dequantization offload accounting (fixes Flux2 OOMs - incl TEs) (#11171)

* make setattr safe for non existent attributes

Handle the case where the attribute doesnt exist by returning a static
sentinel (distinct from None). If the sentinel is passed in as the set
value, del the attr.

* Account for dequantization and type-casts in offload costs

When measuring the cost of offload, identify weights that need a type
change or dequantization and add the size of the conversion result
to the offload cost.

This is mutually exclusive with lowvram patches which already has
a large conservative estimate and wont overlap the dequant cost so\
dont double count.

* Set the compute type on CLIP MPs

So that the loader can know the size of weights for dequant accounting.
This commit is contained in:
rattus
2025-12-09 14:21:31 +10:00
committed by GitHub
parent d50f342c90
commit e136b6dbb0
3 changed files with 22 additions and 8 deletions

View File

@@ -803,12 +803,17 @@ def safetensors_header(safetensors_path, max_size=100*1024*1024):
return None
return f.read(length_of_header)
ATTR_UNSET={}
def set_attr(obj, attr, value):
attrs = attr.split(".")
for name in attrs[:-1]:
obj = getattr(obj, name)
prev = getattr(obj, attrs[-1])
setattr(obj, attrs[-1], value)
prev = getattr(obj, attrs[-1], ATTR_UNSET)
if value is ATTR_UNSET:
delattr(obj, attrs[-1])
else:
setattr(obj, attrs[-1], value)
return prev
def set_attr_param(obj, attr, value):