Merge f0caa15a17 into 3cd7b32f1b

Added max-parallel setting to enforce sequential execution in test jobs.

.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/ldm/flux/math.py

@@ -21,8 +21,15 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     else:
         device = pos.device
 
-    scale = torch.linspace(0, (dim - 2) / dim, steps=dim//2, dtype=torch.float64, device=device)
-    omega = 1.0 / (theta**scale)
+    if device.type == "musa":
+        # XXX (MUSA): Unsupported tensor dtype in Neg: Double
+        scale = torch.linspace(0, (dim - 2) / dim, steps=dim//2, dtype=torch.float32, device=device)
+        if not isinstance(theta, torch.Tensor):
+            theta = torch.tensor(theta, dtype=torch.float32, device=device)
+        omega = torch.exp(-scale * torch.log(theta + 1e-6))
+    else:
+        scale = torch.linspace(0, (dim - 2) / dim, steps=dim//2, dtype=torch.float64, device=device)
+        omega = 1.0 / (theta**scale)
     out = torch.einsum("...n,d->...nd", pos.to(dtype=torch.float32, device=device), omega)
     out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
     out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)

comfy/model_management.py

@@ -139,6 +139,12 @@ try:
 except:
     ixuca_available = False
 
+try:
+    import torchada  # noqa: F401
+    musa_available = hasattr(torch, "musa") and torch.musa.is_available()
+except:
+    musa_available = False
+
 if args.cpu:
     cpu_state = CPUState.CPU
 
@@ -146,27 +152,24 @@ def is_intel_xpu():
     global cpu_state
     global xpu_available
     if cpu_state == CPUState.GPU:
-        if xpu_available:
-            return True
+        return xpu_available
     return False
 
 def is_ascend_npu():
     global npu_available
-    if npu_available:
-        return True
-    return False
+    return npu_available
 
 def is_mlu():
     global mlu_available
-    if mlu_available:
-        return True
-    return False
+    return mlu_available
 
 def is_ixuca():
     global ixuca_available
-    if ixuca_available:
-        return True
-    return False
+    return ixuca_available
+
+def is_musa():
+    global musa_available
+    return musa_available
 
 def get_torch_device():
     global directml_enabled
@@ -311,7 +314,7 @@ def amd_min_version(device=None, min_rdna_version=0):
     return False
 
 MIN_WEIGHT_MEMORY_RATIO = 0.4
-if is_nvidia():
+if is_nvidia() or is_musa():
     MIN_WEIGHT_MEMORY_RATIO = 0.0
 
 ENABLE_PYTORCH_ATTENTION = False
@@ -320,7 +323,7 @@ if args.use_pytorch_cross_attention:
     XFORMERS_IS_AVAILABLE = False
 
 try:
-    if is_nvidia():
+    if is_nvidia() or is_musa():
         if torch_version_numeric[0] >= 2:
             if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
                 ENABLE_PYTORCH_ATTENTION = True
@@ -375,7 +378,7 @@ if ENABLE_PYTORCH_ATTENTION:
 
 PRIORITIZE_FP16 = False  # TODO: remove and replace with something that shows exactly which dtype is faster than the other
 try:
-    if (is_nvidia() or is_amd()) and PerformanceFeature.Fp16Accumulation in args.fast:
+    if (is_nvidia() or is_amd() or is_musa()) and PerformanceFeature.Fp16Accumulation in args.fast:
         torch.backends.cuda.matmul.allow_fp16_accumulation = True
         PRIORITIZE_FP16 = True  # TODO: limit to cards where it actually boosts performance
         logging.info("Enabled fp16 accumulation.")
@@ -1020,7 +1023,7 @@ if args.async_offload is not None:
     NUM_STREAMS = args.async_offload
 else:
     #  Enable by default on Nvidia and AMD
-    if is_nvidia() or is_amd():
+    if is_nvidia() or is_amd() or is_musa():
         NUM_STREAMS = 2
 
 if args.disable_async_offload:
@@ -1117,7 +1120,7 @@ PINNED_MEMORY = {}
 TOTAL_PINNED_MEMORY = 0
 MAX_PINNED_MEMORY = -1
 if not args.disable_pinned_memory:
-    if is_nvidia() or is_amd():
+    if is_nvidia() or is_amd() or is_musa():
         if WINDOWS:
             MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45  # Windows limit is apparently 50%
         else:
@@ -1261,6 +1264,8 @@ def pytorch_attention_flash_attention():
             return True #if you have pytorch attention enabled on AMD it probably supports at least mem efficient attention
         if is_ixuca():
             return True
+        if is_musa():
+            return True
     return False
 
 def force_upcast_attention_dtype():
@@ -1392,6 +1397,9 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     if torch.version.hip:
         return True
 
+    if is_musa():
+        return True
+
     props = torch.cuda.get_device_properties(device)
     if props.major >= 8:
         return True
@@ -1462,6 +1470,9 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
                 return True
             return False
 
+    if is_musa():
+        return True
+
     props = torch.cuda.get_device_properties(device)
 
     if is_mlu():
@@ -1484,25 +1495,27 @@ def supports_fp8_compute(device=None):
     if SUPPORT_FP8_OPS:
         return True
 
-    if not is_nvidia():
-        return False
-
     props = torch.cuda.get_device_properties(device)
-    if props.major >= 9:
-        return True
-    if props.major < 8:
-        return False
-    if props.minor < 9:
-        return False
-
-    if torch_version_numeric < (2, 3):
-        return False
-
-    if WINDOWS:
-        if torch_version_numeric < (2, 4):
+    if is_nvidia():
+        if props.major >= 9:
+            return True
+        if props.major < 8:
+            return False
+        if props.minor < 9:
             return False
 
-    return True
+        if torch_version_numeric < (2, 3):
+            return False
+
+        if WINDOWS:
+            if torch_version_numeric < (2, 4):
+                return False
+
+    elif is_musa():
+        if props.major >= 3:
+            return True
+
+    return False
 
 def supports_nvfp4_compute(device=None):
     if not is_nvidia():
@@ -1553,7 +1566,7 @@ def unload_all_models():
     free_memory(1e30, get_torch_device())
 
 def debug_memory_summary():
-    if is_amd() or is_nvidia():
+    if is_amd() or is_nvidia() or is_musa():
         return torch.cuda.memory.memory_summary()
     return ""
 

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)

comfy/quant_ops.py

@@ -13,6 +13,13 @@ 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")
+
     ck.registry.disable("triton")
     for k, v in ck.list_backends().items():
         logging.info(f"Found comfy_kitchen backend {k}: {v}")

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)
 
@@ -119,12 +119,12 @@ class LTXAVTEModel(torch.nn.Module):
             return self.load_state_dict(sdo, strict=False)
 
 
-def ltxav_te(dtype_llama=None, llama_scaled_fp8=None):
+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()

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__ = "0.8.0"

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.7.0"
+version = "0.8.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -21,10 +21,11 @@ psutil
 alembic
 SQLAlchemy
 av>=14.2.0
-comfy-kitchen>=0.2.2
+comfy-kitchen>=0.2.3
 
 #non essential dependencies:
 kornia>=0.7.1
 spandrel
 pydantic~=2.0
 pydantic-settings~=2.0
+torchada>=0.1.11