torch.compile seems to be broken in wan 2.2 ?

[oioitff]

In the lastest version, torch.compile seems to be broken when I run the wan2.2 14b i2v workflow one more times. It reports errors:

Now the speed is almost the same as not using torch.compile.

After I check the code, this error seems to relate with commit: #9496

If I comment it, the torch.compile works well:

[comfyanonymous]

Works fine here, make sure you are not using easycache.

[oioitff]

Works fine here, make sure you are not using easycache.

@comfyanonymous I'm not using easycache, just official wan2.2 14b i2v workflow + lightx2v lora + torch.compile.
In the lastest 0.3.56 version, the torch.compile reports errors, and the speed is slower than previous versions, sometimes almost the same as not using torch.compile.
After I comment the code relate with commit: #9496, the torch.compile works well, and the speed is as well as previous versions. That means this commit has already broken something. If it's hard to find it out, maybe put this feature to a custom node (just like comfyui-teacache) is a better choice?

[StrongerXi]

@oioitff what's your pytorch version? It's possible that a newer version of pytorch (or nightly) will fix this.

[oioitff]

@oioitff what's your pytorch version? It's possible that a newer version of pytorch (or nightly) will fix this.

Hi @StrongerXi , my pytorch version is 2.7.1+cu128
And just now I update my pytorch version to 2.8.0+cu128 and test again, a terrible result is that they both report errors in torch compile, and the speed become the same slower.
That means in my computer,
pytorch 2.7.1 + previous version comfyui (fastest and not report error)> pytorch 2.7.1 + 0.3.56 version comfyui = pytorch 2.8.0 + 0.3.56 version comfyui = pytorch 2.8.0 + previous version comfyui

(pytorch 2.7.1 + previous version comfyui) does not report error in torch.compile, the speed is the fastest.
The rest three both report errors in torch.compile, their speed are both the same slower.

[jprsyt5]

hmm, you found solutions? @oioitff or maybe you have any idea what should i try @StrongerXi ?

I'm running into a similar issue.

Right now I'm on 0.3.49 with pytorch 2.7.1+cu128. Using Kijai, the workflow runs super fast. But with ComfyUI, it's like 3x slower and sometimes got recompile limit error, and the workflow just stuck.

Then still on 0.3.49, i upgraded to pytorch 2.8 + CU128, now the ComfyUI workflow works fine, but there’s a problem with VRAM usage, blockswap, and torch compile when using the kijaiwrapper. On the first run, VRAM usage always like 98% even with --low-vram, which makes blockswapping really slow. Refer to kijai's comments here

I also tried using the nightly comfyui, with seperate venv, and used pytorch 2.9.0 stable. Kijai workflow runs normal, but now comfyui always runs into OOM errors using the latest commit. I’m running the exact same workflow, kinda frustrating 😂

I forgot which Comfy version I used before, but both the native and Kijai workflows used to run just fine...

[oioitff]

I test again with the latest comfyui 0.12.2 and pytorch 2.10+cu130.
Now the speed is much faster than before even without torch.compile, and torch.compile also works well.
Thanks for your great works!!!

[andypotato]

@oioitff mind sharing your workflow? I am unable to get torch.compile to work on the latest comfyui 0.12.3. Using the same pytorch210+cu130 setup.