Made multigpu deepclone load model from disk to avoid needing to deepclone actual model object, fixed issues with merge, turn off cuda backend as it causes device mismatch issue with rope (and potentially other ops), will investigate

comfy/model_patcher.py

@@ -23,6 +23,7 @@ import inspect
 import logging
 import math
 import uuid
+import copy
 from typing import Callable, Optional
 
 import torch
@@ -274,6 +275,7 @@ class ModelPatcher:
         self.is_clip = False
         self.hook_mode = comfy.hooks.EnumHookMode.MaxSpeed
 
+        self.cached_patcher_init: tuple[Callable, tuple] | tuple[Callable, tuple, int] | None = None
         self.is_multigpu_base_clone = False
         self.clone_base_uuid = uuid.uuid4()
 
@@ -368,6 +370,7 @@ class ModelPatcher:
         n.is_clip = self.is_clip
         n.hook_mode = self.hook_mode
 
+        n.cached_patcher_init = self.cached_patcher_init
         n.is_multigpu_base_clone = self.is_multigpu_base_clone
         n.clone_base_uuid = self.clone_base_uuid
 
@@ -382,12 +385,18 @@ class ModelPatcher:
         # set load device, if present
         if new_load_device is not None:
             n.load_device = new_load_device
+        if self.cached_patcher_init is not None:
+            temp_model_patcher: ModelPatcher | list[ModelPatcher] = self.cached_patcher_init[0](*self.cached_patcher_init[1])
+            if len(self.cached_patcher_init) > 2:
+                temp_model_patcher = temp_model_patcher[self.cached_patcher_init[2]]
+            n.model = temp_model_patcher.model
+        else:
+            n.model = copy.deepcopy(n.model)
         # 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

comfy/quant_ops.py

@@ -20,7 +20,7 @@ try:
         if cuda_version < (13,):
             ck.registry.disable("cuda")
             logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
-
+    ck.registry.disable("cuda") # multigpu will not work rn with comfy-kitchen on cuda backend
     ck.registry.disable("triton")
     for k, v in ck.list_backends().items():
         logging.info(f"Found comfy_kitchen backend {k}: {v}")

comfy/samplers.py

@@ -418,7 +418,7 @@ def _calc_cond_batch_multigpu(model: BaseModel, conds: list[list[dict]], x_in: t
             to_batch_temp.reverse()
             to_batch = to_batch_temp[:1]
 
-            free_memory = model_management.get_free_memory(current_device)
+            free_memory = comfy.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:]
@@ -487,7 +487,7 @@ def _calc_cond_batch_multigpu(model: BaseModel, conds: list[list[dict]], x_in: t
 
                     transformer_options["cond_or_uncond"] = cond_or_uncond[:]
                     transformer_options["uuids"] = uuids[:]
-                    transformer_options["sigmas"] = timestep
+                    transformer_options["sigmas"] = timestep.to(device)
                     transformer_options["sample_sigmas"] = transformer_options["sample_sigmas"].to(device)
                     transformer_options["multigpu_thread_device"] = device
 

comfy/sd.py

@@ -1510,6 +1510,7 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
     out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata)
     if out is None:
         raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(ckpt_path, model_detection_error_hint(ckpt_path, sd)))
+    out[0].cached_patcher_init = (load_checkpoint_guess_config, (ckpt_path, False, False, False, embedding_directory, output_model, model_options, te_model_options), 0)
     return out
 
 def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}, metadata=None):
@@ -1711,6 +1712,7 @@ def load_diffusion_model(unet_path, model_options={}):
     if model is None:
         logging.error("ERROR UNSUPPORTED DIFFUSION MODEL {}".format(unet_path))
         raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(unet_path, model_detection_error_hint(unet_path, sd)))
+    model.cached_patcher_init = (load_diffusion_model, (unet_path, model_options))
     return model
 
 def load_unet(unet_path, dtype=None):