caching: Remove model awareness from RAM caching

Model RAM pressure is now implemented via the DynamicVRAM system.

comfy/model_patcher.py

@@ -300,9 +300,6 @@ class ModelPatcher:
     def model_mmap_residency(self, free=False):
         return comfy.model_management.module_mmap_residency(self.model, free=free)
 
-    def get_ram_usage(self):
-        return self.model_size()
-
     def loaded_size(self):
         return self.model.model_loaded_weight_memory
 

comfy/sd.py

@@ -280,9 +280,6 @@ class CLIP:
         n.apply_hooks_to_conds = self.apply_hooks_to_conds
         return n
 
-    def get_ram_usage(self):
-        return self.patcher.get_ram_usage()
-
     def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
         return self.patcher.add_patches(patches, strength_patch, strength_model)
 
@@ -840,9 +837,6 @@ class VAE:
         self.size = comfy.model_management.module_size(self.first_stage_model)
         return self.size
 
-    def get_ram_usage(self):
-        return self.model_size()
-
     def throw_exception_if_invalid(self):
         if self.first_stage_model is None:
             raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")

comfy_execution/caching.py

@@ -494,10 +494,10 @@ class LRUCache(BasicCache):
 
 RAM_CACHE_HYSTERESIS = 1.1
 
-#This is kinda in GB but not really. It needs to be non-zero for the below heuristic
-#and as long as Multi GB models dwarf this it will approximate OOM scoring OK
+#Small baseline weight used when a cache entry has no measurable CPU tensors.
+#Keeps unknown-sized entries in eviction scoring without dominating tensor-backed entries.
 
-RAM_CACHE_DEFAULT_RAM_USAGE = 0.1
+RAM_CACHE_DEFAULT_RAM_USAGE = 0.05
 
 #Exponential bias towards evicting older workflows so garbage will be taken out
 #in constantly changing setups.
@@ -545,11 +545,7 @@ class RAMPressureCache(LRUCache):
                     if isinstance(output, list):
                         scan_list_for_ram_usage(output)
                     elif isinstance(output, torch.Tensor) and output.device.type == 'cpu':
-                        #score Tensors at a 50% discount for RAM usage as they are likely to
-                        #be high value intermediates
-                        ram_usage += (output.numel() * output.element_size()) * 0.5
-                    elif hasattr(output, "get_ram_usage"):
-                        ram_usage += output.get_ram_usage()
+                        ram_usage += output.numel() * output.element_size()
             scan_list_for_ram_usage(outputs)
 
             oom_score *= ram_usage