Merge 69d3bfa391 into 4c4be1bba5

comfy-aimdo 0.2.12 fixes support for non-ASCII filepaths in the new
mmap helper.

README.md

@@ -38,6 +38,8 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 
 ## Get Started
 
+### Local
+
 #### [Desktop Application](https://www.comfy.org/download)
 - The easiest way to get started.
 - Available on Windows & macOS.
@@ -49,8 +51,13 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 #### [Manual Install](#manual-install-windows-linux)
 Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon, Ascend).
 
-## [Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
-See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
+### Cloud
+
+#### [Comfy Cloud](https://www.comfy.org/cloud)
+- Our official paid cloud version for those who can't afford local hardware.
+
+## Examples
+See what ComfyUI can do with the [newer template workflows](https://comfy.org/workflows) or old [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
 
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.

comfy/comfy_types/node_typing.py

@@ -193,6 +193,8 @@ class HiddenInputTypeDict(TypedDict):
     """EXTRA_PNGINFO is a dictionary that will be copied into the metadata of any .png files saved. Custom nodes can store additional information in this dictionary for saving (or as a way to communicate with a downstream node)."""
     dynprompt: NotRequired[Literal["DYNPROMPT"]]
     """DYNPROMPT is an instance of comfy_execution.graph.DynamicPrompt. It differs from PROMPT in that it may mutate during the course of execution in response to Node Expansion."""
+    prompt_id: NotRequired[Literal["PROMPT_ID"]]
+    """PROMPT_ID is the unique identifier of the current prompt/job being executed. Useful for associating progress updates with specific jobs."""
 
 
 class InputTypeDict(TypedDict):

comfy/memory_management.py

@@ -1,9 +1,68 @@
 import math
+import ctypes
+import threading
+import dataclasses
 import torch
 from typing import NamedTuple
 
 from comfy.quant_ops import QuantizedTensor
 
+
+class TensorFileSlice(NamedTuple):
+    file_ref: object
+    thread_id: int
+    offset: int
+    size: int
+
+
+def read_tensor_file_slice_into(tensor, destination):
+
+    if isinstance(tensor, QuantizedTensor):
+        if not isinstance(destination, QuantizedTensor):
+            return False
+        if tensor._layout_cls != destination._layout_cls:
+            return False
+
+        if not read_tensor_file_slice_into(tensor._qdata, destination._qdata):
+            return False
+
+        dst_orig_dtype = destination._params.orig_dtype
+        destination._params.copy_from(tensor._params, non_blocking=False)
+        destination._params = dataclasses.replace(destination._params, orig_dtype=dst_orig_dtype)
+        return True
+
+    info = getattr(tensor.untyped_storage(), "_comfy_tensor_file_slice", None)
+    if info is None:
+        return False
+
+    file_obj = info.file_ref
+    if (destination.device.type != "cpu"
+            or file_obj is None
+            or threading.get_ident() != info.thread_id
+            or destination.numel() * destination.element_size() < info.size):
+        return False
+
+    if info.size == 0:
+        return True
+
+    buf_type = ctypes.c_ubyte * info.size
+    view = memoryview(buf_type.from_address(destination.data_ptr()))
+
+    try:
+        file_obj.seek(info.offset)
+        done = 0
+        while done < info.size:
+            try:
+                n = file_obj.readinto(view[done:])
+            except OSError:
+                return False
+            if n <= 0:
+                return False
+            done += n
+        return True
+    finally:
+        view.release()
+
 class TensorGeometry(NamedTuple):
     shape: any
     dtype: torch.dtype

comfy/model_management.py

@@ -505,6 +505,28 @@ def module_size(module):
         module_mem += t.nbytes
     return module_mem
 
+def module_mmap_residency(module, free=False):
+    mmap_touched_mem = 0
+    module_mem = 0
+    bounced_mmaps = set()
+    sd = module.state_dict()
+    for k in sd:
+        t = sd[k]
+        module_mem += t.nbytes
+        storage = t._qdata.untyped_storage() if isinstance(t, comfy.quant_ops.QuantizedTensor) else t.untyped_storage()
+        if not getattr(storage, "_comfy_tensor_mmap_touched", False):
+            continue
+        mmap_touched_mem += t.nbytes
+        if not free:
+            continue
+        storage._comfy_tensor_mmap_touched = False
+        mmap_obj = storage._comfy_tensor_mmap_refs[0]
+        if mmap_obj in bounced_mmaps:
+            continue
+        mmap_obj.bounce()
+        bounced_mmaps.add(mmap_obj)
+    return mmap_touched_mem, module_mem
+
 class LoadedModel:
     def __init__(self, model):
         self._set_model(model)
@@ -532,6 +554,9 @@ class LoadedModel:
     def model_memory(self):
         return self.model.model_size()
 
+    def model_mmap_residency(self, free=False):
+        return self.model.model_mmap_residency(free=free)
+
     def model_loaded_memory(self):
         return self.model.loaded_size()
 
@@ -633,7 +658,7 @@ 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=[], for_dynamic=False, ram_required=0):
+def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins_required=0, ram_required=0):
     cleanup_models_gc()
     unloaded_model = []
     can_unload = []
@@ -646,13 +671,14 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
                 can_unload.append((-shift_model.model_offloaded_memory(), sys.getrefcount(shift_model.model), shift_model.model_memory(), i))
                 shift_model.currently_used = False
 
-    for x in sorted(can_unload):
+    can_unload_sorted = sorted(can_unload)
+    for x in can_unload_sorted:
         i = x[-1]
         memory_to_free = 1e32
-        ram_to_free = 1e32
+        pins_to_free = 1e32
         if not DISABLE_SMART_MEMORY:
             memory_to_free = memory_required - get_free_memory(device)
-            ram_to_free = ram_required - get_free_ram()
+            pins_to_free = pins_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.
@@ -661,9 +687,18 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
         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:
+        if pins_to_free > 0:
+            logging.debug(f"PIN Unloading {current_loaded_models[i].model.model.__class__.__name__}")
+            current_loaded_models[i].model.partially_unload_ram(pins_to_free)
+
+    for x in can_unload_sorted:
+        i = x[-1]
+        ram_to_free = ram_required - psutil.virtual_memory().available
+        if ram_to_free <= 0 and i not in unloaded_model:
+            continue
+        resident_memory, _ = current_loaded_models[i].model_mmap_residency(free=True)
+        if resident_memory > 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))
@@ -729,17 +764,27 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
 
 
     total_memory_required = {}
+    total_pins_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
+        device = loaded_model.device
+        total_memory_required[device] = total_memory_required.get(device, 0) + loaded_model.model_memory_required(device)
+        resident_memory, model_memory = loaded_model.model_mmap_residency()
+        pinned_memory = loaded_model.model.pinned_memory_size()
+        #FIXME: This can over-free the pins as it budgets to pin the entire model. We should
+        #make this JIT to keep as much pinned as possible.
+        pins_required = model_memory - pinned_memory
+        ram_required = model_memory - resident_memory
+        total_pins_required[device] = total_pins_required.get(device, 0) + pins_required
+        total_ram_required[device] = total_ram_required.get(device, 0) + ram_required
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
-            free_memory(total_memory_required[device] * 1.1 + extra_mem, device, for_dynamic=free_for_dynamic, ram_required=total_ram_required[device])
+            free_memory(total_memory_required[device] * 1.1 + extra_mem,
+                        device,
+                        for_dynamic=free_for_dynamic,
+                        pins_required=total_pins_required[device],
+                        ram_required=total_ram_required[device])
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
@@ -1225,6 +1270,11 @@ def cast_to_gathered(tensors, r, non_blocking=False, stream=None):
             dest_view = dest_views.pop(0)
             if tensor is None:
                 continue
+            if comfy.memory_management.read_tensor_file_slice_into(tensor, dest_view):
+                continue
+            storage = tensor._qdata.untyped_storage() if isinstance(tensor, comfy.quant_ops.QuantizedTensor) else tensor.untyped_storage()
+            if hasattr(storage, "_comfy_tensor_mmap_touched"):
+                storage._comfy_tensor_mmap_touched = True
             dest_view.copy_(tensor, non_blocking=non_blocking)
 
 

comfy/model_patcher.py

@@ -297,6 +297,9 @@ class ModelPatcher:
         self.size = comfy.model_management.module_size(self.model)
         return self.size
 
+    def model_mmap_residency(self, free=False):
+        return comfy.model_management.module_mmap_residency(self.model, free=free)
+
     def get_ram_usage(self):
         return self.model_size()
 
@@ -1063,6 +1066,10 @@ class ModelPatcher:
 
             return self.model.model_loaded_weight_memory - current_used
 
+    def pinned_memory_size(self):
+        # Pinned memory pressure tracking is only implemented for DynamicVram loading
+        return 0
+
     def partially_unload_ram(self, ram_to_unload):
         pass
 
@@ -1653,6 +1660,16 @@ class ModelPatcherDynamic(ModelPatcher):
 
         return freed
 
+    def pinned_memory_size(self):
+        total = 0
+        loading = self._load_list(for_dynamic=True)
+        for x in loading:
+            _, _, _, _, m, _ = x
+            pin = comfy.pinned_memory.get_pin(m)
+            if pin is not None:
+                total += pin.numel() * pin.element_size()
+        return total
+
     def partially_unload_ram(self, ram_to_unload):
         loading = self._load_list(for_dynamic=True, default_device=self.offload_device)
         for x in loading:

comfy/ops.py

@@ -306,6 +306,33 @@ class CastWeightBiasOp:
     bias_function = []
 
 class disable_weight_init:
+    @staticmethod
+    def _lazy_load_from_state_dict(module, state_dict, prefix, local_metadata,
+                                   missing_keys, unexpected_keys, weight_shape,
+                                   bias_shape=None):
+        assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False)
+        prefix_len = len(prefix)
+        for k, v in state_dict.items():
+            key = k[prefix_len:]
+            if key == "weight":
+                if not assign_to_params_buffers:
+                    v = v.clone()
+                module.weight = torch.nn.Parameter(v, requires_grad=False)
+            elif bias_shape is not None and key == "bias" and v is not None:
+                if not assign_to_params_buffers:
+                    v = v.clone()
+                module.bias = torch.nn.Parameter(v, requires_grad=False)
+            else:
+                unexpected_keys.append(k)
+
+        if module.weight is None:
+            module.weight = torch.nn.Parameter(torch.zeros(weight_shape), requires_grad=False)
+            missing_keys.append(prefix + "weight")
+
+        if bias_shape is not None and module.bias is None and getattr(module, "comfy_need_lazy_init_bias", False):
+            module.bias = torch.nn.Parameter(torch.zeros(bias_shape), requires_grad=False)
+            missing_keys.append(prefix + "bias")
+
     class Linear(torch.nn.Linear, CastWeightBiasOp):
 
         def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
@@ -333,29 +360,16 @@ class disable_weight_init:
             if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
                 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")
+            disable_weight_init._lazy_load_from_state_dict(
+                self,
+                state_dict,
+                prefix,
+                local_metadata,
+                missing_keys,
+                unexpected_keys,
+                weight_shape=(self.in_features, self.out_features),
+                bias_shape=(self.out_features,),
+            )
 
 
         def reset_parameters(self):
@@ -547,6 +561,48 @@ class disable_weight_init:
                 return super().forward(*args, **kwargs)
 
     class Embedding(torch.nn.Embedding, CastWeightBiasOp):
+        def __init__(self, num_embeddings, embedding_dim, padding_idx=None, max_norm=None,
+                     norm_type=2.0, scale_grad_by_freq=False, sparse=False, _weight=None,
+                     _freeze=False, device=None, dtype=None):
+            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
+                super().__init__(num_embeddings, embedding_dim, padding_idx, max_norm,
+                                 norm_type, scale_grad_by_freq, sparse, _weight,
+                                 _freeze, device, dtype)
+                return
+
+            torch.nn.Module.__init__(self)
+            self.num_embeddings = num_embeddings
+            self.embedding_dim = embedding_dim
+            self.padding_idx = padding_idx
+            self.max_norm = max_norm
+            self.norm_type = norm_type
+            self.scale_grad_by_freq = scale_grad_by_freq
+            self.sparse = sparse
+            # Keep shape/dtype visible for module introspection without reserving storage.
+            embedding_dtype = dtype if dtype is not None else torch.get_default_dtype()
+            self.weight = torch.nn.Parameter(
+                torch.empty((num_embeddings, embedding_dim), device="meta", dtype=embedding_dtype),
+                requires_grad=False,
+            )
+            self.bias = None
+            self.weight_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 comfy.memory_management.aimdo_enabled:
+                return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
+                                                     missing_keys, unexpected_keys, error_msgs)
+            disable_weight_init._lazy_load_from_state_dict(
+                self,
+                state_dict,
+                prefix,
+                local_metadata,
+                missing_keys,
+                unexpected_keys,
+                weight_shape=(self.num_embeddings, self.embedding_dim),
+            )
+
         def reset_parameters(self):
             self.bias = None
             return None

comfy/pinned_memory.py

@@ -1,6 +1,7 @@
-import torch
 import comfy.model_management
 import comfy.memory_management
+import comfy_aimdo.host_buffer
+import comfy_aimdo.torch
 
 from comfy.cli_args import args
 
@@ -12,18 +13,31 @@ def pin_memory(module):
         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:
+
+    if comfy.model_management.MAX_PINNED_MEMORY <= 0 or (comfy.model_management.TOTAL_PINNED_MEMORY + size) > comfy.model_management.MAX_PINNED_MEMORY:
         module.pin_failed = True
         return False
+
+    try:
+        hostbuf = comfy_aimdo.host_buffer.HostBuffer(size)
+    except RuntimeError:
+        module.pin_failed = True
+        return False
+
+    module._pin = comfy_aimdo.torch.hostbuf_to_tensor(hostbuf)
+    module._pin_hostbuf = hostbuf
+    comfy.model_management.TOTAL_PINNED_MEMORY += size
     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)
+
+    comfy.model_management.TOTAL_PINNED_MEMORY -= size
+    if comfy.model_management.TOTAL_PINNED_MEMORY < 0:
+        comfy.model_management.TOTAL_PINNED_MEMORY = 0
+
     del module._pin
+    del module._pin_hostbuf
     return size

comfy/utils.py

@@ -20,6 +20,8 @@
 import torch
 import math
 import struct
+import ctypes
+import os
 import comfy.memory_management
 import safetensors.torch
 import numpy as np
@@ -32,7 +34,7 @@ from einops import rearrange
 from comfy.cli_args import args
 import json
 import time
-import mmap
+import threading
 import warnings
 
 MMAP_TORCH_FILES = args.mmap_torch_files
@@ -81,14 +83,17 @@ _TYPES = {
 }
 
 def load_safetensors(ckpt):
-    f = open(ckpt, "rb")
-    mapping = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
-    mv = memoryview(mapping)
+    import comfy_aimdo.model_mmap
 
-    header_size = struct.unpack("<Q", mapping[:8])[0]
-    header = json.loads(mapping[8:8+header_size].decode("utf-8"))
+    f = open(ckpt, "rb", buffering=0)
+    model_mmap = comfy_aimdo.model_mmap.ModelMMAP(ckpt)
+    file_size = os.path.getsize(ckpt)
+    mv = memoryview((ctypes.c_uint8 * file_size).from_address(model_mmap.get()))
 
-    mv = mv[8 + header_size:]
+    header_size = struct.unpack("<Q", mv[:8])[0]
+    header = json.loads(mv[8:8 + header_size].tobytes().decode("utf-8"))
+
+    mv = mv[(data_base_offset := 8 + header_size):]
 
     sd = {}
     for name, info in header.items():
@@ -102,7 +107,14 @@ def load_safetensors(ckpt):
             with warnings.catch_warnings():
                 #We are working with read-only RAM by design
                 warnings.filterwarnings("ignore", message="The given buffer is not writable")
-                sd[name] = torch.frombuffer(mv[start:end], dtype=_TYPES[info["dtype"]]).view(info["shape"])
+                tensor = torch.frombuffer(mv[start:end], dtype=_TYPES[info["dtype"]]).view(info["shape"])
+                storage = tensor.untyped_storage()
+                setattr(storage,
+                        "_comfy_tensor_file_slice",
+                        comfy.memory_management.TensorFileSlice(f, threading.get_ident(), data_base_offset + start, end - start))
+                setattr(storage, "_comfy_tensor_mmap_refs", (model_mmap, mv))
+                setattr(storage, "_comfy_tensor_mmap_touched", False)
+                sd[name] = tensor
 
     return sd, header.get("__metadata__", {}),
 

comfy_api/feature_flags.py

@@ -12,6 +12,7 @@ from comfy.cli_args import args
 # Default server capabilities
 SERVER_FEATURE_FLAGS: dict[str, Any] = {
     "supports_preview_metadata": True,
+    "supports_progress_text_metadata": True,
     "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes
     "extension": {"manager": {"supports_v4": True}},
     "node_replacements": True,

comfy_api/latest/_io.py

@@ -1282,9 +1282,16 @@ class V3Data(TypedDict):
     'When True, the value of the dynamic input will be in the format (value, path_key).'
 
 class HiddenHolder:
+    """Holds hidden input values resolved during node execution.
+
+    Hidden inputs are special values automatically provided by the execution
+    engine (e.g., node ID, prompt data, authentication tokens) rather than
+    being connected by the user in the graph.
+    """
     def __init__(self, unique_id: str, prompt: Any,
                  extra_pnginfo: Any, dynprompt: Any,
-                 auth_token_comfy_org: str, api_key_comfy_org: str, **kwargs):
+                 auth_token_comfy_org: str, api_key_comfy_org: str,
+                 prompt_id: str = None, **kwargs):
         self.unique_id = unique_id
         """UNIQUE_ID is the unique identifier of the node, and matches the id property of the node on the client side. It is commonly used in client-server communications (see messages)."""
         self.prompt = prompt
@@ -1297,6 +1304,8 @@ class HiddenHolder:
         """AUTH_TOKEN_COMFY_ORG is a token acquired from signing into a ComfyOrg account on frontend."""
         self.api_key_comfy_org = api_key_comfy_org
         """API_KEY_COMFY_ORG is an API Key generated by ComfyOrg that allows skipping signing into a ComfyOrg account on frontend."""
+        self.prompt_id = prompt_id
+        """PROMPT_ID is the unique identifier of the current prompt/job being executed."""
 
     def __getattr__(self, key: str):
         '''If hidden variable not found, return None.'''
@@ -1304,6 +1313,14 @@ class HiddenHolder:
 
     @classmethod
     def from_dict(cls, d: dict | None):
+        """Create a HiddenHolder from a dictionary of hidden input values.
+
+        Args:
+            d: Dictionary mapping Hidden enum values to their resolved values.
+
+        Returns:
+            A new HiddenHolder instance with values populated from the dict.
+        """
         if d is None:
             d = {}
         return cls(
@@ -1313,6 +1330,7 @@ class HiddenHolder:
             dynprompt=d.get(Hidden.dynprompt, None),
             auth_token_comfy_org=d.get(Hidden.auth_token_comfy_org, None),
             api_key_comfy_org=d.get(Hidden.api_key_comfy_org, None),
+            prompt_id=d.get(Hidden.prompt_id, None),
         )
 
     @classmethod
@@ -1335,6 +1353,8 @@ class Hidden(str, Enum):
     """AUTH_TOKEN_COMFY_ORG is a token acquired from signing into a ComfyOrg account on frontend."""
     api_key_comfy_org = "API_KEY_COMFY_ORG"
     """API_KEY_COMFY_ORG is an API Key generated by ComfyOrg that allows skipping signing into a ComfyOrg account on frontend."""
+    prompt_id = "PROMPT_ID"
+    """PROMPT_ID is the unique identifier of the current prompt/job being executed. Useful for associating progress updates with specific jobs."""
 
 
 @dataclass

comfy_api_nodes/util/client.py

@@ -17,6 +17,7 @@ from pydantic import BaseModel
 
 from comfy import utils
 from comfy_api.latest import IO
+from comfy_execution.utils import get_executing_context
 from server import PromptServer
 
 from . import request_logger
@@ -440,6 +441,17 @@ def _display_text(
     status: str | int | None = None,
     price: float | None = None,
 ) -> None:
+    """Send a progress text message to the client for display on a node.
+
+    Assembles status, price, and text lines, then sends them via WebSocket.
+    Automatically retrieves the current prompt_id from the execution context.
+
+    Args:
+        node_cls: The ComfyNode class sending the progress text.
+        text: Optional text content to display.
+        status: Optional status string or code to display.
+        price: Optional price in dollars to display as credits.
+    """
     display_lines: list[str] = []
     if status:
         display_lines.append(f"Status: {status.capitalize() if isinstance(status, str) else status}")
@@ -450,7 +462,9 @@ def _display_text(
     if text is not None:
         display_lines.append(text)
     if display_lines:
-        PromptServer.instance.send_progress_text("\n".join(display_lines), get_node_id(node_cls))
+        ctx = get_executing_context()
+        prompt_id = ctx.prompt_id if ctx is not None else None
+        PromptServer.instance.send_progress_text("\n".join(display_lines), get_node_id(node_cls), prompt_id=prompt_id)
 
 
 def _display_time_progress(

comfy_extras/nodes_images.py

@@ -566,7 +566,7 @@ class GetImageSize(IO.ComfyNode):
                 IO.Int.Output(display_name="height"),
                 IO.Int.Output(display_name="batch_size"),
             ],
-            hidden=[IO.Hidden.unique_id],
+            hidden=[IO.Hidden.unique_id, IO.Hidden.prompt_id],
         )
 
     @classmethod
@@ -577,7 +577,7 @@ class GetImageSize(IO.ComfyNode):
 
         # Send progress text to display size on the node
         if cls.hidden.unique_id:
-            PromptServer.instance.send_progress_text(f"width: {width}, height: {height}\n batch size: {batch_size}", cls.hidden.unique_id)
+            PromptServer.instance.send_progress_text(f"width: {width}, height: {height}\n batch size: {batch_size}", cls.hidden.unique_id, prompt_id=cls.hidden.prompt_id)
 
         return IO.NodeOutput(width, height, batch_size)
 

execution.py

@@ -150,7 +150,7 @@ class CacheSet:
 
 SENSITIVE_EXTRA_DATA_KEYS = ("auth_token_comfy_org", "api_key_comfy_org")
 
-def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}):
+def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}, prompt_id=None):
     is_v3 = issubclass(class_def, _ComfyNodeInternal)
     v3_data: io.V3Data = {}
     hidden_inputs_v3 = {}
@@ -197,6 +197,8 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=
                 hidden_inputs_v3[io.Hidden.auth_token_comfy_org] = extra_data.get("auth_token_comfy_org", None)
             if io.Hidden.api_key_comfy_org.name in hidden:
                 hidden_inputs_v3[io.Hidden.api_key_comfy_org] = extra_data.get("api_key_comfy_org", None)
+            if io.Hidden.prompt_id.name in hidden:
+                hidden_inputs_v3[io.Hidden.prompt_id] = prompt_id
     else:
         if "hidden" in valid_inputs:
             h = valid_inputs["hidden"]
@@ -213,6 +215,8 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=
                     input_data_all[x] = [extra_data.get("auth_token_comfy_org", None)]
                 if h[x] == "API_KEY_COMFY_ORG":
                     input_data_all[x] = [extra_data.get("api_key_comfy_org", None)]
+                if h[x] == "PROMPT_ID":
+                    input_data_all[x] = [prompt_id]
     v3_data["hidden_inputs"] = hidden_inputs_v3
     return input_data_all, missing_keys, v3_data
 
@@ -470,7 +474,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
             has_subgraph = False
         else:
             get_progress_state().start_progress(unique_id)
-            input_data_all, missing_keys, v3_data = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data)
+            input_data_all, missing_keys, v3_data = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data, prompt_id=prompt_id)
             if server.client_id is not None:
                 server.last_node_id = display_node_id
                 server.send_sync("executing", { "node": unique_id, "display_node": display_node_id, "prompt_id": prompt_id }, server.client_id)

manager_requirements.txt

@@ -1 +1 @@
-comfyui_manager==4.1b2
+comfyui_manager==4.1b4

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.41.18
+comfyui-frontend-package==1.41.19
 comfyui-workflow-templates==0.9.21
 comfyui-embedded-docs==0.4.3
 torch
@@ -23,7 +23,7 @@ SQLAlchemy
 filelock
 av>=14.2.0
 comfy-kitchen>=0.2.8
-comfy-aimdo>=0.2.10
+comfy-aimdo>=0.2.12
 requests
 simpleeval>=1.0.0
 blake3

server.py

@@ -1230,13 +1230,47 @@ class PromptServer():
         return json_data
 
     def send_progress_text(
-        self, text: Union[bytes, bytearray, str], node_id: str, sid=None
+        self,
+        text: Union[bytes, bytearray, str],
+        node_id: str,
+        prompt_id: Optional[str] = None,
+        sid=None,
     ):
+        """Send a progress text message to the client via WebSocket.
+
+        Encodes the text as a binary message with length-prefixed node_id. When
+        the client supports the ``supports_progress_text_metadata`` feature flag,
+        the prompt_id is always prepended as a length-prefixed field (empty string
+        when None) to ensure consistent binary framing.
+
+        Args:
+            text: The progress text content to send.
+            node_id: The unique identifier of the node sending the progress.
+            prompt_id: Optional prompt/job identifier to associate the message with.
+            sid: Optional session ID to target a specific client.
+        """
         if isinstance(text, str):
             text = text.encode("utf-8")
         node_id_bytes = str(node_id).encode("utf-8")
 
-        # Pack the node_id length as a 4-byte unsigned integer, followed by the node_id bytes
-        message = struct.pack(">I", len(node_id_bytes)) + node_id_bytes + text
+        # Auto-resolve sid to the currently executing client
+        target_sid = sid if sid is not None else self.client_id
 
-        self.send_sync(BinaryEventTypes.TEXT, message, sid)
+        # When client supports the new format, always send
+        # [prompt_id_len][prompt_id][node_id_len][node_id][text]
+        # even when prompt_id is None (encoded as zero-length string)
+        if feature_flags.supports_feature(
+            self.sockets_metadata, target_sid, "supports_progress_text_metadata"
+        ):
+            prompt_id_bytes = (prompt_id or "").encode("utf-8")
+            message = (
+                struct.pack(">I", len(prompt_id_bytes))
+                + prompt_id_bytes
+                + struct.pack(">I", len(node_id_bytes))
+                + node_id_bytes
+                + text
+            )
+        else:
+            message = struct.pack(">I", len(node_id_bytes)) + node_id_bytes + text
+
+        self.send_sync(BinaryEventTypes.TEXT, message, target_sid)