Merge 2c87be5530 into c5d9edacd0

Currently if the graph contains a cycle, the just inifitiate recursions,
hits a catch all then throws a generic error against the output node
that seeded the validation. Instead, fail the offending cycling mode
chain and handlng it as an error in its own right.

Co-authored-by: guill <jacob.e.segal@gmail.com>

comfy/cli_args.py

@@ -92,6 +92,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 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.")
+parser.add_argument("--enable-triton-backend", action="store_true", help="ComfyUI will enable the use of Triton backend in comfy-kitchen. Is disabled at launch by default.")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"

comfy/ldm/sam3/detector.py

@@ -54,7 +54,7 @@ class SplitMHA(nn.Module):
         if mask is not None and mask.ndim == 2:
             mask = mask[:, None, None, :]  # [B, T] -> [B, 1, 1, T] for SDPA broadcast
         dtype = q.dtype  # manual_cast may produce mixed dtypes
-        out = optimized_attention(q, k.to(dtype), v.to(dtype), self.num_heads, mask=mask)
+        out = optimized_attention(q, k.to(dtype), v.to(dtype), self.num_heads, mask=mask, low_precision_attention=False)
         return self.out_proj(out)
 
 

comfy/ldm/sam3/sam.py

@@ -40,7 +40,7 @@ class SAMAttention(nn.Module):
         q = self.q_proj(q)
         k = self.k_proj(k)
         v = self.v_proj(v)
-        return self.out_proj(optimized_attention(q, k, v, self.num_heads))
+        return self.out_proj(optimized_attention(q, k, v, self.num_heads, low_precision_attention=False))
 
 
 class TwoWayAttentionBlock(nn.Module):
@@ -179,7 +179,7 @@ class Attention(nn.Module):
         q, k, v = qkv.permute(2, 0, 3, 1, 4).unbind(dim=0)
         if self.use_rope and freqs_cis is not None:
             q, k = apply_rope(q, k, freqs_cis)
-        return self.proj(optimized_attention(q, k, v, self.num_heads, skip_reshape=True))
+        return self.proj(optimized_attention(q, k, v, self.num_heads, skip_reshape=True, low_precision_attention=False))
 
 
 class Block(nn.Module):

comfy/ldm/sam3/tracker.py

@@ -364,7 +364,7 @@ class SplitAttn(nn.Module):
         v = self.v_proj(v)
         if rope is not None:
             q, k = apply_rope_memory(q, k, rope, self.num_heads, num_k_exclude_rope)
-        out = optimized_attention(q, k, v, self.num_heads)
+        out = optimized_attention(q, k, v, self.num_heads, low_precision_attention=False)
         return self.out_proj(out)
 
 
@@ -657,7 +657,7 @@ class DecoupledMemoryAttnLayer(nn.Module):
         v = self.self_attn_v_proj(normed)
         if rope is not None:
             q, k = apply_rope_memory(q, k, rope, self.num_heads, 0)
-        x = x + self.self_attn_out_proj(optimized_attention(q, k, v, self.num_heads))
+        x = x + self.self_attn_out_proj(optimized_attention(q, k, v, self.num_heads, low_precision_attention=False))
 
         # Decoupled cross-attention: fuse image and memory projections
         normed = self.norm2(x)
@@ -668,7 +668,7 @@ class DecoupledMemoryAttnLayer(nn.Module):
         v = self.cross_attn_v_proj(memory)
         if rope is not None:
             q, k = apply_rope_memory(q, k, rope, self.num_heads, num_k_exclude_rope)
-        x = x + self.cross_attn_out_proj(optimized_attention(q, k, v, self.num_heads))
+        x = x + self.cross_attn_out_proj(optimized_attention(q, k, v, self.num_heads, low_precision_attention=False))
 
         # FFN
         x = x + self.linear2(F.gelu(self.linear1(self.norm3(x))))

comfy/quant_ops.py

@@ -1,6 +1,8 @@
 import torch
 import logging
 
+from comfy.cli_args import args
+
 try:
     import comfy_kitchen as ck
     from comfy_kitchen.tensor import (
@@ -21,7 +23,15 @@ try:
             ck.registry.disable("cuda")
             logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
-    ck.registry.disable("triton")
+    if args.enable_triton_backend:
+        try:
+            import triton
+            logging.info("Found triton %s. Enabling comfy-kitchen triton backend.", triton.__version__)
+        except ImportError as e:
+            logging.error(f"Failed to import triton, Error: {e}, the comfy-kitchen triton backend will not be available.")
+            ck.registry.disable("triton")
+    else:
+        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:

comfy_api_nodes/nodes_kling.py

@@ -276,6 +276,7 @@ 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))
@@ -862,7 +863,7 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                 ),
                 IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]),
                 IO.Int.Input("duration", default=5, min=3, max=15, display_mode=IO.NumberDisplay.slider),
-                IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True),
+                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p", optional=True),
                 IO.DynamicCombo.Input(
                     "storyboards",
                     options=[
@@ -904,12 +905,13 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                 depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution", "model_name", "generate_audio"]),
                 expr="""
                 (
-                  $mode := (widgets.resolution = "720p") ? "std" : "pro";
+                  $res := widgets.resolution;
+                  $mode := $res = "4k" ? "4k" : ($res = "720p" ? "std" : "pro");
                   $isV3 := $contains(widgets.model_name, "v3");
                   $audio := $isV3 and widgets.generate_audio;
                   $rates := $audio
-                    ? {"std": 0.112, "pro": 0.14}
-                    : {"std": 0.084, "pro": 0.112};
+                    ? {"std": 0.112, "pro": 0.14, "4k": 0.42}
+                    : {"std": 0.084, "pro": 0.112, "4k": 0.42};
                   {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration}
                 )
                 """,
@@ -934,6 +936,8 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                 raise ValueError("kling-video-o1 only supports durations of 5 or 10 seconds.")
             if generate_audio:
                 raise ValueError("kling-video-o1 does not support audio generation.")
+            if resolution == "4k":
+                raise ValueError("kling-video-o1 does not support 4k resolution.")
         stories_enabled = storyboards is not None and storyboards["storyboards"] != "disabled"
         if stories_enabled and model_name == "kling-video-o1":
             raise ValueError("kling-video-o1 does not support storyboards.")
@@ -963,6 +967,12 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                     f"must equal the global duration ({duration}s)."
                 )
 
+        if resolution == "4k":
+            mode = "4k"
+        elif resolution == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         response = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"),
@@ -972,7 +982,7 @@ class OmniProTextToVideoNode(IO.ComfyNode):
                 prompt=prompt,
                 aspect_ratio=aspect_ratio,
                 duration=str(duration),
-                mode="pro" if resolution == "1080p" else "std",
+                mode=mode,
                 multi_shot=multi_shot,
                 multi_prompt=multi_prompt_list,
                 shot_type="customize" if multi_shot else None,
@@ -1014,7 +1024,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                     optional=True,
                     tooltip="Up to 6 additional reference images.",
                 ),
-                IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True),
+                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p", optional=True),
                 IO.DynamicCombo.Input(
                     "storyboards",
                     options=[
@@ -1061,12 +1071,13 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                 depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution", "model_name", "generate_audio"]),
                 expr="""
                 (
-                  $mode := (widgets.resolution = "720p") ? "std" : "pro";
+                  $res := widgets.resolution;
+                  $mode := $res = "4k" ? "4k" : ($res = "720p" ? "std" : "pro");
                   $isV3 := $contains(widgets.model_name, "v3");
                   $audio := $isV3 and widgets.generate_audio;
                   $rates := $audio
-                    ? {"std": 0.112, "pro": 0.14}
-                    : {"std": 0.084, "pro": 0.112};
+                    ? {"std": 0.112, "pro": 0.14, "4k": 0.42}
+                    : {"std": 0.084, "pro": 0.112, "4k": 0.42};
                   {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration}
                 )
                 """,
@@ -1093,6 +1104,8 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                 raise ValueError("kling-video-o1 does not support durations greater than 10 seconds.")
             if generate_audio:
                 raise ValueError("kling-video-o1 does not support audio generation.")
+            if resolution == "4k":
+                raise ValueError("kling-video-o1 does not support 4k resolution.")
         stories_enabled = storyboards is not None and storyboards["storyboards"] != "disabled"
         if stories_enabled and model_name == "kling-video-o1":
             raise ValueError("kling-video-o1 does not support storyboards.")
@@ -1161,6 +1174,12 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                 validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1))
             for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference frame(s)"):
                 image_list.append(OmniParamImage(image_url=i))
+        if resolution == "4k":
+            mode = "4k"
+        elif resolution == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         response = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"),
@@ -1170,7 +1189,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                 prompt=prompt,
                 duration=str(duration),
                 image_list=image_list,
-                mode="pro" if resolution == "1080p" else "std",
+                mode=mode,
                 sound="on" if generate_audio else "off",
                 multi_shot=multi_shot,
                 multi_prompt=multi_prompt_list,
@@ -1204,7 +1223,7 @@ class OmniProImageToVideoNode(IO.ComfyNode):
                     "reference_images",
                     tooltip="Up to 7 reference images.",
                 ),
-                IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True),
+                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p", optional=True),
                 IO.DynamicCombo.Input(
                     "storyboards",
                     options=[
@@ -1251,12 +1270,13 @@ class OmniProImageToVideoNode(IO.ComfyNode):
                 depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution", "model_name", "generate_audio"]),
                 expr="""
                 (
-                  $mode := (widgets.resolution = "720p") ? "std" : "pro";
+                  $res := widgets.resolution;
+                  $mode := $res = "4k" ? "4k" : ($res = "720p" ? "std" : "pro");
                   $isV3 := $contains(widgets.model_name, "v3");
                   $audio := $isV3 and widgets.generate_audio;
                   $rates := $audio
-                    ? {"std": 0.112, "pro": 0.14}
-                    : {"std": 0.084, "pro": 0.112};
+                    ? {"std": 0.112, "pro": 0.14, "4k": 0.42}
+                    : {"std": 0.084, "pro": 0.112, "4k": 0.42};
                   {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration}
                 )
                 """,
@@ -1282,6 +1302,8 @@ class OmniProImageToVideoNode(IO.ComfyNode):
                 raise ValueError("kling-video-o1 does not support durations greater than 10 seconds.")
             if generate_audio:
                 raise ValueError("kling-video-o1 does not support audio generation.")
+            if resolution == "4k":
+                raise ValueError("kling-video-o1 does not support 4k resolution.")
         stories_enabled = storyboards is not None and storyboards["storyboards"] != "disabled"
         if stories_enabled and model_name == "kling-video-o1":
             raise ValueError("kling-video-o1 does not support storyboards.")
@@ -1320,6 +1342,12 @@ class OmniProImageToVideoNode(IO.ComfyNode):
         image_list: list[OmniParamImage] = []
         for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"):
             image_list.append(OmniParamImage(image_url=i))
+        if resolution == "4k":
+            mode = "4k"
+        elif resolution == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         response = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"),
@@ -1330,7 +1358,7 @@ class OmniProImageToVideoNode(IO.ComfyNode):
                 aspect_ratio=aspect_ratio,
                 duration=str(duration),
                 image_list=image_list,
-                mode="pro" if resolution == "1080p" else "std",
+                mode=mode,
                 sound="on" if generate_audio else "off",
                 multi_shot=multi_shot,
                 multi_prompt=multi_prompt_list,
@@ -2860,7 +2888,7 @@ class KlingVideoNode(IO.ComfyNode):
                         IO.DynamicCombo.Option(
                             "kling-v3",
                             [
-                                IO.Combo.Input("resolution", options=["1080p", "720p"]),
+                                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p"),
                                 IO.Combo.Input(
                                     "aspect_ratio",
                                     options=["16:9", "9:16", "1:1"],
@@ -2913,7 +2941,11 @@ class KlingVideoNode(IO.ComfyNode):
                 ),
                 expr="""
                 (
-                  $rates := {"1080p": {"off": 0.112, "on": 0.168}, "720p": {"off": 0.084, "on": 0.126}};
+                  $rates := {
+                    "4k": {"off": 0.42, "on": 0.42},
+                    "1080p": {"off": 0.112, "on": 0.168},
+                    "720p": {"off": 0.084, "on": 0.126}
+                  };
                   $res := $lookup(widgets, "model.resolution");
                   $audio := widgets.generate_audio ? "on" : "off";
                   $rate := $lookup($lookup($rates, $res), $audio);
@@ -2943,7 +2975,12 @@ class KlingVideoNode(IO.ComfyNode):
         start_frame: Input.Image | None = None,
     ) -> IO.NodeOutput:
         _ = seed
-        mode = "pro" if model["resolution"] == "1080p" else "std"
+        if model["resolution"] == "4k":
+            mode = "4k"
+        elif model["resolution"] == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         custom_multi_shot = False
         if multi_shot["multi_shot"] == "disabled":
             shot_type = None
@@ -3025,6 +3062,7 @@ 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))
@@ -3057,7 +3095,7 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
                         IO.DynamicCombo.Option(
                             "kling-v3",
                             [
-                                IO.Combo.Input("resolution", options=["1080p", "720p"]),
+                                IO.Combo.Input("resolution", options=["4k", "1080p", "720p"], default="1080p"),
                             ],
                         ),
                     ],
@@ -3089,7 +3127,11 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
                 ),
                 expr="""
                 (
-                  $rates := {"1080p": {"off": 0.112, "on": 0.168}, "720p": {"off": 0.084, "on": 0.126}};
+                  $rates := {
+                    "4k": {"off": 0.42, "on": 0.42},
+                    "1080p": {"off": 0.112, "on": 0.168},
+                    "720p": {"off": 0.084, "on": 0.126}
+                  };
                   $res := $lookup(widgets, "model.resolution");
                   $audio := widgets.generate_audio ? "on" : "off";
                   $rate := $lookup($lookup($rates, $res), $audio);
@@ -3118,6 +3160,12 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
         validate_image_aspect_ratio(end_frame, (1, 2.5), (2.5, 1))
         image_url = await upload_image_to_comfyapi(cls, first_frame, wait_label="Uploading first frame")
         image_tail_url = await upload_image_to_comfyapi(cls, end_frame, wait_label="Uploading end frame")
+        if model["resolution"] == "4k":
+            mode = "4k"
+        elif model["resolution"] == "1080p":
+            mode = "pro"
+        else:
+            mode = "std"
         response = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/kling/v1/videos/image2video", method="POST"),
@@ -3127,7 +3175,7 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
                 image=image_url,
                 image_tail=image_tail_url,
                 prompt=prompt,
-                mode="pro" if model["resolution"] == "1080p" else "std",
+                mode=mode,
                 duration=str(duration),
                 sound="on" if generate_audio else "off",
             ),
@@ -3140,6 +3188,7 @@ 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_extras/nodes_lt.py

@@ -1,6 +1,7 @@
 import nodes
 import node_helpers
 import torch
+import torchaudio
 import comfy.model_management
 import comfy.model_sampling
 import comfy.samplers
@@ -711,7 +712,14 @@ class LTXVReferenceAudio(io.ComfyNode):
     @classmethod
     def execute(cls, model, positive, negative, reference_audio, audio_vae, identity_guidance_scale, start_percent, end_percent) -> io.NodeOutput:
         # Encode reference audio to latents and patchify
-        audio_latents = audio_vae.encode(reference_audio)
+        sample_rate = reference_audio["sample_rate"]
+        vae_sample_rate = getattr(audio_vae, "audio_sample_rate", 44100)
+        if vae_sample_rate != sample_rate:
+            waveform = torchaudio.functional.resample(reference_audio["waveform"], sample_rate, vae_sample_rate)
+        else:
+            waveform = reference_audio["waveform"]
+
+        audio_latents = audio_vae.encode(waveform.movedim(1, -1))
         b, c, t, f = audio_latents.shape
         ref_tokens = audio_latents.permute(0, 2, 1, 3).reshape(b, t, c * f)
         ref_audio = {"tokens": ref_tokens}

comfy_extras/nodes_preview_any.py

@@ -1,5 +1,6 @@
 import json
 from comfy.comfy_types.node_typing import IO
+import torch
 
 # Preview Any - original implement from
 # https://github.com/rgthree/rgthree-comfy/blob/main/py/display_any.py
@@ -19,6 +20,7 @@ class PreviewAny():
     SEARCH_ALIASES = ["show output", "inspect", "debug", "print value", "show text"]
 
     def main(self, source=None):
+        torch.set_printoptions(edgeitems=6)
         value = 'None'
         if isinstance(source, str):
             value = source
@@ -33,6 +35,7 @@ class PreviewAny():
                 except Exception:
                     value = 'source exists, but could not be serialized.'
 
+        torch.set_printoptions()
         return {"ui": {"text": (value,)}, "result": (value,)}
 
 NODE_CLASS_MAPPINGS = {

execution.py

@@ -811,11 +811,30 @@ class PromptExecutor:
             self._notify_prompt_lifecycle("end", prompt_id)
 
 
-async def validate_inputs(prompt_id, prompt, item, validated):
+async def validate_inputs(prompt_id, prompt, item, validated, visiting=None):
+    if visiting is None:
+        visiting = []
+
     unique_id = item
     if unique_id in validated:
         return validated[unique_id]
 
+    if unique_id in visiting:
+        cycle_path_nodes = visiting[visiting.index(unique_id):] + [unique_id]
+        cycle_nodes = list(dict.fromkeys(cycle_path_nodes))
+        cycle_path = " -> ".join(f"{node_id} ({prompt[node_id]['class_type']})" for node_id in cycle_path_nodes)
+        for node_id in cycle_nodes:
+            validated[node_id] = (False, [{
+                "type": "dependency_cycle",
+                "message": "Dependency cycle detected",
+                "details": cycle_path,
+                "extra_info": {
+                    "node_id": node_id,
+                    "cycle_nodes": cycle_nodes,
+                }
+            }], node_id)
+        return validated[unique_id]
+
     inputs = prompt[unique_id]['inputs']
     class_type = prompt[unique_id]['class_type']
     obj_class = nodes.NODE_CLASS_MAPPINGS[class_type]
@@ -899,7 +918,11 @@ async def validate_inputs(prompt_id, prompt, item, validated):
                 errors.append(error)
                 continue
             try:
-                r = await validate_inputs(prompt_id, prompt, o_id, validated)
+                visiting.append(unique_id)
+                try:
+                    r = await validate_inputs(prompt_id, prompt, o_id, validated, visiting)
+                finally:
+                    visiting.pop()
                 if r[0] is False:
                     # `r` will be set in `validated[o_id]` already
                     valid = False
@@ -1048,10 +1071,13 @@ async def validate_inputs(prompt_id, prompt, item, validated):
                     errors.append(error)
                     continue
 
-    if len(errors) > 0 or valid is not True:
-        ret = (False, errors, unique_id)
-    else:
-        ret = (True, [], unique_id)
+    ret = validated.get(unique_id, (True, [], unique_id))
+    # Recursive cycle detection may have already populated an error on us. Join it.
+    ret = (
+        ret[0] and valid is True and not errors,
+        ret[1] + [error for error in errors if error not in ret[1]],
+        unique_id,
+    )
 
     validated[unique_id] = ret
     return ret

requirements.txt

@@ -1,6 +1,6 @@
 comfyui-frontend-package==1.42.14
-comfyui-workflow-templates==0.9.59
-comfyui-embedded-docs==0.4.3
+comfyui-workflow-templates==0.9.62
+comfyui-embedded-docs==0.4.4
 torch
 torchsde
 torchvision
@@ -23,7 +23,7 @@ SQLAlchemy>=2.0
 filelock
 av>=14.2.0
 comfy-kitchen>=0.2.8
-comfy-aimdo>=0.2.12
+comfy-aimdo==0.2.14
 requests
 simpleeval>=1.0.0
 blake3