TorchCompileModelHyVideo

The node tells you exactly where it stands: "This node has been replaced with TorchCompileModelAdvanced node, please use that instead." If you're building a HunyuanVideo workflow today, that's the one to reach for. This page is for anyone who's inherited an older workflow and wants to know what this node used to do before swapping it out.

HunyuanVideo, for context, is the video model kijai - the same person who maintains KJNodes - built a dedicated wrapper for (ComfyUI-HunyuanVideoWrapper), separately from this general-purpose pack. This node was KJNodes' compile helper aimed specifically at that model family, back before compile support got unified into one architecture-agnostic node. It's a good example of the pattern across this pack: per-model compile nodes existed first (Cosmos, Flux, HyVideo each had their own), and they've since been consolidated.

What it was doing

torch.compile traces a model's computation graph once and fuses the operations into optimized CUDA kernels, instead of PyTorch executing them one at a time in eager mode. The trade-off is upfront cost: the first pass through a freshly compiled model is slower, because that's when the tracing and kernel generation happen. Every render after that, at the same resolution and frame count, reuses the compiled kernels and comes out ahead. For video specifically, where you're grinding through many denoising steps across many frames, that per-step saving compounds - but only across a session where the shape stays fixed. A single one-off clip just pays the tax.

Worth knowing before you go looking for this exact combination: people running torch.compile against HunyuanVideo have reported mixed results depending on hardware and setup - occasionally no measurable speedup at all even with everything configured "correctly," and getting it running in the first place usually meant installing Triton and a C compiler toolchain by hand rather than it working out of the box. It was never a guaranteed win, even when it worked.

Inputs and outputs

By the time this node was deprecated, its schema had been reduced to a bare pass-through: one required model input, typed as a wildcard, and one wildcard output mirroring it back. No exposed backend, mode, or block-selection settings remain - whatever configuration it had while actively maintained isn't part of the current registration. Model in, model out.

How to install it

Same pack as everything else covered here. Through ComfyUI Manager: search "KJNodes for ComfyUI," install, restart. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r ComfyUI-KJNodes/requirements.txt (portable Windows build: python_embeded\python.exe -m pip install -r ...), then restart. Note this is KJNodes, not ComfyUI-HunyuanVideoWrapper - the two are separate repos from the same author, and you'll typically want both if you're running HunyuanVideo at all.

Common issues & troubleshooting

Errors after an update, or the node behaving oddly. Don't debug it - it's not maintained anymore. Move to TorchCompileModelAdvanced and carry on.

"Triton not found" trying to get the old compile pipeline working. torch.compile needs Triton installed, and on Linux usually a C compiler too (build-essential if it's missing). This was the main setup obstacle people hit trying to get any compile node - this one included - working with HunyuanVideo at all.

Windows setups struggle more than Linux ones. There's no official Triton build for Windows; community wheels are the usual path in, and it's the same friction people fight with SageAttention.

No visible speedup even after it compiles successfully. Reported by real users on real hardware - this combination didn't reliably deliver a win on every card. If you're chasing speed on HunyuanVideo today, the current recommendation is TorchCompileModelAdvanced, which is the actively fixed version of the same idea.