fix: prevent --cpu flag from allocating GPU memory

Two root causes fixed:

1. soft_empty_cache() and synchronize() in model_management.py lacked a
   cpu_state == CPUState.CPU guard. They fell through to torch.cuda calls
   that initialize a CUDA context (150-500MB VRAM) even in CPU-only mode.

2. comfy_kitchen is imported unconditionally at startup via quant_ops.py.
   The import chain triggers torch.cuda.is_available() -> cuInit, which
   initializes the CUDA driver. Now gated behind args.cpu check.

Also adds missing QuantizedLayout and register_layout_op fallback stubs
that were absent from the original ImportError handler.

Amp-Thread-ID: https://ampcode.com/threads/T-019cbd03-433e-7601-93ff-3887227496b4

comfy/model_management.py

@@ -1666,16 +1666,12 @@ def lora_compute_dtype(device):
     return dtype
 
 def synchronize():
-    if cpu_mode():
-        return
     if is_intel_xpu():
         torch.xpu.synchronize()
     elif torch.cuda.is_available():
         torch.cuda.synchronize()
 
 def soft_empty_cache(force=False):
-    if cpu_mode():
-        return
     global cpu_state
     if cpu_state == CPUState.MPS:
         torch.mps.empty_cache()

comfy/ops.py

@@ -80,21 +80,6 @@ def cast_to_input(weight, input, non_blocking=False, copy=True):
 
 
 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):
-        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)
-
     offload_stream = None
     xfer_dest = None
 
@@ -675,29 +660,23 @@ class fp8_ops(manual_cast):
 
 CUBLAS_IS_AVAILABLE = False
 try:
-    from cublas_ops import CublasLinear, cublas_half_matmul
+    from cublas_ops import CublasLinear
     CUBLAS_IS_AVAILABLE = True
 except ImportError:
     pass
 
 if CUBLAS_IS_AVAILABLE:
-    class cublas_ops(manual_cast):
-        class Linear(CublasLinear, manual_cast.Linear):
+    class cublas_ops(disable_weight_init):
+        class Linear(CublasLinear, disable_weight_init.Linear):
             def reset_parameters(self):
                 return None
 
             def forward_comfy_cast_weights(self, input):
-                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-                x = cublas_half_matmul(input, weight, bias, self._epilogue_str, self.has_bias)
-                uncast_bias_weight(self, weight, bias, offload_stream)
-                return x
+                return super().forward(input)
 
             def forward(self, *args, **kwargs):
-                run_every_op()
-                if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
-                    return self.forward_comfy_cast_weights(*args, **kwargs)
-                else:
-                    return super().forward(*args, **kwargs)
+                return super().forward(*args, **kwargs)
+
 
 # ==============================================================================
 # Mixed Precision Operations

comfy/quant_ops.py

@@ -1,33 +1,38 @@
 import torch
 import logging
+from comfy.cli_args import args
 
-try:
-    import comfy_kitchen as ck
-    from comfy_kitchen.tensor import (
-        QuantizedTensor,
-        QuantizedLayout,
-        TensorCoreFP8Layout as _CKFp8Layout,
-        TensorCoreNVFP4Layout as _CKNvfp4Layout,
-        register_layout_op,
-        register_layout_class,
-        get_layout_class,
-    )
-    _CK_AVAILABLE = True
-    if torch.version.cuda is None:
-        ck.registry.disable("cuda")
-    else:
-        cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
-        if cuda_version < (13,):
-            ck.registry.disable("cuda")
-            logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
-
-    ck.registry.disable("triton")
-    for k, v in ck.list_backends().items():
-        logging.info(f"Found comfy_kitchen backend {k}: {v}")
-except ImportError as e:
-    logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.")
+if args.cpu:
     _CK_AVAILABLE = False
+else:
+    try:
+        import comfy_kitchen as ck
+        from comfy_kitchen.tensor import (
+            QuantizedTensor,
+            QuantizedLayout,
+            TensorCoreFP8Layout as _CKFp8Layout,
+            TensorCoreNVFP4Layout as _CKNvfp4Layout,
+            register_layout_op,
+            register_layout_class,
+            get_layout_class,
+        )
+        _CK_AVAILABLE = True
+        if torch.version.cuda is None:
+            ck.registry.disable("cuda")
+        else:
+            cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
+            if cuda_version < (13,):
+                ck.registry.disable("cuda")
+                logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
+        ck.registry.disable("triton")
+        for k, v in ck.list_backends().items():
+            logging.info(f"Found comfy_kitchen backend {k}: {v}")
+    except ImportError as e:
+        logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.")
+        _CK_AVAILABLE = False
+
+if not _CK_AVAILABLE:
     class QuantizedTensor:
         pass
 

comfy_api_nodes/apis/grok.py

@@ -7,8 +7,7 @@ class ImageGenerationRequest(BaseModel):
     aspect_ratio: str = Field(...)
     n: int = Field(...)
     seed: int = Field(...)
-    response_format: str = Field("url")
-    resolution: str = Field(...)
+    response_for: str = Field("url")
 
 
 class InputUrlObject(BaseModel):
@@ -17,13 +16,12 @@ class InputUrlObject(BaseModel):
 
 class ImageEditRequest(BaseModel):
     model: str = Field(...)
-    images: list[InputUrlObject] = Field(...)
+    image: InputUrlObject = Field(...)
     prompt: str = Field(...)
     resolution: str = Field(...)
     n: int = Field(...)
     seed: int = Field(...)
-    response_format: str = Field("url")
-    aspect_ratio: str | None = Field(...)
+    response_for: str = Field("url")
 
 
 class VideoGenerationRequest(BaseModel):
@@ -49,13 +47,8 @@ class ImageResponseObject(BaseModel):
     revised_prompt: str | None = Field(None)
 
 
-class UsageObject(BaseModel):
-    cost_in_usd_ticks: int | None = Field(None)
-
-
 class ImageGenerationResponse(BaseModel):
     data: list[ImageResponseObject] = Field(...)
-    usage: UsageObject | None = Field(None)
 
 
 class VideoGenerationResponse(BaseModel):
@@ -72,4 +65,3 @@ class VideoStatusResponse(BaseModel):
     status: str | None = Field(None)
     video: VideoResponseObject | None = Field(None)
     model: str | None = Field(None)
-    usage: UsageObject | None = Field(None)

comfy_api_nodes/apis/kling.py

@@ -148,4 +148,3 @@ class MotionControlRequest(BaseModel):
     keep_original_sound: str = Field(...)
     character_orientation: str = Field(...)
     mode: str = Field(..., description="'pro' or 'std'")
-    model_name: str = Field(...)

comfy_api_nodes/nodes_grok.py

@@ -27,12 +27,6 @@ from comfy_api_nodes.util import (
 )
 
 
-def _extract_grok_price(response) -> float | None:
-    if response.usage and response.usage.cost_in_usd_ticks is not None:
-        return response.usage.cost_in_usd_ticks / 10_000_000_000
-    return None
-
-
 class GrokImageNode(IO.ComfyNode):
 
     @classmethod
@@ -43,10 +37,7 @@ class GrokImageNode(IO.ComfyNode):
             category="api node/image/Grok",
             description="Generate images using Grok based on a text prompt",
             inputs=[
-                IO.Combo.Input(
-                    "model",
-                    options=["grok-imagine-image-pro", "grok-imagine-image", "grok-imagine-image-beta"],
-                ),
+                IO.Combo.Input("model", options=["grok-imagine-image-beta"]),
                 IO.String.Input(
                     "prompt",
                     multiline=True,
@@ -90,7 +81,6 @@ class GrokImageNode(IO.ComfyNode):
                     tooltip="Seed to determine if node should re-run; "
                     "actual results are nondeterministic regardless of seed.",
                 ),
-                IO.Combo.Input("resolution", options=["1K", "2K"], optional=True),
             ],
             outputs=[
                 IO.Image.Output(),
@@ -102,13 +92,8 @@ class GrokImageNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images"]),
-                expr="""
-                (
-                  $rate := $contains(widgets.model, "pro") ? 0.07 : 0.02;
-                  {"type":"usd","usd": $rate * widgets.number_of_images}
-                )
-                """,
+                depends_on=IO.PriceBadgeDepends(widgets=["number_of_images"]),
+                expr="""{"type":"usd","usd":0.033 * widgets.number_of_images}""",
             ),
         )
 
@@ -120,7 +105,6 @@ class GrokImageNode(IO.ComfyNode):
         aspect_ratio: str,
         number_of_images: int,
         seed: int,
-        resolution: str = "1K",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
         response = await sync_op(
@@ -132,10 +116,8 @@ class GrokImageNode(IO.ComfyNode):
                 aspect_ratio=aspect_ratio,
                 n=number_of_images,
                 seed=seed,
-                resolution=resolution.lower(),
             ),
             response_model=ImageGenerationResponse,
-            price_extractor=_extract_grok_price,
         )
         if len(response.data) == 1:
             return IO.NodeOutput(await download_url_to_image_tensor(response.data[0].url))
@@ -156,17 +138,14 @@ class GrokImageEditNode(IO.ComfyNode):
             category="api node/image/Grok",
             description="Modify an existing image based on a text prompt",
             inputs=[
-                IO.Combo.Input(
-                    "model",
-                    options=["grok-imagine-image-pro", "grok-imagine-image", "grok-imagine-image-beta"],
-                ),
-                IO.Image.Input("image", display_name="images"),
+                IO.Combo.Input("model", options=["grok-imagine-image-beta"]),
+                IO.Image.Input("image"),
                 IO.String.Input(
                     "prompt",
                     multiline=True,
                     tooltip="The text prompt used to generate the image",
                 ),
-                IO.Combo.Input("resolution", options=["1K", "2K"]),
+                IO.Combo.Input("resolution", options=["1K"]),
                 IO.Int.Input(
                     "number_of_images",
                     default=1,
@@ -187,27 +166,6 @@ class GrokImageEditNode(IO.ComfyNode):
                     tooltip="Seed to determine if node should re-run; "
                     "actual results are nondeterministic regardless of seed.",
                 ),
-                IO.Combo.Input(
-                    "aspect_ratio",
-                    options=[
-                        "auto",
-                        "1:1",
-                        "2:3",
-                        "3:2",
-                        "3:4",
-                        "4:3",
-                        "9:16",
-                        "16:9",
-                        "9:19.5",
-                        "19.5:9",
-                        "9:20",
-                        "20:9",
-                        "1:2",
-                        "2:1",
-                    ],
-                    optional=True,
-                    tooltip="Only allowed when multiple images are connected to the image input.",
-                ),
             ],
             outputs=[
                 IO.Image.Output(),
@@ -219,13 +177,8 @@ class GrokImageEditNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images"]),
-                expr="""
-                (
-                  $rate := $contains(widgets.model, "pro") ? 0.07 : 0.02;
-                  {"type":"usd","usd": 0.002 + $rate * widgets.number_of_images}
-                )
-                """,
+                depends_on=IO.PriceBadgeDepends(widgets=["number_of_images"]),
+                expr="""{"type":"usd","usd":0.002 + 0.033 * widgets.number_of_images}""",
             ),
         )
 
@@ -238,32 +191,22 @@ class GrokImageEditNode(IO.ComfyNode):
         resolution: str,
         number_of_images: int,
         seed: int,
-        aspect_ratio: str = "auto",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
-        if model == "grok-imagine-image-pro":
-            if get_number_of_images(image) > 1:
-                raise ValueError("The pro model supports only 1 input image.")
-        elif get_number_of_images(image) > 3:
-            raise ValueError("A maximum of 3 input images is supported.")
-        if aspect_ratio != "auto" and get_number_of_images(image) == 1:
-            raise ValueError(
-                "Custom aspect ratio is only allowed when multiple images are connected to the image input."
-            )
+        if get_number_of_images(image) != 1:
+            raise ValueError("Only one input image is supported.")
         response = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/xai/v1/images/edits", method="POST"),
             data=ImageEditRequest(
                 model=model,
-                images=[InputUrlObject(url=f"data:image/png;base64,{tensor_to_base64_string(i)}") for i in image],
+                image=InputUrlObject(url=f"data:image/png;base64,{tensor_to_base64_string(image)}"),
                 prompt=prompt,
                 resolution=resolution.lower(),
                 n=number_of_images,
                 seed=seed,
-                aspect_ratio=None if aspect_ratio == "auto" else aspect_ratio,
             ),
             response_model=ImageGenerationResponse,
-            price_extractor=_extract_grok_price,
         )
         if len(response.data) == 1:
             return IO.NodeOutput(await download_url_to_image_tensor(response.data[0].url))
@@ -284,7 +227,7 @@ class GrokVideoNode(IO.ComfyNode):
             category="api node/video/Grok",
             description="Generate video from a prompt or an image",
             inputs=[
-                IO.Combo.Input("model", options=["grok-imagine-video", "grok-imagine-video-beta"]),
+                IO.Combo.Input("model", options=["grok-imagine-video-beta"]),
                 IO.String.Input(
                     "prompt",
                     multiline=True,
@@ -332,11 +275,10 @@ class GrokVideoNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"], inputs=["image"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["duration"], inputs=["image"]),
                 expr="""
                 (
-                  $rate := widgets.resolution = "720p" ? 0.07 : 0.05;
-                  $base := $rate * widgets.duration;
+                  $base := 0.181 * widgets.duration;
                   {"type":"usd","usd": inputs.image.connected ? $base + 0.002 : $base}
                 )
                 """,
@@ -379,7 +321,6 @@ class GrokVideoNode(IO.ComfyNode):
             ApiEndpoint(path=f"/proxy/xai/v1/videos/{initial_response.request_id}"),
             status_extractor=lambda r: r.status if r.status is not None else "complete",
             response_model=VideoStatusResponse,
-            price_extractor=_extract_grok_price,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.video.url))
 
@@ -394,7 +335,7 @@ class GrokVideoEditNode(IO.ComfyNode):
             category="api node/video/Grok",
             description="Edit an existing video based on a text prompt.",
             inputs=[
-                IO.Combo.Input("model", options=["grok-imagine-video", "grok-imagine-video-beta"]),
+                IO.Combo.Input("model", options=["grok-imagine-video-beta"]),
                 IO.String.Input(
                     "prompt",
                     multiline=True,
@@ -423,7 +364,7 @@ class GrokVideoEditNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                expr="""{"type":"usd","usd": 0.06, "format": {"suffix": "/sec", "approximate": true}}""",
+                expr="""{"type":"usd","usd": 0.191, "format": {"suffix": "/sec", "approximate": true}}""",
             ),
         )
 
@@ -457,7 +398,6 @@ class GrokVideoEditNode(IO.ComfyNode):
             ApiEndpoint(path=f"/proxy/xai/v1/videos/{initial_response.request_id}"),
             status_extractor=lambda r: r.status if r.status is not None else "complete",
             response_model=VideoStatusResponse,
-            price_extractor=_extract_grok_price,
         )
         return IO.NodeOutput(await download_url_to_video_output(response.video.url))
 

comfy_api_nodes/nodes_kling.py

@@ -2747,7 +2747,6 @@ class MotionControl(IO.ComfyNode):
                     "but the character orientation matches the reference image (camera/other details via prompt).",
                 ),
                 IO.Combo.Input("mode", options=["pro", "std"]),
-                IO.Combo.Input("model", options=["kling-v3", "kling-v2-6"], optional=True),
             ],
             outputs=[
                 IO.Video.Output(),
@@ -2778,7 +2777,6 @@ class MotionControl(IO.ComfyNode):
         keep_original_sound: bool,
         character_orientation: str,
         mode: str,
-        model: str = "kling-v2-6",
     ) -> IO.NodeOutput:
         validate_string(prompt, max_length=2500)
         validate_image_dimensions(reference_image, min_width=340, min_height=340)
@@ -2799,7 +2797,6 @@ class MotionControl(IO.ComfyNode):
                 keep_original_sound="yes" if keep_original_sound else "no",
                 character_orientation=character_orientation,
                 mode=mode,
-                model_name=model,
             ),
         )
         if response.code:

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.16.2"
+__version__ = "0.15.1"

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.16.2"
+version = "0.15.1"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.39.19
-comfyui-workflow-templates==0.9.10
+comfyui-workflow-templates==0.9.5
 comfyui-embedded-docs==0.4.3
 torch
 torchsde
@@ -22,7 +22,7 @@ alembic
 SQLAlchemy
 av>=14.2.0
 comfy-kitchen>=0.2.7
-comfy-aimdo>=0.2.7
+comfy-aimdo>=0.2.6
 requests
 
 #non essential dependencies: