Merge 49da9c325a into 223364743c

This is using a different layers weight with .to(). Change it to use
the ops caster if the original layer is a comfy weight so that it picks
up dynamic_vram and async_offload functionality in full.

Co-authored-by: Rattus <rattus128@gmail.com>

.dockerignore

@@ -0,0 +1,6 @@
+__pycache__/
+*.py[cod]
+input
+models
+notebooks
+output

Dockerfile

@@ -0,0 +1,70 @@
+# 3.10.11-bullseye,  has python, git, but relatively small (<250MB)
+ARG BASE_IMAGE="python@sha256:88fb365ea5d52ec8f5799f40a4742b9fb3c91dac92f7048eabaae194a25ccc28"
+ARG GPU_MAKE="nvidia"
+ARG UID=1000
+ARG GID=1000
+
+FROM ${BASE_IMAGE}
+ARG GPU_MAKE
+ARG UID
+ARG GID
+
+SHELL [ "/bin/bash", "-uec"]
+RUN \
+  --mount=target=/var/lib/apt/lists,type=cache,sharing=locked \
+  --mount=target=/var/cache/apt,type=cache,sharing=locked \
+<<'EOF'
+  apt-get update
+  apt-get install -yq git-lfs
+  echo "comfyui" >> /etc/hostname
+EOF
+
+# run instructions as user
+USER ${UID}:${GID}
+# run python from future venv
+ENV PATH="/app/venv/bin:${PATH}"
+# copy context to obvious location
+COPY --chown=${UID}:${GID} ./ /app
+# create cache directory *with user permissions*
+WORKDIR /app/.cache
+# default to app directory
+WORKDIR /app
+# set pip cache location
+ENV XDG_CACHE_HOME="/app/.cache/pip"
+# run with mounted cache
+RUN --mount=type=cache,target=/app/.cache,uid=${UID},gid=${GID} <<'EOF'
+  mkdir -p /app/.cache/transformers
+  # choose package index based on chosen hardware
+  if [ "${GPU_MAKE}" = "nvidia" ]; then
+    EXTRA_INDEX_URL="https://download.pytorch.org/whl/cu118"
+    EXTRAS="xformers"
+  elif [ "${GPU_MAKE}" = "amd" ]; then
+    EXTRA_INDEX_URL="https://download.pytorch.org/whl/rocm5.4.2"
+    EXTRAS=""
+  elif [ "${GPU_MAKE}" = "cpu" ]; then
+    EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
+    EXTRAS=""
+  else
+    echo "Unknown GPU_MAKE provided as docker build arg."
+    exit 2
+  fi
+  # create virtual environment to manage packages
+  python -m venv venv
+  # install framework packages
+  pip install \
+    --extra-index-url "${EXTRA_INDEX_URL}" \
+    install \
+    torch \
+    torchvision \
+    torchaudio \
+    ${EXTRAS}
+  pip install -r requirements.txt
+EOF
+
+# default environment variables
+ENV COMFYUI_ADDRESS=0.0.0.0
+ENV COMFYUI_PORT=8188
+ENV COMFYUI_EXTRA_ARGS=""
+ENV TRANSFORMERS_CACHE="/app/.cache/transformers"
+# default start command
+CMD bash -c "python -u main.py --listen ${COMFYUI_ADDRESS} --port ${COMFYUI_PORT} ${COMFYUI_EXTRA_ARGS}"

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/latent_formats.py

@@ -755,6 +755,10 @@ class ACEAudio(LatentFormat):
     latent_channels = 8
     latent_dimensions = 2
 
+class ACEAudio15(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+
 class ChromaRadiance(LatentFormat):
     latent_channels = 3
     spacial_downscale_ratio = 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

@@ -50,6 +50,7 @@ import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
 import comfy.ldm.kandinsky5.model
 import comfy.ldm.anima.model
+import comfy.ldm.ace.ace_step15
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -149,6 +150,8 @@ 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
@@ -299,7 +302,7 @@ 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:
@@ -307,7 +310,7 @@ 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))
 
@@ -322,7 +325,7 @@ 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))
@@ -330,10 +333,7 @@ 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([])
 
@@ -776,8 +776,8 @@ 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]
@@ -1541,6 +1541,47 @@ class ACEStep(BaseModel):
         out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
         return out
 
+class ACEStep15(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.ace_step15.AceStepConditionGenerationModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        device = kwargs["device"]
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
+        if cross_attn is not None:
+            out['lyric_embed'] = comfy.conds.CONDRegular(conditioning_lyrics)
+
+        refer_audio = kwargs.get("reference_audio_timbre_latents", None)
+        if refer_audio is None or len(refer_audio) == 0:
+            refer_audio = torch.tensor([[[-1.3672e-01, -1.5820e-01,  5.8594e-01, -5.7422e-01,  3.0273e-02,
+                                        2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01,
+                                        -2.7710e-02, -1.8066e-01, -2.9688e-01,  1.6016e+00, -2.6719e+00,
+                                        7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00,
+                                        2.4316e-01,  4.7266e-01,  4.6387e-02, -6.6406e-01, -2.1973e-01,
+                                        -6.7578e-01, -1.5723e-01,  9.5312e-01, -2.0020e-01, -1.7109e+00,
+                                        5.8984e-01, -5.7422e-01,  5.1562e-01,  2.8320e-01,  1.4551e-01,
+                                        -1.8750e-01, -5.9814e-02,  3.6719e-01, -1.0059e-01, -1.5723e-01,
+                                        2.0605e-01, -4.3359e-01, -8.2812e-01,  4.5654e-02, -6.6016e-01,
+                                        1.4844e-01,  9.4727e-02,  3.8477e-01, -1.2578e+00, -3.3203e-01,
+                                        -8.5547e-01,  4.3359e-01,  4.2383e-01, -8.9453e-01, -5.0391e-01,
+                                        -5.6152e-02, -2.9219e+00, -2.4658e-02,  5.0391e-01,  9.8438e-01,
+                                        7.2754e-02, -2.1582e-01,  6.3672e-01,  1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, 750)
+        else:
+            refer_audio = refer_audio[-1]
+        out['refer_audio'] = comfy.conds.CONDRegular(refer_audio)
+
+        audio_codes = kwargs.get("audio_codes", None)
+        if audio_codes is not None:
+            out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
+
+        return out
+
 class Omnigen2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel)

comfy/model_detection.py

@@ -655,6 +655,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.')
         return dit_config
 
+    if '{}encoder.lyric_encoder.layers.0.input_layernorm.weight'.format(key_prefix) in state_dict_keys:
+        dit_config = {}
+        dit_config["audio_model"] = "ace1.5"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 

comfy/model_management.py

@@ -19,13 +19,21 @@
 import psutil
 import logging
 from enum import Enum
-from comfy.cli_args import args, PerformanceFeature
+from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
+import threading
 import torch
 import sys
 import platform
 import weakref
 import gc
 import os
+from contextlib import nullcontext
+import comfy.memory_management
+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
@@ -578,9 +586,15 @@ WINDOWS = any(platform.win32_ver())
 
 EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
 if WINDOWS:
+    import comfy.windows
     EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
     if total_vram > (15 * 1024):  # more extra reserved vram on 16GB+ cards
         EXTRA_RESERVED_VRAM += 100 * 1024 * 1024
+    def get_free_ram():
+        return comfy.windows.get_free_ram()
+else:
+    def get_free_ram():
+        return psutil.virtual_memory().available
 
 if args.reserve_vram is not None:
     EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
@@ -592,7 +606,7 @@ 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 +621,23 @@ 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
-        logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
-        if current_loaded_models[i].model_unload(memory_to_free):
+            memory_to_free = memory_required - get_free_memory(device)
+            ram_to_free = ram_required - get_free_ram()
+
+        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()
+            memory_to_free = 0
+        if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free):
+            logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
             unloaded_model.append(i)
+        if ram_to_free > 0:
+            logging.debug(f"RAM Unloading {current_loaded_models[i].model.model.__class__.__name__}")
+            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))
@@ -629,7 +651,7 @@ def free_memory(memory_required, device, keep_loaded=[]):
                 soft_empty_cache()
     return unloaded_models
 
-def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
+def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
     cleanup_models_gc()
     global vram_state
 
@@ -650,7 +672,10 @@ 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 +701,25 @@ 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:
@@ -716,6 +747,26 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         current_loaded_models.insert(0, loaded_model)
     return
 
+def load_models_gpu_thread(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load):
+    with torch.inference_mode():
+        load_models_gpu_orig(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
+        soft_empty_cache()
+
+def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
+    #Deliberately load models outside of the Aimdo mempool so they can be retained accross
+    #nodes. Use a dummy thread to do it as pytorch documents that mempool contexts are
+    #thread local. So exploit that to escape context
+    if enables_dynamic_vram():
+        t = threading.Thread(
+            target=load_models_gpu_thread,
+            args=(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
+        )
+        t.start()
+        t.join()
+    else:
+        load_models_gpu_orig(models, memory_required=memory_required, force_patch_weights=force_patch_weights,
+                             minimum_memory_required=minimum_memory_required, force_full_load=force_full_load)
+
 def load_model_gpu(model):
     return load_models_gpu([model])
 
@@ -732,6 +783,9 @@ 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 +803,11 @@ 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 = []
@@ -792,7 +851,7 @@ def unet_inital_load_device(parameters, dtype):
 
     mem_dev = get_free_memory(torch_dev)
     mem_cpu = get_free_memory(cpu_dev)
-    if mem_dev > mem_cpu and model_size < mem_dev:
+    if mem_dev > mem_cpu and model_size < mem_dev and comfy.memory_management.aimdo_allocator is None:
         return torch_dev
     else:
         return cpu_dev
@@ -1051,6 +1110,51 @@ 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
+            synchronize()
+            del STREAM_CAST_BUFFERS[offload_stream]
+            del cast_buffer
+            #FIXME: This doesn't work in Aimdo because mempool cant clear cache
+            soft_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)
+    for offload_stream in STREAM_CAST_BUFFERS:
+        offload_stream.synchronize()
+    STREAM_CAST_BUFFERS.clear()
+    soft_empty_cache()
+
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
     if NUM_STREAMS == 0:
@@ -1093,7 +1197,62 @@ 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
+
+        cast_geometry = comfy.memory_management.tensors_to_geometries([ weight ])
+
+        if dtype is None:
+            dtype = weight._model_dtype
+
+        r = torch.empty_like(weight, dtype=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(cast_geometry, raw_tensor)[0]
+            if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
+                weight._v_signature = signature
+                #Send it over
+                v_tensor.copy_(weight, non_blocking=non_blocking)
+            #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
+
+        if weight.dtype != r.dtype and weight.dtype != weight._model_dtype:
+            #Offloaded casting could skip this, however it would make the quantizations
+            #inconsistent between loaded and offloaded weights. So force the double casting
+            #that would happen in regular flow to make offload deterministic.
+            cast_buffer = torch.empty_like(weight, dtype=weight._model_dtype, device=device)
+            cast_buffer.copy_(weight, non_blocking=non_blocking)
+            weight = cast_buffer
+        r.copy_(weight, non_blocking=non_blocking)
+
+        return r
+
     if device is None or weight.device == device:
         if not copy:
             if dtype is None or weight.dtype == dtype:
@@ -1112,10 +1271,12 @@ 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
 
@@ -1135,14 +1296,14 @@ 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:
         a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         _ = a + b
-        torch.cuda.synchronize()
+        synchronize()
     except torch.AcceleratorError:
         #Dump it! We already know about it from the synchronous return
         pass
@@ -1546,6 +1707,12 @@ def lora_compute_dtype(device):
     LORA_COMPUTE_DTYPES[device] = dtype
     return dtype
 
+def synchronize():
+    if is_intel_xpu():
+        torch.xpu.synchronize()
+    elif torch.cuda.is_available():
+        torch.cuda.synchronize()
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:
@@ -1557,8 +1724,11 @@ def soft_empty_cache(force=False):
     elif is_mlu():
         torch.mlu.empty_cache()
     elif torch.cuda.is_available():
-        torch.cuda.empty_cache()
-        torch.cuda.ipc_collect()
+        if comfy.memory_management.aimdo_allocator is None:
+            #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist
+            torch.cuda.synchronize()
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
 
 def unload_all_models():
     free_memory(1e30, get_torch_device())
@@ -1568,9 +1738,6 @@ def debug_memory_summary():
         return torch.cuda.memory.memory_summary()
     return ""
 
-#TODO: might be cleaner to put this somewhere else
-import threading
-
 class InterruptProcessingException(Exception):
     pass
 

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()
@@ -123,6 +111,10 @@ def move_weight_functions(m, device):
                 memory += f.move_to(device=device)
     return memory
 
+def string_to_seed(data):
+    logging.warning("WARNING: string_to_seed has moved from comfy.model_patcher to comfy.utils")
+    return comfy.utils.string_to_seed(data)
+
 class LowVramPatch:
     def __init__(self, key, patches, convert_func=None, set_func=None):
         self.key = key
@@ -169,6 +161,11 @@ def get_key_weight(model, key):
 
     return weight, set_func, convert_func
 
+def key_param_name_to_key(key, param):
+    if len(key) == 0:
+        return param
+    return "{}.{}".format(key, param)
+
 class AutoPatcherEjector:
     def __init__(self, model: 'ModelPatcher', skip_and_inject_on_exit_only=False):
         self.model = model
@@ -212,6 +209,27 @@ 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 +287,9 @@ 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 +305,9 @@ 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 +635,14 @@ 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 +655,15 @@ 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 +680,7 @@ 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 +707,8 @@ 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):
@@ -773,7 +800,7 @@ class ModelPatcher:
                         continue
 
                 for param in params:
-                    key = "{}.{}".format(n, param)
+                    key = key_param_name_to_key(n, param)
                     self.unpin_weight(key)
                     self.patch_weight_to_device(key, device_to=device_to)
                 if comfy.model_management.is_device_cuda(device_to):
@@ -789,7 +816,7 @@ class ModelPatcher:
                 n = x[1]
                 params = x[3]
                 for param in params:
-                    self.pin_weight_to_device("{}.{}".format(n, param))
+                    self.pin_weight_to_device(key_param_name_to_key(n, param))
 
             usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
@@ -895,7 +922,7 @@ class ModelPatcher:
                 if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
                     move_weight = True
                     for param in params:
-                        key = "{}.{}".format(n, param)
+                        key = key_param_name_to_key(n, param)
                         bk = self.backup.get(key, None)
                         if bk is not None:
                             if not lowvram_possible:
@@ -946,7 +973,7 @@ class ModelPatcher:
                         logging.debug("freed {}".format(n))
 
                         for param in params:
-                            self.pin_weight_to_device("{}.{}".format(n, param))
+                            self.pin_weight_to_device(key_param_name_to_key(n, param))
 
 
             self.model.model_lowvram = True
@@ -984,6 +1011,9 @@ 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 +1347,10 @@ 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 +1385,249 @@ 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):
+        self.partially_unload_ram(1e32)
+
+    def memory_required(self, input_shape):
+        #Pad this significantly. We are trying to get away from precise estimates. This
+        #estimate is only used when using the ModelPatcherDynamic after ModelPatcher. If you
+        #use all ModelPatcherDynamic this is ignored and its all done dynamically.
+        return super().memory_required(input_shape=input_shape) * 1.3 + (1024 ** 3)
+
+
+    def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False, dirty=False):
+
+        #Force patching doesn't make sense in Dynamic loading, as you dont know what does and
+        #doesn't need to be forced at this stage. The only thing you could do would be patch
+        #it all on CPU which consumes huge RAM.
+        assert not force_patch_weights
+
+        #Full load doesn't make sense as we dont actually have any loader capability here and
+        #now.
+        assert not full_load
+
+        assert device_to == self.load_device
+
+        num_patches = 0
+        allocated_size = 0
+
+        with self.use_ejected():
+            self.unpatch_hooks()
+
+            vbar = self._vbar_get(create=True)
+            if vbar is not None:
+                vbar.prioritize()
+
+            #We 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
+
+                def setup_param(self, m, n, param_key):
+                    nonlocal num_patches
+                    key = key_param_name_to_key(n, param_key)
+
+                    weight_function = []
+
+                    weight, _, _ = get_key_weight(self.model, key)
+                    if weight is None:
+                        return 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 = key_param_name_to_key(n, param)
+                        weight, _, _ = get_key_weight(self.model, key)
+                        weight.seed_key = key
+                        set_dirty(weight, dirty)
+                        geometry = weight
+                        model_dtype = getattr(m, param + "_comfy_model_dtype", 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, 1e32)
+
+    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 +78,115 @@ 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 ]
+
+        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
+
+        if signature is None and pin is None:
+            comfy.pinned_memory.pin_memory(s)
+            pin = comfy.pinned_memory.get_pin(s)
+        else:
+            pin = None
+
+        if pin is not None:
+            comfy.model_management.cast_to_gathered(xfer_source, pin)
+            xfer_source = [ pin ]
+        #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([s.weight, s.bias ], 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
+
+    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
+
+    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
+
+    #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 +201,38 @@ 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 +240,7 @@ 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 +262,20 @@ 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 +286,57 @@ 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
 
@@ -655,8 +843,8 @@ 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
@@ -666,6 +854,8 @@ 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
@@ -684,7 +874,8 @@ 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 +1,29 @@
+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
+    size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ])
+    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

@@ -59,6 +59,7 @@ import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.jina_clip_2
 import comfy.text_encoders.newbie
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 import comfy.model_patcher
 import comfy.lora
@@ -228,8 +229,10 @@ 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
@@ -389,8 +392,18 @@ 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):
@@ -440,6 +453,8 @@ class VAE:
         self.extra_1d_channel = None
         self.crop_input = True
 
+        self.audio_sample_rate = 44100
+
         if config is None:
             if "decoder.mid.block_1.mix_factor" in sd:
                 encoder_config = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
@@ -537,14 +552,25 @@ class VAE:
                                                                     encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig},
                                                                     decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig})
             elif "decoder.layers.1.layers.0.beta" in sd:
-                self.first_stage_model = AudioOobleckVAE()
+                config = {}
+                param_key = None
+                if "decoder.layers.2.layers.1.weight_v" in sd:
+                    param_key = "decoder.layers.2.layers.1.weight_v"
+                if "decoder.layers.2.layers.1.parametrizations.weight.original1" in sd:
+                    param_key = "decoder.layers.2.layers.1.parametrizations.weight.original1"
+                if param_key is not None:
+                    if sd[param_key].shape[-1] == 12:
+                        config["strides"] = [2, 4, 4, 6, 10]
+                        self.audio_sample_rate = 48000
+
+                self.first_stage_model = AudioOobleckVAE(**config)
                 self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)
                 self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype)
                 self.latent_channels = 64
                 self.output_channels = 2
                 self.pad_channel_value = "replicate"
                 self.upscale_ratio = 2048
-                self.downscale_ratio =  2048
+                self.downscale_ratio = 2048
                 self.latent_dim = 1
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
@@ -765,12 +791,7 @@ 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()
@@ -782,7 +803,18 @@ 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()
 
@@ -897,7 +929,7 @@ 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)
 
@@ -971,7 +1003,7 @@ 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
@@ -1409,6 +1441,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 clip_data_jina = clip_data[0]
             tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
             tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
+        elif clip_type == CLIPType.ACE:
+            clip_target.clip = comfy.text_encoders.ace15.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.ace15.ACE15Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -1432,7 +1467,7 @@ 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)
@@ -1520,7 +1555,8 @@ 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)
@@ -1563,7 +1599,6 @@ 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)
@@ -1655,13 +1690,14 @@ 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)
@@ -1692,9 +1728,9 @@ 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

@@ -155,6 +155,8 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.execution_device = options.get("execution_device", self.execution_device)
         if isinstance(self.layer, list) or self.layer == "all":
             pass
+        elif isinstance(layer_idx, list):
+            self.layer = layer_idx
         elif layer_idx is None or abs(layer_idx) > self.num_layers:
             self.layer = "last"
         else:
@@ -297,7 +299,7 @@ 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/supported_models.py

@@ -24,6 +24,7 @@ import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.z_image
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 from . import supported_models_base
 from . import latent_formats
@@ -1596,6 +1597,38 @@ class Kandinsky5Image(Kandinsky5):
         return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
+class ACEStep15(supported_models_base.BASE):
+    unet_config = {
+        "audio_model": "ace1.5",
+    }
+
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 3.0,
+    }
+
+    latent_format = comfy.latent_formats.ACEAudio15
+
+    memory_usage_factor = 4.7
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.ACEStep15(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**hunyuan_detect))
+
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
 
 models += [SVD_img2vid]

comfy/text_encoders/ace15.py

@@ -0,0 +1,223 @@
+from .anima import Qwen3Tokenizer
+import comfy.text_encoders.llama
+from comfy import sd1_clip
+import torch
+import math
+import comfy.utils
+
+
+def sample_manual_loop_no_classes(
+    model,
+    ids=None,
+    paddings=[],
+    execution_dtype=None,
+    cfg_scale: float = 2.0,
+    temperature: float = 0.85,
+    top_p: float = 0.9,
+    top_k: int = None,
+    seed: int = 1,
+    min_tokens: int = 1,
+    max_new_tokens: int = 2048,
+    audio_start_id: int = 151669,  # The cutoff ID for audio codes
+    eos_token_id: int = 151645,
+):
+    device = model.execution_device
+
+    if execution_dtype is None:
+        if comfy.model_management.should_use_bf16(device):
+            execution_dtype = torch.bfloat16
+        else:
+            execution_dtype = torch.float32
+
+    embeds, attention_mask, num_tokens, embeds_info = model.process_tokens(ids, device)
+    for i, t in enumerate(paddings):
+        attention_mask[i, :t] = 0
+        attention_mask[i, t:] = 1
+
+    output_audio_codes = []
+    past_key_values = []
+    generator = torch.Generator(device=device)
+    generator.manual_seed(seed)
+    model_config = model.transformer.model.config
+
+    for x in range(model_config.num_hidden_layers):
+        past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), 0))
+
+    progress_bar = comfy.utils.ProgressBar(max_new_tokens)
+
+    for step in range(max_new_tokens):
+        outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values)
+        next_token_logits = model.transformer.logits(outputs[0])[:, -1]
+        past_key_values = outputs[2]
+
+        cond_logits = next_token_logits[0:1]
+        uncond_logits = next_token_logits[1:2]
+        cfg_logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits)
+
+        if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
+            eos_score = cfg_logits[:, eos_token_id].clone()
+
+        remove_logit_value = torch.finfo(cfg_logits.dtype).min
+        # Only generate audio tokens
+        cfg_logits[:, :audio_start_id] = remove_logit_value
+
+        if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
+            cfg_logits[:, eos_token_id] = eos_score
+
+        if top_k is not None and top_k > 0:
+            top_k_vals, _ = torch.topk(cfg_logits, top_k)
+            min_val = top_k_vals[..., -1, None]
+            cfg_logits[cfg_logits < min_val] = remove_logit_value
+
+        if top_p is not None and top_p < 1.0:
+            sorted_logits, sorted_indices = torch.sort(cfg_logits, descending=True)
+            cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+            sorted_indices_to_remove[..., 0] = 0
+            indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+            cfg_logits[indices_to_remove] = remove_logit_value
+
+        if temperature > 0:
+            cfg_logits = cfg_logits / temperature
+            next_token = torch.multinomial(torch.softmax(cfg_logits, dim=-1), num_samples=1, generator=generator).squeeze(1)
+        else:
+            next_token = torch.argmax(cfg_logits, dim=-1)
+
+        token = next_token.item()
+
+        if token == eos_token_id:
+            break
+
+        embed, _, _, _ = model.process_tokens([[token]], device)
+        embeds = embed.repeat(2, 1, 1)
+        attention_mask = torch.cat([attention_mask, torch.ones((2, 1), device=device, dtype=attention_mask.dtype)], dim=1)
+
+        output_audio_codes.append(token - audio_start_id)
+        progress_bar.update_absolute(step)
+
+    return output_audio_codes
+
+
+def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0):
+    cfg_scale = 2.0
+
+    positive = [[token for token, _ in inner_list] for inner_list in positive]
+    negative = [[token for token, _ in inner_list] for inner_list in negative]
+    positive = positive[0]
+    negative = negative[0]
+
+    neg_pad = 0
+    if len(negative) < len(positive):
+        neg_pad = (len(positive) - len(negative))
+        negative = [model.special_tokens["pad"]] * neg_pad + negative
+
+    pos_pad = 0
+    if len(negative) > len(positive):
+        pos_pad = (len(negative) - len(positive))
+        positive = [model.special_tokens["pad"]] * pos_pad + positive
+
+    paddings = [pos_pad, neg_pad]
+    return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens)
+
+
+class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_06b", tokenizer=Qwen3Tokenizer)
+
+    def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
+        out = {}
+        lyrics = kwargs.get("lyrics", "")
+        bpm = kwargs.get("bpm", 120)
+        duration = kwargs.get("duration", 120)
+        keyscale = kwargs.get("keyscale", "C major")
+        timesignature = kwargs.get("timesignature", 2)
+        language = kwargs.get("language", "en")
+        seed = kwargs.get("seed", 0)
+
+        duration = math.ceil(duration)
+        meta_lm = 'bpm: {}\nduration: {}\nkeyscale: {}\ntimesignature: {}'.format(bpm, duration, keyscale, timesignature)
+        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n{}\n<|im_end|>\n<|im_start|>assistant\n<think>\n{}\n</think>\n\n<|im_end|>\n"
+
+        meta_cap = '- bpm: {}\n- timesignature: {}\n- keyscale: {}\n- duration: {}\n'.format(bpm, timesignature, keyscale, duration)
+        out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, meta_lm), disable_weights=True)
+        out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, ""), disable_weights=True)
+
+        out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric{}<|endoftext|><|endoftext|>".format(language, lyrics), return_word_ids, disable_weights=True, **kwargs)
+        out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}# Metas\n{}<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs)
+        out["lm_metadata"] = {"min_tokens": duration * 5, "seed": seed}
+        return out
+
+
+class Qwen3_06BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_06B_ACE15, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class Qwen3_2B_ACE15(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_2B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class ACE15TEModel(torch.nn.Module):
+    def __init__(self, device="cpu", dtype=None, dtype_llama=None, model_options={}):
+        super().__init__()
+        if dtype_llama is None:
+            dtype_llama = dtype
+
+        self.qwen3_06b = Qwen3_06BModel(device=device, dtype=dtype, model_options=model_options)
+        self.qwen3_2b = Qwen3_2B_ACE15(device=device, dtype=dtype_llama, model_options=model_options)
+        self.dtypes = set([dtype, dtype_llama])
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_base = token_weight_pairs["qwen3_06b"]
+        token_weight_pairs_lyrics = token_weight_pairs["lyrics"]
+
+        self.qwen3_06b.set_clip_options({"layer": None})
+        base_out, _, extra = self.qwen3_06b.encode_token_weights(token_weight_pairs_base)
+        self.qwen3_06b.set_clip_options({"layer": [0]})
+        lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics)
+
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        audio_codes = generate_audio_codes(self.qwen3_2b, token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"])
+
+        return base_out, None, {"conditioning_lyrics": lyrics_embeds[:, 0], "audio_codes": [audio_codes]}
+
+    def set_clip_options(self, options):
+        self.qwen3_06b.set_clip_options(options)
+        self.qwen3_2b.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.qwen3_06b.reset_clip_options()
+        self.qwen3_2b.reset_clip_options()
+
+    def load_sd(self, sd):
+        if "model.layers.0.post_attention_layernorm.weight" in sd:
+            shape = sd["model.layers.0.post_attention_layernorm.weight"].shape
+            if shape[0] == 1024:
+                return self.qwen3_06b.load_sd(sd)
+            else:
+                return self.qwen3_2b.load_sd(sd)
+
+    def memory_estimation_function(self, token_weight_pairs, device=None):
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        constant = 0.4375
+        if comfy.model_management.should_use_bf16(device):
+            constant *= 0.5
+
+        token_weight_pairs = token_weight_pairs.get("lm_prompt", [])
+        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
+        num_tokens += lm_metadata['min_tokens']
+        return num_tokens * constant * 1024 * 1024
+
+def te(dtype_llama=None, llama_quantization_metadata=None):
+    class ACE15TEModel_(ACE15TEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+            super().__init__(device=device, dtype_llama=dtype_llama, dtype=dtype, model_options=model_options)
+    return ACE15TEModel_

comfy/text_encoders/anima.py

@@ -8,7 +8,7 @@ import torch
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):

comfy/text_encoders/flux.py

@@ -118,7 +118,7 @@ class MistralTokenizerClass:
 class Mistral3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         self.tekken_data = tokenizer_data.get("tekken_model", None)
-        super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
+        super().__init__("", pad_with_end=False, embedding_directory=embedding_directory, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"tekken_model": self.tekken_data}
@@ -176,12 +176,12 @@ def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False):
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class Qwen3Tokenizer8B(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class KleinTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_4b"):

comfy/text_encoders/llama.py

@@ -1,11 +1,12 @@
 import torch
 import torch.nn as nn
 from dataclasses import dataclass
-from typing import Optional, Any
+from typing import Optional, Any, Tuple
 import math
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
+import comfy.ops
 import comfy.ldm.common_dit
 import comfy.clip_model
 
@@ -32,6 +33,7 @@ class Llama2Config:
     k_norm = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Mistral3Small24BConfig:
@@ -54,6 +56,7 @@ class Mistral3Small24BConfig:
     k_norm = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen25_3BConfig:
@@ -76,6 +79,7 @@ class Qwen25_3BConfig:
     k_norm = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen3_06BConfig:
@@ -98,6 +102,53 @@ class Qwen3_06BConfig:
     k_norm = "gemma3"
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
+
+@dataclass
+class Qwen3_06B_ACE15_Config:
+    vocab_size: int = 151669
+    hidden_size: int = 1024
+    intermediate_size: int = 3072
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 32768
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
+@dataclass
+class Qwen3_2B_ACE15_lm_Config:
+    vocab_size: int = 217204
+    hidden_size: int = 2048
+    intermediate_size: int = 6144
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 40960
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen3_4BConfig:
@@ -120,6 +171,7 @@ class Qwen3_4BConfig:
     k_norm = "gemma3"
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen3_8BConfig:
@@ -142,6 +194,7 @@ class Qwen3_8BConfig:
     k_norm = "gemma3"
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Ovis25_2BConfig:
@@ -164,6 +217,7 @@ class Ovis25_2BConfig:
     k_norm = "gemma3"
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Qwen25_7BVLI_Config:
@@ -186,6 +240,7 @@ class Qwen25_7BVLI_Config:
     k_norm = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Gemma2_2B_Config:
@@ -209,6 +264,7 @@ class Gemma2_2B_Config:
     sliding_attention = None
     rope_scale = None
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Gemma3_4B_Config:
@@ -232,6 +288,7 @@ class Gemma3_4B_Config:
     sliding_attention = [1024, 1024, 1024, 1024, 1024, False]
     rope_scale = [8.0, 1.0]
     final_norm: bool = True
+    lm_head: bool = False
 
 @dataclass
 class Gemma3_12B_Config:
@@ -255,6 +312,7 @@ class Gemma3_12B_Config:
     sliding_attention = [1024, 1024, 1024, 1024, 1024, False]
     rope_scale = [8.0, 1.0]
     final_norm: bool = True
+    lm_head: bool = False
     vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14}
     mm_tokens_per_image = 256
 
@@ -356,6 +414,7 @@ class Attention(nn.Module):
         attention_mask: Optional[torch.Tensor] = None,
         freqs_cis: Optional[torch.Tensor] = None,
         optimized_attention=None,
+        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ):
         batch_size, seq_length, _ = hidden_states.shape
         xq = self.q_proj(hidden_states)
@@ -373,11 +432,30 @@ class Attention(nn.Module):
 
         xq, xk = apply_rope(xq, xk, freqs_cis=freqs_cis)
 
+        present_key_value = None
+        if past_key_value is not None:
+            index = 0
+            num_tokens = xk.shape[2]
+            if len(past_key_value) > 0:
+                past_key, past_value, index = past_key_value
+                if past_key.shape[2] >= (index + num_tokens):
+                    past_key[:, :, index:index + xk.shape[2]] = xk
+                    past_value[:, :, index:index + xv.shape[2]] = xv
+                    xk = past_key[:, :, :index + xk.shape[2]]
+                    xv = past_value[:, :, :index + xv.shape[2]]
+                    present_key_value = (past_key, past_value, index + num_tokens)
+                else:
+                    xk = torch.cat((past_key[:, :, :index], xk), dim=2)
+                    xv = torch.cat((past_value[:, :, :index], xv), dim=2)
+                    present_key_value = (xk, xv, index + num_tokens)
+            else:
+                present_key_value = (xk, xv, index + num_tokens)
+
         xk = xk.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
         xv = xv.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
 
         output = optimized_attention(xq, xk, xv, self.num_heads, mask=attention_mask, skip_reshape=True)
-        return self.o_proj(output)
+        return self.o_proj(output), present_key_value
 
 class MLP(nn.Module):
     def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
@@ -408,15 +486,17 @@ class TransformerBlock(nn.Module):
         attention_mask: Optional[torch.Tensor] = None,
         freqs_cis: Optional[torch.Tensor] = None,
         optimized_attention=None,
+        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ):
         # Self Attention
         residual = x
         x = self.input_layernorm(x)
-        x = self.self_attn(
+        x, present_key_value = self.self_attn(
             hidden_states=x,
             attention_mask=attention_mask,
             freqs_cis=freqs_cis,
             optimized_attention=optimized_attention,
+            past_key_value=past_key_value,
         )
         x = residual + x
 
@@ -426,7 +506,7 @@ class TransformerBlock(nn.Module):
         x = self.mlp(x)
         x = residual + x
 
-        return x
+        return x, present_key_value
 
 class TransformerBlockGemma2(nn.Module):
     def __init__(self, config: Llama2Config, index, device=None, dtype=None, ops: Any = None):
@@ -451,6 +531,7 @@ class TransformerBlockGemma2(nn.Module):
         attention_mask: Optional[torch.Tensor] = None,
         freqs_cis: Optional[torch.Tensor] = None,
         optimized_attention=None,
+        past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ):
         if self.transformer_type == 'gemma3':
             if self.sliding_attention:
@@ -468,11 +549,12 @@ class TransformerBlockGemma2(nn.Module):
         # Self Attention
         residual = x
         x = self.input_layernorm(x)
-        x = self.self_attn(
+        x, present_key_value = self.self_attn(
             hidden_states=x,
             attention_mask=attention_mask,
             freqs_cis=freqs_cis,
             optimized_attention=optimized_attention,
+            past_key_value=past_key_value,
         )
 
         x = self.post_attention_layernorm(x)
@@ -485,7 +567,7 @@ class TransformerBlockGemma2(nn.Module):
         x = self.post_feedforward_layernorm(x)
         x = residual + x
 
-        return x
+        return x, present_key_value
 
 class Llama2_(nn.Module):
     def __init__(self, config, device=None, dtype=None, ops=None):
@@ -516,9 +598,10 @@ class Llama2_(nn.Module):
         else:
             self.norm = None
 
-        # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
+        if config.lm_head:
+            self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
 
-    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[]):
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None):
         if embeds is not None:
             x = embeds
         else:
@@ -527,8 +610,13 @@ class Llama2_(nn.Module):
         if self.normalize_in:
             x *= self.config.hidden_size ** 0.5
 
+        seq_len = x.shape[1]
+        past_len = 0
+        if past_key_values is not None and len(past_key_values) > 0:
+            past_len = past_key_values[0][2]
+
         if position_ids is None:
-            position_ids = torch.arange(0, x.shape[1], device=x.device).unsqueeze(0)
+            position_ids = torch.arange(past_len, past_len + seq_len, device=x.device).unsqueeze(0)
 
         freqs_cis = precompute_freqs_cis(self.config.head_dim,
                                          position_ids,
@@ -539,14 +627,16 @@ class Llama2_(nn.Module):
 
         mask = None
         if attention_mask is not None:
-            mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
+            mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1])
             mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
 
-        causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
-        if mask is not None:
-            mask += causal_mask
-        else:
-            mask = causal_mask
+        if seq_len > 1:
+            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
+            if mask is not None:
+                mask += causal_mask
+            else:
+                mask = causal_mask
+
         optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None, small_input=True)
 
         intermediate = None
@@ -562,16 +652,27 @@ class Llama2_(nn.Module):
             elif intermediate_output < 0:
                 intermediate_output = len(self.layers) + intermediate_output
 
+        next_key_values = []
         for i, layer in enumerate(self.layers):
             if all_intermediate is not None:
                 if only_layers is None or (i in only_layers):
                     all_intermediate.append(x.unsqueeze(1).clone())
-            x = layer(
+
+            past_kv = None
+            if past_key_values is not None:
+                past_kv = past_key_values[i] if len(past_key_values) > 0 else []
+
+            x, current_kv = layer(
                 x=x,
                 attention_mask=mask,
                 freqs_cis=freqs_cis,
                 optimized_attention=optimized_attention,
+                past_key_value=past_kv,
             )
+
+            if current_kv is not None:
+                next_key_values.append(current_kv)
+
             if i == intermediate_output:
                 intermediate = x.clone()
 
@@ -588,7 +689,10 @@ class Llama2_(nn.Module):
         if intermediate is not None and final_layer_norm_intermediate and self.norm is not None:
             intermediate = self.norm(intermediate)
 
-        return x, intermediate
+        if len(next_key_values) > 0:
+            return x, intermediate, next_key_values
+        else:
+            return x, intermediate
 
 
 class Gemma3MultiModalProjector(torch.nn.Module):
@@ -672,6 +776,39 @@ class Qwen3_06B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_06B_ACE15_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_2B_ACE15_lm_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+    def logits(self, x):
+        input = x[:, -1:]
+        module = self.model.embed_tokens
+
+        offload_stream = None
+        if module.comfy_cast_weights:
+            weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True)
+        else:
+            weight = self.model.embed_tokens.weight.to(x)
+
+        x = torch.nn.functional.linear(input, weight, None)
+
+        comfy.ops.uncast_bias_weight(module, weight, None, offload_stream)
+        return x
+
 class Qwen3_4B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy/text_encoders/lt.py

@@ -125,7 +125,7 @@ class LTXAVTEModel(torch.nn.Module):
             for prefix, component in [("text_embedding_projection.", self.text_embedding_projection), ("video_embeddings_connector.", self.video_embeddings_connector), ("audio_embeddings_connector.", self.audio_embeddings_connector)]:
                 component_sd = {k.replace(prefix, ""): v for k, v in sdo.items() if k.startswith(prefix)}
                 if component_sd:
-                    missing, unexpected = component.load_state_dict(component_sd, strict=False)
+                    missing, unexpected = component.load_state_dict(component_sd, strict=False, assign=getattr(self, "can_assign_sd", False))
                     missing_all.extend([f"{prefix}{k}" for k in missing])
                     unexpected_all.extend([f"{prefix}{k}" for k in unexpected])
 

comfy/text_encoders/z_image.py

@@ -6,7 +6,7 @@ import os
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
 
 
 class ZImageTokenizer(sd1_clip.SD1Tokenizer):

comfy/utils.py

@@ -28,9 +28,11 @@ 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 warnings
 
 MMAP_TORCH_FILES = args.mmap_torch_files
 DISABLE_MMAP = args.disable_mmap
@@ -56,21 +58,70 @@ 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,
+
+    "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"))
+
+    with warnings.catch_warnings():
+        #We are working with read-only RAM by design
+        warnings.filterwarnings("ignore", message="The given buffer is not writable")
+        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")
     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 +1359,16 @@ 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/windows.py

@@ -0,0 +1,52 @@
+import ctypes
+import logging
+import psutil
+from ctypes import wintypes
+
+import comfy_aimdo.control
+
+psapi = ctypes.WinDLL("psapi")
+kernel32 = ctypes.WinDLL("kernel32")
+
+class PERFORMANCE_INFORMATION(ctypes.Structure):
+    _fields_ = [
+        ("cb", wintypes.DWORD),
+        ("CommitTotal", ctypes.c_size_t),
+        ("CommitLimit", ctypes.c_size_t),
+        ("CommitPeak", ctypes.c_size_t),
+        ("PhysicalTotal", ctypes.c_size_t),
+        ("PhysicalAvailable", ctypes.c_size_t),
+        ("SystemCache", ctypes.c_size_t),
+        ("KernelTotal", ctypes.c_size_t),
+        ("KernelPaged", ctypes.c_size_t),
+        ("KernelNonpaged", ctypes.c_size_t),
+        ("PageSize", ctypes.c_size_t),
+        ("HandleCount", wintypes.DWORD),
+        ("ProcessCount", wintypes.DWORD),
+        ("ThreadCount", wintypes.DWORD),
+    ]
+
+def get_free_ram():
+    #Windows is way too conservative and chalks recently used uncommitted model RAM
+    #as "in-use". So, calculate free RAM for the sake of general use as the greater of:
+    #
+    #1: What psutil says
+    #2: Total Memory - (Committed Memory - VRAM in use)
+    #
+    #We have to subtract VRAM in use from the comitted memory as WDDM creates a naked
+    #commit charge for all VRAM used just incase it wants to page it all out. This just
+    #isn't realistic so "overcommit" on our calculations by just subtracting it off.
+
+    pi = PERFORMANCE_INFORMATION()
+    pi.cb = ctypes.sizeof(pi)
+
+    if not psapi.GetPerformanceInfo(ctypes.byref(pi), pi.cb):
+        logging.warning("WARNING: Failed to query windows performance info. RAM usage may be sub optimal")
+        return psutil.virtual_memory().available
+
+    committed = pi.CommitTotal * pi.PageSize
+    total = pi.PhysicalTotal * pi.PageSize
+
+    return max(psutil.virtual_memory().available,
+               total - (committed - comfy_aimdo.control.get_total_vram_usage()))
+

comfy_api/latest/_io.py

@@ -1248,6 +1248,7 @@ class Hidden(str, Enum):
 class NodeInfoV1:
     input: dict=None
     input_order: dict[str, list[str]]=None
+    is_input_list: bool=None
     output: list[str]=None
     output_is_list: list[bool]=None
     output_name: list[str]=None
@@ -1474,6 +1475,7 @@ class Schema:
         info = NodeInfoV1(
             input=input,
             input_order={key: list(value.keys()) for (key, value) in input.items()},
+            is_input_list=self.is_input_list,
             output=output,
             output_is_list=output_is_list,
             output_name=output_name,

comfy_api_nodes/apis/hitpaw.py

@@ -0,0 +1,51 @@
+from typing import TypedDict
+
+from pydantic import BaseModel, Field
+
+
+class InputVideoModel(TypedDict):
+    model: str
+    resolution: str
+
+
+class ImageEnhanceTaskCreateRequest(BaseModel):
+    model_name: str = Field(...)
+    img_url: str = Field(...)
+    extension: str = Field(".png")
+    exif: bool = Field(False)
+    DPI: int | None = Field(None)
+
+
+class VideoEnhanceTaskCreateRequest(BaseModel):
+    video_url: str = Field(...)
+    extension: str = Field(".mp4")
+    model_name: str | None = Field(...)
+    resolution: list[int] = Field(..., description="Target resolution [width, height]")
+    original_resolution: list[int] = Field(..., description="Original video resolution [width, height]")
+
+
+class TaskCreateDataResponse(BaseModel):
+    job_id: str = Field(...)
+    consume_coins: int | None = Field(None)
+
+
+class TaskStatusPollRequest(BaseModel):
+    job_id: str = Field(...)
+
+
+class TaskCreateResponse(BaseModel):
+    code: int = Field(...)
+    message: str = Field(...)
+    data: TaskCreateDataResponse | None = Field(None)
+
+
+class TaskStatusDataResponse(BaseModel):
+    job_id: str = Field(...)
+    status: str = Field(...)
+    res_url: str = Field("")
+
+
+class TaskStatusResponse(BaseModel):
+    code: int = Field(...)
+    message: str = Field(...)
+    data: TaskStatusDataResponse = Field(...)

comfy_api_nodes/nodes_hitpaw.py

@@ -0,0 +1,342 @@
+import math
+
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.apis.hitpaw import (
+    ImageEnhanceTaskCreateRequest,
+    InputVideoModel,
+    TaskCreateDataResponse,
+    TaskCreateResponse,
+    TaskStatusPollRequest,
+    TaskStatusResponse,
+    VideoEnhanceTaskCreateRequest,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    download_url_to_image_tensor,
+    download_url_to_video_output,
+    downscale_image_tensor,
+    get_image_dimensions,
+    poll_op,
+    sync_op,
+    upload_image_to_comfyapi,
+    upload_video_to_comfyapi,
+    validate_video_duration,
+)
+
+VIDEO_MODELS_MODELS_MAP = {
+    "Portrait Restore Model (1x)": "portrait_restore_1x",
+    "Portrait Restore Model (2x)": "portrait_restore_2x",
+    "General Restore Model (1x)": "general_restore_1x",
+    "General Restore Model (2x)": "general_restore_2x",
+    "General Restore Model (4x)": "general_restore_4x",
+    "Ultra HD Model (2x)": "ultrahd_restore_2x",
+    "Generative Model (1x)": "generative_1x",
+}
+
+# Resolution name to target dimension (shorter side) in pixels
+RESOLUTION_TARGET_MAP = {
+    "720p": 720,
+    "1080p": 1080,
+    "2K/QHD": 1440,
+    "4K/UHD": 2160,
+    "8K": 4320,
+}
+
+# Square (1:1) resolutions use standard square dimensions
+RESOLUTION_SQUARE_MAP = {
+    "720p": 720,
+    "1080p": 1080,
+    "2K/QHD": 1440,
+    "4K/UHD": 2048,  # DCI 4K square
+    "8K": 4096,  # DCI 8K square
+}
+
+# Models with limited resolution support (no 8K)
+LIMITED_RESOLUTION_MODELS = {"Generative Model (1x)"}
+
+# Resolution options for different model types
+RESOLUTIONS_LIMITED = ["original", "720p", "1080p", "2K/QHD", "4K/UHD"]
+RESOLUTIONS_FULL = ["original", "720p", "1080p", "2K/QHD", "4K/UHD", "8K"]
+
+# Maximum output resolution in pixels
+MAX_PIXELS_GENERATIVE = 32_000_000
+MAX_MP_GENERATIVE = MAX_PIXELS_GENERATIVE // 1_000_000
+
+
+class HitPawGeneralImageEnhance(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="HitPawGeneralImageEnhance",
+            display_name="HitPaw General Image Enhance",
+            category="api node/image/HitPaw",
+            description="Upscale low-resolution images to super-resolution, eliminate artifacts and noise. "
+            f"Maximum output: {MAX_MP_GENERATIVE} megapixels.",
+            inputs=[
+                IO.Combo.Input("model", options=["generative_portrait", "generative"]),
+                IO.Image.Input("image"),
+                IO.Combo.Input("upscale_factor", options=[1, 2, 4]),
+                IO.Boolean.Input(
+                    "auto_downscale",
+                    default=False,
+                    tooltip="Automatically downscale input image if output would exceed the limit.",
+                ),
+            ],
+            outputs=[
+                IO.Image.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model"]),
+                expr="""
+                (
+                  $prices := {
+                    "generative_portrait": {"min": 0.02, "max": 0.06},
+                    "generative": {"min": 0.05, "max": 0.15}
+                  };
+                  $price := $lookup($prices, widgets.model);
+                  {
+                    "type": "range_usd",
+                    "min_usd": $price.min,
+                    "max_usd": $price.max
+                  }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        image: Input.Image,
+        upscale_factor: int,
+        auto_downscale: bool,
+    ) -> IO.NodeOutput:
+        height, width = get_image_dimensions(image)
+        requested_scale = upscale_factor
+        output_pixels = height * width * requested_scale * requested_scale
+        if output_pixels > MAX_PIXELS_GENERATIVE:
+            if auto_downscale:
+                input_pixels = width * height
+                scale = 1
+                max_input_pixels = MAX_PIXELS_GENERATIVE
+
+                for candidate in [4, 2, 1]:
+                    if candidate > requested_scale:
+                        continue
+                    scale_output_pixels = input_pixels * candidate * candidate
+                    if scale_output_pixels <= MAX_PIXELS_GENERATIVE:
+                        scale = candidate
+                        max_input_pixels = None
+                        break
+                    # Check if we can downscale input by at most 2x to fit
+                    downscale_ratio = math.sqrt(scale_output_pixels / MAX_PIXELS_GENERATIVE)
+                    if downscale_ratio <= 2.0:
+                        scale = candidate
+                        max_input_pixels = MAX_PIXELS_GENERATIVE // (candidate * candidate)
+                        break
+
+                if max_input_pixels is not None:
+                    image = downscale_image_tensor(image, total_pixels=max_input_pixels)
+                upscale_factor = scale
+            else:
+                output_width = width * requested_scale
+                output_height = height * requested_scale
+                raise ValueError(
+                    f"Output size ({output_width}x{output_height} = {output_pixels:,} pixels) "
+                    f"exceeds maximum allowed size of {MAX_PIXELS_GENERATIVE:,} pixels ({MAX_MP_GENERATIVE}MP). "
+                    f"Enable auto_downscale or use a smaller input image or a lower upscale factor."
+                )
+
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/photo-enhancer", method="POST"),
+            response_model=TaskCreateResponse,
+            data=ImageEnhanceTaskCreateRequest(
+                model_name=f"{model}_{upscale_factor}x",
+                img_url=await upload_image_to_comfyapi(cls, image, total_pixels=None),
+            ),
+            wait_label="Creating task",
+            final_label_on_success="Task created",
+        )
+        if initial_res.code != 200:
+            raise ValueError(f"Task creation failed with code {initial_res.code}: {initial_res.message}")
+        request_price = initial_res.data.consume_coins / 1000
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/task-status", method="POST"),
+            data=TaskCreateDataResponse(job_id=initial_res.data.job_id),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda x: x.data.status,
+            price_extractor=lambda x: request_price,
+            poll_interval=10.0,
+            max_poll_attempts=480,
+        )
+        return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.res_url))
+
+
+class HitPawVideoEnhance(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        model_options = []
+        for model_name in VIDEO_MODELS_MODELS_MAP:
+            if model_name in LIMITED_RESOLUTION_MODELS:
+                resolutions = RESOLUTIONS_LIMITED
+            else:
+                resolutions = RESOLUTIONS_FULL
+            model_options.append(
+                IO.DynamicCombo.Option(
+                    model_name,
+                    [IO.Combo.Input("resolution", options=resolutions)],
+                )
+            )
+
+        return IO.Schema(
+            node_id="HitPawVideoEnhance",
+            display_name="HitPaw Video Enhance",
+            category="api node/video/HitPaw",
+            description="Upscale low-resolution videos to high resolution, eliminate artifacts and noise. "
+            "Prices shown are per second of video.",
+            inputs=[
+                IO.DynamicCombo.Input("model", options=model_options),
+                IO.Video.Input("video"),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.resolution"]),
+                expr="""
+                (
+                  $m := $lookup(widgets, "model");
+                  $res := $lookup(widgets, "model.resolution");
+                  $standard_model_prices := {
+                    "original": {"min": 0.01, "max": 0.198},
+                    "720p": {"min": 0.01, "max": 0.06},
+                    "1080p": {"min": 0.015, "max": 0.09},
+                    "2k/qhd": {"min": 0.02, "max": 0.117},
+                    "4k/uhd": {"min": 0.025, "max": 0.152},
+                    "8k": {"min": 0.033, "max": 0.198}
+                  };
+                  $ultra_hd_model_prices := {
+                    "original": {"min": 0.015, "max": 0.264},
+                    "720p": {"min": 0.015, "max": 0.092},
+                    "1080p": {"min": 0.02, "max": 0.12},
+                    "2k/qhd": {"min": 0.026, "max": 0.156},
+                    "4k/uhd": {"min": 0.034, "max": 0.203},
+                    "8k": {"min": 0.044, "max": 0.264}
+                  };
+                  $generative_model_prices := {
+                    "original": {"min": 0.015, "max": 0.338},
+                    "720p": {"min": 0.008, "max": 0.090},
+                    "1080p": {"min": 0.05, "max": 0.15},
+                    "2k/qhd": {"min": 0.038, "max": 0.225},
+                    "4k/uhd": {"min": 0.056, "max": 0.338}
+                  };
+                  $prices := $contains($m, "ultra hd") ? $ultra_hd_model_prices :
+                             $contains($m, "generative") ? $generative_model_prices :
+                             $standard_model_prices;
+                  $price := $lookup($prices, $res);
+                  {
+                    "type": "range_usd",
+                    "min_usd": $price.min,
+                    "max_usd": $price.max,
+                    "format": {"approximate": true, "suffix": "/second"}
+                  }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: InputVideoModel,
+        video: Input.Video,
+    ) -> IO.NodeOutput:
+        validate_video_duration(video, min_duration=0.5, max_duration=60 * 60)
+        resolution = model["resolution"]
+        src_width, src_height = video.get_dimensions()
+
+        if resolution == "original":
+            output_width = src_width
+            output_height = src_height
+        else:
+            if src_width == src_height:
+                target_size = RESOLUTION_SQUARE_MAP[resolution]
+                if target_size < src_width:
+                    raise ValueError(
+                        f"Selected resolution {resolution} ({target_size}x{target_size}) is smaller than "
+                        f"the input video ({src_width}x{src_height}). Please select a higher resolution or 'original'."
+                    )
+                output_width = target_size
+                output_height = target_size
+            else:
+                min_dimension = min(src_width, src_height)
+                target_size = RESOLUTION_TARGET_MAP[resolution]
+                if target_size < min_dimension:
+                    raise ValueError(
+                        f"Selected resolution {resolution} ({target_size}p) is smaller than "
+                        f"the input video's shorter dimension ({min_dimension}p). "
+                        f"Please select a higher resolution or 'original'."
+                    )
+                if src_width > src_height:
+                    output_height = target_size
+                    output_width = int(target_size * (src_width / src_height))
+                else:
+                    output_width = target_size
+                    output_height = int(target_size * (src_height / src_width))
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/video-enhancer", method="POST"),
+            response_model=TaskCreateResponse,
+            data=VideoEnhanceTaskCreateRequest(
+                video_url=await upload_video_to_comfyapi(cls, video),
+                resolution=[output_width, output_height],
+                original_resolution=[src_width, src_height],
+                model_name=VIDEO_MODELS_MODELS_MAP[model["model"]],
+            ),
+            wait_label="Creating task",
+            final_label_on_success="Task created",
+        )
+        request_price = initial_res.data.consume_coins / 1000
+        if initial_res.code != 200:
+            raise ValueError(f"Task creation failed with code {initial_res.code}: {initial_res.message}")
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/task-status", method="POST"),
+            data=TaskStatusPollRequest(job_id=initial_res.data.job_id),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda x: x.data.status,
+            price_extractor=lambda x: request_price,
+            poll_interval=10.0,
+            max_poll_attempts=320,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(final_response.data.res_url))
+
+
+class HitPawExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            HitPawGeneralImageEnhance,
+            HitPawVideoEnhance,
+        ]
+
+
+async def comfy_entrypoint() -> HitPawExtension:
+    return HitPawExtension()

comfy_api_nodes/util/upload_helpers.py

@@ -94,7 +94,7 @@ async def upload_image_to_comfyapi(
     *,
     mime_type: str | None = None,
     wait_label: str | None = "Uploading",
-    total_pixels: int = 2048 * 2048,
+    total_pixels: int | None = 2048 * 2048,
 ) -> str:
     """Uploads a single image to ComfyUI API and returns its download URL."""
     return (

comfy_extras/nodes_ace.py

@@ -28,12 +28,39 @@ class TextEncodeAceStepAudio(io.ComfyNode):
         conditioning = node_helpers.conditioning_set_values(conditioning, {"lyrics_strength": lyrics_strength})
         return io.NodeOutput(conditioning)
 
+class TextEncodeAceStepAudio15(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncodeAceStepAudio1.5",
+            category="conditioning",
+            inputs=[
+                io.Clip.Input("clip"),
+                io.String.Input("tags", multiline=True, dynamic_prompts=True),
+                io.String.Input("lyrics", multiline=True, dynamic_prompts=True),
+                io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True),
+                io.Int.Input("bpm", default=120, min=10, max=300),
+                io.Float.Input("duration", default=120.0, min=0.0, max=2000.0, step=0.1),
+                io.Combo.Input("timesignature", options=['2', '3', '4', '6']),
+                io.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]),
+                io.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]),
+            ],
+            outputs=[io.Conditioning.Output()],
+        )
+
+    @classmethod
+    def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale) -> io.NodeOutput:
+        tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed)
+        conditioning = clip.encode_from_tokens_scheduled(tokens)
+        return io.NodeOutput(conditioning)
+
 
 class EmptyAceStepLatentAudio(io.ComfyNode):
     @classmethod
     def define_schema(cls):
         return io.Schema(
             node_id="EmptyAceStepLatentAudio",
+            display_name="Empty Ace Step 1.0 Latent Audio",
             category="latent/audio",
             inputs=[
                 io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.1),
@@ -51,12 +78,60 @@ class EmptyAceStepLatentAudio(io.ComfyNode):
         return io.NodeOutput({"samples": latent, "type": "audio"})
 
 
+class EmptyAceStep15LatentAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="EmptyAceStep1.5LatentAudio",
+            display_name="Empty Ace Step 1.5 Latent Audio",
+            category="latent/audio",
+            inputs=[
+                io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.01),
+                io.Int.Input(
+                    "batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."
+                ),
+            ],
+            outputs=[io.Latent.Output()],
+        )
+
+    @classmethod
+    def execute(cls, seconds, batch_size) -> io.NodeOutput:
+        length = round((seconds * 48000 / 1920))
+        latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": latent, "type": "audio"})
+
+class ReferenceTimbreAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="ReferenceTimbreAudio",
+            category="advanced/conditioning/audio",
+            is_experimental=True,
+            description="This node sets the reference audio for timbre (for ace step 1.5)",
+            inputs=[
+                io.Conditioning.Input("conditioning"),
+                io.Latent.Input("latent", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, conditioning, latent=None) -> io.NodeOutput:
+        if latent is not None:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_audio_timbre_latents": [latent["samples"]]}, append=True)
+        return io.NodeOutput(conditioning)
+
 class AceExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
         return [
             TextEncodeAceStepAudio,
             EmptyAceStepLatentAudio,
+            TextEncodeAceStepAudio15,
+            EmptyAceStep15LatentAudio,
+            ReferenceTimbreAudio,
         ]
 
 async def comfy_entrypoint() -> AceExtension:

comfy_extras/nodes_audio.py

@@ -82,13 +82,14 @@ class VAEEncodeAudio(IO.ComfyNode):
     @classmethod
     def execute(cls, vae, audio) -> IO.NodeOutput:
         sample_rate = audio["sample_rate"]
-        if 44100 != sample_rate:
-            waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, 44100)
+        vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+        if vae_sample_rate != sample_rate:
+            waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, vae_sample_rate)
         else:
             waveform = audio["waveform"]
 
         t = vae.encode(waveform.movedim(1, -1))
-        return IO.NodeOutput({"samples":t})
+        return IO.NodeOutput({"samples": t})
 
     encode = execute  # TODO: remove
 
@@ -114,7 +115,8 @@ class VAEDecodeAudio(IO.ComfyNode):
         std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0
         std[std < 1.0] = 1.0
         audio /= std
-        return IO.NodeOutput({"waveform": audio, "sample_rate": 44100 if "sample_rate" not in samples else samples["sample_rate"]})
+        vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+        return IO.NodeOutput({"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]})
 
     decode = execute  # TODO: remove
 

comfy_extras/nodes_model_patch.py

@@ -267,9 +267,9 @@ class ModelPatchLoader:
                     device=comfy.model_management.unet_offload_device(),
                     operations=comfy.ops.manual_cast)
 
-        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:

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.11.1"
+__version__ = "0.12.0"

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"

docker-compose.yaml

@@ -0,0 +1,20 @@
+version: "3.9"
+
+services:
+  comfyui:
+    user: "1000:1000"
+    build: .
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    ports:
+      - "8188:8188"
+    volumes:
+      - "./models:/app/models"
+      - "./input:/app/input"
+      - "./temp:/app/output/temp"
+      - "./output:/app/output"

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 +517,19 @@ 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())):
+                try:
+                    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)
+                finally:
+                    if allocator is not None:
+                        comfy.model_management.reset_cast_buffers()
+                        torch.cuda.synchronize()
+
             if has_pending_tasks:
                 pending_async_nodes[unique_id] = output_data
                 unblock = execution_list.add_external_block(unique_id)
@@ -1000,22 +1014,34 @@ async def validate_prompt(prompt_id, prompt, partial_execution_list: Union[list[
     outputs = set()
     for x in prompt:
         if 'class_type' not in prompt[x]:
+            node_data = prompt[x]
+            node_title = node_data.get('_meta', {}).get('title')
             error = {
-                "type": "invalid_prompt",
-                "message": "Cannot execute because a node is missing the class_type property.",
+                "type": "missing_node_type",
+                "message": f"Node '{node_title or f'ID #{x}'}' has no class_type. The workflow may be corrupted or a custom node is missing.",
                 "details": f"Node ID '#{x}'",
-                "extra_info": {}
+                "extra_info": {
+                    "node_id": x,
+                    "class_type": None,
+                    "node_title": node_title
+                }
             }
             return (False, error, [], {})
 
         class_type = prompt[x]['class_type']
         class_ = nodes.NODE_CLASS_MAPPINGS.get(class_type, None)
         if class_ is None:
+            node_data = prompt[x]
+            node_title = node_data.get('_meta', {}).get('title', class_type)
             error = {
-                "type": "invalid_prompt",
-                "message": f"Cannot execute because node {class_type} does not exist.",
+                "type": "missing_node_type",
+                "message": f"Node '{node_title}' not found. The custom node may not be installed.",
                 "details": f"Node ID '#{x}'",
-                "extra_info": {}
+                "extra_info": {
+                    "node_id": x,
+                    "class_type": class_type,
+                    "node_title": node_title
+                }
             }
             return (False, error, [], {})
 

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)

nodes.py

@@ -1001,7 +1001,7 @@ class DualCLIPLoader:
     def INPUT_TYPES(s):
         return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ),
                               "clip_name2": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie"], ),
+                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie", "ace"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.11.1"
+version = "0.12.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.37.11
-comfyui-workflow-templates==0.8.27
+comfyui-workflow-templates==0.8.31
 comfyui-embedded-docs==0.4.0
 torch
 torchsde
@@ -22,6 +22,7 @@ alembic
 SQLAlchemy
 av>=14.2.0
 comfy-kitchen>=0.2.7
+comfy-aimdo>=0.1.7
 requests
 
 #non essential dependencies:

server.py

@@ -656,6 +656,7 @@ class PromptServer():
             info = {}
             info['input'] = obj_class.INPUT_TYPES()
             info['input_order'] = {key: list(value.keys()) for (key, value) in obj_class.INPUT_TYPES().items()}
+            info['is_input_list'] = getattr(obj_class, "INPUT_IS_LIST", False)
             info['output'] = obj_class.RETURN_TYPES
             info['output_is_list'] = obj_class.OUTPUT_IS_LIST if hasattr(obj_class, 'OUTPUT_IS_LIST') else [False] * len(obj_class.RETURN_TYPES)
             info['output_name'] = obj_class.RETURN_NAMES if hasattr(obj_class, 'RETURN_NAMES') else info['output']