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.

Author: rattus128

comfy/model_patcher.py

@@ -35,6 +35,7 @@
 import comfy.patcher_extension
 import comfy.utils
 from comfy.comfy_types import UnetWrapperFunction
+from comfy.quant_ops import QuantizedTensor
 from comfy.patcher_extension import CallbacksMP, PatcherInjection, WrappersMP
 
 
@@ -665,12 +666,18 @@ def _load_list(self):
                 module_mem = comfy.model_management.module_size(m)
                 module_offload_mem = module_mem
                 if hasattr(m, "comfy_cast_weights"):
-                    weight_key = "{}.weight".format(n)
-                    bias_key = "{}.bias".format(n)
-                    if weight_key in self.patches:
-                        module_offload_mem += low_vram_patch_estimate_vram(self.model, weight_key)
-                    if bias_key in self.patches:
-                        module_offload_mem += low_vram_patch_estimate_vram(self.model, bias_key)
+                    def check_module_offload_mem(key):
+                        if key in self.patches:
+                            return low_vram_patch_estimate_vram(self.model, key)
+                        model_dtype = getattr(self.model, "manual_cast_dtype", None)
+                        weight, _, _ = get_key_weight(self.model, key)
+                        if model_dtype is None or weight is None:
+                            return 0
+                        if (weight.dtype != model_dtype or isinstance(weight, QuantizedTensor)):
+                            return weight.numel() * model_dtype.itemsize
+                        return 0
+                    module_offload_mem += check_module_offload_mem("{}.weight".format(n))
+                    module_offload_mem += check_module_offload_mem("{}.bias".format(n))
                 loading.append((module_offload_mem, module_mem, n, m, params))
         return loading
 

comfy/sd.py

@@ -127,6 +127,8 @@ def __init__(self, target=None, embedding_directory=None, no_init=False, tokeniz
 
         self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
         self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
+        #Match torch.float32 hardcode upcast in TE implemention
+        self.patcher.set_model_compute_dtype(torch.float32)
         self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram
         self.patcher.is_clip = True
         self.apply_hooks_to_conds = None

comfy/utils.py

@@ -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):