Documentation for socket activation

Also: move github actions out of the way to prevent involuntary execution.

Dockerfile

@@ -0,0 +1,7 @@
+FROM docker.io/mixa3607/pytorch-gfx906:v2.9.0-rocm-7.0.2-20251102214219
+
+WORKDIR /comfyui
+COPY ./requirements.txt ./requirements.txt
+RUN pip install -r requirements.txt
+COPY ./ ./
+ENTRYPOINT ["python", "main.py"] 

README.md

@@ -1,3 +1,10 @@
+# Fork of comfy UI adding systemd socket activation support and a ROCM docker file
+
+**The socket activation support is the result of vibe coding and not sufficiently validated tobe ready for general usage or contributing back upstream. Use at your own risk!**
+
+Instructions for socket activation in [utils/socket_activation](utils/socket_activation/README.md).
+Original readme starts below:
+
 <div align="center">
 
 # ComfyUI

app/user_manager.py

@@ -424,7 +424,7 @@ class UserManager():
                 return source
 
             dest = get_user_data_path(request, check_exists=False, param="dest")
-            if not isinstance(source, str):
+            if not isinstance(dest, str):
                 return dest
 
             overwrite = request.query.get("overwrite", 'true') != "false"

comfy_api_nodes/nodes_kling.py

@@ -4,8 +4,6 @@ For source of truth on the allowed permutations of request fields, please refere
 - [Compatibility Table](https://app.klingai.com/global/dev/document-api/apiReference/model/skillsMap)
 """
 
-from __future__ import annotations
-from typing import Optional, TypeVar
 import math
 import logging
 
@@ -66,9 +64,7 @@ from comfy_api_nodes.util import (
     poll_op,
 )
 from comfy_api.input_impl import VideoFromFile
-from comfy_api.input.basic_types import AudioInput
+from comfy_api.latest import ComfyExtension, IO, Input
-from comfy_api.input.video_types import VideoInput
-from comfy_api.latest import ComfyExtension, IO
 
 KLING_API_VERSION = "v1"
 PATH_TEXT_TO_VIDEO = f"/proxy/kling/{KLING_API_VERSION}/videos/text2video"
@@ -94,8 +90,6 @@ AVERAGE_DURATION_IMAGE_GEN = 32
 AVERAGE_DURATION_VIDEO_EFFECTS = 320
 AVERAGE_DURATION_VIDEO_EXTEND = 320
 
-R = TypeVar("R")
-
 
 MODE_TEXT2VIDEO = {
     "standard mode / 5s duration / kling-v1": ("std", "5", "kling-v1"),
@@ -130,6 +124,8 @@ MODE_START_END_FRAME = {
     "pro mode / 10s duration / kling-v1-6": ("pro", "10", "kling-v1-6"),
     "pro mode / 5s duration / kling-v2-1": ("pro", "5", "kling-v2-1"),
     "pro mode / 10s duration / kling-v2-1": ("pro", "10", "kling-v2-1"),
+    "pro mode / 5s duration / kling-v2-5-turbo": ("pro", "5", "kling-v2-5-turbo"),
+    "pro mode / 10s duration / kling-v2-5-turbo": ("pro", "10", "kling-v2-5-turbo"),
 }
 """
 Returns a mapping of mode strings to their corresponding (mode, duration, model_name) tuples.
@@ -296,7 +292,7 @@ def get_video_from_response(response) -> KlingVideoResult:
     return video
 
 
-def get_video_url_from_response(response) -> Optional[str]:
+def get_video_url_from_response(response) -> str | None:
     """Returns the first video url from the Kling video generation task result.
     Will not raise an error if the response is not valid.
     """
@@ -315,7 +311,7 @@ def get_images_from_response(response) -> list[KlingImageResult]:
     return images
 
 
-def get_images_urls_from_response(response) -> Optional[str]:
+def get_images_urls_from_response(response) -> str | None:
     """Returns the list of image urls from the Kling image generation task result.
     Will not raise an error if the response is not valid. If there is only one image, returns the url as a string. If there are multiple images, returns a list of urls.
     """
@@ -349,7 +345,7 @@ async def execute_text2video(
     model_mode: str,
     duration: str,
     aspect_ratio: str,
-    camera_control: Optional[KlingCameraControl] = None,
+    camera_control: KlingCameraControl | None = None,
 ) -> IO.NodeOutput:
     validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_T2V)
     task_creation_response = await sync_op(
@@ -394,8 +390,8 @@ async def execute_image2video(
     model_mode: str,
     aspect_ratio: str,
     duration: str,
-    camera_control: Optional[KlingCameraControl] = None,
+    camera_control: KlingCameraControl | None = None,
-    end_frame: Optional[torch.Tensor] = None,
+    end_frame: torch.Tensor | None = None,
 ) -> IO.NodeOutput:
     validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_I2V)
     validate_input_image(start_frame)
@@ -451,8 +447,8 @@ async def execute_video_effect(
     model_name: str,
     duration: KlingVideoGenDuration,
     image_1: torch.Tensor,
-    image_2: Optional[torch.Tensor] = None,
+    image_2: torch.Tensor | None = None,
-    model_mode: Optional[KlingVideoGenMode] = None,
+    model_mode: KlingVideoGenMode | None = None,
 ) -> tuple[VideoFromFile, str, str]:
     if dual_character:
         request_input_field = KlingDualCharacterEffectInput(
@@ -499,13 +495,13 @@ async def execute_video_effect(
 
 async def execute_lipsync(
     cls: type[IO.ComfyNode],
-    video: VideoInput,
+    video: Input.Video,
-    audio: Optional[AudioInput] = None,
+    audio: Input.Audio | None = None,
-    voice_language: Optional[str] = None,
+    voice_language: str | None = None,
-    model_mode: Optional[str] = None,
+    model_mode: str | None = None,
-    text: Optional[str] = None,
+    text: str | None = None,
-    voice_speed: Optional[float] = None,
+    voice_speed: float | None = None,
-    voice_id: Optional[str] = None,
+    voice_id: str | None = None,
 ) -> IO.NodeOutput:
     if text:
         validate_string(text, field_name="Text", max_length=MAX_PROMPT_LENGTH_LIP_SYNC)
@@ -787,7 +783,7 @@ class KlingCameraControlT2VNode(IO.ComfyNode):
         negative_prompt: str,
         cfg_scale: float,
         aspect_ratio: str,
-        camera_control: Optional[KlingCameraControl] = None,
+        camera_control: KlingCameraControl | None = None,
     ) -> IO.NodeOutput:
         return await execute_text2video(
             cls,
@@ -854,8 +850,8 @@ class KlingImage2VideoNode(IO.ComfyNode):
         mode: str,
         aspect_ratio: str,
         duration: str,
-        camera_control: Optional[KlingCameraControl] = None,
+        camera_control: KlingCameraControl | None = None,
-        end_frame: Optional[torch.Tensor] = None,
+        end_frame: torch.Tensor | None = None,
     ) -> IO.NodeOutput:
         return await execute_image2video(
             cls,
@@ -965,15 +961,11 @@ class KlingStartEndFrameNode(IO.ComfyNode):
                 IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"),
                 IO.String.Input("negative_prompt", multiline=True, tooltip="Negative text prompt"),
                 IO.Float.Input("cfg_scale", default=0.5, min=0.0, max=1.0),
-                IO.Combo.Input(
+                IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]),
-                    "aspect_ratio",
-                    options=[i.value for i in KlingVideoGenAspectRatio],
-                    default="16:9",
-                ),
                 IO.Combo.Input(
                     "mode",
                     options=modes,
-                    default=modes[2],
+                    default=modes[8],
                     tooltip="The configuration to use for the video generation following the format: mode / duration / model_name.",
                 ),
             ],
@@ -1254,8 +1246,8 @@ class KlingLipSyncAudioToVideoNode(IO.ComfyNode):
     @classmethod
     async def execute(
         cls,
-        video: VideoInput,
+        video: Input.Video,
-        audio: AudioInput,
+        audio: Input.Audio,
         voice_language: str,
     ) -> IO.NodeOutput:
         return await execute_lipsync(
@@ -1314,7 +1306,7 @@ class KlingLipSyncTextToVideoNode(IO.ComfyNode):
     @classmethod
     async def execute(
         cls,
-        video: VideoInput,
+        video: Input.Video,
         text: str,
         voice: str,
         voice_speed: float,
@@ -1471,7 +1463,7 @@ class KlingImageGenerationNode(IO.ComfyNode):
         human_fidelity: float,
         n: int,
         aspect_ratio: KlingImageGenAspectRatio,
-        image: Optional[torch.Tensor] = None,
+        image: torch.Tensor | None = None,
     ) -> IO.NodeOutput:
         validate_string(prompt, field_name="prompt", min_length=1, max_length=MAX_PROMPT_LENGTH_IMAGE_GEN)
         validate_string(negative_prompt, field_name="negative_prompt", max_length=MAX_PROMPT_LENGTH_IMAGE_GEN)

server.py

@@ -37,13 +37,42 @@ from app.user_manager import UserManager
 from app.model_manager import ModelFileManager
 from app.custom_node_manager import CustomNodeManager
 from app.subgraph_manager import SubgraphManager
-from typing import Optional, Union
+from typing import Optional, Union, List, Tuple, Any
 from api_server.routes.internal.internal_routes import InternalRoutes
 from protocol import BinaryEventTypes
 
 # Import cache control middleware
 from middleware.cache_middleware import cache_control
 
+class NoListenSocket(socket.socket):
+    """
+    A socket that pretends to be already listening.
+    The overridden ``listen`` simply returns without calling the kernel.
+    """
+    def listen(self, backlog: int = socket.SOMAXCONN, *args: Any, **kwargs: Any) -> None:   # type: ignore[override]
+        # If the socket is already in LISTEN state the kernel will reject a
+        # second listen() with EPERM (which is what triggered the SELinux
+        # denial).  By turning it into a no‑op we avoid that system call.
+        # The socket is still usable by asyncio because it was created by
+        # systemd with ``listen`` already performed.
+        logging.debug("NoListenSocket.listen() called – no‑op (socket already listening)")
+        return None
+    
+# Helper --------------------------------------------------------------------
+def _fd_to_socket(fd: int) -> socket.socket:
+    """
+    Convert a file‑descriptor received from systemd into a socket that aiohttp
+    can use *without* performing a second listen().
+
+    """
+    try:
+        sock = NoListenSocket(fileno=fd)
+        sock.setblocking(False)
+        return sock
+    except OSError as exc:
+        raise RuntimeError(f"Could not wrap fd {fd} as a NoListenSocket: {exc}") from exc
+
+
 async def send_socket_catch_exception(function, message):
     try:
         await function(message)
@@ -1030,38 +1059,92 @@ class PromptServer():
         await self.start_multi_address([(address, port)], call_on_start=call_on_start)
 
     async def start_multi_address(self, addresses, call_on_start=None, verbose=True):
+        """
+        Starts the aiohttp server.  If systemd activation is detected the
+        provided ``addresses`` are ignored and the sockets returned by
+        ``_get_systemd_sockets`` are used instead.
+        """
         runner = web.AppRunner(self.app, access_log=None)
         await runner.setup()
         ssl_ctx = None
         scheme = "http"
         if args.tls_keyfile and args.tls_certfile:
-            ssl_ctx = ssl.SSLContext(protocol=ssl.PROTOCOL_TLS_SERVER, verify_mode=ssl.CERT_NONE)
+            ssl_ctx = ssl.SSLContext(protocol=ssl.PROTOCOL_TLS_SERVER,
+                                      verify_mode=ssl.CERT_NONE)
             ssl_ctx.load_cert_chain(certfile=args.tls_certfile,
                                     keyfile=args.tls_keyfile)
             scheme = "https"
 
+        # -----------------------------------------------------------------
+        # Systemd activation ------------------------------------------------
+        systemd_sockets = self._get_systemd_sockets()
+        if systemd_sockets:
             if verbose:
-            logging.info("Starting server\n")
+                logging.info("Systemd socket activation detected – using supplied socket(s)")
-        for addr in addresses:
+
-            address = addr[0]
+            for sock in systemd_sockets:
-            port = addr[1]
+                # ``sock.getsockname()`` can be 2‑tuple (IPv4) or 4‑tuple (IPv6)
+                # or a string (Unix).  Normalise for logging.
+                try:
+                    raw_name = sock.getsockname()
+                    if isinstance(raw_name, tuple):
+                        host = raw_name[0]
+                        port = raw_name[1]
+                        if sock.family == socket.AF_INET6:
+                            host = f"[{host}]"
+                    else:
+                        # Unix domain socket – just show the path.
+                        host, port = raw_name, ""
+                except Exception:
+                    host, port = "unknown", "unknown"
+
+                if verbose:
+                    logging.info(f"GUI reachable at: {scheme}://{host}:{port}")
+
+                site = web.SockSite(runner, sock, ssl_context=ssl_ctx)
+                await site.start()
+
+            # Keep a reference – useful for debugging / graceful shutdown
+            self.systemd_sockets = systemd_sockets
+        else:
+            # -----------------------------------------------------------------
+            # Classic TCPSite fallback -----------------------------------------
+            if verbose:
+                logging.info("Systemd activation not detected – falling back to manual bind")
+            for address, port in addresses:
                 site = web.TCPSite(runner, address, port, ssl_context=ssl_ctx)
                 await site.start()
 
-            if not hasattr(self, 'address'):
+                if not hasattr(self, "address"):
-                self.address = address #TODO: remove this
+                    self.address = address
                     self.port = port
 
-            if ':' in address:
+                # Nicely format IPv6 literals for the log line.
-                address_print = "[{}]".format(address)
+                address_print = f"[{address}]" if ":" in address else address
-            else:
-                address_print = address
-
                 if verbose:
-                logging.info("To see the GUI go to: {}://{}:{}".format(scheme, address_print, port))
+                    logging.info(f"GUI reachable at: {scheme}://{address_print}:{port}")
 
-        if call_on_start is not None:
+        # -----------------------------------------------------------------
-            call_on_start(scheme, self.address, self.port)
+        if call_on_start:
+            await call_on_start()
+    
+    def _get_systemd_sockets(self) -> List[socket.socket]:
+        sockets = []
+        if "LISTEN_FDS" not in os.environ or "LISTEN_PID" not in os.environ:
+            return sockets
+        listen_fds = int(os.getenv("LISTEN_FDS", "0"))
+        listen_pid = int(os.getenv("LISTEN_PID", "0"))
+        if listen_pid != os.getpid():
+            return sockets
+        for i in range(listen_fds):
+            fd = 3 + i
+            try:
+                sockets.append(_fd_to_socket(fd))
+            except Exception as e:
+                logging.error(f"Failed to convert fd {fd}: {e}")
+        os.unsetenv("LISTEN_FDS")
+        os.unsetenv("LISTEN_PID")
+        return sockets
 
     def add_on_prompt_handler(self, handler):
         self.on_prompt_handlers.append(handler)

utils/socket_activation/README.md

@@ -0,0 +1,40 @@
+# Rootless podman container with Systemd Socket activation
+
+## Idea
+By passing in the socket from systemd we minimize resource use when not in use.
+Since no other network access is required for operation, we can configure the container
+with network=none and minimize the risk of the AI escaping. 
+
+## Set up 
+
+Optional, if you want to run this as a separate user
+```
+sudo useradd comfy
+sudo machinectl shell comfy@
+```
+
+Check out this repository, navigate to its root directory and build the comfy 
+container with
+```
+podman build -t localhost/comfy:latest .
+```
+
+Place comfy.socket in ` ~/.config/systemd/user`, adjust ports and interfaces if needed.
+Place comfy.container in `~/.config/containers/systemd`. Adjust paths for models and config if desired.
+The files are in `utils/socket_activation`, next to this readme.
+
+Put model files into the models directory (`~/models`).
+
+Start the socket:
+```
+systemctl --user daemon-reload
+systemctl --user enable --now comfy.socket
+```
+
+If you want to run the service also when the user is not logged in, enable lingering:
+```
+sudo loginctl enable-linger <user>
+```
+
+Check that you can access comfy in browser. For troubleshooting, use, e. g., `journalctl -xe`.
+

utils/socket_activation/comfy.container

@@ -0,0 +1,24 @@
+[Unit]
+Description=Comfy ui in a ROCM container
+After=network-online.target
+Wants=network-online.target
+
+[Container]
+Image=localhost/comfy:latest
+#AutoRemove=yes
+#PublishPort=8080:8080
+Environment=ROCR_VISIBLE_DEVICES="GGPU-3b0c81617337ec1b"
+Network=none
+Volume=%h/comfy/models:/comfyui/models:ro,z
+Volume=%h/comfy/input:/comfyui/input:Z
+Volume=%h/comfy/output:/comfyui/output:Z
+Volume=%h/comfy/user:/comfyui/user:Z
+AddDevice=/dev/dri
+AddDevice=/dev/kfd
+Exec= 
+
+#[Service]
+#Restart=always
+
+[Install]
+WantedBy=default.target

utils/socket_activation/comfy.socket

@@ -0,0 +1,8 @@
+[Unit]
+Description=llama socket
+
+[Socket]
+ListenStream=0.0.0.0:8095
+
+[Install]
+WantedBy=default.target

utils/socket_activation/comfy2.container

@@ -0,0 +1,24 @@
+[Unit]
+Description=Comfy ui in a ROCM container
+After=network-online.target
+Wants=network-online.target
+
+[Container]
+Image=localhost/comfy:latest
+#AutoRemove=yes
+#PublishPort=8080:8080
+Network=none
+Environment=ROCR_VISIBLE_DEVICES="GPU-3ce290c173497dfb"
+Volume=%h/comfy/models:/comfyui/models:ro,z
+Volume=%h/comfy/input2:/comfyui/input:Z
+Volume=%h/comfy/output2:/comfyui/output:Z
+Volume=%h/comfy/user2:/comfyui/user:Z
+AddDevice=/dev/dri
+AddDevice=/dev/kfd
+Exec= 
+
+#[Service]
+#Restart=always
+
+[Install]
+WantedBy=default.target

utils/socket_activation/comfy2.socket

@@ -0,0 +1,8 @@
+[Unit]
+Description=Comfy second instance socket
+
+[Socket]
+ListenStream=0.0.0.0:8096
+
+[Install]
+WantedBy=default.target