Merge branch 'master' into worksplit-multigpu

comfy/model_patcher.py

@@ -19,7 +19,6 @@
 from __future__ import annotations
 
 import collections
-import copy
 import inspect
 import logging
 import math
@@ -35,21 +34,10 @@ import comfy.model_management
 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
 
-
-def string_to_seed(data):
-    crc = 0xFFFFFFFF
-    for byte in data:
-        if isinstance(byte, str):
-            byte = ord(byte)
-        crc ^= byte
-        for _ in range(8):
-            if crc & 1:
-                crc = (crc >> 1) ^ 0xEDB88320
-            else:
-                crc >>= 1
-    return crc ^ 0xFFFFFFFF
+import comfy_aimdo.model_vbar
 
 def set_model_options_patch_replace(model_options, patch, name, block_name, number, transformer_index=None):
     to = model_options["transformer_options"].copy()
@@ -125,31 +113,31 @@ def move_weight_functions(m, device):
                 memory += f.move_to(device=device)
     return memory
 
+def string_to_seed(data):
+    logging.warning("WARNING: string_to_seed has moved from comfy.model_patcher to comfy.utils")
+    return comfy.utils.string_to_seed(data)
+
 class LowVramPatch:
     def __init__(self, key, patches, convert_func=None, set_func=None):
         self.key = key
         self.patches = patches
-        self.convert_func = convert_func
+        self.convert_func = convert_func # TODO: remove
         self.set_func = set_func
 
     def __call__(self, weight):
-        intermediate_dtype = weight.dtype
-        if self.convert_func is not None:
-            weight = self.convert_func(weight.to(dtype=torch.float32, copy=True), inplace=True)
+        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
 
-        if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops
-            intermediate_dtype = torch.float32
-            out = comfy.lora.calculate_weight(self.patches[self.key], weight.to(intermediate_dtype), self.key, intermediate_dtype=intermediate_dtype)
-            if self.set_func is None:
-                return comfy.float.stochastic_rounding(out, weight.dtype, seed=string_to_seed(self.key))
-            else:
-                return self.set_func(out, seed=string_to_seed(self.key), return_weight=True)
+LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2
 
-        out = comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype)
-        if self.set_func is not None:
-            return self.set_func(out, seed=string_to_seed(self.key), return_weight=True).to(dtype=intermediate_dtype)
-        else:
-            return out
+def low_vram_patch_estimate_vram(model, key):
+    weight, set_func, convert_func = get_key_weight(model, key)
+    if weight is None:
+        return 0
+    model_dtype = getattr(model, "manual_cast_dtype", torch.float32)
+    if model_dtype is None:
+        model_dtype = weight.dtype
+
+    return weight.numel() * model_dtype.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR
 
 def get_key_weight(model, key):
     set_func = None
@@ -175,6 +163,11 @@ def get_key_weight(model, key):
 
     return weight, set_func, convert_func
 
+def key_param_name_to_key(key, param):
+    if len(key) == 0:
+        return param
+    return "{}.{}".format(key, param)
+
 class AutoPatcherEjector:
     def __init__(self, model: 'ModelPatcher', skip_and_inject_on_exit_only=False):
         self.model = model
@@ -218,6 +211,27 @@ class MemoryCounter:
     def decrement(self, used: int):
         self.value -= used
 
+CustomTorchDevice = collections.namedtuple("FakeDevice", ["type", "index"])("comfy-lazy-caster", 0)
+
+class LazyCastingParam(torch.nn.Parameter):
+    def __new__(cls, model, key, tensor):
+        return super().__new__(cls, tensor)
+
+    def __init__(self, model, key, tensor):
+        self.model = model
+        self.key = key
+
+    @property
+    def device(self):
+        return CustomTorchDevice
+
+    #safetensors will .to() us to the cpu which we catch here to cast on demand. The returned tensor is
+    #then just a short lived thing in the safetensors serialization logic inside its big for loop over
+    #all weights getting garbage collected per-weight
+    def to(self, *args, **kwargs):
+        return self.model.patch_weight_to_device(self.key, device_to=self.model.load_device, return_weight=True).to("cpu")
+
+
 class ModelPatcher:
     def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
         self.size = size
@@ -234,13 +248,13 @@ class ModelPatcher:
         self.object_patches_backup = {}
         self.weight_wrapper_patches = {}
         self.model_options = {"transformer_options":{}}
-        self.model_size()
         self.load_device = load_device
         self.offload_device = offload_device
         self.weight_inplace_update = weight_inplace_update
         self.force_cast_weights = False
         self.patches_uuid = uuid.uuid4()
         self.parent = None
+        self.pinned = set()
 
         self.attachments: dict[str] = {}
         self.additional_models: dict[str, list[ModelPatcher]] = {}
@@ -275,20 +289,32 @@ class ModelPatcher:
         if not hasattr(self.model, 'current_weight_patches_uuid'):
             self.model.current_weight_patches_uuid = None
 
+        if not hasattr(self.model, 'model_offload_buffer_memory'):
+            self.model.model_offload_buffer_memory = 0
+
+    def is_dynamic(self):
+        return False
+
     def model_size(self):
         if self.size > 0:
             return self.size
         self.size = comfy.model_management.module_size(self.model)
         return self.size
 
+    def get_ram_usage(self):
+        return self.model_size()
+
     def loaded_size(self):
         return self.model.model_loaded_weight_memory
 
     def lowvram_patch_counter(self):
         return self.model.lowvram_patch_counter
 
+    def get_free_memory(self, device):
+        return comfy.model_management.get_free_memory(device)
+
     def clone(self):
-        n = self.__class__(self.model, self.load_device, self.offload_device, self.size, weight_inplace_update=self.weight_inplace_update)
+        n = self.__class__(self.model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
         for k in self.patches:
             n.patches[k] = self.patches[k][:]
@@ -296,10 +322,11 @@ class ModelPatcher:
 
         n.object_patches = self.object_patches.copy()
         n.weight_wrapper_patches = self.weight_wrapper_patches.copy()
-        n.model_options = copy.deepcopy(self.model_options)
+        n.model_options = comfy.utils.deepcopy_list_dict(self.model_options)
         n.backup = self.backup
         n.object_patches_backup = self.object_patches_backup
         n.parent = self
+        n.pinned = self.pinned
 
         n.force_cast_weights = self.force_cast_weights
 
@@ -459,13 +486,16 @@ class ModelPatcher:
     def memory_required(self, input_shape):
         return self.model.memory_required(input_shape=input_shape)
 
+    def disable_model_cfg1_optimization(self):
+        self.model_options["disable_cfg1_optimization"] = True
+
     def set_model_sampler_cfg_function(self, sampler_cfg_function, disable_cfg1_optimization=False):
         if len(inspect.signature(sampler_cfg_function).parameters) == 3:
             self.model_options["sampler_cfg_function"] = lambda args: sampler_cfg_function(args["cond"], args["uncond"], args["cond_scale"]) #Old way
         else:
             self.model_options["sampler_cfg_function"] = sampler_cfg_function
         if disable_cfg1_optimization:
-            self.model_options["disable_cfg1_optimization"] = True
+            self.disable_model_cfg1_optimization()
 
     def set_model_sampler_post_cfg_function(self, post_cfg_function, disable_cfg1_optimization=False):
         self.model_options = set_model_options_post_cfg_function(self.model_options, post_cfg_function, disable_cfg1_optimization)
@@ -530,6 +560,22 @@ class ModelPatcher:
     def set_model_post_input_patch(self, patch):
         self.set_model_patch(patch, "post_input")
 
+    def set_model_noise_refiner_patch(self, patch):
+        self.set_model_patch(patch, "noise_refiner")
+
+    def set_model_rope_options(self, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t, **kwargs):
+        rope_options = self.model_options["transformer_options"].get("rope_options", {})
+        rope_options["scale_x"] = scale_x
+        rope_options["scale_y"] = scale_y
+        rope_options["scale_t"] = scale_t
+
+        rope_options["shift_x"] = shift_x
+        rope_options["shift_y"] = shift_y
+        rope_options["shift_t"] = shift_t
+
+        self.model_options["transformer_options"]["rope_options"] = rope_options
+
+
     def add_object_patch(self, name, obj):
         self.object_patches[name] = obj
 
@@ -671,46 +717,81 @@ class ModelPatcher:
                         sd.pop(k)
             return sd
 
-    def patch_weight_to_device(self, key, device_to=None, inplace_update=False):
-        if key not in self.patches:
-            return
-
+    def patch_weight_to_device(self, key, device_to=None, inplace_update=False, return_weight=False):
         weight, set_func, convert_func = get_key_weight(self.model, key)
+        if key not in self.patches:
+            return weight
+
         inplace_update = self.weight_inplace_update or inplace_update
 
-        if key not in self.backup:
+        if key not in self.backup and not return_weight:
             self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
 
+        temp_dtype = comfy.model_management.lora_compute_dtype(device_to)
         if device_to is not None:
-            temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True)
+            temp_weight = comfy.model_management.cast_to_device(weight, device_to, temp_dtype, copy=True)
         else:
-            temp_weight = weight.to(torch.float32, copy=True)
+            temp_weight = weight.to(temp_dtype, copy=True)
         if convert_func is not None:
             temp_weight = convert_func(temp_weight, inplace=True)
 
         out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key)
         if set_func is None:
-            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
-            if inplace_update:
+            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
+            if return_weight:
+                return out_weight
+            elif inplace_update:
                 comfy.utils.copy_to_param(self.model, key, out_weight)
             else:
                 comfy.utils.set_attr_param(self.model, key, out_weight)
         else:
-            set_func(out_weight, inplace_update=inplace_update, seed=string_to_seed(key))
+            return set_func(out_weight, inplace_update=inplace_update, seed=comfy.utils.string_to_seed(key), return_weight=return_weight)
 
-    def _load_list(self):
+    def pin_weight_to_device(self, key):
+        weight, set_func, convert_func = get_key_weight(self.model, key)
+        if comfy.model_management.pin_memory(weight):
+            self.pinned.add(key)
+
+    def unpin_weight(self, key):
+        if key in self.pinned:
+            weight, set_func, convert_func = get_key_weight(self.model, key)
+            comfy.model_management.unpin_memory(weight)
+            self.pinned.remove(key)
+
+    def unpin_all_weights(self):
+        for key in list(self.pinned):
+            self.unpin_weight(key)
+
+    def _load_list(self, prio_comfy_cast_weights=False, default_device=None):
         loading = []
         for n, m in self.model.named_modules():
-            params = []
-            skip = False
-            for name, param in m.named_parameters(recurse=False):
-                params.append(name)
+            default = False
+            params = { name: param for name, param in m.named_parameters(recurse=False) }
             for name, param in m.named_parameters(recurse=True):
                 if name not in params:
-                    skip = True # skip random weights in non leaf modules
+                    default = True # default random weights in non leaf modules
                     break
-            if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
-                loading.append((comfy.model_management.module_size(m), n, m, params))
+            if default and default_device is not None:
+                for param in params.values():
+                    param.data = param.data.to(device=default_device)
+            if not default and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
+                module_mem = comfy.model_management.module_size(m)
+                module_offload_mem = module_mem
+                if hasattr(m, "comfy_cast_weights"):
+                    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))
+                prepend = (not hasattr(m, "comfy_cast_weights"),) if prio_comfy_cast_weights else ()
+                loading.append(prepend + (module_offload_mem, module_mem, n, m, params))
         return loading
 
     def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False):
@@ -719,29 +800,35 @@ class ModelPatcher:
             mem_counter = 0
             patch_counter = 0
             lowvram_counter = 0
+            lowvram_mem_counter = 0
             loading = self._load_list()
 
             load_completely = []
+            offloaded = []
+            offload_buffer = 0
             loading.sort(reverse=True)
-            for x in loading:
-                n = x[1]
-                m = x[2]
-                params = x[3]
-                module_mem = x[0]
+            for i, x in enumerate(loading):
+                module_offload_mem, module_mem, n, m, params = x
 
                 lowvram_weight = False
 
+                potential_offload = max(offload_buffer, module_offload_mem + sum([ x1[1] for x1 in loading[i+1:i+1+comfy.model_management.NUM_STREAMS]]))
+                lowvram_fits = mem_counter + module_mem + potential_offload < lowvram_model_memory
+
                 weight_key = "{}.weight".format(n)
                 bias_key = "{}.bias".format(n)
 
                 if not full_load and hasattr(m, "comfy_cast_weights"):
-                    if mem_counter + module_mem >= lowvram_model_memory:
+                    if not lowvram_fits:
+                        offload_buffer = potential_offload
                         lowvram_weight = True
                         lowvram_counter += 1
+                        lowvram_mem_counter += module_mem
                         if hasattr(m, "prev_comfy_cast_weights"): #Already lowvramed
                             continue
 
                 cast_weight = self.force_cast_weights
+                m.comfy_force_cast_weights = self.force_cast_weights
                 if lowvram_weight:
                     if hasattr(m, "comfy_cast_weights"):
                         m.weight_function = []
@@ -763,13 +850,16 @@ class ModelPatcher:
                             patch_counter += 1
 
                     cast_weight = True
+                    offloaded.append((module_mem, n, m, params))
                 else:
                     if hasattr(m, "comfy_cast_weights"):
                         wipe_lowvram_weight(m)
 
-                    if full_load or mem_counter + module_mem < lowvram_model_memory:
+                    if full_load or lowvram_fits:
                         mem_counter += module_mem
                         load_completely.append((module_mem, n, m, params))
+                    else:
+                        offload_buffer = potential_offload
 
                 if cast_weight and hasattr(m, "comfy_cast_weights"):
                     m.prev_comfy_cast_weights = m.comfy_cast_weights
@@ -793,7 +883,11 @@ class ModelPatcher:
                         continue
 
                 for param in params:
-                    self.patch_weight_to_device("{}.{}".format(n, param), device_to=device_to)
+                    key = key_param_name_to_key(n, param)
+                    self.unpin_weight(key)
+                    self.patch_weight_to_device(key, device_to=device_to)
+                if comfy.model_management.is_device_cuda(device_to):
+                    torch.cuda.synchronize()
 
                 logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
                 m.comfy_patched_weights = True
@@ -801,11 +895,18 @@ class ModelPatcher:
             for x in load_completely:
                 x[2].to(device_to)
 
+            for x in offloaded:
+                n = x[1]
+                params = x[3]
+                for param in params:
+                    self.pin_weight_to_device(key_param_name_to_key(n, param))
+
+            usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
-                logging.info("loaded partially {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), patch_counter))
+                logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
-                logging.info("loaded completely {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
+                logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, mem_counter / (1024 * 1024), full_load))
                 self.model.model_lowvram = False
                 if full_load:
                     self.model.to(device_to)
@@ -814,6 +915,7 @@ class ModelPatcher:
             self.model.lowvram_patch_counter += patch_counter
             self.model.device = device_to
             self.model.model_loaded_weight_memory = mem_counter
+            self.model.model_offload_buffer_memory = offload_buffer
             self.model.current_weight_patches_uuid = self.patches_uuid
 
             for callback in self.get_all_callbacks(CallbacksMP.ON_LOAD):
@@ -842,6 +944,7 @@ class ModelPatcher:
         self.eject_model()
         if unpatch_weights:
             self.unpatch_hooks()
+            self.unpin_all_weights()
             if self.model.model_lowvram:
                 for m in self.model.modules():
                     move_weight_functions(m, device_to)
@@ -866,6 +969,7 @@ class ModelPatcher:
                 self.model.to(device_to)
                 self.model.device = device_to
             self.model.model_loaded_weight_memory = 0
+            self.model.model_offload_buffer_memory = 0
 
             for m in self.model.modules():
                 if hasattr(m, "comfy_patched_weights"):
@@ -877,26 +981,31 @@ class ModelPatcher:
 
         self.object_patches_backup.clear()
 
-    def partially_unload(self, device_to, memory_to_free=0):
+    def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False):
         with self.use_ejected():
             hooks_unpatched = False
             memory_freed = 0
             patch_counter = 0
             unload_list = self._load_list()
             unload_list.sort()
+
+            offload_buffer = self.model.model_offload_buffer_memory
+            if len(unload_list) > 0:
+                NS = comfy.model_management.NUM_STREAMS
+                offload_weight_factor = [ min(offload_buffer / (NS + 1), unload_list[0][1]) ] * NS
+
             for unload in unload_list:
-                if memory_to_free < memory_freed:
+                if memory_to_free + offload_buffer - self.model.model_offload_buffer_memory < memory_freed:
                     break
-                module_mem = unload[0]
-                n = unload[1]
-                m = unload[2]
-                params = unload[3]
+                module_offload_mem, module_mem, n, m, params = unload
+
+                potential_offload = module_offload_mem + sum(offload_weight_factor)
 
                 lowvram_possible = hasattr(m, "comfy_cast_weights")
                 if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
                     move_weight = True
                     for param in params:
-                        key = "{}.{}".format(n, param)
+                        key = key_param_name_to_key(n, param)
                         bk = self.backup.get(key, None)
                         if bk is not None:
                             if not lowvram_possible:
@@ -921,25 +1030,40 @@ class ModelPatcher:
                         module_mem += move_weight_functions(m, device_to)
                         if lowvram_possible:
                             if weight_key in self.patches:
-                                _, set_func, convert_func = get_key_weight(self.model, weight_key)
-                                m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func))
-                                patch_counter += 1
+                                if force_patch_weights:
+                                    self.patch_weight_to_device(weight_key)
+                                else:
+                                    _, set_func, convert_func = get_key_weight(self.model, weight_key)
+                                    m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func))
+                                    patch_counter += 1
                             if bias_key in self.patches:
-                                _, set_func, convert_func = get_key_weight(self.model, bias_key)
-                                m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func))
-                                patch_counter += 1
+                                if force_patch_weights:
+                                    self.patch_weight_to_device(bias_key)
+                                else:
+                                    _, set_func, convert_func = get_key_weight(self.model, bias_key)
+                                    m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func))
+                                    patch_counter += 1
                             cast_weight = True
 
-                        if cast_weight:
+                        if cast_weight and hasattr(m, "comfy_cast_weights"):
                             m.prev_comfy_cast_weights = m.comfy_cast_weights
                             m.comfy_cast_weights = True
                         m.comfy_patched_weights = False
                         memory_freed += module_mem
+                        offload_buffer = max(offload_buffer, potential_offload)
+                        offload_weight_factor.append(module_mem)
+                        offload_weight_factor.pop(0)
                         logging.debug("freed {}".format(n))
 
+                        for param in params:
+                            self.pin_weight_to_device(key_param_name_to_key(n, param))
+
+
             self.model.model_lowvram = True
             self.model.lowvram_patch_counter += patch_counter
             self.model.model_loaded_weight_memory -= memory_freed
+            self.model.model_offload_buffer_memory = offload_buffer
+            logging.info("Unloaded partially: {:.2f} MB freed, {:.2f} MB remains loaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(memory_freed / (1024 * 1024), self.model.model_loaded_weight_memory / (1024 * 1024), offload_buffer / (1024 * 1024), self.model.lowvram_patch_counter))
             return memory_freed
 
     def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
@@ -952,6 +1076,9 @@ class ModelPatcher:
                 extra_memory += (used - self.model.model_loaded_weight_memory)
 
             self.patch_model(load_weights=False)
+            if extra_memory < 0 and not unpatch_weights:
+                self.partially_unload(self.offload_device, -extra_memory, force_patch_weights=force_patch_weights)
+                return 0
             full_load = False
             if self.model.model_lowvram == False and self.model.model_loaded_weight_memory > 0:
                 self.apply_hooks(self.forced_hooks, force_apply=True)
@@ -967,6 +1094,9 @@ class ModelPatcher:
 
             return self.model.model_loaded_weight_memory - current_used
 
+    def partially_unload_ram(self, ram_to_unload):
+        pass
+
     def detach(self, unpatch_all=True):
         self.eject_model()
         self.model_patches_to(self.offload_device)
@@ -1332,10 +1462,10 @@ class ModelPatcher:
                                                  key, original_weights=original_weights)
         del original_weights[key]
         if set_func is None:
-            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
+            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
             comfy.utils.copy_to_param(self.model, key, out_weight)
         else:
-            set_func(out_weight, inplace_update=True, seed=string_to_seed(key))
+            set_func(out_weight, inplace_update=True, seed=comfy.utils.string_to_seed(key))
         if self.hook_mode == comfy.hooks.EnumHookMode.MaxSpeed:
             # TODO: disable caching if not enough system RAM to do so
             target_device = self.offload_device
@@ -1370,6 +1500,275 @@ class ModelPatcher:
         self.unpatch_hooks()
         self.clear_cached_hook_weights()
 
+    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+        unet_state_dict = self.model.diffusion_model.state_dict()
+        for k, v in unet_state_dict.items():
+            op_keys = k.rsplit('.', 1)
+            if (len(op_keys) < 2) or op_keys[1] not in ["weight", "bias"]:
+                continue
+            try:
+                op = comfy.utils.get_attr(self.model.diffusion_model, op_keys[0])
+            except:
+                continue
+            if not op or not hasattr(op, "comfy_cast_weights") or \
+                (hasattr(op, "comfy_patched_weights") and op.comfy_patched_weights == True):
+                continue
+            key = "diffusion_model." + k
+            unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key))
+        return self.model.state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
+
     def __del__(self):
+        self.unpin_all_weights()
         self.detach(unpatch_all=False)
 
+class ModelPatcherDynamic(ModelPatcher):
+
+    def __new__(cls, model=None, load_device=None, offload_device=None, size=0, weight_inplace_update=False):
+        if load_device is not None and comfy.model_management.is_device_cpu(load_device):
+            #reroute to default MP for CPUs
+            return ModelPatcher(model, load_device, offload_device, size, weight_inplace_update)
+        return super().__new__(cls)
+
+    def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
+        super().__init__(model, load_device, offload_device, size, weight_inplace_update)
+        #this is now way more dynamic and we dont support the same base model for both Dynamic
+        #and non-dynamic patchers.
+        if hasattr(self.model, "model_loaded_weight_memory"):
+            del self.model.model_loaded_weight_memory
+        if not hasattr(self.model, "dynamic_vbars"):
+            self.model.dynamic_vbars = {}
+        assert load_device is not None
+
+    def is_dynamic(self):
+        return True
+
+    def _vbar_get(self, create=False):
+        if self.load_device == torch.device("cpu"):
+            return None
+        vbar = self.model.dynamic_vbars.get(self.load_device, None)
+        if create and vbar is None:
+            # x10. We dont know what model defined type casts we have in the vbar, but virtual address
+            # space is pretty free. This will cover someone casting an entire model from FP4 to FP32
+            # with some left over.
+            vbar = comfy_aimdo.model_vbar.ModelVBAR(self.model_size() * 10, self.load_device.index)
+            self.model.dynamic_vbars[self.load_device] = vbar
+        return vbar
+
+    def loaded_size(self):
+        vbar = self._vbar_get()
+        if vbar is None:
+            return 0
+        return vbar.loaded_size()
+
+    def get_free_memory(self, device):
+        #NOTE: on high condition / batch counts, estimate should have already vacated
+        #all non-dynamic models so this is safe even if its not 100% true that this
+        #would all be avaiable for inference use.
+        return comfy.model_management.get_total_memory(device) - self.model_size()
+
+    #Pinning is deferred to ops time. Assert against this API to avoid pin leaks.
+
+    def pin_weight_to_device(self, key):
+        raise RuntimeError("pin_weight_to_device invalid for dymamic weight loading")
+
+    def unpin_weight(self, key):
+        raise RuntimeError("unpin_weight invalid for dymamic weight loading")
+
+    def unpin_all_weights(self):
+        self.partially_unload_ram(1e32)
+
+    def memory_required(self, input_shape):
+        #Pad this significantly. We are trying to get away from precise estimates. This
+        #estimate is only used when using the ModelPatcherDynamic after ModelPatcher. If you
+        #use all ModelPatcherDynamic this is ignored and its all done dynamically.
+        return super().memory_required(input_shape=input_shape) * 1.3 + (1024 ** 3)
+
+
+    def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False, dirty=False):
+
+        #Force patching doesn't make sense in Dynamic loading, as you dont know what does and
+        #doesn't need to be forced at this stage. The only thing you could do would be patch
+        #it all on CPU which consumes huge RAM.
+        assert not force_patch_weights
+
+        #Full load doesn't make sense as we dont actually have any loader capability here and
+        #now.
+        assert not full_load
+
+        assert device_to == self.load_device
+
+        num_patches = 0
+        allocated_size = 0
+
+        with self.use_ejected():
+            self.unpatch_hooks()
+
+            vbar = self._vbar_get(create=True)
+            if vbar is not None:
+                vbar.prioritize()
+
+            #We force reserve VRAM for the non comfy-weight so we dont have to deal
+            #with pin and unpin syncrhonization which can be expensive for small weights
+            #with a high layer rate (e.g. autoregressive LLMs).
+            #prioritize the non-comfy weights (note the order reverse).
+            loading = self._load_list(prio_comfy_cast_weights=True, default_device=device_to)
+            loading.sort(reverse=True)
+
+            for x in loading:
+                _, _, _, n, m, params = x
+
+                def set_dirty(item, dirty):
+                    if dirty or not hasattr(item, "_v_signature"):
+                        item._v_signature = None
+
+                def setup_param(self, m, n, param_key):
+                    nonlocal num_patches
+                    key = key_param_name_to_key(n, param_key)
+
+                    weight_function = []
+
+                    weight, _, _ = get_key_weight(self.model, key)
+                    if weight is None:
+                        return (False, 0)
+                    if key in self.patches:
+                        if comfy.lora.calculate_shape(self.patches[key], weight, key) != weight.shape:
+                            return (True, 0)
+                        setattr(m, param_key + "_lowvram_function", LowVramPatch(key, self.patches))
+                        num_patches += 1
+                    else:
+                        setattr(m, param_key + "_lowvram_function", None)
+
+                    if key in self.weight_wrapper_patches:
+                        weight_function.extend(self.weight_wrapper_patches[key])
+                    setattr(m, param_key + "_function", weight_function)
+                    geometry = weight
+                    if not isinstance(weight, QuantizedTensor):
+                        model_dtype = getattr(m, param_key + "_comfy_model_dtype", None) or weight.dtype
+                        weight._model_dtype = model_dtype
+                        geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype)
+                    return (False, comfy.memory_management.vram_aligned_size(geometry))
+
+                def force_load_param(self, param_key, device_to):
+                    key = key_param_name_to_key(n, param_key)
+                    if key in self.backup:
+                        comfy.utils.set_attr_param(self.model, key, self.backup[key].weight)
+                    self.patch_weight_to_device(key, device_to=device_to)
+
+                if hasattr(m, "comfy_cast_weights"):
+                    m.comfy_cast_weights = True
+                    m.pin_failed = False
+                    m.seed_key = n
+                    set_dirty(m, dirty)
+
+                    force_load, v_weight_size = setup_param(self, m, n, "weight")
+                    force_load_bias, v_weight_bias = setup_param(self, m, n, "bias")
+                    force_load = force_load or force_load_bias
+                    v_weight_size += v_weight_bias
+
+                    if force_load:
+                        logging.info(f"Module {n} has resizing Lora - force loading")
+                        force_load_param(self, "weight", device_to)
+                        force_load_param(self, "bias", device_to)
+                    else:
+                        if vbar is not None and not hasattr(m, "_v"):
+                            m._v = vbar.alloc(v_weight_size)
+                        allocated_size += v_weight_size
+
+                else:
+                    for param in params:
+                        key = key_param_name_to_key(n, param)
+                        weight, _, _ = get_key_weight(self.model, key)
+                        weight.seed_key = key
+                        set_dirty(weight, dirty)
+                        geometry = weight
+                        model_dtype = getattr(m, param + "_comfy_model_dtype", None) or weight.dtype
+                        geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype)
+                        weight_size = geometry.numel() * geometry.element_size()
+                        if vbar is not None and not hasattr(weight, "_v"):
+                            weight._v = vbar.alloc(weight_size)
+                            weight._model_dtype = model_dtype
+                        allocated_size += weight_size
+                    vbar.set_watermark_limit(allocated_size)
+
+                move_weight_functions(m, device_to)
+
+            logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.")
+
+            self.model.device = device_to
+            self.model.current_weight_patches_uuid = self.patches_uuid
+
+            for callback in self.get_all_callbacks(CallbacksMP.ON_LOAD):
+                #These are all super dangerous. Who knows what the custom nodes actually do here...
+                callback(self, device_to, lowvram_model_memory, force_patch_weights, full_load)
+
+            self.apply_hooks(self.forced_hooks, force_apply=True)
+
+    def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False):
+        assert not force_patch_weights #See above
+        assert self.load_device != torch.device("cpu")
+
+        vbar = self._vbar_get()
+        return 0 if vbar is None else vbar.free_memory(memory_to_free)
+
+    def partially_unload_ram(self, ram_to_unload):
+        loading = self._load_list(prio_comfy_cast_weights=True, default_device=self.offload_device)
+        for x in loading:
+            _, _, _, _, m, _ = x
+            ram_to_unload -= comfy.pinned_memory.unpin_memory(m)
+            if ram_to_unload <= 0:
+                return
+
+    def patch_model(self, device_to=None, lowvram_model_memory=0, load_weights=True, force_patch_weights=False):
+        #This isn't used by the core at all and can only be to load a model out of
+        #the control of proper model_managment. If you are a custom node author reading
+        #this, the correct pattern is to call load_models_gpu() to get a proper
+        #managed load of your model.
+        assert not load_weights
+        return super().patch_model(load_weights=load_weights, force_patch_weights=force_patch_weights)
+
+    def unpatch_model(self, device_to=None, unpatch_weights=True):
+        super().unpatch_model(device_to=None, unpatch_weights=False)
+
+        if unpatch_weights:
+            self.partially_unload_ram(1e32)
+            self.partially_unload(None, 1e32)
+            for m in self.model.modules():
+                move_weight_functions(m, device_to)
+
+            keys = list(self.backup.keys())
+            for k in keys:
+                bk = self.backup[k]
+                comfy.utils.set_attr_param(self.model, k, bk.weight)
+
+    def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
+        assert not force_patch_weights #See above
+        with self.use_ejected(skip_and_inject_on_exit_only=True):
+            dirty = self.model.current_weight_patches_uuid is not None and (self.model.current_weight_patches_uuid != self.patches_uuid)
+
+            self.unpatch_model(self.offload_device, unpatch_weights=False)
+            self.patch_model(load_weights=False)
+
+            try:
+                self.load(device_to, dirty=dirty)
+            except Exception as e:
+                self.detach()
+                raise e
+            #ModelPatcher::partially_load returns a number on what got loaded but
+            #nothing in core uses this and we have no data in the Dynamic world. Hit
+            #the custom node devs with a None rather than a 0 that would mislead any
+            #logic they might have.
+            return None
+
+    def patch_cached_hook_weights(self, cached_weights: dict, key: str, memory_counter: MemoryCounter):
+        assert False #Should be unreachable - we dont ever cache in the new implementation
+
+    def patch_hook_weight_to_device(self, hooks: comfy.hooks.HookGroup, combined_patches: dict, key: str, original_weights: dict, memory_counter: MemoryCounter):
+        if key not in combined_patches:
+            return
+
+        raise RuntimeError("Hooks not implemented in ModelPatcherDynamic. Please remove --fast arguments form ComfyUI startup")
+
+    def unpatch_hooks(self, whitelist_keys_set: set[str]=None) -> None:
+        pass
+
+CoreModelPatcher = ModelPatcher