:19:03 -04:00
rattus128
95ca2e56c8
WAN2.2: Fix cache VRAM leak on error (#10308)
Same change pattern as 7e8dd275c2
applied to WAN2.2

If this suffers an exception (such as a VRAM oom) it will leave the
encode() and decode() methods which skips the cleanup of the WAN
feature cache. The comfy node cache then ultimately keeps a reference
this object which is in turn reffing large tensors from the failed
execution.

The feature cache is currently setup at a class variable on the
encoder/decoder however, the encode and decode functions always clear
it on both entry and exit of normal execution.

Its likely the design intent is this is usable as a streaming encoder
where the input comes in batches, however the functions as they are
today don't support that.

So simplify by bringing the cache back to local variable, so that if
it does VRAM OOM the cache itself is properly garbage when the
encode()/decode() functions dissappear from the stack.
2025-10-13 15:23:11 -04:00
comfyanonymous
e693e4db6a
Always set diffusion model to eval() mode. (#10331) 2025-10-13 14:57:27 -04:00
comfyanonymous
a125cd84b0
Improve AMD performance. (#10302)
I honestly have no idea why this improves things but it does.
2025-10-12 00:28:01 -04:00
comfyanonymous
84e9ce32c6
Implement the mmaudio VAE. (#10300) 2025-10-11 22:57:23 -04:00
comfyanonymous
f1dd6e50f8
Fix bug with applying loras on fp8 scaled without fp8 ops. (#10279) 2025-10-09 19:02:40 -04:00
comfyanonymous
139addd53c
More surgical fix for #10267 (#10276) 2025-10-09 16:37:35 -04:00
comfyanonymous
6e59934089
Refactor model sampling sigmas code. (#10250) 2025-10-08 17:49:02 -04:00
comfyanonymous
8aea746212
Implement gemma 3 as a text encoder. (#10241)
Not useful yet.
2025-10-06 22:08:08 -04:00
comfyanonymous
195e0b0639
Remove useless code. (#10223) 2025-10-05 15:41:19 -04:00
Finn-Hecker
93d859cfaa
Fix type annotation syntax in MotionEncoder_tc __init__ (#10186)
## Summary
Fixed incorrect type hint syntax in `MotionEncoder_tc.__init__()` parameter list.

## Changes
- Line 647: Changed `num_heads=int` to `num_heads: int` 
- This corrects the parameter annotation from a default value assignment to proper type hint syntax

## Details
The parameter was using assignment syntax (`=`) instead of type annotation syntax (`:`), which would incorrectly set the default value to the `int` class itself rather than annotating the expected type.
2025-10-03 14:32:19 -07:00
rattus128
4965c0e2ac
WAN: Fix cache VRAM leak on error (#10141)
If this suffers an exception (such as a VRAM oom) it will leave the
encode() and decode() methods which skips the cleanup of the WAN
feature cache. The comfy node cache then ultimately keeps a reference
this object which is in turn reffing large tensors from the failed
execution.

The feature cache is currently setup at a class variable on the
encoder/decoder however, the encode and decode functions always clear
it on both entry and exit of normal execution.

Its likely the design intent is this is usable as a streaming encoder
where the input comes in batches, however the functions as they are
today don't support that.

So simplify by bringing the cache back to local variable, so that if
it does VRAM OOM the cache itself is properly garbage when the
encode()/decode() functions dissappear from the stack.
2025-10-01 18:42:16 -04:00