comfyanonymous/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-03-09 15:20:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

model_management: lazy-cache aimdo_tensor

Browse Source 
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.
This commit is contained in:
Rattus
2026-02-11 23:23:47 +10:00
parent 89dc4a8df8
commit f7aebddcf6
2 changed files with 13 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
comfy/ops.py
View File

@@ -86,9 +86,13 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
if signature is not None:
xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
if signature is not None:
if resident:
xfer_dest = s._v_tensor
else:
xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
s._v_tensor = xfer_dest
if not resident:
cast_dest = None