Merge 6517b60172 into 783782d5d7

* mm: Use Aimdo raw allocator for cast buffers

pytorch manages allocation of growing buffers on streams poorly. Pyt
has no windows support for the expandable segments allocator (which is
the right tool for this job), while also segmenting the memory by
stream such that it can be generally re-used. So kick the problem to
aimdo which can just grow a virtual region thats freed per stream.

* plan

* ops: move cpu handler up to the caller

* ops: split up prefetch from weight prep block prefetching API

Split up the casting and weight formating/lora stuff in prep for
arbitrary prefetch support.

* ops: implement block prefetching API

allow a model to construct a prefetch list and operate it for increased
async offload.

* ltxv2: Implement block prefetching

* Implement lora async offload

Implement async offload of loras.

.ci/windows_amd_base_files/run_amd_gpu_enable_dynamic_vram.bat

@@ -1,2 +1,2 @@
-.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-smart-memory
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --enable-dynamic-vram
 pause

CODEOWNERS

@@ -1,2 +1,2 @@
 # Admins
-* @comfyanonymous @kosinkadink @guill @alexisrolland @rattus128
+* @comfyanonymous @kosinkadink @guill @alexisrolland @rattus128 @kijai

README.md

@@ -193,13 +193,15 @@ If you have trouble extracting it, right click the file -> properties -> unblock
 
 The portable above currently comes with python 3.13 and pytorch cuda 13.0. Update your Nvidia drivers if it doesn't start.
 
-#### Alternative Downloads:
+#### All Official Portable Downloads:
 
 [Portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
 
-[Experimental portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z)
+[Portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z)
 
-[Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
+[Portable for Nvidia GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z) (supports 20 series and above).
+
+[Portable for Nvidia GPUs with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
 
 #### How do I share models between another UI and ComfyUI?
 

app/user_manager.py

@@ -28,8 +28,8 @@ def get_file_info(path: str, relative_to: str) -> FileInfo:
     return {
         "path": os.path.relpath(path, relative_to).replace(os.sep, '/'),
         "size": os.path.getsize(path),
-        "modified": os.path.getmtime(path),
-        "created": os.path.getctime(path)
+        "modified": int(os.path.getmtime(path) * 1000),
+        "created": int(os.path.getctime(path) * 1000),
     }
 
 

comfy/cli_args.py

@@ -90,7 +90,6 @@ parser.add_argument("--force-channels-last", action="store_true", help="Force ch
 parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")
 
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
-parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
 parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
 
 class LatentPreviewMethod(enum.Enum):

comfy/ldm/lightricks/av_model.py

@@ -16,6 +16,7 @@ from comfy.ldm.lightricks.model import (
 from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
 from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
 import comfy.ldm.common_dit
+import comfy.model_prefetch
 
 class CompressedTimestep:
     """Store video timestep embeddings in compressed form using per-frame indexing."""
@@ -907,9 +908,11 @@ class LTXAVModel(LTXVModel):
         """Process transformer blocks for LTXAV."""
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        prefetch_queue = comfy.model_prefetch.make_prefetch_queue(list(self.transformer_blocks), vx.device, transformer_options)
 
         # Process transformer blocks
         for i, block in enumerate(self.transformer_blocks):
+            comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, block)
             if ("double_block", i) in blocks_replace:
 
                 def block_wrap(args):
@@ -982,6 +985,8 @@ class LTXAVModel(LTXVModel):
                     a_prompt_timestep=a_prompt_timestep,
                 )
 
+        comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, None)
+
         return [vx, ax]
 
     def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):

comfy/lora.py

@@ -17,6 +17,7 @@
 """
 
 from __future__ import annotations
+import comfy.memory_management
 import comfy.utils
 import comfy.model_management
 import comfy.model_base
@@ -473,3 +474,17 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
             weight = old_weight
 
     return weight
+
+def prefetch_prepared_value(value, allocate_buffer, stream):
+    if isinstance(value, torch.Tensor):
+        dest = allocate_buffer(comfy.memory_management.vram_aligned_size(value))
+        comfy.model_management.cast_to_gathered([value], dest, non_blocking=True, stream=stream)
+        return comfy.memory_management.interpret_gathered_like([value], dest)[0]
+    elif isinstance(value, weight_adapter.WeightAdapterBase):
+        return type(value)(value.loaded_keys, prefetch_prepared_value(value.weights, allocate_buffer, stream))
+    elif isinstance(value, tuple):
+        return tuple(prefetch_prepared_value(item, allocate_buffer, stream) for item in value)
+    elif isinstance(value, list):
+        return [prefetch_prepared_value(item, allocate_buffer, stream) for item in value]
+
+    return value

comfy/model_base.py

@@ -214,6 +214,11 @@ class BaseModel(torch.nn.Module):
         if "latent_shapes" in extra_conds:
             xc = utils.unpack_latents(xc, extra_conds.pop("latent_shapes"))
 
+        transformer_options = transformer_options.copy()
+        transformer_options["prefetch_dynamic_vbars"] = (
+            self.current_patcher is not None and self.current_patcher.is_dynamic()
+        )
+
         model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds)
         if len(model_output) > 1 and not torch.is_tensor(model_output):
             model_output, _ = utils.pack_latents(model_output)

comfy/model_management.py

@@ -31,6 +31,7 @@ from contextlib import nullcontext
 import comfy.memory_management
 import comfy.utils
 import comfy.quant_ops
+import comfy_aimdo.vram_buffer
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -112,10 +113,6 @@ if args.directml is not None:
     # torch_directml.disable_tiled_resources(True)
     lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
 
-try:
-    import intel_extension_for_pytorch as ipex  # noqa: F401
-except:
-    pass
 
 try:
     _ = torch.xpu.device_count()
@@ -583,9 +580,6 @@ class LoadedModel:
 
         real_model = self.model.model
 
-        if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None:
-            with torch.no_grad():
-                real_model = ipex.optimize(real_model.eval(), inplace=True, graph_mode=True, concat_linear=True)
 
         self.real_model = weakref.ref(real_model)
         self.model_finalizer = weakref.finalize(real_model, cleanup_models)
@@ -1182,6 +1176,10 @@ stream_counters = {}
 
 STREAM_CAST_BUFFERS = {}
 LARGEST_CASTED_WEIGHT = (None, 0)
+STREAM_AIMDO_CAST_BUFFERS = {}
+LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+
+DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE = 16 * 1024 ** 3
 
 def get_cast_buffer(offload_stream, device, size, ref):
     global LARGEST_CASTED_WEIGHT
@@ -1215,13 +1213,26 @@ def get_cast_buffer(offload_stream, device, size, ref):
 
     return cast_buffer
 
+def get_aimdo_cast_buffer(offload_stream, device):
+    cast_buffer = STREAM_AIMDO_CAST_BUFFERS.get(offload_stream, None)
+    if cast_buffer is None:
+        cast_buffer = comfy_aimdo.vram_buffer.VRAMBuffer(DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE, device.index)
+        STREAM_AIMDO_CAST_BUFFERS[offload_stream] = cast_buffer
+
+    return cast_buffer
 def reset_cast_buffers():
     global LARGEST_CASTED_WEIGHT
+    global LARGEST_AIMDO_CASTED_WEIGHT
+
     LARGEST_CASTED_WEIGHT = (None, 0)
-    for offload_stream in STREAM_CAST_BUFFERS:
-        offload_stream.synchronize()
+    LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+    for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS):
+        if offload_stream is not None:
+            offload_stream.synchronize()
     synchronize()
+
     STREAM_CAST_BUFFERS.clear()
+    STREAM_AIMDO_CAST_BUFFERS.clear()
     soft_empty_cache()
 
 def get_offload_stream(device):
@@ -1581,10 +1592,7 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        if torch_version_numeric < (2, 3):
-            return True
-        else:
-            return torch.xpu.get_device_properties(device).has_fp16
+        return torch.xpu.get_device_properties(device).has_fp16
 
     if is_ascend_npu():
         return True
@@ -1650,10 +1658,7 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        if torch_version_numeric < (2, 3):
-            return True
-        else:
-            return torch.xpu.is_bf16_supported()
+        return torch.xpu.is_bf16_supported()
 
     if is_ascend_npu():
         return True
@@ -1784,6 +1789,7 @@ def soft_empty_cache(force=False):
     if cpu_state == CPUState.MPS:
         torch.mps.empty_cache()
     elif is_intel_xpu():
+        torch.xpu.synchronize()
         torch.xpu.empty_cache()
     elif is_ascend_npu():
         torch.npu.empty_cache()

comfy/model_patcher.py

@@ -121,9 +121,20 @@ class LowVramPatch:
         self.patches = patches
         self.convert_func = convert_func # TODO: remove
         self.set_func = set_func
+        self.prepared_patches = None
+
+    def prepare(self, allocate_buffer, stream):
+        self.prepared_patches = [
+            (patch[0], comfy.lora.prefetch_prepared_value(patch[1], allocate_buffer, stream), patch[2], patch[3], patch[4])
+            for patch in self.patches[self.key]
+        ]
+
+    def clear_prepared(self):
+        self.prepared_patches = None
 
     def __call__(self, weight):
-        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
+        patches = self.prepared_patches if self.prepared_patches is not None else self.patches[self.key]
+        return comfy.lora.calculate_weight(patches, weight, self.key, intermediate_dtype=weight.dtype)
 
 LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2
 

comfy/model_prefetch.py

@@ -0,0 +1,65 @@
+import comfy_aimdo.model_vbar
+import comfy.model_management
+import comfy.ops
+
+PREFETCH_QUEUES = []
+
+def cleanup_prefetched_modules(comfy_modules):
+    for s in comfy_modules:
+        prefetch = getattr(s, "_prefetch", None)
+        if prefetch is None:
+            continue
+        for param_key in ("weight", "bias"):
+            lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+            if lowvram_fn is not None:
+                lowvram_fn.clear_prepared()
+        if prefetch["signature"] is not None:
+            comfy_aimdo.model_vbar.vbar_unpin(s._v)
+        delattr(s, "_prefetch")
+
+def cleanup_prefetch_queues():
+    global PREFETCH_QUEUES
+
+    for queue in PREFETCH_QUEUES:
+        for entry in queue:
+            if entry is None or not isinstance(entry, tuple):
+                continue
+            _, prefetch_state = entry
+            comfy_modules = prefetch_state[1]
+            if comfy_modules is not None:
+                cleanup_prefetched_modules(comfy_modules)
+    PREFETCH_QUEUES = []
+
+def prefetch_queue_pop(queue, device, module):
+    if queue is None:
+        return
+
+    consumed = queue.pop(0)
+    if consumed is not None:
+        offload_stream, prefetch_state = consumed
+        offload_stream.wait_stream(comfy.model_management.current_stream(device))
+        _, comfy_modules = prefetch_state
+        if comfy_modules is not None:
+            cleanup_prefetched_modules(comfy_modules)
+
+    prefetch = queue[0]
+    if prefetch is not None:
+        comfy_modules = []
+        for s in prefetch.modules():
+            if hasattr(s, "_v"):
+                comfy_modules.append(s)
+
+        offload_stream = comfy.ops.cast_modules_with_vbar(comfy_modules, None, device, None, True)
+        comfy.model_management.sync_stream(device, offload_stream)
+        queue[0] = (offload_stream, (prefetch, comfy_modules))
+
+def make_prefetch_queue(queue, device, transformer_options):
+    if (not transformer_options.get("prefetch_dynamic_vbars", False)
+        or comfy.model_management.NUM_STREAMS == 0
+        or comfy.model_management.is_device_cpu(device)
+        or not comfy.model_management.device_supports_non_blocking(device)):
+        return None
+
+    queue = [None] + queue + [None]
+    PREFETCH_QUEUES.append(queue)
+    return queue

comfy/ops.py

@@ -86,38 +86,61 @@ def materialize_meta_param(s, param_keys):
             setattr(s, param_key, torch.nn.Parameter(torch.zeros(param.shape, dtype=param.dtype), requires_grad=param.requires_grad))
 
 
-def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant):
-    #vbar doesn't support CPU weights, but some custom nodes have weird paths
-    #that might switch the layer to the CPU and expect it to work. We have to take
-    #a clone conservatively as we are mmapped and some SFT files are packed misaligned
-    #If you are a custom node author reading this, please move your layer to the GPU
-    #or declare your ModelPatcher as CPU in the first place.
-    if comfy.model_management.is_device_cpu(device):
-        materialize_meta_param(s, ["weight", "bias"])
-        weight = s.weight.to(dtype=dtype, copy=True)
-        if isinstance(weight, QuantizedTensor):
-            weight = weight.dequantize()
-        bias = None
-        if s.bias is not None:
-            bias = s.bias.to(dtype=bias_dtype, copy=True)
-        return weight, bias, (None, None, None)
-
+# FIXME: add n=1 cache hit fast path
+def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blocking):
     offload_stream = None
-    xfer_dest = None
+    cast_buffer = None
+    cast_buffer_offset = 0
+
+    def ensure_offload_stream(module, required_size, check_largest):
+        nonlocal offload_stream
+        nonlocal cast_buffer
+
+        if offload_stream is None:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+        if offload_stream is None or not check_largest or len(comfy_modules) != 1:
+            return
+
+        current_size = 0 if cast_buffer is None else cast_buffer.size()
+        if current_size < required_size and module is comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[0]:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+            cast_buffer = None
+        if required_size > comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[1]:
+            comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT = (module, required_size)
+
+    def get_cast_buffer(buffer_size):
+        nonlocal offload_stream
+        nonlocal cast_buffer
+        nonlocal cast_buffer_offset
+
+        if buffer_size == 0:
+            return None
+
+        if offload_stream is None:
+            return torch.empty((buffer_size,), dtype=torch.uint8, device=device)
+
+        cast_buffer = comfy.model_management.get_aimdo_cast_buffer(offload_stream, device)
+        buffer = comfy_aimdo.torch.aimdo_to_tensor(cast_buffer.get(buffer_size, cast_buffer_offset), device)
+        cast_buffer_offset += buffer_size
+        return buffer
+
+    for s in comfy_modules:
+        signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
+        resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
+        prefetch = {
+            "signature": signature,
+            "resident": resident,
+        }
 
-    signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
-    resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
-    if signature is not None:
         if resident:
-            weight = s._v_weight
-            bias = s._v_bias
-        else:
-            xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
+            s._prefetch = prefetch
+            continue
 
-    if not resident:
         materialize_meta_param(s, ["weight", "bias"])
+        xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device) if signature is not None else None
         cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
         cast_dest = None
+        needs_cast = False
 
         xfer_source = [ s.weight, s.bias ]
 
@@ -129,22 +152,15 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             if data is None:
                 continue
             if data.dtype != geometry.dtype:
+                needs_cast = True
                 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)
+        ensure_offload_stream(s, dest_size if xfer_dest is None else 0, True)
         if xfer_dest is None:
-            xfer_dest = torch.empty((dest_size,), dtype=torch.uint8, device=device)
-            offload_stream = None
+            xfer_dest = get_cast_buffer(dest_size)
 
         if signature is None and pin is None:
             comfy.pinned_memory.pin_memory(s)
@@ -157,27 +173,54 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             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 param_key in ("weight", "bias"):
+            lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+            if lowvram_fn is not None:
+                ensure_offload_stream(s, cast_buffer_offset, False)
+                lowvram_fn.prepare(lambda size: get_cast_buffer(size), offload_stream)
+
+        prefetch["xfer_dest"] = xfer_dest
+        prefetch["cast_dest"] = cast_dest
+        prefetch["cast_geometry"] = cast_geometry
+        prefetch["needs_cast"] = needs_cast
+        s._prefetch = prefetch
+
+    return offload_stream
+
+
+def resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant):
+
+    prefetch = getattr(s, "_prefetch", None)
+
+    if prefetch["resident"]:
+        weight = s._v_weight
+        bias = s._v_bias
+    else:
+        xfer_dest = prefetch["xfer_dest"]
+        if prefetch["needs_cast"]:
+            cast_dest = prefetch["cast_dest"] if prefetch["cast_dest"] is not None else torch.empty((comfy.memory_management.vram_aligned_size(prefetch["cast_geometry"]),), dtype=torch.uint8, device=device)
             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)):
+                                           comfy.memory_management.interpret_gathered_like(prefetch["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)
+        params = comfy.memory_management.interpret_gathered_like(prefetch["cast_geometry"], xfer_dest)
         weight = params[0]
         bias = params[1]
-        if signature is not None:
+        if prefetch["signature"] is not None:
             s._v_weight = weight
             s._v_bias = bias
-        s._v_signature=signature
+        s._v_signature = prefetch["signature"]
 
     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", [])
 
+        if x is None:
+            return None
+
         orig = x
 
         def to_dequant(tensor, dtype):
@@ -205,14 +248,12 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             x = f(x)
         return x
 
-    update_weight = signature is not None
+    update_weight = prefetch["signature"] is not None
+    weight = post_cast(s, "weight", weight, dtype, prefetch["resident"], update_weight)
+    if bias is not None:
+        bias = post_cast(s, "bias", bias, bias_dtype, prefetch["resident"], update_weight)
 
-    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)
-
-    #FIXME: weird offload return protocol
-    return weight, bias, (offload_stream, device if signature is not None else None, None)
+    return weight, bias
 
 
 def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None, want_requant=False):
@@ -230,10 +271,46 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         if device is None:
             device = input.device
 
+    def format_return(result, offloadable):
+        weight, bias, offload_stream = result
+        return (weight, bias, offload_stream) if offloadable else (weight, bias)
+
     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, want_requant)
+
+        #vbar doesn't support CPU weights, but some custom nodes have weird paths
+        #that might switch the layer to the CPU and expect it to work. We have to take
+        #a clone conservatively as we are mmapped and some SFT files are packed misaligned
+        #If you are a custom node author reading this, please move your layer to the GPU
+        #or declare your ModelPatcher as CPU in the first place.
+        if comfy.model_management.is_device_cpu(device):
+            materialize_meta_param(s, ["weight", "bias"])
+            weight = s.weight.to(dtype=dtype, copy=True)
+            if isinstance(weight, QuantizedTensor):
+                weight = weight.dequantize()
+            bias = s.bias.to(dtype=bias_dtype, copy=True) if s.bias is not None else None
+            return format_return((weight, bias, (None, None, None)), offloadable)
+
+        prefetched = hasattr(s, "_prefetch")
+        offload_stream = None
+        offload_device = None
+        if not prefetched:
+            offload_stream = cast_modules_with_vbar([s], dtype, device, bias_dtype, non_blocking)
+            comfy.model_management.sync_stream(device, offload_stream)
+
+        weight, bias = resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant)
+
+        if not prefetched:
+            if getattr(s, "_prefetch")["signature"] is not None:
+                offload_device = device
+            for param_key in ("weight", "bias"):
+                lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+                if lowvram_fn is not None:
+                    lowvram_fn.clear_prepared()
+            delattr(s, "_prefetch")
+        return format_return((weight, bias, (offload_stream, offload_device, None)), offloadable)
+
 
     if offloadable and (device != s.weight.device or
                         (s.bias is not None and device != s.bias.device)):
@@ -280,11 +357,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         for f in s.weight_function:
             weight = f(weight)
 
-    if offloadable:
-        return weight, bias, (offload_stream, weight_a, bias_a)
-    else:
-        #Legacy function signature
-        return weight, bias
+    return format_return((weight, bias, (offload_stream, weight_a, bias_a)), offloadable)
 
 
 def uncast_bias_weight(s, weight, bias, offload_stream):

comfy_api_nodes/nodes_bytedance.py

@@ -1403,7 +1403,6 @@ class ByteDance2TextToVideoNode(IO.ComfyNode):
             status_extractor=lambda r: r.status,
             price_extractor=_seedance2_price_extractor(model_id, has_video_input=False),
             poll_interval=9,
-            max_poll_attempts=180,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
 
@@ -1585,7 +1584,6 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
             status_extractor=lambda r: r.status,
             price_extractor=_seedance2_price_extractor(model_id, has_video_input=False),
             poll_interval=9,
-            max_poll_attempts=180,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
 
@@ -1907,7 +1905,6 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
             status_extractor=lambda r: r.status,
             price_extractor=_seedance2_price_extractor(model_id, has_video_input=has_video_input),
             poll_interval=9,
-            max_poll_attempts=180,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
 

comfy_api_nodes/nodes_hitpaw.py

@@ -178,7 +178,6 @@ class HitPawGeneralImageEnhance(IO.ComfyNode):
             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))
 
@@ -324,7 +323,6 @@ class HitPawVideoEnhance(IO.ComfyNode):
             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))
 

comfy_api_nodes/nodes_kling.py

@@ -276,7 +276,6 @@ async def finish_omni_video_task(cls: type[IO.ComfyNode], response: TaskStatusRe
         cls,
         ApiEndpoint(path=f"/proxy/kling/v1/videos/omni-video/{response.data.task_id}"),
         response_model=TaskStatusResponse,
-        max_poll_attempts=280,
         status_extractor=lambda r: (r.data.task_status if r.data else None),
     )
     return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url))
@@ -3062,7 +3061,6 @@ class KlingVideoNode(IO.ComfyNode):
             cls,
             ApiEndpoint(path=poll_path),
             response_model=TaskStatusResponse,
-            max_poll_attempts=280,
             status_extractor=lambda r: (r.data.task_status if r.data else None),
         )
         return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url))
@@ -3188,7 +3186,6 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
             cls,
             ApiEndpoint(path=f"/proxy/kling/v1/videos/image2video/{response.data.task_id}"),
             response_model=TaskStatusResponse,
-            max_poll_attempts=280,
             status_extractor=lambda r: (r.data.task_status if r.data else None),
         )
         return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url))

comfy_api_nodes/nodes_magnific.py

@@ -230,7 +230,6 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode):
             status_extractor=lambda x: x.status,
             price_extractor=lambda _: price_usd,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 
@@ -391,7 +390,6 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode):
             status_extractor=lambda x: x.status,
             price_extractor=lambda _: price_usd,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 
@@ -541,7 +539,6 @@ class MagnificImageStyleTransferNode(IO.ComfyNode):
             response_model=TaskResponse,
             status_extractor=lambda x: x.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 
@@ -782,7 +779,6 @@ class MagnificImageRelightNode(IO.ComfyNode):
             response_model=TaskResponse,
             status_extractor=lambda x: x.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 
@@ -924,7 +920,6 @@ class MagnificImageSkinEnhancerNode(IO.ComfyNode):
             response_model=TaskResponse,
             status_extractor=lambda x: x.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0]))
 

comfy_api_nodes/nodes_topaz.py

@@ -453,7 +453,6 @@ class TopazVideoEnhance(IO.ComfyNode):
             progress_extractor=lambda x: getattr(x, "progress", 0),
             price_extractor=lambda x: (x.estimates.cost[0] * 0.08 if x.estimates and x.estimates.cost[0] else None),
             poll_interval=10.0,
-            max_poll_attempts=320,
         )
         return IO.NodeOutput(await download_url_to_video_output(final_response.download.url))
 

comfy_api_nodes/nodes_vidu.py

@@ -38,7 +38,7 @@ async def execute_task(
     cls: type[IO.ComfyNode],
     vidu_endpoint: str,
     payload: TaskCreationRequest | TaskExtendCreationRequest | TaskMultiFrameCreationRequest,
-    max_poll_attempts: int = 320,
+    max_poll_attempts: int = 480,
 ) -> list[TaskResult]:
     task_creation_response = await sync_op(
         cls,
@@ -1097,7 +1097,6 @@ class ViduExtendVideoNode(IO.ComfyNode):
                 video_url=await upload_video_to_comfyapi(cls, video, wait_label="Uploading video"),
                 images=[image_url] if image_url else None,
             ),
-            max_poll_attempts=480,
         )
         return IO.NodeOutput(await download_url_to_video_output(results[0].url))
 

comfy_api_nodes/nodes_wan.py

@@ -818,7 +818,6 @@ class WanReferenceVideoApi(IO.ComfyNode):
             response_model=VideoTaskStatusResponse,
             status_extractor=lambda x: x.output.task_status,
             poll_interval=6,
-            max_poll_attempts=280,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.output.video_url))
 

comfy_api_nodes/nodes_wavespeed.py

@@ -84,7 +84,6 @@ class WavespeedFlashVSRNode(IO.ComfyNode):
             response_model=TaskResultResponse,
             status_extractor=lambda x: "failed" if x.data is None else x.data.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         if final_response.code != 200:
             raise ValueError(
@@ -156,7 +155,6 @@ class WavespeedImageUpscaleNode(IO.ComfyNode):
             response_model=TaskResultResponse,
             status_extractor=lambda x: "failed" if x.data is None else x.data.status,
             poll_interval=10.0,
-            max_poll_attempts=480,
         )
         if final_response.code != 200:
             raise ValueError(

comfy_api_nodes/util/client.py

@@ -148,7 +148,7 @@ async def poll_op(
     queued_statuses: list[str | int] | None = None,
     data: BaseModel | None = None,
     poll_interval: float = 5.0,
-    max_poll_attempts: int = 160,
+    max_poll_attempts: int = 480,
     timeout_per_poll: float = 120.0,
     max_retries_per_poll: int = 10,
     retry_delay_per_poll: float = 1.0,
@@ -254,7 +254,7 @@ async def poll_op_raw(
     queued_statuses: list[str | int] | None = None,
     data: dict[str, Any] | BaseModel | None = None,
     poll_interval: float = 5.0,
-    max_poll_attempts: int = 160,
+    max_poll_attempts: int = 480,
     timeout_per_poll: float = 120.0,
     max_retries_per_poll: int = 10,
     retry_delay_per_poll: float = 1.0,

comfy_extras/nodes_sdpose.py

@@ -459,27 +459,23 @@ class SDPoseKeypointExtractor(io.ComfyNode):
         total_images = image.shape[0]
         captured_feat = None
 
-        model_h = int(head.heatmap_size[0]) * 4   # e.g. 192 * 4 = 768
-        model_w = int(head.heatmap_size[1]) * 4   # e.g. 256 * 4 = 1024
+        model_w = int(head.heatmap_size[0]) * 4   # 192 * 4 = 768
+        model_h = int(head.heatmap_size[1]) * 4   # 256 * 4 = 1024
 
         def _resize_to_model(imgs):
-            """Aspect-preserving resize + zero-pad BHWC images to (model_h, model_w). Returns (resized_bhwc, scale, pad_top, pad_left)."""
+            """Stretch BHWC images to (model_h, model_w), model expects no aspect preservation."""
             h, w = imgs.shape[-3], imgs.shape[-2]
-            scale = min(model_h / h, model_w / w)
-            sh, sw = int(round(h * scale)), int(round(w * scale))
-            pt, pl = (model_h - sh) // 2, (model_w - sw) // 2
+            method = "area" if (model_h <= h and model_w <= w) else "bilinear"
             chw = imgs.permute(0, 3, 1, 2).float()
-            scaled = comfy.utils.common_upscale(chw, sw, sh, upscale_method="bilinear", crop="disabled")
-            padded = torch.zeros(scaled.shape[0], scaled.shape[1], model_h, model_w, dtype=scaled.dtype, device=scaled.device)
-            padded[:, :, pt:pt + sh, pl:pl + sw] = scaled
-            return padded.permute(0, 2, 3, 1), scale, pt, pl
+            scaled = comfy.utils.common_upscale(chw, model_w, model_h, upscale_method=method, crop="disabled")
+            return scaled.permute(0, 2, 3, 1), model_w / w, model_h / h
 
-        def _remap_keypoints(kp, scale, pad_top, pad_left, offset_x=0, offset_y=0):
+        def _remap_keypoints(kp, scale_x, scale_y, offset_x=0, offset_y=0):
             """Remap keypoints from model space back to original image space."""
             kp = kp.copy() if isinstance(kp, np.ndarray) else np.array(kp, dtype=np.float32)
             invalid = kp[..., 0] < 0
-            kp[..., 0] = (kp[..., 0] - pad_left) / scale + offset_x
-            kp[..., 1] = (kp[..., 1] - pad_top)  / scale + offset_y
+            kp[..., 0] = kp[..., 0] / scale_x + offset_x
+            kp[..., 1] = kp[..., 1] / scale_y + offset_y
             kp[invalid] = -1
             return kp
 
@@ -529,18 +525,18 @@ class SDPoseKeypointExtractor(io.ComfyNode):
                             continue
 
                         crop = img[:, y1:y2, x1:x2, :]  # (1, crop_h, crop_w, C)
-                        crop_resized, scale, pad_top, pad_left = _resize_to_model(crop)
+                        crop_resized, sx, sy = _resize_to_model(crop)
 
                         latent_crop = vae.encode(crop_resized)
                         kp_batch, sc_batch = _run_on_latent(latent_crop)
-                        kp = _remap_keypoints(kp_batch[0], scale, pad_top, pad_left, x1, y1)
+                        kp = _remap_keypoints(kp_batch[0], sx, sy, x1, y1)
                         img_keypoints.append(kp)
                         img_scores.append(sc_batch[0])
                 else:
-                    img_resized, scale, pad_top, pad_left = _resize_to_model(img)
+                    img_resized, sx, sy = _resize_to_model(img)
                     latent_img = vae.encode(img_resized)
                     kp_batch, sc_batch = _run_on_latent(latent_img)
-                    img_keypoints.append(_remap_keypoints(kp_batch[0], scale, pad_top, pad_left))
+                    img_keypoints.append(_remap_keypoints(kp_batch[0], sx, sy))
                     img_scores.append(sc_batch[0])
 
                 all_keypoints.append(img_keypoints)
@@ -549,12 +545,12 @@ class SDPoseKeypointExtractor(io.ComfyNode):
 
         else: # full-image mode, batched
             for batch_start in tqdm(range(0, total_images, batch_size), desc="Extracting keypoints"):
-                batch_resized, scale, pad_top, pad_left = _resize_to_model(image[batch_start:batch_start + batch_size])
+                batch_resized, sx, sy = _resize_to_model(image[batch_start:batch_start + batch_size])
                 latent_batch = vae.encode(batch_resized)
                 kp_batch, sc_batch = _run_on_latent(latent_batch)
 
                 for kp, sc in zip(kp_batch, sc_batch):
-                    all_keypoints.append([_remap_keypoints(kp, scale, pad_top, pad_left)])
+                    all_keypoints.append([_remap_keypoints(kp, sx, sy)])
                     all_scores.append([sc])
 
                 pbar.update(len(kp_batch))
@@ -727,13 +723,13 @@ class CropByBBoxes(io.ComfyNode):
                 scale = min(output_width / crop_w, output_height / crop_h)
                 scaled_w = int(round(crop_w * scale))
                 scaled_h = int(round(crop_h * scale))
-                scaled = comfy.utils.common_upscale(crop_chw, scaled_w, scaled_h, upscale_method="bilinear", crop="disabled")
+                scaled = comfy.utils.common_upscale(crop_chw, scaled_w, scaled_h, upscale_method="area", crop="disabled")
                 pad_left = (output_width  - scaled_w) // 2
                 pad_top  = (output_height - scaled_h) // 2
                 resized = torch.zeros(1, num_ch, output_height, output_width, dtype=image.dtype, device=image.device)
                 resized[:, :, pad_top:pad_top + scaled_h, pad_left:pad_left + scaled_w] = scaled
             else:  # "stretch"
-                resized = comfy.utils.common_upscale(crop_chw, output_width, output_height, upscale_method="bilinear", crop="disabled")
+                resized = comfy.utils.common_upscale(crop_chw, output_width, output_height, upscale_method="area", crop="disabled")
             crops.append(resized)
 
         if not crops:

execution.py

@@ -15,6 +15,7 @@ import torch
 from comfy.cli_args import args
 import comfy.memory_management
 import comfy.model_management
+import comfy.model_prefetch
 import comfy_aimdo.model_vbar
 
 from latent_preview import set_preview_method
@@ -537,6 +538,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
                     if args.verbose == "DEBUG":
                         comfy_aimdo.control.analyze()
                     comfy.model_management.reset_cast_buffers()
+                    comfy.model_prefetch.cleanup_prefetch_queues()
                     comfy_aimdo.model_vbar.vbars_reset_watermark_limits()
 
             if has_pending_tasks:

tests-unit/prompt_server_test/user_manager_test.py

@@ -69,7 +69,11 @@ async def test_listuserdata_full_info(aiohttp_client, app, tmp_path):
     assert len(result) == 1
     assert result[0]["path"] == "file1.txt"
     assert "size" in result[0]
-    assert "modified" in result[0]
+    assert isinstance(result[0]["modified"], int)
+    assert isinstance(result[0]["created"], int)
+    # Verify millisecond magnitude (timestamps after year 2000 in ms are > 946684800000)
+    assert result[0]["modified"] > 946684800000
+    assert result[0]["created"] > 946684800000
 
 
 async def test_listuserdata_split_path(aiohttp_client, app, tmp_path):