Merge 3988f37386 into ac12f77bed

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.

.github/workflows/stable-release.yml

@@ -117,7 +117,7 @@ jobs:
           ./python.exe get-pip.py
           ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/*
 
-          grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
+          grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
           ./python.exe -s -m pip install -r requirements_comfyui.txt
           rm requirements_comfyui.txt
 

.github/workflows/test-ci.yml

@@ -20,6 +20,7 @@ jobs:
   test-stable:
     strategy:
       fail-fast: false
+      max-parallel: 1  # This forces sequential execution
       matrix:
         # os: [macos, linux, windows]
         # os: [macos, linux]
@@ -74,6 +75,7 @@ jobs:
   test-unix-nightly:
     strategy:
       fail-fast: false
+      max-parallel: 1  # This forces sequential execution
       matrix:
         # os: [macos, linux]
         os: [linux]

comfy/model_management.py

@@ -1241,7 +1241,7 @@ def pytorch_attention_enabled():
     return ENABLE_PYTORCH_ATTENTION
 
 def pytorch_attention_enabled_vae():
-    if is_amd():
+    if is_amd() and not amd_min_version(device=None, min_rdna_version=4):
         return False  # enabling pytorch attention on AMD currently causes crash when doing high res
     return pytorch_attention_enabled()
 

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

@@ -427,12 +427,12 @@ def fp8_linear(self, input):
     input = torch.clamp(input, min=-448, max=448, out=input)
     input_fp8 = input.to(dtype).contiguous()
     layout_params_input = TensorCoreFP8Layout.Params(scale=scale_input, orig_dtype=input_dtype, orig_shape=tuple(input_fp8.shape))
-    quantized_input = QuantizedTensor(input_fp8, TensorCoreFP8Layout, layout_params_input)
+    quantized_input = QuantizedTensor(input_fp8, "TensorCoreFP8Layout", layout_params_input)
 
     # Wrap weight in QuantizedTensor - this enables unified dispatch
     # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py!
     layout_params_weight = TensorCoreFP8Layout.Params(scale=scale_weight, orig_dtype=input_dtype, orig_shape=tuple(w.shape))
-    quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight)
+    quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight)
     o = torch.nn.functional.linear(quantized_input, quantized_weight, bias)
 
     uncast_bias_weight(self, w, bias, offload_stream)
@@ -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

@@ -13,6 +13,14 @@ try:
         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}")

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):

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

@@ -36,10 +36,10 @@ class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer):
 
 class Gemma3_12BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
-        llama_scaled_fp8 = model_options.get("gemma_scaled_fp8", None)
-        if llama_scaled_fp8 is not None:
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
             model_options = model_options.copy()
-            model_options["scaled_fp8"] = llama_scaled_fp8
+            model_options["quantization_metadata"] = llama_quantization_metadata
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
@@ -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)
@@ -118,13 +121,21 @@ 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
 
-def ltxav_te(dtype_llama=None, llama_scaled_fp8=None):
+        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):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
+            if llama_quantization_metadata is not None:
                 model_options = model_options.copy()
-                model_options["llama_scaled_fp8"] = llama_scaled_fp8
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
             if dtype_llama is not None:
                 dtype = dtype_llama
             super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)

comfy_api_nodes/nodes_wan.py

@@ -13,7 +13,9 @@ from comfy_api_nodes.util import (
     poll_op,
     sync_op,
     tensor_to_base64_string,
+    upload_video_to_comfyapi,
     validate_audio_duration,
+    validate_video_duration,
 )
 
 
@@ -41,6 +43,12 @@ class Image2VideoInputField(BaseModel):
     audio_url: str | None = Field(None)
 
 
+class Reference2VideoInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    reference_video_urls: list[str] = Field(...)
+
+
 class Txt2ImageParametersField(BaseModel):
     size: str = Field(...)
     n: int = Field(1, description="Number of images to generate.")  # we support only value=1
@@ -76,6 +84,14 @@ class Image2VideoParametersField(BaseModel):
     shot_type: str = Field("single")
 
 
+class Reference2VideoParametersField(BaseModel):
+    size: str = Field(...)
+    duration: int = Field(5, ge=5, le=15)
+    shot_type: str = Field("single")
+    seed: int = Field(..., ge=0, le=2147483647)
+    watermark: bool = Field(False)
+
+
 class Text2ImageTaskCreationRequest(BaseModel):
     model: str = Field(...)
     input: Text2ImageInputField = Field(...)
@@ -100,6 +116,12 @@ class Image2VideoTaskCreationRequest(BaseModel):
     parameters: Image2VideoParametersField = Field(...)
 
 
+class Reference2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Reference2VideoInputField = Field(...)
+    parameters: Reference2VideoParametersField = Field(...)
+
+
 class TaskCreationOutputField(BaseModel):
     task_id: str = Field(...)
     task_status: str = Field(...)
@@ -721,6 +743,143 @@ class WanImageToVideoApi(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_video_output(response.output.video_url))
 
 
+class WanReferenceVideoApi(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="WanReferenceVideoApi",
+            display_name="Wan Reference to Video",
+            category="api node/video/Wan",
+            description="Use the character and voice from input videos, combined with a prompt, "
+            "to generate a new video that maintains character consistency.",
+            inputs=[
+                IO.Combo.Input("model", options=["wan2.6-r2v"]),
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Prompt describing the elements and visual features. Supports English and Chinese. "
+                    "Use identifiers such as `character1` and `character2` to refer to the reference characters.",
+                ),
+                IO.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Negative prompt describing what to avoid.",
+                ),
+                IO.Autogrow.Input(
+                    "reference_videos",
+                    template=IO.Autogrow.TemplateNames(
+                        IO.Video.Input("reference_video"),
+                        names=["character1", "character2", "character3"],
+                        min=1,
+                    ),
+                ),
+                IO.Combo.Input(
+                    "size",
+                    options=[
+                        "720p: 1:1 (960x960)",
+                        "720p: 16:9 (1280x720)",
+                        "720p: 9:16 (720x1280)",
+                        "720p: 4:3 (1088x832)",
+                        "720p: 3:4 (832x1088)",
+                        "1080p: 1:1 (1440x1440)",
+                        "1080p: 16:9 (1920x1080)",
+                        "1080p: 9:16 (1080x1920)",
+                        "1080p: 4:3 (1632x1248)",
+                        "1080p: 3:4 (1248x1632)",
+                    ],
+                ),
+                IO.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=10,
+                    step=5,
+                    display_mode=IO.NumberDisplay.slider,
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                ),
+                IO.Combo.Input(
+                    "shot_type",
+                    options=["single", "multi"],
+                    tooltip="Specifies the shot type for the generated video, that is, whether the video is a "
+                    "single continuous shot or multiple shots with cuts.",
+                ),
+                IO.Boolean.Input(
+                    "watermark",
+                    default=False,
+                    tooltip="Whether to add an AI-generated watermark to the result.",
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        prompt: str,
+        negative_prompt: str,
+        reference_videos: IO.Autogrow.Type,
+        size: str,
+        duration: int,
+        seed: int,
+        shot_type: str,
+        watermark: bool,
+    ):
+        reference_video_urls = []
+        for i in reference_videos:
+            validate_video_duration(reference_videos[i], min_duration=2, max_duration=30)
+        for i in reference_videos:
+            reference_video_urls.append(await upload_video_to_comfyapi(cls, reference_videos[i]))
+        width, height = RES_IN_PARENS.search(size).groups()
+        initial_response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis", method="POST"),
+            response_model=TaskCreationResponse,
+            data=Reference2VideoTaskCreationRequest(
+                model=model,
+                input=Reference2VideoInputField(
+                    prompt=prompt, negative_prompt=negative_prompt, reference_video_urls=reference_video_urls
+                ),
+                parameters=Reference2VideoParametersField(
+                    size=f"{width}*{height}",
+                    duration=duration,
+                    shot_type=shot_type,
+                    watermark=watermark,
+                    seed=seed,
+                ),
+            ),
+        )
+        if not initial_response.output:
+            raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}")
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"),
+            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))
+
+
 class WanApiExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -729,6 +888,7 @@ class WanApiExtension(ComfyExtension):
             WanImageToImageApi,
             WanTextToVideoApi,
             WanImageToVideoApi,
+            WanReferenceVideoApi,
         ]
 
 

comfy_api_nodes/util/upload_helpers.py

@@ -119,7 +119,7 @@ async def upload_video_to_comfyapi(
             raise ValueError(f"Could not verify video duration from source: {e}") from e
 
     upload_mime_type = f"video/{container.value.lower()}"
-    filename = f"uploaded_video.{container.value.lower()}"
+    filename = f"{uuid.uuid4()}.{container.value.lower()}"
 
     # Convert VideoInput to BytesIO using specified container/codec
     video_bytes_io = BytesIO()

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__ = "0.7.0"
+__version__ = "0.8.1"

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.7.0"
+version = "0.8.1"
 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>=0.2.2
+comfy-kitchen>=0.2.5
 
 #non essential dependencies:
 kornia>=0.7.1