Troubleshooting

Every entry here was hit for real. They are ordered by how badly the symptom misdirects you.

pkill killed my SSH session, exit 255

Symptom: you run a stop command over SSH and get bare exit code 255 with no message. ComfyUI is still running.

Cause: pkill -f "main.py --listen 127.0.0.1 --port 8188" matches any process whose command line contains that string — including the shell running the very script that contains it. It kills itself. Exit 255 is SSH reporting that its remote command died.

Fix: stop by pidfile, or by the owner of the port:

ss -lptnH 'sport = :8188'

The RAG replica did not come back

Symptom: after comfy-down, nvidia-smi shows ~5200 MiB instead of the ~7350 MiB you had before. It looks like a half-restore.

Cause: none. The API loads its embedding and reranking models lazily — ~1.3 GB at startup, ~3.0 GB after the first real query.

Fix: trust the health endpoint, not the memory number. comfy-down already polls it and fails loudly if it never answers.

Everything CUDA “works” but generation fails deep in the model

Symptom: torch.cuda.is_available() is True, the device name is right, and then a kernel error appears partway through a run.

Cause: the installed PyTorch has no kernels compiled for your card’s compute capability. Availability is not capability.

Fix:

python -c "import torch; print(torch.cuda.get_arch_list())"   # must contain your sm_XX

torch quietly stopped being the CUDA build

Symptom: things were fine, you updated ComfyUI, now it is slow or CPU-only.

Cause: requirements.txt lists torch unpinned. A resolver that decides to reinstall pulls the generic PyPI build over your CUDA build. Nothing warns you.

Fix: check the version string after any dependency operation. 2.11.0+cu130 is right; a bare 2.11.0 is the generic wheel.

Out of memory partway through, with plenty of VRAM “free”

Cause: free memory at load time is not free memory at peak. Attention working set scales with resolution and batch, and peaks well above the resting figure.

Fix: generate smaller and upscale, drop batch size, or run comfy-up to evict the residents. See The VRAM budget .

The benchmark says the GPU is terrible

Cause: you measured a cold run. The first timed iteration includes cuBLAS initialisation and autotuning.

Fix: warm up ten iterations before timing. On this box that is the difference between a reported 14.1 and 36.6 TFLOPS.

A package manager wants to import a GPG key

Symptom: installing a tool from the distribution repositories stops at an interactive key-import prompt from a vendor unrelated to what you asked for.

Fix: do not accept it to get unblocked. Use the tool’s official self-contained installer into your own home directory instead, which is what the toolchain page does for uv.

ComfyUI is unreachable in the browser

Check in this order:

1. Is the tunnel still up? It dies silently with the terminal that owns it.
2. Is ComfyUI actually listening? ss -lptnH 'sport = :8188' on the GPU box.
3. Did you browse to 127.0.0.1:8188 on your own machine — not the server’s hostname? It binds loopback deliberately and is not reachable any other way.