Merge ccd5c01e5a into 22ad513c72

* add dependency aware cache that removed a cached node as soon as all of its decendents have executed. This allows users with lower RAM to run workflows they would otherwise not be able to run. The downside is that every workflow will fully run each time even if no nodes have changed.

* remove test code

* tidy code

comfy/cli_args.py

@@ -49,7 +49,7 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Co
 parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.")
 parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
 parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.")
-parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.")
+parser.add_argument("--cuda-device", type=str, default=None, metavar="DEVICE_ID", help="Set the ids of cuda devices this instance will use.")
 cm_group = parser.add_mutually_exclusive_group()
 cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")
 cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.")
@@ -101,6 +101,7 @@ parser.add_argument("--preview-size", type=int, default=512, help="Sets the maxi
 cache_group = parser.add_mutually_exclusive_group()
 cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
 cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
+cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")
 
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")

comfy/controlnet.py

@@ -15,13 +15,14 @@
     You should have received a copy of the GNU General Public License
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 """
-
+from __future__ import annotations
 
 import torch
 from enum import Enum
 import math
 import os
 import logging
+import copy
 import comfy.utils
 import comfy.model_management
 import comfy.model_detection
@@ -36,7 +37,7 @@ import comfy.cldm.mmdit
 import comfy.ldm.hydit.controlnet
 import comfy.ldm.flux.controlnet
 import comfy.cldm.dit_embedder
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Union
 if TYPE_CHECKING:
     from comfy.hooks import HookGroup
 
@@ -63,6 +64,18 @@ class StrengthType(Enum):
     CONSTANT = 1
     LINEAR_UP = 2
 
+class ControlIsolation:
+    '''Temporarily set a ControlBase object's previous_controlnet to None to prevent cascading calls.'''
+    def __init__(self, control: ControlBase):
+        self.control = control
+        self.orig_previous_controlnet = control.previous_controlnet
+
+    def __enter__(self):
+        self.control.previous_controlnet = None
+
+    def __exit__(self, *args):
+        self.control.previous_controlnet = self.orig_previous_controlnet
+
 class ControlBase:
     def __init__(self):
         self.cond_hint_original = None
@@ -76,7 +89,7 @@ class ControlBase:
         self.compression_ratio = 8
         self.upscale_algorithm = 'nearest-exact'
         self.extra_args = {}
-        self.previous_controlnet = None
+        self.previous_controlnet: Union[ControlBase, None] = None
         self.extra_conds = []
         self.strength_type = StrengthType.CONSTANT
         self.concat_mask = False
@@ -84,6 +97,7 @@ class ControlBase:
         self.extra_concat = None
         self.extra_hooks: HookGroup = None
         self.preprocess_image = lambda a: a
+        self.multigpu_clones: dict[torch.device, ControlBase] = {}
 
     def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None, extra_concat=[]):
         self.cond_hint_original = cond_hint
@@ -110,17 +124,38 @@ class ControlBase:
     def cleanup(self):
         if self.previous_controlnet is not None:
             self.previous_controlnet.cleanup()
-
+        for device_cnet in self.multigpu_clones.values():
+            with ControlIsolation(device_cnet):
+                device_cnet.cleanup()
         self.cond_hint = None
         self.extra_concat = None
         self.timestep_range = None
 
     def get_models(self):
         out = []
+        for device_cnet in self.multigpu_clones.values():
+            out += device_cnet.get_models_only_self()
         if self.previous_controlnet is not None:
             out += self.previous_controlnet.get_models()
         return out
 
+    def get_models_only_self(self):
+        'Calls get_models, but temporarily sets previous_controlnet to None.'
+        with ControlIsolation(self):
+            return self.get_models()
+
+    def get_instance_for_device(self, device):
+        'Returns instance of this Control object intended for selected device.'
+        return self.multigpu_clones.get(device, self)
+
+    def deepclone_multigpu(self, load_device, autoregister=False):
+        '''
+        Create deep clone of Control object where model(s) is set to other devices.
+
+        When autoregister is set to True, the deep clone is also added to multigpu_clones dict.
+        '''
+        raise NotImplementedError("Classes inheriting from ControlBase should define their own deepclone_multigpu funtion.")
+
     def get_extra_hooks(self):
         out = []
         if self.extra_hooks is not None:
@@ -129,7 +164,7 @@ class ControlBase:
             out += self.previous_controlnet.get_extra_hooks()
         return out
 
-    def copy_to(self, c):
+    def copy_to(self, c: ControlBase):
         c.cond_hint_original = self.cond_hint_original
         c.strength = self.strength
         c.timestep_percent_range = self.timestep_percent_range
@@ -280,6 +315,14 @@ class ControlNet(ControlBase):
         self.copy_to(c)
         return c
 
+    def deepclone_multigpu(self, load_device, autoregister=False):
+        c = self.copy()
+        c.control_model = copy.deepcopy(c.control_model)
+        c.control_model_wrapped = comfy.model_patcher.ModelPatcher(c.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
+        if autoregister:
+            self.multigpu_clones[load_device] = c
+        return c
+
     def get_models(self):
         out = super().get_models()
         out.append(self.control_model_wrapped)
@@ -804,6 +847,14 @@ class T2IAdapter(ControlBase):
         self.copy_to(c)
         return c
 
+    def deepclone_multigpu(self, load_device, autoregister=False):
+        c = self.copy()
+        c.t2i_model = copy.deepcopy(c.t2i_model)
+        c.device = load_device
+        if autoregister:
+            self.multigpu_clones[load_device] = c
+        return c
+
 def load_t2i_adapter(t2i_data, model_options={}): #TODO: model_options
     compression_ratio = 8
     upscale_algorithm = 'nearest-exact'

comfy/model_management.py

@@ -15,6 +15,7 @@
     You should have received a copy of the GNU General Public License
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 """
+from __future__ import annotations
 
 import psutil
 import logging
@@ -26,6 +27,10 @@ import platform
 import weakref
 import gc
 
+from typing import TYPE_CHECKING
+if TYPE_CHECKING:
+    from comfy.model_patcher import ModelPatcher
+
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
     NO_VRAM = 1     #Very low vram: enable all the options to save vram
@@ -171,6 +176,25 @@ def get_torch_device():
         else:
             return torch.device(torch.cuda.current_device())
 
+def get_all_torch_devices(exclude_current=False):
+    global cpu_state
+    devices = []
+    if cpu_state == CPUState.GPU:
+        if is_nvidia():
+            for i in range(torch.cuda.device_count()):
+                devices.append(torch.device(i))
+        elif is_intel_xpu():
+            for i in range(torch.xpu.device_count()):
+                devices.append(torch.device(i))
+        elif is_ascend_npu():
+            for i in range(torch.npu.device_count()):
+                devices.append(torch.device(i))
+    else:
+        devices.append(get_torch_device())
+    if exclude_current:
+        devices.remove(get_torch_device())
+    return devices
+
 def get_total_memory(dev=None, torch_total_too=False):
     global directml_enabled
     if dev is None:
@@ -382,9 +406,13 @@ try:
     logging.info("Device: {}".format(get_torch_device_name(get_torch_device())))
 except:
     logging.warning("Could not pick default device.")
+try:
+    for device in get_all_torch_devices(exclude_current=True):
+        logging.info("Device: {}".format(get_torch_device_name(device)))
+except:
+    pass
 
-
-current_loaded_models = []
+current_loaded_models: list[LoadedModel] = []
 
 def module_size(module):
     module_mem = 0
@@ -395,7 +423,7 @@ def module_size(module):
     return module_mem
 
 class LoadedModel:
-    def __init__(self, model):
+    def __init__(self, model: ModelPatcher):
         self._set_model(model)
         self.device = model.load_device
         self.real_model = None
@@ -403,7 +431,7 @@ class LoadedModel:
         self.model_finalizer = None
         self._patcher_finalizer = None
 
-    def _set_model(self, model):
+    def _set_model(self, model: ModelPatcher):
         self._model = weakref.ref(model)
         if model.parent is not None:
             self._parent_model = weakref.ref(model.parent)
@@ -1246,6 +1274,31 @@ def soft_empty_cache(force=False):
 def unload_all_models():
     free_memory(1e30, get_torch_device())
 
+def unload_model_and_clones(model: ModelPatcher, unload_additional_models=True, all_devices=False):
+    'Unload only model and its clones - primarily for multigpu cloning purposes.'
+    initial_keep_loaded: list[LoadedModel] = current_loaded_models.copy()
+    additional_models = []
+    if unload_additional_models:
+        additional_models = model.get_nested_additional_models()
+    keep_loaded = []
+    for loaded_model in initial_keep_loaded:
+        if loaded_model.model is not None:
+            if model.clone_base_uuid == loaded_model.model.clone_base_uuid:
+                continue
+            # check additional models if they are a match
+            skip = False
+            for add_model in additional_models:
+                if add_model.clone_base_uuid == loaded_model.model.clone_base_uuid:
+                    skip = True
+                    break
+            if skip:
+                continue
+        keep_loaded.append(loaded_model)
+    if not all_devices:
+        free_memory(1e30, get_torch_device(), keep_loaded)
+    else:
+        for device in get_all_torch_devices():
+            free_memory(1e30, device, keep_loaded)
 
 #TODO: might be cleaner to put this somewhere else
 import threading

comfy/model_patcher.py

@@ -84,12 +84,15 @@ def set_model_options_pre_cfg_function(model_options, pre_cfg_function, disable_
 def create_model_options_clone(orig_model_options: dict):
     return comfy.patcher_extension.copy_nested_dicts(orig_model_options)
 
-def create_hook_patches_clone(orig_hook_patches):
+def create_hook_patches_clone(orig_hook_patches, copy_tuples=False):
     new_hook_patches = {}
     for hook_ref in orig_hook_patches:
         new_hook_patches[hook_ref] = {}
         for k in orig_hook_patches[hook_ref]:
             new_hook_patches[hook_ref][k] = orig_hook_patches[hook_ref][k][:]
+            if copy_tuples:
+                for i in range(len(new_hook_patches[hook_ref][k])):
+                    new_hook_patches[hook_ref][k][i] = tuple(new_hook_patches[hook_ref][k][i])
     return new_hook_patches
 
 def wipe_lowvram_weight(m):
@@ -240,6 +243,9 @@ class ModelPatcher:
         self.is_clip = False
         self.hook_mode = comfy.hooks.EnumHookMode.MaxSpeed
 
+        self.is_multigpu_base_clone = False
+        self.clone_base_uuid = uuid.uuid4()
+
         if not hasattr(self.model, 'model_loaded_weight_memory'):
             self.model.model_loaded_weight_memory = 0
 
@@ -318,18 +324,92 @@ class ModelPatcher:
         n.is_clip = self.is_clip
         n.hook_mode = self.hook_mode
 
+        n.is_multigpu_base_clone = self.is_multigpu_base_clone
+        n.clone_base_uuid = self.clone_base_uuid
+
         for callback in self.get_all_callbacks(CallbacksMP.ON_CLONE):
             callback(self, n)
         return n
 
+    def deepclone_multigpu(self, new_load_device=None, models_cache: dict[uuid.UUID,ModelPatcher]=None):
+        logging.info(f"Creating deepclone of {self.model.__class__.__name__} for {new_load_device if new_load_device else self.load_device}.")
+        comfy.model_management.unload_model_and_clones(self)
+        n = self.clone()
+        # set load device, if present
+        if new_load_device is not None:
+            n.load_device = new_load_device
+        # unlike for normal clone, backup dicts that shared same ref should not;
+        # otherwise, patchers that have deep copies of base models will erroneously influence each other.
+        n.backup = copy.deepcopy(n.backup)
+        n.object_patches_backup = copy.deepcopy(n.object_patches_backup)
+        n.hook_backup = copy.deepcopy(n.hook_backup)
+        n.model = copy.deepcopy(n.model)
+        # multigpu clone should not have multigpu additional_models entry
+        n.remove_additional_models("multigpu")
+        # multigpu_clone all stored additional_models; make sure circular references are properly handled
+        if models_cache is None:
+            models_cache = {}
+        for key, model_list in n.additional_models.items():
+            for i in range(len(model_list)):
+                add_model = n.additional_models[key][i]
+                if add_model.clone_base_uuid not in models_cache:
+                    models_cache[add_model.clone_base_uuid] = add_model.deepclone_multigpu(new_load_device=new_load_device, models_cache=models_cache)
+                n.additional_models[key][i] = models_cache[add_model.clone_base_uuid]
+        for callback in self.get_all_callbacks(CallbacksMP.ON_DEEPCLONE_MULTIGPU):
+            callback(self, n)
+        return n
+
+    def match_multigpu_clones(self):
+        multigpu_models = self.get_additional_models_with_key("multigpu")
+        if len(multigpu_models) > 0:
+            new_multigpu_models = []
+            for mm in multigpu_models:
+                # clone main model, but bring over relevant props from existing multigpu clone
+                n = self.clone()
+                n.load_device = mm.load_device
+                n.backup = mm.backup
+                n.object_patches_backup = mm.object_patches_backup
+                n.hook_backup = mm.hook_backup
+                n.model = mm.model
+                n.is_multigpu_base_clone = mm.is_multigpu_base_clone
+                n.remove_additional_models("multigpu")
+                orig_additional_models: dict[str, list[ModelPatcher]] = comfy.patcher_extension.copy_nested_dicts(n.additional_models)
+                n.additional_models = comfy.patcher_extension.copy_nested_dicts(mm.additional_models)
+                # figure out which additional models are not present in multigpu clone
+                models_cache = {}
+                for mm_add_model in mm.get_additional_models():
+                    models_cache[mm_add_model.clone_base_uuid] = mm_add_model
+                remove_models_uuids = set(list(models_cache.keys()))
+                for key, model_list in orig_additional_models.items():
+                    for orig_add_model in model_list:
+                        if orig_add_model.clone_base_uuid not in models_cache:
+                            models_cache[orig_add_model.clone_base_uuid] = orig_add_model.deepclone_multigpu(new_load_device=n.load_device, models_cache=models_cache)
+                            existing_list = n.get_additional_models_with_key(key)
+                            existing_list.append(models_cache[orig_add_model.clone_base_uuid])
+                            n.set_additional_models(key, existing_list)
+                        if orig_add_model.clone_base_uuid in remove_models_uuids:
+                            remove_models_uuids.remove(orig_add_model.clone_base_uuid)
+                # remove duplicate additional models
+                for key, model_list in n.additional_models.items():
+                    new_model_list = [x for x in model_list if x.clone_base_uuid not in remove_models_uuids]
+                    n.set_additional_models(key, new_model_list)
+                for callback in self.get_all_callbacks(CallbacksMP.ON_MATCH_MULTIGPU_CLONES):
+                    callback(self, n)
+                new_multigpu_models.append(n)
+            self.set_additional_models("multigpu", new_multigpu_models)
+
     def is_clone(self, other):
         if hasattr(other, 'model') and self.model is other.model:
             return True
         return False
 
-    def clone_has_same_weights(self, clone: 'ModelPatcher'):
-        if not self.is_clone(clone):
-            return False
+    def clone_has_same_weights(self, clone: ModelPatcher, allow_multigpu=False):
+        if allow_multigpu:
+            if self.clone_base_uuid != clone.clone_base_uuid:
+                return False
+        else:
+            if not self.is_clone(clone):
+                return False
 
         if self.current_hooks != clone.current_hooks:
             return False
@@ -929,7 +1009,7 @@ class ModelPatcher:
         return self.additional_models.get(key, [])
 
     def get_additional_models(self):
-        all_models = []
+        all_models: list[ModelPatcher] = []
         for models in self.additional_models.values():
             all_models.extend(models)
         return all_models
@@ -983,9 +1063,13 @@ class ModelPatcher:
         for callback in self.get_all_callbacks(CallbacksMP.ON_PRE_RUN):
             callback(self)
 
-    def prepare_state(self, timestep):
+    def prepare_state(self, timestep, model_options, ignore_multigpu=False):
         for callback in self.get_all_callbacks(CallbacksMP.ON_PREPARE_STATE):
-            callback(self, timestep)
+            callback(self, timestep, model_options, ignore_multigpu)
+        if not ignore_multigpu and "multigpu_clones" in model_options:
+            for p in model_options["multigpu_clones"].values():
+                p: ModelPatcher
+                p.prepare_state(timestep, model_options, ignore_multigpu=True)
 
     def restore_hook_patches(self):
         if self.hook_patches_backup is not None:
@@ -998,12 +1082,18 @@ class ModelPatcher:
     def prepare_hook_patches_current_keyframe(self, t: torch.Tensor, hook_group: comfy.hooks.HookGroup, model_options: dict[str]):
         curr_t = t[0]
         reset_current_hooks = False
+        multigpu_kf_changed_cache = None
         transformer_options = model_options.get("transformer_options", {})
         for hook in hook_group.hooks:
             changed = hook.hook_keyframe.prepare_current_keyframe(curr_t=curr_t, transformer_options=transformer_options)
             # if keyframe changed, remove any cached HookGroups that contain hook with the same hook_ref;
             # this will cause the weights to be recalculated when sampling
             if changed:
+                # cache changed for multigpu usage
+                if "multigpu_clones" in model_options:
+                    if multigpu_kf_changed_cache is None:
+                        multigpu_kf_changed_cache = []
+                    multigpu_kf_changed_cache.append(hook)
                 # reset current_hooks if contains hook that changed
                 if self.current_hooks is not None:
                     for current_hook in self.current_hooks.hooks:
@@ -1015,6 +1105,28 @@ class ModelPatcher:
                         self.cached_hook_patches.pop(cached_group)
         if reset_current_hooks:
             self.patch_hooks(None)
+        if "multigpu_clones" in model_options:
+            for p in model_options["multigpu_clones"].values():
+                p: ModelPatcher
+                p._handle_changed_hook_keyframes(multigpu_kf_changed_cache)
+
+    def _handle_changed_hook_keyframes(self, kf_changed_cache: list[comfy.hooks.Hook]):
+        'Used to handle multigpu behavior inside prepare_hook_patches_current_keyframe.'
+        if kf_changed_cache is None:
+            return
+        reset_current_hooks = False
+        # reset current_hooks if contains hook that changed
+        for hook in kf_changed_cache:
+            if self.current_hooks is not None:
+                for current_hook in self.current_hooks.hooks:
+                    if current_hook == hook:
+                        reset_current_hooks = True
+                        break
+            for cached_group in list(self.cached_hook_patches.keys()):
+                if cached_group.contains(hook):
+                    self.cached_hook_patches.pop(cached_group)
+        if reset_current_hooks:
+            self.patch_hooks(None)
 
     def register_all_hook_patches(self, hooks: comfy.hooks.HookGroup, target_dict: dict[str], model_options: dict=None,
                                   registered: comfy.hooks.HookGroup = None):

comfy/multigpu.py

@@ -0,0 +1,167 @@
+from __future__ import annotations
+import torch
+import logging
+
+from collections import namedtuple
+from typing import TYPE_CHECKING
+if TYPE_CHECKING:
+    from comfy.model_patcher import ModelPatcher
+import comfy.utils
+import comfy.patcher_extension
+import comfy.model_management
+
+
+class GPUOptions:
+    def __init__(self, device_index: int, relative_speed: float):
+        self.device_index = device_index
+        self.relative_speed = relative_speed
+
+    def clone(self):
+        return GPUOptions(self.device_index, self.relative_speed)
+
+    def create_dict(self):
+        return {
+            "relative_speed": self.relative_speed
+        }
+
+class GPUOptionsGroup:
+    def __init__(self):
+        self.options: dict[int, GPUOptions] = {}
+
+    def add(self, info: GPUOptions):
+        self.options[info.device_index] = info
+
+    def clone(self):
+        c = GPUOptionsGroup()
+        for opt in self.options.values():
+            c.add(opt)
+        return c
+
+    def register(self, model: ModelPatcher):
+        opts_dict = {}
+        # get devices that are valid for this model
+        devices: list[torch.device] = [model.load_device]
+        for extra_model in model.get_additional_models_with_key("multigpu"):
+            extra_model: ModelPatcher
+            devices.append(extra_model.load_device)
+        # create dictionary with actual device mapped to its GPUOptions
+        device_opts_list: list[GPUOptions] = []
+        for device in devices:
+            device_opts = self.options.get(device.index, GPUOptions(device_index=device.index, relative_speed=1.0))
+            opts_dict[device] = device_opts.create_dict()
+            device_opts_list.append(device_opts)
+        # make relative_speed relative to 1.0
+        min_speed = min([x.relative_speed for x in device_opts_list])
+        for value in opts_dict.values():
+            value['relative_speed'] /= min_speed
+        model.model_options['multigpu_options'] = opts_dict
+
+
+def create_multigpu_deepclones(model: ModelPatcher, max_gpus: int, gpu_options: GPUOptionsGroup=None, reuse_loaded=False):
+    'Prepare ModelPatcher to contain deepclones of its BaseModel and related properties.'
+    model = model.clone()
+    # check if multigpu is already prepared - get the load devices from them if possible to exclude
+    skip_devices = set()
+    multigpu_models = model.get_additional_models_with_key("multigpu")
+    if len(multigpu_models) > 0:
+        for mm in multigpu_models:
+            skip_devices.add(mm.load_device)
+    skip_devices = list(skip_devices)
+
+    full_extra_devices = comfy.model_management.get_all_torch_devices(exclude_current=True)
+    limit_extra_devices = full_extra_devices[:max_gpus-1]
+    extra_devices = limit_extra_devices.copy()
+    # exclude skipped devices
+    for skip in skip_devices:
+        if skip in extra_devices:
+            extra_devices.remove(skip)
+    # create new deepclones
+    if len(extra_devices) > 0:
+        for device in extra_devices:
+            device_patcher = None
+            if reuse_loaded:
+                # check if there are any ModelPatchers currently loaded that could be referenced here after a clone
+                loaded_models: list[ModelPatcher] = comfy.model_management.loaded_models()
+                for lm in loaded_models:
+                    if lm.model is not None and lm.clone_base_uuid == model.clone_base_uuid and lm.load_device == device:
+                        device_patcher = lm.clone()
+                        logging.info(f"Reusing loaded deepclone of {device_patcher.model.__class__.__name__} for {device}")
+                        break
+            if device_patcher is None:
+                device_patcher = model.deepclone_multigpu(new_load_device=device)
+                device_patcher.is_multigpu_base_clone = True
+            multigpu_models = model.get_additional_models_with_key("multigpu")
+            multigpu_models.append(device_patcher)
+            model.set_additional_models("multigpu", multigpu_models)
+        model.match_multigpu_clones()
+        if gpu_options is None:
+            gpu_options = GPUOptionsGroup()
+        gpu_options.register(model)
+    else:
+        logging.info("No extra torch devices need initialization, skipping initializing MultiGPU Work Units.")
+    # TODO: only keep model clones that don't go 'past' the intended max_gpu count
+    # multigpu_models = model.get_additional_models_with_key("multigpu")
+    # new_multigpu_models = []
+    # for m in multigpu_models:
+    #     if m.load_device in limit_extra_devices:
+    #         new_multigpu_models.append(m)
+    # model.set_additional_models("multigpu", new_multigpu_models)
+    # persist skip_devices for use in sampling code
+    # if len(skip_devices) > 0 or "multigpu_skip_devices" in model.model_options:
+    #     model.model_options["multigpu_skip_devices"] = skip_devices
+    return model
+
+
+LoadBalance = namedtuple('LoadBalance', ['work_per_device', 'idle_time'])
+def load_balance_devices(model_options: dict[str], total_work: int, return_idle_time=False, work_normalized: int=None):
+    'Optimize work assigned to different devices, accounting for their relative speeds and splittable work.'
+    opts_dict = model_options['multigpu_options']
+    devices = list(model_options['multigpu_clones'].keys())
+    speed_per_device = []
+    work_per_device = []
+    # get sum of each device's relative_speed
+    total_speed = 0.0
+    for opts in opts_dict.values():
+        total_speed += opts['relative_speed']
+    # get relative work for each device;
+    # obtained by w = (W*r)/R
+    for device in devices:
+        relative_speed = opts_dict[device]['relative_speed']
+        relative_work = (total_work*relative_speed) / total_speed
+        speed_per_device.append(relative_speed)
+        work_per_device.append(relative_work)
+    # relative work must be expressed in whole numbers, but likely is a decimal;
+    # perform rounding while maintaining total sum equal to total work (sum of relative works)
+    work_per_device = round_preserved(work_per_device)
+    dict_work_per_device = {}
+    for device, relative_work in zip(devices, work_per_device):
+        dict_work_per_device[device] = relative_work
+    if not return_idle_time:
+        return LoadBalance(dict_work_per_device, None)
+    # divide relative work by relative speed to get estimated completion time of said work by each device;
+    # time here is relative and does not correspond to real-world units
+    completion_time = [w/r for w,r in zip(work_per_device, speed_per_device)]
+    # calculate relative time spent by the devices waiting on each other after their work is completed
+    idle_time = abs(min(completion_time) - max(completion_time))
+    # if need to compare work idle time, need to normalize to a common total work
+    if work_normalized:
+        idle_time *= (work_normalized/total_work)
+
+    return LoadBalance(dict_work_per_device, idle_time)
+
+def round_preserved(values: list[float]):
+    'Round all values in a list, preserving the combined sum of values.'
+    # get floor of values; casting to int does it too
+    floored = [int(x) for x in values]
+    total_floored = sum(floored)
+    # get remainder to distribute
+    remainder = round(sum(values)) - total_floored
+    # pair values with fractional portions
+    fractional = [(i, x-floored[i]) for i, x in enumerate(values)]
+    # sort by fractional part in descending order
+    fractional.sort(key=lambda x: x[1], reverse=True)
+    # distribute the remainder
+    for i in range(remainder):
+        index = fractional[i][0]
+        floored[index] += 1
+    return floored

comfy/patcher_extension.py

@@ -3,6 +3,8 @@ from typing import Callable
 
 class CallbacksMP:
     ON_CLONE = "on_clone"
+    ON_DEEPCLONE_MULTIGPU = "on_deepclone_multigpu"
+    ON_MATCH_MULTIGPU_CLONES = "on_match_multigpu_clones"
     ON_LOAD = "on_load_after"
     ON_DETACH = "on_detach_after"
     ON_CLEANUP = "on_cleanup"

comfy/sampler_helpers.py

@@ -1,7 +1,9 @@
 from __future__ import annotations
+import torch
 import uuid
 import comfy.model_management
 import comfy.conds
+import comfy.model_patcher
 import comfy.utils
 import comfy.hooks
 import comfy.patcher_extension
@@ -104,6 +106,46 @@ def cleanup_additional_models(models):
         if hasattr(m, 'cleanup'):
             m.cleanup()
 
+def preprocess_multigpu_conds(conds: dict[str, list[dict[str]]], model: ModelPatcher, model_options: dict[str]):
+    '''If multigpu acceleration required, creates deepclones of ControlNets and GLIGEN per device.'''
+    multigpu_models: list[ModelPatcher] = model.get_additional_models_with_key("multigpu")
+    if len(multigpu_models) == 0:
+        return
+    extra_devices = [x.load_device for x in multigpu_models]
+    # handle controlnets
+    controlnets: set[ControlBase] = set()
+    for k in conds:
+        for kk in conds[k]:
+            if 'control' in kk:
+                controlnets.add(kk['control'])
+    if len(controlnets) > 0:
+        # first, unload all controlnet clones
+        for cnet in list(controlnets):
+            cnet_models = cnet.get_models()
+            for cm in cnet_models:
+                comfy.model_management.unload_model_and_clones(cm, unload_additional_models=True)
+
+        # next, make sure each controlnet has a deepclone for all relevant devices
+        for cnet in controlnets:
+            curr_cnet = cnet
+            while curr_cnet is not None:
+                for device in extra_devices:
+                    if device not in curr_cnet.multigpu_clones:
+                        curr_cnet.deepclone_multigpu(device, autoregister=True)
+                curr_cnet = curr_cnet.previous_controlnet
+        # since all device clones are now present, recreate the linked list for cloned cnets per device
+        for cnet in controlnets:
+            curr_cnet = cnet
+            while curr_cnet is not None:
+                prev_cnet = curr_cnet.previous_controlnet
+                for device in extra_devices:
+                    device_cnet = curr_cnet.get_instance_for_device(device)
+                    prev_device_cnet = None
+                    if prev_cnet is not None:
+                        prev_device_cnet = prev_cnet.get_instance_for_device(device)
+                    device_cnet.set_previous_controlnet(prev_device_cnet)
+                curr_cnet = prev_cnet
+    # potentially handle gligen - since not widely used, ignored for now
 
 def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
@@ -113,14 +155,15 @@ def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None
     return executor.execute(model, noise_shape, conds, model_options=model_options)
 
 def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
-    real_model: BaseModel = None
+    model.match_multigpu_clones()
+    preprocess_multigpu_conds(conds, model, model_options)
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)
     models += model.get_nested_additional_models()  # TODO: does this require inference_memory update?
     memory_required = model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory
     minimum_memory_required = model.memory_required([noise_shape[0]] + list(noise_shape[1:])) + inference_memory
     comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required)
-    real_model = model.model
+    real_model: BaseModel = model.model
 
     return real_model, conds, models
 
@@ -133,7 +176,7 @@ def cleanup_models(conds, models):
 
     cleanup_additional_models(set(control_cleanup))
 
-def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
+def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict):
     '''
     Registers hooks from conds.
     '''
@@ -166,3 +209,18 @@ def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
         comfy.patcher_extension.merge_nested_dicts(to_load_options.setdefault(wc_name, {}), model_options["transformer_options"][wc_name],
                                                     copy_dict1=False)
     return to_load_options
+
+def prepare_model_patcher_multigpu_clones(model_patcher: ModelPatcher, loaded_models: list[ModelPatcher], model_options: dict):
+    '''
+    In case multigpu acceleration is enabled, prep ModelPatchers for each device.
+    '''
+    multigpu_patchers: list[ModelPatcher] = [x for x in loaded_models if x.is_multigpu_base_clone]
+    if len(multigpu_patchers) > 0:
+        multigpu_dict: dict[torch.device, ModelPatcher] = {}
+        multigpu_dict[model_patcher.load_device] = model_patcher
+        for x in multigpu_patchers:
+            x.hook_patches = comfy.model_patcher.create_hook_patches_clone(model_patcher.hook_patches, copy_tuples=True)
+            x.hook_mode = model_patcher.hook_mode # match main model's hook_mode
+            multigpu_dict[x.load_device] = x
+        model_options["multigpu_clones"] = multigpu_dict
+    return multigpu_patchers

comfy/samplers.py

@@ -1,4 +1,6 @@
 from __future__ import annotations
+
+import comfy.model_management
 from .k_diffusion import sampling as k_diffusion_sampling
 from .extra_samplers import uni_pc
 from typing import TYPE_CHECKING, Callable, NamedTuple
@@ -18,6 +20,7 @@ import comfy.patcher_extension
 import comfy.hooks
 import scipy.stats
 import numpy
+import threading
 
 
 def add_area_dims(area, num_dims):
@@ -140,7 +143,7 @@ def can_concat_cond(c1, c2):
 
     return cond_equal_size(c1.conditioning, c2.conditioning)
 
-def cond_cat(c_list):
+def cond_cat(c_list, device=None):
     temp = {}
     for x in c_list:
         for k in x:
@@ -152,6 +155,8 @@ def cond_cat(c_list):
     for k in temp:
         conds = temp[k]
         out[k] = conds[0].concat(conds[1:])
+        if device is not None and hasattr(out[k], 'to'):
+            out[k] = out[k].to(device)
 
     return out
 
@@ -205,7 +210,9 @@ def calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Ten
     )
     return executor.execute(model, conds, x_in, timestep, model_options)
 
-def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+    if 'multigpu_clones' in model_options:
+        return _calc_cond_batch_multigpu(model, conds, x_in, timestep, model_options)
     out_conds = []
     out_counts = []
     # separate conds by matching hooks
@@ -237,7 +244,7 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
     if has_default_conds:
         finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options)
 
-    model.current_patcher.prepare_state(timestep)
+    model.current_patcher.prepare_state(timestep, model_options)
 
     # run every hooked_to_run separately
     for hooks, to_run in hooked_to_run.items():
@@ -339,6 +346,190 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
 
     return out_conds
 
+def _calc_cond_batch_multigpu(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+    out_conds = []
+    out_counts = []
+    # separate conds by matching hooks
+    hooked_to_run: dict[comfy.hooks.HookGroup,list[tuple[tuple,int]]] = {}
+    default_conds = []
+    has_default_conds = False
+
+    output_device = x_in.device
+
+    for i in range(len(conds)):
+        out_conds.append(torch.zeros_like(x_in))
+        out_counts.append(torch.ones_like(x_in) * 1e-37)
+
+        cond = conds[i]
+        default_c = []
+        if cond is not None:
+            for x in cond:
+                if 'default' in x:
+                    default_c.append(x)
+                    has_default_conds = True
+                    continue
+                p = get_area_and_mult(x, x_in, timestep)
+                if p is None:
+                    continue
+                if p.hooks is not None:
+                    model.current_patcher.prepare_hook_patches_current_keyframe(timestep, p.hooks, model_options)
+                hooked_to_run.setdefault(p.hooks, list())
+                hooked_to_run[p.hooks] += [(p, i)]
+        default_conds.append(default_c)
+
+    if has_default_conds:
+        finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options)
+
+    model.current_patcher.prepare_state(timestep, model_options)
+
+    devices = [dev_m for dev_m in model_options['multigpu_clones'].keys()]
+    device_batched_hooked_to_run: dict[torch.device, list[tuple[comfy.hooks.HookGroup, tuple]]] = {}
+
+    total_conds = 0
+    for to_run in hooked_to_run.values():
+        total_conds += len(to_run)
+    conds_per_device = max(1, math.ceil(total_conds//len(devices)))
+    index_device = 0
+    current_device = devices[index_device]
+    # run every hooked_to_run separately
+    for hooks, to_run in hooked_to_run.items():
+        while len(to_run) > 0:
+            current_device = devices[index_device % len(devices)]
+            batched_to_run = device_batched_hooked_to_run.setdefault(current_device, [])
+            # keep track of conds currently scheduled onto this device
+            batched_to_run_length = 0
+            for btr in batched_to_run:
+                batched_to_run_length += len(btr[1])
+
+            first = to_run[0]
+            first_shape = first[0][0].shape
+            to_batch_temp = []
+            # make sure not over conds_per_device limit when creating temp batch
+            for x in range(len(to_run)):
+                if can_concat_cond(to_run[x][0], first[0]) and len(to_batch_temp) < (conds_per_device - batched_to_run_length):
+                    to_batch_temp += [x]
+
+            to_batch_temp.reverse()
+            to_batch = to_batch_temp[:1]
+
+            free_memory = model_management.get_free_memory(current_device)
+            for i in range(1, len(to_batch_temp) + 1):
+                batch_amount = to_batch_temp[:len(to_batch_temp)//i]
+                input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
+                if model.memory_required(input_shape) * 1.5 < free_memory:
+                    to_batch = batch_amount
+                    break
+            conds_to_batch = []
+            for x in to_batch:
+                conds_to_batch.append(to_run.pop(x))
+            batched_to_run_length += len(conds_to_batch)
+
+            batched_to_run.append((hooks, conds_to_batch))
+            if batched_to_run_length >= conds_per_device:
+                index_device += 1
+
+    thread_result = collections.namedtuple('thread_result', ['output', 'mult', 'area', 'batch_chunks', 'cond_or_uncond'])
+    def _handle_batch(device: torch.device, batch_tuple: tuple[comfy.hooks.HookGroup, tuple], results: list[thread_result]):
+        model_current: BaseModel = model_options["multigpu_clones"][device].model
+        # run every hooked_to_run separately
+        with torch.no_grad():
+            for hooks, to_batch in batch_tuple:
+                input_x = []
+                mult = []
+                c = []
+                cond_or_uncond = []
+                uuids = []
+                area = []
+                control: ControlBase = None
+                patches = None
+                for x in to_batch:
+                    o = x
+                    p = o[0]
+                    input_x.append(p.input_x)
+                    mult.append(p.mult)
+                    c.append(p.conditioning)
+                    area.append(p.area)
+                    cond_or_uncond.append(o[1])
+                    uuids.append(p.uuid)
+                    control = p.control
+                    patches = p.patches
+
+                batch_chunks = len(cond_or_uncond)
+                input_x = torch.cat(input_x).to(device)
+                c = cond_cat(c, device=device)
+                timestep_ = torch.cat([timestep.to(device)] * batch_chunks)
+
+                transformer_options = model_current.current_patcher.apply_hooks(hooks=hooks)
+                if 'transformer_options' in model_options:
+                    transformer_options = comfy.patcher_extension.merge_nested_dicts(transformer_options,
+                                                                                    model_options['transformer_options'],
+                                                                                    copy_dict1=False)
+
+                if patches is not None:
+                    # TODO: replace with merge_nested_dicts function
+                    if "patches" in transformer_options:
+                        cur_patches = transformer_options["patches"].copy()
+                        for p in patches:
+                            if p in cur_patches:
+                                cur_patches[p] = cur_patches[p] + patches[p]
+                            else:
+                                cur_patches[p] = patches[p]
+                        transformer_options["patches"] = cur_patches
+                    else:
+                        transformer_options["patches"] = patches
+
+                transformer_options["cond_or_uncond"] = cond_or_uncond[:]
+                transformer_options["uuids"] = uuids[:]
+                transformer_options["sigmas"] = timestep
+                transformer_options["sample_sigmas"] = transformer_options["sample_sigmas"].to(device)
+                transformer_options["multigpu_thread_device"] = device
+
+                cast_transformer_options(transformer_options, device=device)
+                c['transformer_options'] = transformer_options
+
+                if control is not None:
+                    device_control = control.get_instance_for_device(device)
+                    c['control'] = device_control.get_control(input_x, timestep_, c, len(cond_or_uncond), transformer_options)
+
+                if 'model_function_wrapper' in model_options:
+                    output = model_options['model_function_wrapper'](model_current.apply_model, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).to(output_device).chunk(batch_chunks)
+                else:
+                    output = model_current.apply_model(input_x, timestep_, **c).to(output_device).chunk(batch_chunks)
+                results.append(thread_result(output, mult, area, batch_chunks, cond_or_uncond))
+
+
+    results: list[thread_result] = []
+    threads: list[threading.Thread] = []
+    for device, batch_tuple in device_batched_hooked_to_run.items():
+        new_thread = threading.Thread(target=_handle_batch, args=(device, batch_tuple, results))
+        threads.append(new_thread)
+        new_thread.start()
+
+    for thread in threads:
+        thread.join()
+
+    for output, mult, area, batch_chunks, cond_or_uncond in results:
+        for o in range(batch_chunks):
+            cond_index = cond_or_uncond[o]
+            a = area[o]
+            if a is None:
+                out_conds[cond_index] += output[o] * mult[o]
+                out_counts[cond_index] += mult[o]
+            else:
+                out_c = out_conds[cond_index]
+                out_cts = out_counts[cond_index]
+                dims = len(a) // 2
+                for i in range(dims):
+                    out_c = out_c.narrow(i + 2, a[i + dims], a[i])
+                    out_cts = out_cts.narrow(i + 2, a[i + dims], a[i])
+                out_c += output[o] * mult[o]
+                out_cts += mult[o]
+
+    for i in range(len(out_conds)):
+        out_conds[i] /= out_counts[i]
+
+    return out_conds
+
 def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): #TODO: remove
     logging.warning("WARNING: The comfy.samplers.calc_cond_uncond_batch function is deprecated please use the calc_cond_batch one instead.")
     return tuple(calc_cond_batch(model, [cond, uncond], x_in, timestep, model_options))
@@ -636,6 +827,8 @@ def pre_run_control(model, conds):
         percent_to_timestep_function = lambda a: s.percent_to_sigma(a)
         if 'control' in x:
             x['control'].pre_run(model, percent_to_timestep_function)
+            for device_cnet in x['control'].multigpu_clones.values():
+                device_cnet.pre_run(model, percent_to_timestep_function)
 
 def apply_empty_x_to_equal_area(conds, uncond, name, uncond_fill_func):
     cond_cnets = []
@@ -878,7 +1071,9 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
     to_load_options = model_options.get("to_load_options", None)
     if to_load_options is None:
         return
+    cast_transformer_options(to_load_options, device, dtype)
 
+def cast_transformer_options(transformer_options: dict[str], device=None, dtype=None):
     casts = []
     if device is not None:
         casts.append(device)
@@ -887,18 +1082,17 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
     # if nothing to apply, do nothing
     if len(casts) == 0:
         return
-
     # try to call .to on patches
-    if "patches" in to_load_options:
-        patches = to_load_options["patches"]
+    if "patches" in transformer_options:
+        patches = transformer_options["patches"]
         for name in patches:
             patch_list = patches[name]
             for i in range(len(patch_list)):
                 if hasattr(patch_list[i], "to"):
                     for cast in casts:
                         patch_list[i] = patch_list[i].to(cast)
-    if "patches_replace" in to_load_options:
-        patches = to_load_options["patches_replace"]
+    if "patches_replace" in transformer_options:
+        patches = transformer_options["patches_replace"]
         for name in patches:
             patch_list = patches[name]
             for k in patch_list:
@@ -908,8 +1102,8 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
     # try to call .to on any wrappers/callbacks
     wrappers_and_callbacks = ["wrappers", "callbacks"]
     for wc_name in wrappers_and_callbacks:
-        if wc_name in to_load_options:
-            wc: dict[str, list] = to_load_options[wc_name]
+        if wc_name in transformer_options:
+            wc: dict[str, list] = transformer_options[wc_name]
             for wc_dict in wc.values():
                 for wc_list in wc_dict.values():
                     for i in range(len(wc_list)):
@@ -917,7 +1111,6 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
                             for cast in casts:
                                 wc_list[i] = wc_list[i].to(cast)
 
-
 class CFGGuider:
     def __init__(self, model_patcher: ModelPatcher):
         self.model_patcher = model_patcher
@@ -963,6 +1156,8 @@ class CFGGuider:
         self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options)
         device = self.model_patcher.load_device
 
+        multigpu_patchers = comfy.sampler_helpers.prepare_model_patcher_multigpu_clones(self.model_patcher, self.loaded_models, self.model_options)
+
         if denoise_mask is not None:
             denoise_mask = comfy.sampler_helpers.prepare_mask(denoise_mask, noise.shape, device)
 
@@ -973,9 +1168,13 @@ class CFGGuider:
 
         try:
             self.model_patcher.pre_run()
+            for multigpu_patcher in multigpu_patchers:
+                multigpu_patcher.pre_run()
             output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
         finally:
             self.model_patcher.cleanup()
+            for multigpu_patcher in multigpu_patchers:
+                multigpu_patcher.cleanup()
 
         comfy.sampler_helpers.cleanup_models(self.conds, self.loaded_models)
         del self.inner_model

comfy_execution/caching.py

@@ -316,3 +316,156 @@ class LRUCache(BasicCache):
             self.children[cache_key].append(self.cache_key_set.get_data_key(child_id))
         return self
 
+
+class DependencyAwareCache(BasicCache):
+    """
+    A cache implementation that tracks dependencies between nodes and manages
+    their execution and caching accordingly. It extends the BasicCache class.
+    Nodes are removed from this cache once all of their descendants have been
+    executed.
+    """
+
+    def __init__(self, key_class):
+        """
+        Initialize the DependencyAwareCache.
+
+        Args:
+            key_class: The class used for generating cache keys.
+        """
+        super().__init__(key_class)
+        self.descendants = {}  # Maps node_id -> set of descendant node_ids
+        self.ancestors = {}    # Maps node_id -> set of ancestor node_ids
+        self.executed_nodes = set()  # Tracks nodes that have been executed
+
+    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+        """
+        Clear the entire cache and rebuild the dependency graph.
+
+        Args:
+            dynprompt: The dynamic prompt object containing node information.
+            node_ids: List of node IDs to initialize the cache for.
+            is_changed_cache: Flag indicating if the cache has changed.
+        """
+        # Clear all existing cache data
+        self.cache.clear()
+        self.subcaches.clear()
+        self.descendants.clear()
+        self.ancestors.clear()
+        self.executed_nodes.clear()
+
+        # Call the parent method to initialize the cache with the new prompt
+        super().set_prompt(dynprompt, node_ids, is_changed_cache)
+
+        # Rebuild the dependency graph
+        self._build_dependency_graph(dynprompt, node_ids)
+
+    def _build_dependency_graph(self, dynprompt, node_ids):
+        """
+        Build the dependency graph for all nodes.
+
+        Args:
+            dynprompt: The dynamic prompt object containing node information.
+            node_ids: List of node IDs to build the graph for.
+        """
+        self.descendants.clear()
+        self.ancestors.clear()
+        for node_id in node_ids:
+            self.descendants[node_id] = set()
+            self.ancestors[node_id] = set()
+
+        for node_id in node_ids:
+            inputs = dynprompt.get_node(node_id)["inputs"]
+            for input_data in inputs.values():
+                if is_link(input_data):  # Check if the input is a link to another node
+                    ancestor_id = input_data[0]
+                    self.descendants[ancestor_id].add(node_id)
+                    self.ancestors[node_id].add(ancestor_id)
+
+    def set(self, node_id, value):
+        """
+        Mark a node as executed and store its value in the cache.
+
+        Args:
+            node_id: The ID of the node to store.
+            value: The value to store for the node.
+        """
+        self._set_immediate(node_id, value)
+        self.executed_nodes.add(node_id)
+        self._cleanup_ancestors(node_id)
+
+    def get(self, node_id):
+        """
+        Retrieve the cached value for a node.
+
+        Args:
+            node_id: The ID of the node to retrieve.
+
+        Returns:
+            The cached value for the node.
+        """
+        return self._get_immediate(node_id)
+
+    def ensure_subcache_for(self, node_id, children_ids):
+        """
+        Ensure a subcache exists for a node and update dependencies.
+
+        Args:
+            node_id: The ID of the parent node.
+            children_ids: List of child node IDs to associate with the parent node.
+
+        Returns:
+            The subcache object for the node.
+        """
+        subcache = super()._ensure_subcache(node_id, children_ids)
+        for child_id in children_ids:
+            self.descendants[node_id].add(child_id)
+            self.ancestors[child_id].add(node_id)
+        return subcache
+
+    def _cleanup_ancestors(self, node_id):
+        """
+        Check if ancestors of a node can be removed from the cache.
+
+        Args:
+            node_id: The ID of the node whose ancestors are to be checked.
+        """
+        for ancestor_id in self.ancestors.get(node_id, []):
+            if ancestor_id in self.executed_nodes:
+                # Remove ancestor if all its descendants have been executed
+                if all(descendant in self.executed_nodes for descendant in self.descendants[ancestor_id]):
+                    self._remove_node(ancestor_id)
+
+    def _remove_node(self, node_id):
+        """
+        Remove a node from the cache.
+
+        Args:
+            node_id: The ID of the node to remove.
+        """
+        cache_key = self.cache_key_set.get_data_key(node_id)
+        if cache_key in self.cache:
+            del self.cache[cache_key]
+        subcache_key = self.cache_key_set.get_subcache_key(node_id)
+        if subcache_key in self.subcaches:
+            del self.subcaches[subcache_key]
+
+    def clean_unused(self):
+        """
+        Clean up unused nodes. This is a no-op for this cache implementation.
+        """
+        pass
+
+    def recursive_debug_dump(self):
+        """
+        Dump the cache and dependency graph for debugging.
+
+        Returns:
+            A list containing the cache state and dependency graph.
+        """
+        result = super().recursive_debug_dump()
+        result.append({
+            "descendants": self.descendants,
+            "ancestors": self.ancestors,
+            "executed_nodes": list(self.executed_nodes),
+        })
+        return result

comfy_extras/nodes_multigpu.py

@@ -0,0 +1,86 @@
+from __future__ import annotations
+from inspect import cleandoc
+
+from typing import TYPE_CHECKING
+if TYPE_CHECKING:
+    from comfy.model_patcher import ModelPatcher
+import comfy.multigpu
+
+
+class MultiGPUWorkUnitsNode:
+    """
+    Prepares model to have sampling accelerated via splitting work units.
+
+    Should be placed after nodes that modify the model object itself, such as compile or attention-switch nodes.
+
+    Other than those exceptions, this node can be placed in any order.
+    """
+
+    NodeId = "MultiGPU_WorkUnits"
+    NodeName = "MultiGPU Work Units"
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model": ("MODEL",),
+                "max_gpus" : ("INT", {"default": 8, "min": 1, "step": 1}),
+            },
+            "optional": {
+                "gpu_options": ("GPU_OPTIONS",)
+            }
+        }
+
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "init_multigpu"
+    CATEGORY = "advanced/multigpu"
+    DESCRIPTION = cleandoc(__doc__)
+
+    def init_multigpu(self, model: ModelPatcher, max_gpus: int, gpu_options: comfy.multigpu.GPUOptionsGroup=None):
+        model = comfy.multigpu.create_multigpu_deepclones(model, max_gpus, gpu_options, reuse_loaded=True)
+        return (model,)
+
+class MultiGPUOptionsNode:
+    """
+    Select the relative speed of GPUs in the special case they have significantly different performance from one another.
+    """
+
+    NodeId = "MultiGPU_Options"
+    NodeName = "MultiGPU Options"
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "device_index": ("INT", {"default": 0, "min": 0, "max": 64}),
+                "relative_speed": ("FLOAT", {"default": 1.0, "min": 0.0, "step": 0.01})
+            },
+            "optional": {
+                "gpu_options": ("GPU_OPTIONS",)
+            }
+        }
+
+    RETURN_TYPES = ("GPU_OPTIONS",)
+    FUNCTION = "create_gpu_options"
+    CATEGORY = "advanced/multigpu"
+    DESCRIPTION = cleandoc(__doc__)
+
+    def create_gpu_options(self, device_index: int, relative_speed: float, gpu_options: comfy.multigpu.GPUOptionsGroup=None):
+        if not gpu_options:
+            gpu_options = comfy.multigpu.GPUOptionsGroup()
+        gpu_options.clone()
+
+        opt = comfy.multigpu.GPUOptions(device_index=device_index, relative_speed=relative_speed)
+        gpu_options.add(opt)
+
+        return (gpu_options,)
+
+
+node_list = [
+    MultiGPUWorkUnitsNode,
+    MultiGPUOptionsNode
+]
+NODE_CLASS_MAPPINGS = {}
+NODE_DISPLAY_NAME_MAPPINGS = {}
+
+for node in node_list:
+    NODE_CLASS_MAPPINGS[node.NodeId] = node
+    NODE_DISPLAY_NAME_MAPPINGS[node.NodeId] = node.NodeName

execution.py

@@ -15,7 +15,7 @@ import nodes
 import comfy.model_management
 from comfy_execution.graph import get_input_info, ExecutionList, DynamicPrompt, ExecutionBlocker
 from comfy_execution.graph_utils import is_link, GraphBuilder
-from comfy_execution.caching import HierarchicalCache, LRUCache, CacheKeySetInputSignature, CacheKeySetID
+from comfy_execution.caching import HierarchicalCache, LRUCache, DependencyAwareCache, CacheKeySetInputSignature, CacheKeySetID
 from comfy_execution.validation import validate_node_input
 
 class ExecutionResult(Enum):
@@ -59,20 +59,27 @@ class IsChangedCache:
             self.is_changed[node_id] = node["is_changed"]
         return self.is_changed[node_id]
 
-class CacheSet:
-    def __init__(self, lru_size=None):
-        if lru_size is None or lru_size == 0:
-            self.init_classic_cache()
-        else:
-            self.init_lru_cache(lru_size)
-        self.all = [self.outputs, self.ui, self.objects]
 
-    # Useful for those with ample RAM/VRAM -- allows experimenting without
-    # blowing away the cache every time
-    def init_lru_cache(self, cache_size):
-        self.outputs = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
-        self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
-        self.objects = HierarchicalCache(CacheKeySetID)
+class CacheType(Enum):
+    CLASSIC = 0
+    LRU = 1
+    DEPENDENCY_AWARE = 2
+
+
+class CacheSet:
+    def __init__(self, cache_type=None, cache_size=None):
+        if cache_type == CacheType.DEPENDENCY_AWARE:
+            self.init_dependency_aware_cache()
+            logging.info("Disabling intermediate node cache.")
+        elif cache_type == CacheType.LRU:
+            if cache_size is None:
+                cache_size = 0
+            self.init_lru_cache(cache_size)
+            logging.info("Using LRU cache")
+        else:
+            self.init_classic_cache()
+
+        self.all = [self.outputs, self.ui, self.objects]
 
     # Performs like the old cache -- dump data ASAP
     def init_classic_cache(self):
@@ -80,6 +87,17 @@ class CacheSet:
         self.ui = HierarchicalCache(CacheKeySetInputSignature)
         self.objects = HierarchicalCache(CacheKeySetID)
 
+    def init_lru_cache(self, cache_size):
+        self.outputs = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+        self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+        self.objects = HierarchicalCache(CacheKeySetID)
+
+    # only hold cached items while the decendents have not executed
+    def init_dependency_aware_cache(self):
+        self.outputs = DependencyAwareCache(CacheKeySetInputSignature)
+        self.ui = DependencyAwareCache(CacheKeySetInputSignature)
+        self.objects = DependencyAwareCache(CacheKeySetID)
+
     def recursive_debug_dump(self):
         result = {
             "outputs": self.outputs.recursive_debug_dump(),
@@ -414,13 +432,14 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
     return (ExecutionResult.SUCCESS, None, None)
 
 class PromptExecutor:
-    def __init__(self, server, lru_size=None):
-        self.lru_size = lru_size
+    def __init__(self, server, cache_type=False, cache_size=None):
+        self.cache_size = cache_size
+        self.cache_type = cache_type
         self.server = server
         self.reset()
 
     def reset(self):
-        self.caches = CacheSet(self.lru_size)
+        self.caches = CacheSet(cache_type=self.cache_type, cache_size=self.cache_size)
         self.status_messages = []
         self.success = True
 

main.py

@@ -156,7 +156,13 @@ def cuda_malloc_warning():
 
 def prompt_worker(q, server_instance):
     current_time: float = 0.0
-    e = execution.PromptExecutor(server_instance, lru_size=args.cache_lru)
+    cache_type = execution.CacheType.CLASSIC
+    if args.cache_lru > 0:
+        cache_type = execution.CacheType.LRU
+    elif args.cache_none:
+        cache_type = execution.CacheType.DEPENDENCY_AWARE
+
+    e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_size=args.cache_lru)
     last_gc_collect = 0
     need_gc = False
     gc_collect_interval = 10.0

nodes.py

@@ -2270,6 +2270,7 @@ def init_builtin_extra_nodes():
         "nodes_mahiro.py",
         "nodes_lt.py",
         "nodes_hooks.py",
+        "nodes_multigpu.py",
         "nodes_load_3d.py",
         "nodes_cosmos.py",
         "nodes_video.py",

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.14.6
+comfyui-frontend-package==1.15.13
 torch
 torchsde
 torchvision

server.py

@@ -48,7 +48,7 @@ async def send_socket_catch_exception(function, message):
 @web.middleware
 async def cache_control(request: web.Request, handler):
     response: web.Response = await handler(request)
-    if request.path.endswith('.js') or request.path.endswith('.css'):
+    if request.path.endswith('.js') or request.path.endswith('.css') or request.path.endswith('index.json'):
         response.headers.setdefault('Cache-Control', 'no-cache')
     return response