Merge f39c9caa55 into 2e9d51680a

This logic was checking comfy_cast_weights, and going straight to
to the forward_comfy_cast_weights implementation without
attempting to downscale input to fp8 in the event comfy_cast_weights
is set.

The main reason comfy_cast_weights would be set would be for async
offload, which is not a good reason to nix FP8MM.

So instead, and together the underlying exclusions for FP8MM which
are:

* having a weight_function (usually LowVramPatch)
* force_cast_weights (compute dtype override)
* the weight is not Quantized
* the input is already quantized
* the model or layer has MM explictily disabled.

If you get past all of those exclusions, quantize the input tensor.
Then hand the new input, quantized or not off to
forward_comfy_cast_weights to handle it. If the weight is offloaded
but input is quantized you will get an offloaded MM8.

comfy/model_patcher.py

@@ -718,6 +718,7 @@ class ModelPatcher:
                             continue
 
                 cast_weight = self.force_cast_weights
+                m.comfy_force_cast_weights = self.force_cast_weights
                 if lowvram_weight:
                     if hasattr(m, "comfy_cast_weights"):
                         m.weight_function = []
@@ -790,11 +791,12 @@ class ModelPatcher:
                 for param in params:
                     self.pin_weight_to_device("{}.{}".format(n, param))
 
+            usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
-                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
+                logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
-                logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
+                logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, mem_counter / (1024 * 1024), full_load))
                 self.model.model_lowvram = False
                 if full_load:
                     self.model.to(device_to)

comfy/ops.py

@@ -654,29 +654,29 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 run_every_op()
 
                 input_shape = input.shape
-                tensor_3d = input.ndim == 3
-
-                if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
-                    return self.forward_comfy_cast_weights(input, *args, **kwargs)
+                reshaped_3d = False
 
                 if (getattr(self, 'layout_type', None) is not None and
-                    not isinstance(input, QuantizedTensor)):
+                    not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
+                    not getattr(self, 'comfy_force_cast_weights', False) and
+                    len(self.weight_function) == 0 and len(self.bias_function) == 0):
 
                     # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others)
-                    if tensor_3d:
-                        input = input.reshape(-1, input_shape[2])
+                    input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input
 
-                    if input.ndim != 2:
-                        # Fall back to comfy_cast_weights for non-2D tensors
-                        return self.forward_comfy_cast_weights(input.reshape(input_shape), *args, **kwargs)
+                    # Fall back to non-quantized for non-2D tensors
+                    if input_reshaped.ndim == 2:
+                        reshaped_3d = input.ndim == 3
+                        # dtype is now implicit in the layout class
+                        scale = getattr(self, 'input_scale', None)
+                        if scale is not None:
+                            scale = comfy.model_management.cast_to_device(scale, input.device, None)
+                        input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale)
 
-                    # dtype is now implicit in the layout class
-                    input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None))
-
-                output = self._forward(input, self.weight, self.bias)
+                output = self.forward_comfy_cast_weights(input)
 
                 # Reshape output back to 3D if input was 3D
-                if tensor_3d:
+                if reshaped_3d:
                     output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0]))
 
                 return output

comfy/quant_ops.py

@@ -19,6 +19,7 @@ try:
         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():

comfy/sd.py

@@ -218,7 +218,7 @@ class CLIP:
             if unprojected:
                 self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-            self.load_model()
+            self.load_model(tokens)
             self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
             all_hooks.reset()
             self.patcher.patch_hooks(None)
@@ -266,7 +266,7 @@ class CLIP:
         if return_pooled == "unprojected":
             self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-        self.load_model()
+        self.load_model(tokens)
         self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
         o = self.cond_stage_model.encode_token_weights(tokens)
         cond, pooled = o[:2]
@@ -299,8 +299,11 @@ class CLIP:
             sd_clip[k] = sd_tokenizer[k]
         return sd_clip
 
-    def load_model(self):
-        model_management.load_model_gpu(self.patcher)
+    def load_model(self, tokens={}):
+        memory_used = 0
+        if hasattr(self.cond_stage_model, "memory_estimation_function"):
+            memory_used = self.cond_stage_model.memory_estimation_function(tokens, device=self.patcher.load_device)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
         return self.patcher
 
     def get_key_patches(self):
@@ -476,8 +479,8 @@ class VAE:
                 self.first_stage_model = comfy.ldm.lightricks.vae.causal_video_autoencoder.VideoVAE(version=version, config=vae_config)
                 self.latent_channels = 128
                 self.latent_dim = 3
-                self.memory_used_decode = lambda shape, dtype: (900 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
-                self.memory_used_encode = lambda shape, dtype: (70 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (1200 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (80 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 32, 32)
                 self.upscale_index_formula = (8, 32, 32)
                 self.downscale_ratio = (lambda a: max(0, math.floor((a + 7) / 8)), 32, 32)

comfy/supported_models.py

@@ -845,7 +845,7 @@ class LTXAV(LTXV):
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
-        self.memory_usage_factor = 0.055  # TODO
+        self.memory_usage_factor = 0.061  # TODO
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.LTXAV(self, device=device)

comfy/text_encoders/lt.py

@@ -98,10 +98,13 @@ class LTXAVTEModel(torch.nn.Module):
 
         out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)
         out_device = out.device
+        if comfy.model_management.should_use_bf16(self.execution_device):
+            out = out.to(device=self.execution_device, dtype=torch.bfloat16)
         out = out.movedim(1, -1).to(self.execution_device)
         out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6)
         out = out.reshape((out.shape[0], out.shape[1], -1))
         out = self.text_embedding_projection(out)
+        out = out.float()
         out_vid = self.video_embeddings_connector(out)[0]
         out_audio = self.audio_embeddings_connector(out)[0]
         out = torch.concat((out_vid, out_audio), dim=-1)
@@ -118,6 +121,14 @@ class LTXAVTEModel(torch.nn.Module):
 
             return self.load_state_dict(sdo, strict=False)
 
+    def memory_estimation_function(self, token_weight_pairs, device=None):
+        constant = 6.0
+        if comfy.model_management.should_use_bf16(device):
+            constant /= 2.0
+
+        token_weight_pairs = token_weight_pairs.get("gemma3_12b", [])
+        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
+        return num_tokens * constant * 1024 * 1024
 
 def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
     class LTXAVTEModel_(LTXAVTEModel):

comfy_api/latest/_io.py

@@ -1113,6 +1113,18 @@ class DynamicSlot(ComfyTypeI):
             out_dict[input_type][finalized_id] = value
             out_dict["dynamic_paths"][finalized_id] = finalize_prefix(curr_prefix, curr_prefix[-1])
 
+@comfytype(io_type="IMAGECOMPARE")
+class ImageCompare(ComfyTypeI):
+  Type = dict
+
+  class Input(WidgetInput):
+      def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None,
+                   socketless: bool=True):
+          super().__init__(id, display_name, optional, tooltip, None, None, socketless)
+
+      def as_dict(self):
+          return super().as_dict()
+
 DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {}
 def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]):
     DYNAMIC_INPUT_LOOKUP[io_type] = func
@@ -1958,4 +1970,5 @@ __all__ = [
     "add_to_dict_v1",
     "add_to_dict_v3",
     "V3Data",
+    "ImageCompare",
 ]

comfy_extras/nodes_image_compare.py

@@ -0,0 +1,53 @@
+import nodes
+
+from typing_extensions import override
+from comfy_api.latest import IO, ComfyExtension
+
+
+class ImageCompare(IO.ComfyNode):
+    """Compares two images with a slider interface."""
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="ImageCompare",
+            display_name="Image Compare",
+            description="Compares two images side by side with a slider.",
+            category="image",
+            is_experimental=True,
+            is_output_node=True,
+            inputs=[
+                IO.Image.Input("image_a", optional=True),
+                IO.Image.Input("image_b", optional=True),
+                IO.ImageCompare.Input("compare_view"),
+            ],
+            outputs=[],
+        )
+
+    @classmethod
+    def execute(cls, image_a=None, image_b=None, compare_view=None) -> IO.NodeOutput:
+        result = {"a_images": [], "b_images": []}
+
+        preview_node = nodes.PreviewImage()
+
+        if image_a is not None and len(image_a) > 0:
+            saved = preview_node.save_images(image_a, "comfy.compare.a")
+            result["a_images"] = saved["ui"]["images"]
+
+        if image_b is not None and len(image_b) > 0:
+            saved = preview_node.save_images(image_b, "comfy.compare.b")
+            result["b_images"] = saved["ui"]["images"]
+
+        return IO.NodeOutput(ui=result)
+
+
+class ImageCompareExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            ImageCompare,
+        ]
+
+
+async def comfy_entrypoint() -> ImageCompareExtension:
+    return ImageCompareExtension()

comfy_extras/nodes_lt_audio.py

@@ -185,6 +185,10 @@ class LTXAVTextEncoderLoader(io.ComfyNode):
                 io.Combo.Input(
                     "ckpt_name",
                     options=folder_paths.get_filename_list("checkpoints"),
+                ),
+                io.Combo.Input(
+                    "device",
+                    options=["default", "cpu"],
                 )
             ],
             outputs=[io.Clip.Output()],
@@ -197,7 +201,11 @@ class LTXAVTextEncoderLoader(io.ComfyNode):
         clip_path1 = folder_paths.get_full_path_or_raise("text_encoders", text_encoder)
         clip_path2 = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name)
 
-        clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type)
+        model_options = {}
+        if device == "cpu":
+            model_options["load_device"] = model_options["offload_device"] = torch.device("cpu")
+
+        clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type, model_options=model_options)
         return io.NodeOutput(clip)
 
 

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.8.0"
+__version__ = "0.8.2"

nodes.py

@@ -2369,6 +2369,7 @@ async def init_builtin_extra_nodes():
         "nodes_nop.py",
         "nodes_kandinsky5.py",
         "nodes_wanmove.py",
+        "nodes_image_compare.py",
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.8.0"
+version = "0.8.2"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.35.9
-comfyui-workflow-templates==0.7.67
+comfyui-workflow-templates==0.7.69
 comfyui-embedded-docs==0.3.1
 torch
 torchsde
@@ -21,7 +21,7 @@ psutil
 alembic
 SQLAlchemy
 av>=14.2.0
-comfy-kitchen>=0.2.3
+comfy-kitchen>=0.2.5
 
 #non essential dependencies:
 kornia>=0.7.1