Merge 2d96b2fdf1 into 451af70154

comfy/audio_encoders/audio_encoders.py

@@ -25,11 +25,11 @@ class AudioEncoderModel():
         elif model_type == "whisper3":
             self.model = WhisperLargeV3(**model_config)
         self.model.eval()
-        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
         self.model_sample_rate = 16000
 
     def load_sd(self, sd):
-        return self.model.load_state_dict(sd, strict=False)
+        return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
 
     def get_sd(self):
         return self.model.state_dict()

comfy/cli_args.py

@@ -159,6 +159,7 @@ class PerformanceFeature(enum.Enum):
     Fp8MatrixMultiplication = "fp8_matrix_mult"
     CublasOps = "cublas_ops"
     AutoTune = "autotune"
+    DynamicVRAM = "dynamic_vram"
 
 parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
 
@@ -257,3 +258,6 @@ elif args.fast == []:
 # '--fast' is provided with a list of performance features, use that list
 else:
     args.fast = set(args.fast)
+
+def enables_dynamic_vram():
+    return PerformanceFeature.DynamicVRAM in args.fast and not args.highvram and not args.gpu_only

comfy/clip_vision.py

@@ -47,10 +47,10 @@ class ClipVisionModel():
         self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast)
         self.model.eval()
 
-        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
 
     def load_sd(self, sd):
-        return self.model.load_state_dict(sd, strict=False)
+        return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
 
     def get_sd(self):
         return self.model.state_dict()

comfy/controlnet.py

@@ -203,7 +203,7 @@ class ControlNet(ControlBase):
         self.control_model = control_model
         self.load_device = load_device
         if control_model is not None:
-            self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
+            self.control_model_wrapped = comfy.model_patcher.CoreModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
 
         self.compression_ratio = compression_ratio
         self.global_average_pooling = global_average_pooling

comfy/k_diffusion/sampling.py

@@ -1,11 +1,12 @@
 import math
+import time
 from functools import partial
 
 from scipy import integrate
 import torch
 from torch import nn
 import torchsde
-from tqdm.auto import trange, tqdm
+from tqdm.auto import trange as trange_, tqdm
 
 from . import utils
 from . import deis
@@ -13,6 +14,36 @@ from . import sa_solver
 import comfy.model_patcher
 import comfy.model_sampling
 
+import comfy.memory_management
+
+
+def trange(*args, **kwargs):
+    if comfy.memory_management.aimdo_allocator is None:
+        return trange_(*args, **kwargs)
+
+    pbar = trange_(*args, **kwargs, smoothing=1.0)
+    pbar._i = 0
+    pbar.set_postfix_str("  Model Initializing ...  ")
+
+    _update = pbar.update
+
+    def warmup_update(n=1):
+        pbar._i += 1
+        if pbar._i == 1:
+            pbar.i1_time = time.time()
+            pbar.set_postfix_str(" Model Initialization complete!  ")
+        elif pbar._i == 2:
+            #bring forward the effective start time based the the diff between first and second iteration
+            #to attempt to remove load overhead from the final step rate estimate.
+            pbar.start_t = pbar.i1_time - (time.time() - pbar.i1_time)
+            pbar.set_postfix_str("")
+
+        _update(n)
+
+    pbar.update = warmup_update
+    return pbar
+
+
 def append_zero(x):
     return torch.cat([x, x.new_zeros([1])])
 

comfy/ldm/hunyuan_video/upsampler.py

@@ -109,10 +109,10 @@ class HunyuanVideo15SRModel():
         self.model_class = UPSAMPLERS.get(model_type)
         self.model = self.model_class(**config).eval()
 
-        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
 
     def load_sd(self, sd):
-        return self.model.load_state_dict(sd, strict=True)
+        return self.model.load_state_dict(sd, strict=True, assign=self.patcher.is_dynamic())
 
     def get_sd(self):
         return self.model.state_dict()

comfy/memory_management.py

@@ -0,0 +1,81 @@
+import math
+import torch
+from typing import NamedTuple
+
+from comfy.quant_ops import QuantizedTensor
+
+class TensorGeometry(NamedTuple):
+    shape: any
+    dtype: torch.dtype
+
+    def element_size(self):
+        info = torch.finfo(self.dtype) if self.dtype.is_floating_point else torch.iinfo(self.dtype)
+        return info.bits // 8
+
+    def numel(self):
+        return math.prod(self.shape)
+
+def tensors_to_geometries(tensors, dtype=None):
+    geometries = []
+    for t in tensors:
+        if t is None or isinstance(t, QuantizedTensor):
+            geometries.append(t)
+            continue
+        tdtype = t.dtype
+        if hasattr(t, "_model_dtype"):
+            tdtype = t._model_dtype
+        if dtype is not None:
+            tdtype = dtype
+        geometries.append(TensorGeometry(shape=t.shape, dtype=tdtype))
+    return geometries
+
+def vram_aligned_size(tensor):
+    if isinstance(tensor, list):
+        return sum([vram_aligned_size(t) for t in tensor])
+
+    if isinstance(tensor, QuantizedTensor):
+        inner_tensors, _ = tensor.__tensor_flatten__()
+        return vram_aligned_size([ getattr(tensor, attr) for attr in inner_tensors ])
+
+    if tensor is None:
+        return 0
+
+    size = tensor.numel() * tensor.element_size()
+    aligment_req = 1024
+    return (size + aligment_req - 1) // aligment_req * aligment_req
+
+def interpret_gathered_like(tensors, gathered):
+    offset = 0
+    dest_views = []
+
+    if gathered.dim() != 1 or gathered.element_size() != 1:
+        raise ValueError(f"Buffer must be 1D and single-byte (got {gathered.dim()}D {gathered.dtype})")
+
+    for tensor in tensors:
+
+        if tensor is None:
+            dest_views.append(None)
+            continue
+
+        if isinstance(tensor, QuantizedTensor):
+            inner_tensors, qt_ctx = tensor.__tensor_flatten__()
+            templates = { attr: getattr(tensor, attr) for attr in inner_tensors }
+        else:
+            templates = { "data": tensor }
+
+        actuals = {}
+        for attr, template in templates.items():
+            size = template.numel() * template.element_size()
+            if offset + size > gathered.numel():
+                raise ValueError(f"Buffer too small: needs {offset + size} bytes, but only has {gathered.numel()}. ")
+            actuals[attr] = gathered[offset:offset+size].view(dtype=template.dtype).view(template.shape)
+            offset += vram_aligned_size(template)
+
+        if isinstance(tensor, QuantizedTensor):
+            dest_views.append(QuantizedTensor.__tensor_unflatten__(actuals, qt_ctx, 0, 0))
+        else:
+            dest_views.append(actuals["data"])
+
+    return dest_views
+
+aimdo_allocator = None

comfy/model_base.py

@@ -148,6 +148,8 @@ class BaseModel(torch.nn.Module):
         self.model_type = model_type
         self.model_sampling = model_sampling(model_config, model_type)
 
+        comfy.model_management.archive_model_dtypes(self.diffusion_model)
+
         self.adm_channels = unet_config.get("adm_in_channels", None)
         if self.adm_channels is None:
             self.adm_channels = 0
@@ -298,7 +300,7 @@ class BaseModel(torch.nn.Module):
 
         return out
 
-    def load_model_weights(self, sd, unet_prefix=""):
+    def load_model_weights(self, sd, unet_prefix="", assign=False):
         to_load = {}
         keys = list(sd.keys())
         for k in keys:
@@ -306,7 +308,7 @@ class BaseModel(torch.nn.Module):
                 to_load[k[len(unet_prefix):]] = sd.pop(k)
 
         to_load = self.model_config.process_unet_state_dict(to_load)
-        m, u = self.diffusion_model.load_state_dict(to_load, strict=False)
+        m, u = self.diffusion_model.load_state_dict(to_load, strict=False, assign=assign)
         if len(m) > 0:
             logging.warning("unet missing: {}".format(m))
 
@@ -321,7 +323,7 @@ class BaseModel(torch.nn.Module):
     def process_latent_out(self, latent):
         return self.latent_format.process_out(latent)
 
-    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+    def state_dict_for_saving(self, unet_state_dict, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
         extra_sds = []
         if clip_state_dict is not None:
             extra_sds.append(self.model_config.process_clip_state_dict_for_saving(clip_state_dict))
@@ -329,10 +331,7 @@ class BaseModel(torch.nn.Module):
             extra_sds.append(self.model_config.process_vae_state_dict_for_saving(vae_state_dict))
         if clip_vision_state_dict is not None:
             extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict))
-
-        unet_state_dict = self.diffusion_model.state_dict()
         unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
-
         if self.model_type == ModelType.V_PREDICTION:
             unet_state_dict["v_pred"] = torch.tensor([])
 
@@ -775,8 +774,8 @@ class StableAudio1(BaseModel):
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
 
-    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
-        sd = super().state_dict_for_saving(clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
+    def state_dict_for_saving(self, unet_state_dict, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+        sd = super().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)
         d = {"conditioner.conditioners.seconds_start.": self.seconds_start_embedder.state_dict(), "conditioner.conditioners.seconds_total.": self.seconds_total_embedder.state_dict()}
         for k in d:
             s = d[k]

comfy/model_management.py

@@ -26,6 +26,12 @@ import platform
 import weakref
 import gc
 import os
+from contextlib import nullcontext
+import comfy.utils
+import comfy.quant_ops
+
+import comfy_aimdo.torch
+import comfy_aimdo.model_vbar
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -592,7 +598,7 @@ def extra_reserved_memory():
 def minimum_inference_memory():
     return (1024 * 1024 * 1024) * 0.8 + extra_reserved_memory()
 
-def free_memory(memory_required, device, keep_loaded=[]):
+def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_required=0):
     cleanup_models_gc()
     unloaded_model = []
     can_unload = []
@@ -607,15 +613,22 @@ def free_memory(memory_required, device, keep_loaded=[]):
 
     for x in sorted(can_unload):
         i = x[-1]
-        memory_to_free = None
+        memory_to_free = 1e32
+        ram_to_free = 1e32
         if not DISABLE_SMART_MEMORY:
-            free_mem = get_free_memory(device)
-            if free_mem > memory_required:
-                break
-            memory_to_free = memory_required - free_mem
+            memory_to_free = memory_required - get_free_memory(device)
+            ram_to_free = ram_required - psutil.virtual_memory().available
+
+        if current_loaded_models[i].model.is_dynamic() and for_dynamic:
+            #don't actually unload dynamic models for the sake of other dynamic models
+            #as that works on-demand.
+            memory_required -= current_loaded_models[i].model.loaded_size()
+            continue
         logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
-        if current_loaded_models[i].model_unload(memory_to_free):
+        if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free):
             unloaded_model.append(i)
+        if ram_to_free > 0:
+            current_loaded_models[i].model.partially_unload_ram(ram_to_free)
 
     for i in sorted(unloaded_model, reverse=True):
         unloaded_models.append(current_loaded_models.pop(i))
@@ -650,7 +663,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
 
     models_to_load = []
 
+    free_for_dynamic=True
     for x in models:
+        if not x.is_dynamic():
+            free_for_dynamic = False
         loaded_model = LoadedModel(x)
         try:
             loaded_model_index = current_loaded_models.index(loaded_model)
@@ -676,19 +692,25 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             model_to_unload.model.detach(unpatch_all=False)
             model_to_unload.model_finalizer.detach()
 
+
     total_memory_required = {}
+    total_ram_required = {}
     for loaded_model in models_to_load:
         total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
+        #x2, one to make sure the OS can fit the model for loading in disk cache, and for us to do any pinning we
+        #want to do.
+        #FIXME: This should subtract off the to_load current pin consumption.
+        total_ram_required[loaded_model.device] = total_ram_required.get(loaded_model.device, 0) + loaded_model.model_memory() * 2
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
-            free_memory(total_memory_required[device] * 1.1 + extra_mem, device)
+            free_memory(total_memory_required[device] * 1.1 + extra_mem, device, for_dynamic=free_for_dynamic, ram_required=total_ram_required[device])
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
             free_mem = get_free_memory(device)
             if free_mem < minimum_memory_required:
-                models_l = free_memory(minimum_memory_required, device)
+                models_l = free_memory(minimum_memory_required, device, for_dynamic=free_for_dynamic)
                 logging.info("{} models unloaded.".format(len(models_l)))
 
     for loaded_model in models_to_load:
@@ -732,6 +754,9 @@ def loaded_models(only_currently_used=False):
 
 def cleanup_models_gc():
     do_gc = False
+
+    reset_cast_buffers()
+
     for i in range(len(current_loaded_models)):
         cur = current_loaded_models[i]
         if cur.is_dead():
@@ -749,6 +774,11 @@ def cleanup_models_gc():
                 logging.warning("WARNING, memory leak with model {}. Please make sure it is not being referenced from somewhere.".format(cur.real_model().__class__.__name__))
 
 
+def archive_model_dtypes(model):
+    for name, module in model.named_modules():
+        for param_name, param in module.named_parameters(recurse=False):
+            setattr(module, f"{param_name}_comfy_model_dtype", param.dtype)
+
 
 def cleanup_models():
     to_delete = []
@@ -1051,6 +1081,49 @@ def current_stream(device):
         return None
 
 stream_counters = {}
+
+STREAM_CAST_BUFFERS = {}
+LARGEST_CASTED_WEIGHT = (None, 0)
+
+def get_cast_buffer(offload_stream, device, size, ref):
+    global LARGEST_CASTED_WEIGHT
+
+    if offload_stream is not None:
+        wf_context = offload_stream
+        if hasattr(wf_context, "as_context"):
+            wf_context = wf_context.as_context(offload_stream)
+    else:
+        wf_context = nullcontext()
+
+    cast_buffer = STREAM_CAST_BUFFERS.get(offload_stream, None)
+    if cast_buffer is None or cast_buffer.numel() < size:
+        if ref is LARGEST_CASTED_WEIGHT[0]:
+            #If there is one giant weight we do not want both streams to
+            #allocate a buffer for it. It's up to the caster to get the other
+            #offload stream in this corner case
+            return None
+        if cast_buffer is not None and cast_buffer.numel() > 50 * (1024 ** 2):
+            #I want my wrongly sized 50MB+ of VRAM back from the caching allocator right now
+            torch.cuda.synchronize()
+            del STREAM_CAST_BUFFERS[offload_stream]
+            del cast_buffer
+            torch.cuda.empty_cache()
+        with wf_context:
+            cast_buffer = torch.empty((size), dtype=torch.int8, device=device)
+            STREAM_CAST_BUFFERS[offload_stream] = cast_buffer
+
+        if  size > LARGEST_CASTED_WEIGHT[1]:
+            LARGEST_CASTED_WEIGHT = (ref, size)
+
+    return cast_buffer
+
+def reset_cast_buffers():
+    global LARGEST_CASTED_WEIGHT
+    LARGEST_CASTED_WEIGHT = (None, 0)
+    torch.cuda.synchronize()
+    STREAM_CAST_BUFFERS.clear()
+    torch.cuda.empty_cache()
+
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
     if NUM_STREAMS == 0:
@@ -1093,7 +1166,52 @@ def sync_stream(device, stream):
         return
     current_stream(device).wait_stream(stream)
 
-def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None):
+
+def cast_to_gathered(tensors, r, non_blocking=False, stream=None):
+    wf_context = nullcontext()
+    if stream is not None:
+       wf_context = stream
+       if hasattr(wf_context, "as_context"):
+           wf_context = wf_context.as_context(stream)
+
+    dest_views = comfy.memory_management.interpret_gathered_like(tensors, r)
+    with wf_context:
+        for tensor in tensors:
+            dest_view = dest_views.pop(0)
+            if tensor is None:
+                continue
+            dest_view.copy_(tensor, non_blocking=non_blocking)
+
+
+def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None, r=None):
+    if hasattr(weight, "_v"):
+        #Unexpected usage patterns. There is no reason these don't work but they
+        #have no testing and no callers do this.
+        assert r is None
+        assert stream is None
+
+        r = torch.empty_like(weight, dtype=weight._model_dtype, device=device)
+
+        signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
+        if signature is not None:
+            raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device)
+            v_tensor = comfy.memory_management.interpret_gathered_like([r], raw_tensor)[0]
+
+        if comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
+            #always take a deep copy even if _v is good, as we have no reasonable point to unpin
+            #a non comfy weight
+            r.copy_(v_tensor)
+            comfy_aimdo.model_vbar.vbar_unpin(weight._v)
+            return r
+
+        r.copy_(weight, non_blocking=non_blocking)
+
+        if signature is not None:
+            v_tensor.copy_(r)
+            comfy_aimdo.model_vbar.vbar_unpin(weight._v)
+
+        return r
+
     if device is None or weight.device == device:
         if not copy:
             if dtype is None or weight.dtype == dtype:
@@ -1112,10 +1230,12 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
         if hasattr(wf_context, "as_context"):
             wf_context = wf_context.as_context(stream)
         with wf_context:
-            r = torch.empty_like(weight, dtype=dtype, device=device)
+            if r is None:
+                r = torch.empty_like(weight, dtype=dtype, device=device)
             r.copy_(weight, non_blocking=non_blocking)
     else:
-        r = torch.empty_like(weight, dtype=dtype, device=device)
+        if r is None:
+            r = torch.empty_like(weight, dtype=dtype, device=device)
         r.copy_(weight, non_blocking=non_blocking)
     return r
 
@@ -1135,7 +1255,7 @@ if not args.disable_pinned_memory:
             MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95
         logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024)))
 
-PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"])
+PINNING_ALLOWED_TYPES = set(["Tensor", "Parameter", "QuantizedTensor"])
 
 def discard_cuda_async_error():
     try:
@@ -1557,6 +1677,7 @@ def soft_empty_cache(force=False):
     elif is_mlu():
         torch.mlu.empty_cache()
     elif torch.cuda.is_available():
+        torch.cuda.synchronize()
         torch.cuda.empty_cache()
         torch.cuda.ipc_collect()
 

comfy/model_patcher.py

@@ -38,19 +38,7 @@ 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()
@@ -212,6 +200,27 @@ 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
@@ -269,6 +278,9 @@ class ModelPatcher:
         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
@@ -284,6 +296,9 @@ class ModelPatcher:
     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.model_size(), weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
@@ -611,14 +626,14 @@ 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)
@@ -631,13 +646,15 @@ class ModelPatcher:
 
         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 pin_weight_to_device(self, key):
         weight, set_func, convert_func = get_key_weight(self.model, key)
@@ -654,7 +671,7 @@ class ModelPatcher:
         for key in list(self.pinned):
             self.unpin_weight(key)
 
-    def _load_list(self):
+    def _load_list(self, prio_comfy_cast_weights=False):
         loading = []
         for n, m in self.model.named_modules():
             params = []
@@ -681,7 +698,8 @@ class ModelPatcher:
                         return 0
                     module_offload_mem += check_module_offload_mem("{}.weight".format(n))
                     module_offload_mem += check_module_offload_mem("{}.bias".format(n))
-                loading.append((module_offload_mem, module_mem, n, m, params))
+                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):
@@ -984,6 +1002,9 @@ 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)
@@ -1317,10 +1338,10 @@ 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
@@ -1355,7 +1376,251 @@ 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)
+
     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):
+        pass
+
+    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 have way more tools for acceleration on comfy weight offloading, so always
+            #prioritize the non-comfy weights (note the order reverse).
+            loading = self._load_list(prio_comfy_cast_weights=True)
+            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
+                    if dirty:
+                        comfy.pinned_memory.unpin_memory(item)
+
+                def setup_param(self, m, n, param_key):
+                    nonlocal num_patches
+                    key = "{}.{}".format(n, param_key)
+
+                    weight_function = []
+
+                    weight, _, _ = get_key_weight(self.model, key)
+                    if weight is None:
+                        return 0
+                    if key in self.patches:
+                        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", weight.dtype)
+                        weight._model_dtype = model_dtype
+                        geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype)
+                    return comfy.memory_management.vram_aligned_size(geometry)
+
+                if hasattr(m, "comfy_cast_weights"):
+                    m.comfy_cast_weights = True
+                    m.pin_failed = False
+                    m.seed_key = n
+                    set_dirty(m, dirty)
+
+                    v_weight_size = 0
+                    v_weight_size += setup_param(self, m, n, "weight")
+                    v_weight_size += setup_param(self, m, n, "bias")
+
+                    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 = "{}.{}".format(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", 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
+
+            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)
+        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)
+
+    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

comfy/ops.py

@@ -19,10 +19,16 @@
 import torch
 import logging
 import comfy.model_management
-from comfy.cli_args import args, PerformanceFeature
+from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
 import comfy.float
 import comfy.rmsnorm
 import json
+import comfy.memory_management
+import comfy.pinned_memory
+import comfy.utils
+
+import comfy_aimdo.model_vbar
+import comfy_aimdo.torch
 
 def run_every_op():
     if torch.compiler.is_compiling():
@@ -72,7 +78,128 @@ def cast_to_input(weight, input, non_blocking=False, copy=True):
     return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
 
-def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False):
+def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype):
+    offload_stream = None
+    xfer_dest = None
+    cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
+
+    signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
+    if signature is not None:
+        xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
+    resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
+
+    if not resident:
+        cast_dest = None
+
+        xfer_source = [ s.weight, s.bias ]
+
+        pin = comfy.pinned_memory.get_pin(s)
+        if pin is not None:
+            xfer_source = [ pin ]
+            resident = True #If pinned data exists, it always has LowVram already applied
+        else:
+            for data, geometry in zip([ s.weight, s.bias ], cast_geometry):
+                if data is None:
+                    continue
+                if data.dtype != geometry.dtype:
+                    cast_dest = xfer_dest
+                    if cast_dest is None:
+                        cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device)
+                    xfer_dest = None
+                    break
+
+        dest_size = comfy.memory_management.vram_aligned_size(xfer_source)
+        offload_stream = comfy.model_management.get_offload_stream(device)
+        if xfer_dest is None and offload_stream is not None:
+                xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
+                if xfer_dest is None:
+                    offload_stream = comfy.model_management.get_offload_stream(device)
+                    xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
+        if xfer_dest is None:
+            xfer_dest = torch.empty((dest_size,), dtype=torch.uint8, device=device)
+            offload_stream = None
+
+        #send it over
+        comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=offload_stream)
+        comfy.model_management.sync_stream(device, offload_stream)
+
+        if cast_dest is not None:
+            for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like(xfer_source, xfer_dest),
+                                           comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)):
+                if post_cast is not None:
+                    post_cast.copy_(pre_cast)
+            xfer_dest = cast_dest
+
+    pin = None
+    if signature is not None:
+        #If we are able to increase our load level (e.g. user reduces resolution or batch number)
+        #reclaim the pin previously used for offload.
+        comfy.pinned_memory.unpin_memory(s)
+    elif not resident:
+        #prepare a new pin
+        assert comfy.pinned_memory.get_pin(s) is None
+        comfy.pinned_memory.pin_memory(s)
+        pin = comfy.pinned_memory.get_pin(s)
+
+    params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest)
+    weight = params[0]
+    bias = params[1]
+
+    def post_cast(s, param_key, x, dtype, resident, update_weight):
+        lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+        fns = getattr(s, param_key + "_function", [])
+
+        orig = x
+
+        def to_dequant(tensor, dtype):
+            tensor = tensor.to(dtype=dtype)
+            if isinstance(tensor, QuantizedTensor):
+                tensor = tensor.dequantize()
+            return tensor
+
+        if orig.dtype != dtype or len(fns) > 0:
+            x = to_dequant(x, dtype)
+        if not resident and lowvram_fn is not None:
+            x = to_dequant(x, dtype if compute_dtype is None else compute_dtype)
+            #FIXME: this is not accurate, we need to be sensitive to the compute dtype
+            x = lowvram_fn(x)
+            if (isinstance(orig, QuantizedTensor) and
+                (orig.dtype == dtype and len(fns) == 0 or update_weight)):
+                seed = comfy.utils.string_to_seed(s.seed_key)
+                y = QuantizedTensor.from_float(x, s.layout_type, scale="recalculate", stochastic_rounding=seed)
+                if orig.dtype == dtype and len(fns) == 0:
+                    #The layer actually wants our freshly saved QT
+                    x = y
+            else:
+                y = x
+            if update_weight:
+                orig.copy_(y)
+        for f in fns:
+            x = f(x)
+        return x
+
+    update_weight = signature is not None or pin is not None
+
+    weight = post_cast(s, "weight", weight, dtype, resident, update_weight)
+    if s.bias is not None:
+        bias = post_cast(s, "bias", bias, bias_dtype, resident, update_weight)
+    s._v_signature=signature
+
+    if pin is not None:
+        xfer_dest = comfy.memory_management.interpret_gathered_like([ pin ], xfer_dest)[0]
+        #FIXME: This might be the wrong thing to do. Some reading suggests the DMA engine
+        #is posted writes and the compute stream could just fire and forget here. That
+        #would save this sync and some stalling on the offload stream that is better off
+        #running ahead to the next layer to read.
+        if offload_stream is not None:
+            offload_stream.wait_stream(comfy.model_management.current_stream(device))
+        comfy.model_management.cast_to(xfer_dest, device=pin.device, non_blocking=non_blocking, stream=offload_stream, r=pin)
+
+    #FIXME: weird offload return protocol
+    return weight, bias, (offload_stream, device if signature is not None else None, None)
+
+
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None):
     # NOTE: offloadable=False is a a legacy and if you are a custom node author reading this please pass
     # offloadable=True and call uncast_bias_weight() after your last usage of the weight/bias. This
     # will add async-offload support to your cast and improve performance.
@@ -87,22 +214,38 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         if device is None:
             device = input.device
 
+    non_blocking = comfy.model_management.device_supports_non_blocking(device)
+
+    if hasattr(s, "_v"):
+        return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype)
+
     if offloadable and (device != s.weight.device or
                         (s.bias is not None and device != s.bias.device)):
         offload_stream = comfy.model_management.get_offload_stream(device)
     else:
         offload_stream = None
 
-    non_blocking = comfy.model_management.device_supports_non_blocking(device)
+    bias = None
+    weight = None
+
+    if offload_stream is not None and not args.cuda_malloc:
+        cast_buffer_size = comfy.memory_management.vram_aligned_size([ s.weight, s.bias ])
+        cast_buffer = comfy.model_management.get_cast_buffer(offload_stream, device, cast_buffer_size, s)
+        #The streams can be uneven in buffer capability and reject us. Retry to get the other stream
+        if cast_buffer is None:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+            cast_buffer = comfy.model_management.get_cast_buffer(offload_stream, device, cast_buffer_size, s)
+        params = comfy.memory_management.interpret_gathered_like([ s.weight, s.bias ], cast_buffer)
+        weight = params[0]
+        bias = params[1]
 
     weight_has_function = len(s.weight_function) > 0
     bias_has_function = len(s.bias_function) > 0
 
-    weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream)
+    weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream, r=weight)
 
-    bias = None
     if s.bias is not None:
-        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream)
+        bias = comfy.model_management.cast_to(s.bias, None, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream, r=bias)
 
     comfy.model_management.sync_stream(device, offload_stream)
 
@@ -110,6 +253,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     weight_a = weight
 
     if s.bias is not None:
+        bias = bias.to(dtype=bias_dtype)
         for f in s.bias_function:
             bias = f(bias)
 
@@ -131,14 +275,20 @@ def uncast_bias_weight(s, weight, bias, offload_stream):
     if offload_stream is None:
         return
     os, weight_a, bias_a = offload_stream
+    device=None
+    #FIXME: This is not good RTTI
+    if not isinstance(weight_a, torch.Tensor):
+        comfy_aimdo.model_vbar.vbar_unpin(s._v)
+        device = weight_a
     if os is None:
         return
-    if weight_a is not None:
-        device = weight_a.device
-    else:
-        if bias_a is None:
-            return
-        device = bias_a.device
+    if device is None:
+        if weight_a is not None:
+            device = weight_a.device
+        else:
+            if bias_a is None:
+                return
+            device = bias_a.device
     os.wait_stream(comfy.model_management.current_stream(device))
 
 
@@ -149,6 +299,57 @@ class CastWeightBiasOp:
 
 class disable_weight_init:
     class Linear(torch.nn.Linear, CastWeightBiasOp):
+
+        def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
+            if not comfy.model_management.WINDOWS or not enables_dynamic_vram():
+                super().__init__(in_features, out_features, bias, device, dtype)
+                return
+
+            # Issue is with `torch.empty` still reserving the full memory for the layer.
+            # Windows doesn't over-commit memory so without this, We are momentarily commit
+            # charged for the weight even though we might zero-copy it when we load the
+            # state dict. If the commit charge exceeds the ceiling we can destabilize the
+            # system.
+            torch.nn.Module.__init__(self)
+            self.in_features = in_features
+            self.out_features = out_features
+            self.weight = None
+            self.bias = None
+            self.comfy_need_lazy_init_bias=bias
+            self.weight_comfy_model_dtype = dtype
+            self.bias_comfy_model_dtype = dtype
+
+        def _load_from_state_dict(self, state_dict, prefix, local_metadata,
+                                strict, missing_keys, unexpected_keys, error_msgs):
+
+            if not comfy.model_management.WINDOWS or not enables_dynamic_vram():
+                return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
+                                                     missing_keys, unexpected_keys, error_msgs)
+            assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False)
+            prefix_len = len(prefix)
+            for k,v in state_dict.items():
+                if k[prefix_len:] == "weight":
+                    if not assign_to_params_buffers:
+                        v = v.clone()
+                    self.weight = torch.nn.Parameter(v, requires_grad=False)
+                elif k[prefix_len:] == "bias" and v is not None:
+                    if not assign_to_params_buffers:
+                        v = v.clone()
+                    self.bias = torch.nn.Parameter(v, requires_grad=False)
+                else:
+                    unexpected_keys.append(k)
+
+            #Reconcile default construction of the weight if its missing.
+            if self.weight is None:
+                v = torch.zeros(self.in_features, self.out_features)
+                self.weight = torch.nn.Parameter(v, requires_grad=False)
+                missing_keys.append(prefix+"weight")
+            if self.bias is None and self.comfy_need_lazy_init_bias:
+                v = torch.zeros(self.out_features,)
+                self.bias = torch.nn.Parameter(v, requires_grad=False)
+                missing_keys.append(prefix+"bias")
+
+
         def reset_parameters(self):
             return None
 
@@ -653,8 +854,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
             def _forward(self, input, weight, bias):
                 return torch.nn.functional.linear(input, weight, bias)
 
-            def forward_comfy_cast_weights(self, input):
-                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            def forward_comfy_cast_weights(self, input, compute_dtype=None):
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True, compute_dtype=compute_dtype)
                 x = self._forward(input, weight, bias)
                 uncast_bias_weight(self, weight, bias, offload_stream)
                 return x
@@ -664,6 +865,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 
                 input_shape = input.shape
                 reshaped_3d = False
+                #If cast needs to apply lora, it should be done in the compute dtype
+                compute_dtype = input.dtype
 
                 if (getattr(self, 'layout_type', None) is not None and
                     not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
@@ -682,7 +885,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                             scale = comfy.model_management.cast_to_device(scale, input.device, None)
                         input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale)
 
-                output = self.forward_comfy_cast_weights(input)
+
+                output = self.forward_comfy_cast_weights(input, compute_dtype)
 
                 # Reshape output back to 3D if input was 3D
                 if reshaped_3d:

comfy/pinned_memory.py

@@ -0,0 +1,30 @@
+import torch
+import comfy.model_management
+import comfy.memory_management
+
+from comfy.cli_args import args
+
+def get_pin(module):
+    return getattr(module, "_pin", None)
+
+def pin_memory(module):
+    if module.pin_failed or args.disable_pinned_memory or get_pin(module) is not None:
+        return
+    #FIXME: This is a RAM cache trigger event
+    params = comfy.memory_management.tensors_to_geometries([ module.weight, module.bias ])
+    size = comfy.memory_management.vram_aligned_size(params)
+    pin = torch.empty((size,), dtype=torch.uint8)
+    if comfy.model_management.pin_memory(pin):
+        module._pin = pin
+    else:
+        module.pin_failed = True
+        return False
+    return True
+
+def unpin_memory(module):
+    if get_pin(module) is None:
+        return 0
+    size = module._pin.numel() * module._pin.element_size()
+    comfy.model_management.unpin_memory(module._pin)
+    del module._pin
+    return size

comfy/samplers.py

@@ -9,7 +9,6 @@ if TYPE_CHECKING:
 import torch
 from functools import partial
 import collections
-from comfy import model_management
 import math
 import logging
 import comfy.sampler_helpers
@@ -260,7 +259,7 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens
             to_batch_temp.reverse()
             to_batch = to_batch_temp[:1]
 
-            free_memory = model_management.get_free_memory(x_in.device)
+            free_memory = model.current_patcher.get_free_memory(x_in.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:]

comfy/sd.py

@@ -127,8 +127,10 @@ class CLIP:
                     self.cond_stage_model.to(offload_device)
                     logging.warning("Had to shift TE back.")
 
+        model_management.archive_model_dtypes(self.cond_stage_model)
+
         self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
-        self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
         #Match torch.float32 hardcode upcast in TE implemention
         self.patcher.set_model_compute_dtype(torch.float32)
         self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram
@@ -288,8 +290,18 @@ class CLIP:
 
     def load_sd(self, sd, full_model=False):
         if full_model:
-            return self.cond_stage_model.load_state_dict(sd, strict=False)
+            return self.cond_stage_model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
         else:
+            can_assign = self.patcher.is_dynamic()
+            self.cond_stage_model.can_assign_sd = can_assign
+
+            # The CLIP models are a pretty complex web of wrappers and its
+            # a bit of an API change to plumb this all the way through.
+            # So spray paint the model with this flag that the loading
+            # nn.Module can then inspect for itself.
+            for m in self.cond_stage_model.modules():
+                m.can_assign_sd = can_assign
+
             return self.cond_stage_model.load_sd(sd)
 
     def get_sd(self):
@@ -665,12 +677,7 @@ class VAE:
             self.first_stage_model = AutoencoderKL(**(config['params']))
         self.first_stage_model = self.first_stage_model.eval()
 
-        m, u = self.first_stage_model.load_state_dict(sd, strict=False)
-        if len(m) > 0:
-            logging.warning("Missing VAE keys {}".format(m))
-
-        if len(u) > 0:
-            logging.debug("Leftover VAE keys {}".format(u))
+        model_management.archive_model_dtypes(self.first_stage_model)
 
         if device is None:
             device = model_management.vae_device()
@@ -682,7 +689,18 @@ class VAE:
         self.first_stage_model.to(self.vae_dtype)
         self.output_device = model_management.intermediate_device()
 
-        self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device)
+        mp = comfy.model_patcher.CoreModelPatcher
+        if self.disable_offload:
+            mp = comfy.model_patcher.ModelPatcher
+        self.patcher = mp(self.first_stage_model, load_device=self.device, offload_device=offload_device)
+
+        m, u = self.first_stage_model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
+        if len(m) > 0:
+            logging.warning("Missing VAE keys {}".format(m))
+
+        if len(u) > 0:
+            logging.debug("Leftover VAE keys {}".format(u))
+
         logging.info("VAE load device: {}, offload device: {}, dtype: {}".format(self.device, offload_device, self.vae_dtype))
         self.model_size()
 
@@ -797,7 +815,7 @@ class VAE:
         try:
             memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
             model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
-            free_memory = model_management.get_free_memory(self.device)
+            free_memory = self.patcher.get_free_memory(self.device)
             batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
 
@@ -816,6 +834,7 @@ class VAE:
             do_tile = True
 
         if do_tile:
+            torch.cuda.empty_cache()
             dims = samples_in.ndim - 2
             if dims == 1 or self.extra_1d_channel is not None:
                 pixel_samples = self.decode_tiled_1d(samples_in)
@@ -871,7 +890,7 @@ class VAE:
         try:
             memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
             model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
-            free_memory = model_management.get_free_memory(self.device)
+            free_memory = self.patcher.get_free_memory(self.device)
             batch_number = int(free_memory / max(1, memory_used))
             batch_number = max(1, batch_number)
             samples = None
@@ -891,6 +910,7 @@ class VAE:
             do_tile = True
 
         if do_tile:
+            torch.cuda.empty_cache()
             if self.latent_dim == 3:
                 tile = 256
                 overlap = tile // 4
@@ -1326,7 +1346,7 @@ def load_gligen(ckpt_path):
     model = gligen.load_gligen(data)
     if model_management.should_use_fp16():
         model = model.half()
-    return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
+    return comfy.model_patcher.CoreModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
 
 def model_detection_error_hint(path, state_dict):
     filename = os.path.basename(path)
@@ -1414,7 +1434,8 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     if output_model:
         inital_load_device = model_management.unet_inital_load_device(parameters, unet_dtype)
         model = model_config.get_model(sd, diffusion_model_prefix, device=inital_load_device)
-        model.load_model_weights(sd, diffusion_model_prefix)
+        model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())
+        model.load_model_weights(sd, diffusion_model_prefix, assign=model_patcher.is_dynamic())
 
     if output_vae:
         vae_sd = comfy.utils.state_dict_prefix_replace(sd, {k: "" for k in model_config.vae_key_prefix}, filter_keys=True)
@@ -1457,7 +1478,6 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
         logging.debug("left over keys: {}".format(left_over))
 
     if output_model:
-        model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())
         if inital_load_device != torch.device("cpu"):
             logging.info("loaded diffusion model directly to GPU")
             model_management.load_models_gpu([model_patcher], force_full_load=True)
@@ -1549,13 +1569,14 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
         model_config.optimizations["fp8"] = True
 
     model = model_config.get_model(new_sd, "")
-    model = model.to(offload_device)
-    model.load_model_weights(new_sd, "")
+    model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=load_device, offload_device=offload_device)
+    if not model_management.is_device_cpu(offload_device):
+        model.to(offload_device)
+    model.load_model_weights(new_sd, "", assign=model_patcher.is_dynamic())
     left_over = sd.keys()
     if len(left_over) > 0:
         logging.info("left over keys in diffusion model: {}".format(left_over))
-    return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device)
-
+    return model_patcher
 
 def load_diffusion_model(unet_path, model_options={}):
     sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True)
@@ -1586,9 +1607,9 @@ def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, m
     if metadata is None:
         metadata = {}
 
-    model_management.load_models_gpu(load_models, force_patch_weights=True)
+    model_management.load_models_gpu(load_models)
     clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None
-    sd = model.model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd)
+    sd = model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd)
     for k in extra_keys:
         sd[k] = extra_keys[k]
 

comfy/sd1_clip.py

@@ -297,7 +297,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         return self(tokens)
 
     def load_sd(self, sd):
-        return self.transformer.load_state_dict(sd, strict=False)
+        return self.transformer.load_state_dict(sd, strict=False, assign=getattr(self, "can_assign_sd", False))
 
 def parse_parentheses(string):
     result = []

comfy/text_encoders/lt.py

@@ -118,7 +118,7 @@ class LTXAVTEModel(torch.nn.Module):
             sdo = comfy.utils.state_dict_prefix_replace(sd, {"text_embedding_projection.aggregate_embed.weight": "text_embedding_projection.weight", "model.diffusion_model.video_embeddings_connector.": "video_embeddings_connector.", "model.diffusion_model.audio_embeddings_connector.": "audio_embeddings_connector."}, filter_keys=True)
             if len(sdo) == 0:
                 sdo = sd
-            missing, unexpected = self.load_state_dict(sdo, strict=False)
+            missing, unexpected = self.load_state_dict(sdo, strict=False, assign=getattr(self, "can_assign_sd", False))
             missing = [k for k in missing if not k.startswith("gemma3_12b.")] # filter out keys that belong to the main gemma model
             return (missing, unexpected)
 

comfy/utils.py

@@ -28,9 +28,12 @@ import logging
 import itertools
 from torch.nn.functional import interpolate
 from einops import rearrange
-from comfy.cli_args import args
+from comfy.cli_args import args, enables_dynamic_vram
 import json
 import time
+import mmap
+
+import packaging
 
 MMAP_TORCH_FILES = args.mmap_torch_files
 DISABLE_MMAP = args.disable_mmap
@@ -56,21 +59,73 @@ if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in
 else:
     logging.warning("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended as older versions of pytorch are no longer supported.")
 
+# Current as of safetensors 0.7.0
+_TYPES = {
+    "F64": torch.float64,
+    "F32": torch.float32,
+    "F16": torch.float16,
+    "BF16": torch.bfloat16,
+    "I64": torch.int64,
+    "I32": torch.int32,
+    "I16": torch.int16,
+    "I8": torch.int8,
+    "U8": torch.uint8,
+    "BOOL": torch.bool,
+    "F8_E4M3": torch.float8_e4m3fn,
+    "F8_E5M2": torch.float8_e5m2,
+    "C64": torch.complex64,
+}
+if packaging.version.Version(torch.__version__) >= packaging.version.Version("2.3.0"):
+    _TYPES.update(
+        {
+            "U64": torch.uint64,
+            "U32": torch.uint32,
+            "U16": torch.uint16,
+        }
+    )
+
+def load_safetensors(ckpt):
+    f = open(ckpt, "rb")
+    mapping = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
+
+    header_size = struct.unpack("<Q", mapping[:8])[0]
+    header = json.loads(mapping[8:8+header_size].decode("utf-8"))
+
+    data_area = torch.frombuffer(mapping, dtype=torch.uint8)[8 + header_size:]
+
+    sd = {}
+    for name, info in header.items():
+        if name == "__metadata__":
+            continue
+
+        start, end = info["data_offsets"]
+        sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"])
+
+    return sd, header.get("__metadata__", {}),
+
+
 def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
     if device is None:
         device = torch.device("cpu")
+    else:
+        assert False
     metadata = None
     if ckpt.lower().endswith(".safetensors") or ckpt.lower().endswith(".sft"):
         try:
-            with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
-                sd = {}
-                for k in f.keys():
-                    tensor = f.get_tensor(k)
-                    if DISABLE_MMAP:  # TODO: Not sure if this is the best way to bypass the mmap issues
-                        tensor = tensor.to(device=device, copy=True)
-                    sd[k] = tensor
-                if return_metadata:
-                    metadata = f.metadata()
+            if enables_dynamic_vram():
+                sd, metadata = load_safetensors(ckpt)
+                if not return_metadata:
+                    metadata = None
+            else:
+                with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
+                    sd = {}
+                    for k in f.keys():
+                        tensor = f.get_tensor(k)
+                        if DISABLE_MMAP:  # TODO: Not sure if this is the best way to bypass the mmap issues
+                            tensor = tensor.to(device=device, copy=True)
+                        sd[k] = tensor
+                    if return_metadata:
+                        metadata = f.metadata()
         except Exception as e:
             if len(e.args) > 0:
                 message = e.args[0]
@@ -1308,3 +1363,16 @@ def convert_old_quants(state_dict, model_prefix="", metadata={}):
             state_dict["{}.comfy_quant".format(k)] = torch.tensor(list(json.dumps(v).encode('utf-8')), dtype=torch.uint8)
 
     return state_dict, metadata
+
+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

comfy_extras/nodes_model_patch.py

@@ -258,9 +258,9 @@ class ModelPatchLoader:
                         config['broken'] = True
             model = comfy.ldm.lumina.controlnet.ZImage_Control(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast, **config)
 
-        model.load_state_dict(sd)
-        model = comfy.model_patcher.ModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device())
-        return (model,)
+        model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device())
+        model.load_state_dict(sd, assign=model_patcher.is_dynamic())
+        return (model_patcher,)
 
 
 class DiffSynthCnetPatch:

cuda_malloc.py

@@ -1,8 +1,10 @@
 import os
 import importlib.util
-from comfy.cli_args import args, PerformanceFeature
+from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
 import subprocess
 
+import comfy_aimdo.control
+
 #Can't use pytorch to get the GPU names because the cuda malloc has to be set before the first import.
 def get_gpu_names():
     if os.name == 'nt':
@@ -85,8 +87,14 @@ if not args.cuda_malloc:
     except:
         pass
 
+if enables_dynamic_vram() and comfy_aimdo.control.init():
+    args.cuda_malloc = False
+    os.environ['PYTORCH_CUDA_ALLOC_CONF'] = ""
 
-if args.cuda_malloc and not args.disable_cuda_malloc:
+if args.disable_cuda_malloc:
+    args.cuda_malloc = False
+
+if args.cuda_malloc:
     env_var = os.environ.get('PYTORCH_CUDA_ALLOC_CONF', None)
     if env_var is None:
         env_var = "backend:cudaMallocAsync"

execution.py

@@ -9,9 +9,11 @@ import traceback
 from enum import Enum
 from typing import List, Literal, NamedTuple, Optional, Union
 import asyncio
+from contextlib import nullcontext
 
 import torch
 
+import comfy.memory_management
 import comfy.model_management
 from latent_preview import set_preview_method
 import nodes
@@ -515,7 +517,20 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
             def pre_execute_cb(call_index):
                 # TODO - How to handle this with async functions without contextvars (which requires Python 3.12)?
                 GraphBuilder.set_default_prefix(unique_id, call_index, 0)
-            output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
+
+            #Do comfy_aimdo mempool chunking here on the per-node level. Multi-model workflows
+            #will cause all sorts of incompatible memory shapes to fragment the pytorch alloc
+            #that we just want to cull out each model run.
+            allocator = comfy.memory_management.aimdo_allocator
+            with nullcontext() if allocator is None else torch.cuda.use_mem_pool(torch.cuda.MemPool(allocator.allocator())):
+                output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
+                torch.cuda.synchronize()
+            if allocator is not None:
+                #FIXME: this is probably a little zealous
+                # Torch code comments says some stuff about not actually freeing tensors on mempool
+                #context release. Explicitly garbage collect now.
+                torch.cuda.empty_cache()
+
             if has_pending_tasks:
                 pending_async_nodes[unique_id] = output_data
                 unblock = execution_list.add_external_block(unique_id)

main.py

@@ -5,7 +5,7 @@ import os
 import importlib.util
 import folder_paths
 import time
-from comfy.cli_args import args
+from comfy.cli_args import args, enables_dynamic_vram
 from app.logger import setup_logger
 from app.assets.scanner import seed_assets
 import itertools
@@ -173,6 +173,7 @@ import gc
 if 'torch' in sys.modules:
     logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.")
 
+
 import comfy.utils
 
 import execution
@@ -184,6 +185,33 @@ import comfyui_version
 import app.logger
 import hook_breaker_ac10a0
 
+import comfy.memory_management
+import comfy.model_patcher
+
+import comfy_aimdo.control
+import comfy_aimdo.torch
+
+if enables_dynamic_vram():
+    if comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index):
+        if args.verbose == 'DEBUG':
+            comfy_aimdo.control.set_log_debug()
+        elif args.verbose == 'CRITICAL':
+            comfy_aimdo.control.set_log_critical()
+        elif args.verbose == 'ERROR':
+            comfy_aimdo.control.set_log_error()
+        elif args.verbose == 'WARNING':
+            comfy_aimdo.control.set_log_warning()
+        else: #INFO
+            comfy_aimdo.control.set_log_info()
+
+        comfy.model_patcher.CoreModelPatcher = comfy.model_patcher.ModelPatcherDynamic
+        comfy.memory_management.aimdo_allocator = comfy_aimdo.torch.get_torch_allocator()
+        logging.info("DynamicVRAM support detected and enabled")
+    else:
+        logging.info("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
+        comfy.memory_management.aimdo_allocator = None
+
+
 def cuda_malloc_warning():
     device = comfy.model_management.get_torch_device()
     device_name = comfy.model_management.get_torch_device_name(device)

requirements.txt

@@ -22,6 +22,7 @@ alembic
 SQLAlchemy
 av>=14.2.0
 comfy-kitchen>=0.2.7
+comfy-aimdo>=0.1.2
 
 #non essential dependencies:
 kornia>=0.7.1