[pull] master from comfyanonymous:master

README.md

@@ -39,7 +39,7 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 ## Get Started
 
 #### [Desktop Application](https://www.comfy.org/download)
-- The easiest way to get started. 
+- The easiest way to get started.
 - Available on Windows & macOS.
 
 #### [Windows Portable Package](#installing)
@@ -211,27 +211,19 @@ This is the command to install the nightly with ROCm 6.4 which might have some p
 
 ### Intel GPUs (Windows and Linux)
 
-(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip (currently available in PyTorch nightly builds). More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
-
-1. To install PyTorch nightly, use the following command:
+(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu```
-
-2. Launch ComfyUI by running `python main.py`
+1. To install PyTorch xpu, use the following command:
 
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu```
 
-(Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance.
-
-1. For Intel® Arc™ A-Series Graphics utilizing IPEX, create a conda environment and use the commands below:
+This is the command to install the Pytorch xpu nightly which might have some performance improvements:
 
-```
-conda install libuv
-pip install torch==2.3.1.post0+cxx11.abi torchvision==0.18.1.post0+cxx11.abi torchaudio==2.3.1.post0+cxx11.abi intel-extension-for-pytorch==2.3.110.post0+xpu --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/
-```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu```
 
-For other supported Intel GPUs with IPEX, visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
+(Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance.
 
-Additional discussion and help can be found [here](https://github.com/comfyanonymous/ComfyUI/discussions/476).
+1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
 
 ### NVIDIA
 
@@ -352,7 +344,7 @@ Generate a self-signed certificate (not appropriate for shared/production use) a
 
 Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app will now be accessible with `https://...` instead of `http://...`.
 
-> Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above. 
+> Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above.
 <br/><br/>If you use a container, note that the volume mount `-v` can be a relative path so `... -v ".\:/openssl-certs" ...` would create the key & cert files in the current directory of your command prompt or powershell terminal.
 
 ## Support and dev channel

comfy/cli_args.py

@@ -132,6 +132,8 @@ class LatentPreviewMethod(enum.Enum):
 
 parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
 
+parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
+
 parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
 
 parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")

comfy/model_management.py

@@ -78,7 +78,6 @@ def get_supported_float8_types():
     torch_version = torch.version.__version__
     temp = torch_version.split(".")
     torch_version_numeric = (int(temp[0]), int(temp[1]))
-    xpu_available = (torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] <= 4)) and torch.xpu.is_available()
 except:
     pass
 
@@ -102,10 +101,14 @@ def get_supported_float8_types():
 
 try:
     import intel_extension_for_pytorch as ipex  # noqa: F401
+except:
+    pass
+
+try:
     _ = torch.xpu.device_count()
-    xpu_available = xpu_available or torch.xpu.is_available()
+    xpu_available = torch.xpu.is_available()
 except:
-    xpu_available = xpu_available or (hasattr(torch, "xpu") and torch.xpu.is_available())
+    xpu_available = False
 
 try:
     if torch.backends.mps.is_available():
@@ -946,10 +949,12 @@ def pick_weight_dtype(dtype, fallback_dtype, device=None):
     return dtype
 
 def device_supports_non_blocking(device):
+    if args.force_non_blocking:
+        return True
     if is_device_mps(device):
         return False #pytorch bug? mps doesn't support non blocking
-    if is_intel_xpu():
-        return True
+    if is_intel_xpu(): #xpu does support non blocking but it is slower on iGPUs for some reason so disable by default until situation changes
+        return False
     if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews)
         return False
     if directml_enabled:
@@ -1282,10 +1287,10 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        if torch_version_numeric < (2, 6):
+        if torch_version_numeric < (2, 3):
             return True
         else:
-            return torch.xpu.get_device_capability(device)['has_bfloat16_conversions']
+            return torch.xpu.is_bf16_supported()
 
     if is_ascend_npu():
         return True

comfy/sampler_helpers.py

@@ -149,7 +149,7 @@ def cleanup_models(conds, models):
 
     cleanup_additional_models(set(control_cleanup))
 
-def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
+def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict):
     '''
     Registers hooks from conds.
     '''
@@ -158,8 +158,8 @@ def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
     for k in conds:
         get_hooks_from_cond(conds[k], hooks)
     # add wrappers and callbacks from ModelPatcher to transformer_options
-    model_options["transformer_options"]["wrappers"] = comfy.patcher_extension.copy_nested_dicts(model.wrappers)
-    model_options["transformer_options"]["callbacks"] = comfy.patcher_extension.copy_nested_dicts(model.callbacks)
+    comfy.patcher_extension.merge_nested_dicts(model_options["transformer_options"].setdefault("wrappers", {}), model.wrappers, copy_dict1=False)
+    comfy.patcher_extension.merge_nested_dicts(model_options["transformer_options"].setdefault("callbacks", {}), model.callbacks, copy_dict1=False)
     # begin registering hooks
     registered = comfy.hooks.HookGroup()
     target_dict = comfy.hooks.create_target_dict(comfy.hooks.EnumWeightTarget.Model)

comfy/samplers.py

@@ -16,6 +16,7 @@
 import comfy.model_patcher
 import comfy.patcher_extension
 import comfy.hooks
+import comfy.context_windows
 import scipy.stats
 import numpy
 
@@ -198,14 +199,20 @@ def finalize_default_conds(model: 'BaseModel', hooked_to_run: dict[comfy.hooks.H
             hooked_to_run.setdefault(p.hooks, list())
             hooked_to_run[p.hooks] += [(p, i)]
 
-def calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options: dict[str]):
+    handler: comfy.context_windows.ContextHandlerABC = model_options.get("context_handler", None)
+    if handler is None or not handler.should_use_context(model, conds, x_in, timestep, model_options):
+        return _calc_cond_batch_outer(model, conds, x_in, timestep, model_options)
+    return handler.execute(_calc_cond_batch_outer, model, conds, x_in, timestep, model_options)
+
+def _calc_cond_batch_outer(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
         _calc_cond_batch,
         comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.CALC_COND_BATCH, model_options, is_model_options=True)
     )
     return executor.execute(model, conds, x_in, timestep, model_options)
 
-def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
     out_conds = []
     out_counts = []
     # separate conds by matching hooks

comfy_extras/nodes_context_windows.py

@@ -0,0 +1,89 @@
+from __future__ import annotations
+from comfy_api.latest import ComfyExtension, io
+import comfy.context_windows
+import nodes
+
+
+class ContextWindowsManualNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="ContextWindowsManual",
+            display_name="Context Windows (Manual)",
+            category="context",
+            description="Manually set context windows.",
+            inputs=[
+                io.Model.Input("model", tooltip="The model to apply context windows to during sampling."),
+                io.Int.Input("context_length", min=1, default=16, tooltip="The length of the context window."),
+                io.Int.Input("context_overlap", min=0, default=4, tooltip="The overlap of the context window."),
+                io.Combo.Input("context_schedule", options=[
+                    comfy.context_windows.ContextSchedules.STATIC_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_LOOPED,
+                    comfy.context_windows.ContextSchedules.BATCHED,
+                    ], tooltip="The stride of the context window."),
+                io.Int.Input("context_stride", min=1, default=1, tooltip="The stride of the context window; only applicable to uniform schedules."),
+                io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."),
+                io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."),
+                io.Int.Input("dim", min=0, max=5, default=0, tooltip="The dimension to apply the context windows to."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with context windows applied during sampling."),
+            ],
+            is_experimental=True,
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str, dim: int) -> io.Model:
+        model = model.clone()
+        model.model_options["context_handler"] = comfy.context_windows.IndexListContextHandler(
+            context_schedule=comfy.context_windows.get_matching_context_schedule(context_schedule),
+            fuse_method=comfy.context_windows.get_matching_fuse_method(fuse_method),
+            context_length=context_length,
+            context_overlap=context_overlap,
+            context_stride=context_stride,
+            closed_loop=closed_loop,
+            dim=dim)
+        # make memory usage calculation only take into account the context window latents
+        comfy.context_windows.create_prepare_sampling_wrapper(model)
+        return io.NodeOutput(model)
+
+class WanContextWindowsManualNode(ContextWindowsManualNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        schema = super().define_schema()
+        schema.node_id = "WanContextWindowsManual"
+        schema.display_name = "WAN Context Windows (Manual)"
+        schema.description = "Manually set context windows for WAN-like models (dim=2)."
+        schema.inputs = [
+            io.Model.Input("model", tooltip="The model to apply context windows to during sampling."),
+                io.Int.Input("context_length", min=1, max=nodes.MAX_RESOLUTION, step=4, default=81, tooltip="The length of the context window."),
+                io.Int.Input("context_overlap", min=0, default=30, tooltip="The overlap of the context window."),
+                io.Combo.Input("context_schedule", options=[
+                    comfy.context_windows.ContextSchedules.STATIC_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_LOOPED,
+                    comfy.context_windows.ContextSchedules.BATCHED,
+                    ], tooltip="The stride of the context window."),
+                io.Int.Input("context_stride", min=1, default=1, tooltip="The stride of the context window; only applicable to uniform schedules."),
+                io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."),
+                io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."),
+        ]
+        return schema
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str) -> io.Model:
+        context_length = max(((context_length - 1) // 4) + 1, 1)  # at least length 1
+        context_overlap = max(((context_overlap - 1) // 4) + 1, 0)  # at least overlap 0
+        return super().execute(model, context_length, context_overlap, context_schedule, context_stride, closed_loop, fuse_method, dim=2)
+
+
+class ContextWindowsExtension(ComfyExtension):
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            ContextWindowsManualNode,
+            WanContextWindowsManualNode,
+        ]
+
+def comfy_entrypoint():
+    return ContextWindowsExtension()

nodes.py

@@ -2320,6 +2320,7 @@ async def init_builtin_extra_nodes():
         "nodes_camera_trajectory.py",
         "nodes_edit_model.py",
         "nodes_tcfg.py",
+        "nodes_context_windows.py",
     ]
 
     import_failed = []