I got "RuntimeError: Module MetricsHead failed to start" error

[maczolaci]

Describe the bug
Comfyui halt at "Ray Actor Init" node with this error message:

"VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
2025-11-30 14:31:56,832 WARNING worker.py:1686 -- SIGTERM handler is not set because current thread is not the main thread.
2025-11-30 14:33:08,080 ERROR services.py:1342 -- Failed to start the dashboard
2025-11-30 14:33:08,081 ERROR services.py:1367 -- Error should be written to 'dashboard.log' or 'dashboard.err'. We are printing the last 20 lines for you. See 'https://docs.ray.io/en/master/ray-observability/user-guides/configure-logging.html#logging-directory-structure' to find where the log file is.
2025-11-30 14:33:08,082 ERROR services.py:1411 --
The last 20 lines of C:\WINDOWS\TEMP\ray\session_2025-11-30_14-31-57_173499_4204\logs\dashboard.log (it contains the error message from the dashboard):
2025-11-30 14:33:10,572 INFO worker.py:2023 -- Started a local Ray instance.
2025-11-30 14:33:20,437 INFO packaging.py:588 -- Creating a file package for local module '\KITT_P_drive\ComfyUI_windows_portable_nvidia_cu128\ComfyUI\custom_nodes\raylight\src\raylight'.
2025-11-30 14:33:38,499 INFO packaging.py:380 -- Pushing file package 'gcs://_ray_pkg_89f60d63f2921fa8.zip' (0.51MiB) to Ray cluster...
2025-11-30 14:33:38,504 INFO packaging.py:393 -- Successfully pushed file package 'gcs://_ray_pkg_89f60d63f2921fa8.zip'."

When I checked the the dashboard.log file, I found this error:

"2025-11-30 14:33:57,071 ERROR dashboard.py:308 -- The dashboard on node KITT failed with the following error:
Traceback (most recent call last):
File "p:\ComfyUI_windows_portable_nvidia_cu128\python_embeded\Lib\site-packages\ray\dashboard\dashboard.py", line 297, in
loop.run_until_complete(dashboard.run())
File "asyncio\base_events.py", line 691, in run_until_complete
File "p:\ComfyUI_windows_portable_nvidia_cu128\python_embeded\Lib\site-packages\ray\dashboard\dashboard.py", line 98, in run
await self.dashboard_head.run()
File "p:\ComfyUI_windows_portable_nvidia_cu128\python_embeded\Lib\site-packages\ray\dashboard\head.py", line 405, in run
handle.wait_for_module_ready()
File "p:\ComfyUI_windows_portable_nvidia_cu128\python_embeded\Lib\site-packages\ray\dashboard\subprocesses\handle.py", line 153, in wait_for_module_ready
raise RuntimeError(
RuntimeError: Module MetricsHead failed to start. Timeout after 30.0 seconds."

System

• Win 10
• PyTorch 2.7 and its CUDA 12.8 Version.
• 2 3090 GPU and 64GB System RAM

[komikndr]

There are two issues, and one of them is easy to solve.

Your Ray cluster is trying to start a local Ray dashboard.
You can fix this by installing: pip install -U "ray[default]".
The dashboard is not an essential component and is not used by Raylight. It is only for metrics, so the system can run without it.

You are running on Windows, which Raylight does not officially support.
NCCL only works on Linux. On Windows, Ray falls back to GLOO, which is much slower. This also causes the RayCOMM tester to stall because it cannot communicate properly across GPUs.
You can test using WSL instead, where NCCL works and communication should be correct.

[maczolaci]

Thank you for your response.

I updated the ray package. Nothing changed. And when the raylight doesn't use dashboard, why I got this error?

Alright I will try WSL, but running raylight on windows also doesn't cause halting the Ray Init Actor node. I think. It should be run slower due to missing NCCL support on windows.