comfyanonymous/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-03-07 22:30:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

dynamic_vram: Fix windows Aimdo crash + Fix LLM performance (#12408)

Browse Source 
* model_management: lazy-cache aimdo_tensor

These tensors cosntructed from aimdo-allocations are CPU expensive to
make on the pytorch side. Add a cache version that will be valid with
signature match to fast path past whatever torch is doing.

* dynamic_vram: Minimize fast path CPU work

Move as much as possible inside the not resident if block and cache
the formed weight and bias rather than the flat intermediates. In
extreme layer weight rates this adds up.
This commit is contained in:
rattus
2026-02-11 11:50:16 -08:00
committed by GitHub
parent 2b7cc7e3b6
commit d297a749a2
3 changed files with 20 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
comfy/model_management.py
View File

@@ -1213,8 +1213,12 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
if signature is not None:
v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, weight._v_tensor)[0]
if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
if comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
v_tensor = weight._v_tensor
else:
raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device)
v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0]
weight._v_tensor = v_tensor
weight._v_signature = signature
#Send it over
v_tensor.copy_(weight, non_blocking=non_blocking)