AetherScale • Runtime

Every enhancement node in AetherScale quietly depends on a private NVIDIA runtime that installs itself - NVIDIA's nvidia-vfx wheel, bootstrapped into the pack's own vendor/ folder on first use so your global Python environment never gets touched. That's elegant until something goes wrong, and when it does, you want this node. Runtime is the management console for that private installation: inspect it, install or update it, repair it, clear its caches, or wipe it entirely. One input, one action dropdown, one JSON status back.

The actions you'll actually use

• status (default) - a full report: is the runtime ready, what version is installed versus requested, where it lives, plus torch/CUDA versions, GPU info from nvidia-smi, effect-cache entry count, available capabilities, and the DLSS5 neural-rendering state. This is your first diagnostic move, always.
• install_or_update - forces a check and pulls the pinned runtime if it's missing. The thing to run when a node error says the runtime isn't there but you thought it was.
• repair - clears the effect cache and force-reinstalls the runtime. The fix for the classic failure mode: you update your NVIDIA driver, and a VFX node starts crashing or producing garbage because its cached effect objects were built against the old runtime.
• clear_effect_cache - just wipes cached effect instances. Cheaper than a full repair; try this when colors or outputs look stale after an upgrade.
• install_dlss5_bridge - sets up the MIT bridge/caller used by the legacy direct Neural Rendering diagnostic backend.
• shutdown_dlss5 - shuts down the DLSS5 worker process cleanly; useful before you close ComfyUI or swap GPUs.
• clear_runtime - nukes everything. The nuclear option: you'll re-download on the next node run, so only use it when the installation is genuinely corrupted.

How to use it in practice

It's a fire-and-forget node - wire nothing, just drop it on the canvas, pick an action, and read the status JSON output. Because it's a plain node, you can keep a Runtime node sitting in a debugging workflow and queue it with different actions while you troubleshoot, rather than hunting through folders.

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-AetherScale.git

That's the install, plus a ComfyUI restart. The pack is Windows/NVIDIA-only, and the runtime download needs internet the first time - this node is how you check whether that first-time bootstrap actually succeeded (runtime_ready: true in the status JSON is the all-clear).

Where people get confused

The most common stumble is expecting the runtime to live with the rest of your Python packages. It doesn't - it's inside the custom node's vendor/ directory, deliberately, so a ComfyUI venv rebuild won't break it and a broken runtime can't poison your other nodes. The second stumble: repair is not a magic "make my GPU faster" button. Run it when a VFX node errors or misbehaves after a driver change; for "will this even run," that's what the Diagnostics node is for. Start with status, escalate to repair, and keep clear_runtime in your back pocket for the corrupted-install day.