troubleshooting/custom-node-issues

[giscus[bot]]

troubleshooting/custom-node-issues

Fix ComfyUI custom node issues: disable all custom nodes with --disable-all-custom-nodes, then binary-search to isolate, fix, or remove the broken node.

https://docs.comfy.org/troubleshooting/custom-node-issues

[barry7007-max]

ComfyUI Error Report ## Error Details - Node ID: 320:296 - Node Type: LTXVImgToVideoInplace - Exception Type: torch.AcceleratorError - Exception Message: torch.AcceleratorError: CUDA error: operation not supported Search for `cudaErrorNotSupported' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. For more detailed error information, run with CUDA_LOG_FILE=stderr

[lin-bot23]

Hi @barry7007-max, thanks for the error report.

cudaErrorNotSupported is not typically a broken custom node — it means an operation (or allocation mode) that the installed PyTorch/CUDA build tried to use is not supported by your GPU or driver. A few known causes for this exact error in ComfyUI:

1. Outdated NVIDIA driver. Newer PyTorch builds (cu12x wheels) require a relatively recent driver. Updating to the latest NVIDIA driver is the first thing to try.
2. The async CUDA memory allocator. On some GPU/driver combinations ComfyUI's default --cuda-malloc triggers this error; launching with --disable-cuda-malloc has fixed it for other users (see Default arg "cuda-malloc" causes CUDA error: operation not supported on GTX 960M GPU ComfyUI#940).
3. Older GPU architecture. If your card is Pascal (GTX 10xx) or older, it lacks bf16 support and some ops used by recent video models fail with exactly this error. In that case this class of model may not be runnable locally, regardless of the node.

There is also a related open issue on the node's repo: Lightricks/ComfyUI-LTXVideo#337 (CUDA errors in LTXVImgToVideoInplace).

To narrow it down, could you share:

• GPU model + VRAM and driver version (output of nvidia-smi)
• How you launch ComfyUI (Portable / Desktop / manual) and the full console log from startup through the error
• ComfyUI and PyTorch versions (shown in the startup log)

If it turns out the node itself is at fault (after driver/malloc are ruled out), the best place to report it is https://github.com/Lightricks/ComfyUI-LTXVideo/issues so the node author can fix it.

[barry7007-max]

Hi,Thank you for technical support. I have a follow-up question. My graphics card is an NVIDIA GeForce GTX 860M. Is there any version that supports it?Thank you very much, and best wishes for your health and happiness.lin-bot23 ***@***.***> 於 2026年8月30日 21:29 寫道： Hi @barry7007-max, thanks for the error report. cudaErrorNotSupported is not typically a broken custom node — it means an operation (or allocation mode) that the installed PyTorch/CUDA build tried to use is not supported by your GPU or driver. A few known causes for this exact error in ComfyUI: Outdated NVIDIA driver. Newer PyTorch builds (cu12x wheels) require a relatively recent driver. Updating to the latest NVIDIA driver is the first thing to try. The async CUDA memory allocator. On some GPU/driver combinations ComfyUI's default --cuda-malloc triggers this error; launching with --disable-cuda-malloc has fixed it for other users (see Comfy-Org/ComfyUI#940). Older GPU architecture. If your card is Pascal (GTX 10xx) or older, it lacks bf16 support and some ops used by recent video models fail with exactly this error. In that case this class of model may not be runnable locally, regardless of the node. There is also a related open issue on the node's repo: Lightricks/ComfyUI-LTXVideo#337 (CUDA errors in LTXVImgToVideoInplace). To narrow it down, could you share: GPU model + VRAM and driver version (output of nvidia-smi) How you launch ComfyUI (Portable / Desktop / manual) and the full console log from startup through the error ComfyUI and PyTorch versions (shown in the startup log) If it turns out the node itself is at fault (after driver/malloc are ruled out), the best place to report it is https://github.com/Lightricks/ComfyUI-LTXVideo/issues so the node author can fix it. —Reply to this email directly, view it on GitHub, or unsubscribe.Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS and Android. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[lin-bot23]

Hi @barry7007-max, thanks for the follow-up and the kind wishes!

I have some bad news, unfortunately: the GTX 860M is a Kepler-generation chip (2014), and there is no PyTorch/CUDA version you could use with it today:

• NVIDIA ended driver support for Kepler cards in 2021 (the last supported branch is 470.x).
• The newest PyTorch build that still shipped Kepler (sm_3x) kernels was CUDA 10.2; modern ComfyUI needs far newer CUDA/PyTorch than that.
• Newer ComfyUI releases also require capabilities Kepler simply does not have (the card also only has 2-4 GB VRAM, which is far below what current video/image models need).

So cudaErrorNotSupported was exactly what it looked like: your GPU is outside the range of anything ComfyUI can realistically run. No node fix or launch flag will change that.

If you want to keep using ComfyUI on that laptop, your realistic options are:

1. CPU mode (launch with --cpu): it will start, but generation will be extremely slow — minutes per low-resolution image, and most current video models are not practical at all.
2. Comfy Cloud (https://cloud.comfy.org): runs everything in the browser on modern GPUs, nothing depends on your local hardware. This is what I would recommend for the video workflows you were trying.
3. A hardware upgrade whenever that becomes an option: any GTX 16xx / RTX 20xx or newer card changes the situation completely.

Sorry there is not a happier answer for the 860M itself.

[barry7007-max]

Hi lin-bot23 , thanks for the follow-up Regarding the "2.Comfy Cloud" method, does this require paying for points? 4.Do you have any suggestions for teaching workflow? lin-bot23 ***@***.***> 於 2026年9月1日週二 上午5:29寫道：

…

Quick verification note on the [break] fix (commit faffcda <faffcda>): the syntax now matches the Fish Audio official docs (Fine-grained Control <https://docs.fish.audio/developer-guide/core-features/fine-grained-control>), which list (break) under parenthesized special effects. However, the same docs page also uses (break) directly in the phoneme-combining example ("I am, um, an (break) <|phoneme_start|>..."), so the two官方 sources are not fully consistent. We went with the square-bracket form to stay consistent with the rest of this guide, which documents the S2/S2.1-Pro control format. No further changes needed, resolving. — Reply to this email directly, view it on GitHub <#1560?email_source=notifications&email_token=CM6IJ47KBEREEX5ZOZNGIWD5MXUZZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOBSGI2DGNRTUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-18224363>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/CM6IJ4ZIJ7EKTQHGB37MM435MXUZZAVCNFSNUABIKJSXA33TNF2G64TZHM3TONZUGMZDENBRHNCGS43DOVZXG2LPNY5TCMBXGEZTQNRWUF3AE> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/CM6IJ4ZTFCZCSEP5GSJ57YD5MXUZZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOBSGI2DGNRTUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSVGM33PORSXEX3JN5ZQ> and Android <https://github.com/notifications/mobile/android/CM6IJ463ZTT473OLAVEXZX35MXUZZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOBSGI2DGNRTUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSXGM33PORSXEX3BNZSHE33JMQ>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[lin-bot23]

Hi @barry7007-max, happy to clarify both:

1. Does Comfy Cloud require paying for points (credits)?

Not to start with. There are two ways credits come into play:

• Free tier: you get 5 free GPU runs on real GPUs, no credit card required (pricing page). That's enough to try the LTX video workflow you were attempting a few times and see if it works for you.
• Paid plans: after that, you'd need a plan (Standard/Creator/Pro) or top-up credits. Credits are only spent on active GPU time — building or editing workflows is free, and a reference 5-second video on the Wan 2.2 I2V template costs roughly 11 credits (Standard's 4,200 monthly credits ≈ 380 such videos). Credits also cover Partner Nodes.

So: try the 5 free runs first, and only subscribe if Cloud fits your needs.

2. Suggestions for learning/teaching workflows

For someone starting from scratch on modern GPUs (or Cloud), this is the path I'd suggest:

1. Start with the basic tutorials to learn the node/graph mental model: Image to Image and Inpainting are good first steps.
2. Read the core concepts pages (nodes, links, workflow) — understanding why data flows left→right through typed connections makes every template readable.
3. Pick one template and trace it end-to-end: load a workflow template in ComfyUI, and follow one signal (e.g. the image) from loader → conditioning → sampler → VAE decode → save. The LTX-2 video tutorial is a good candidate since it matches what you were trying.
4. Change one variable at a time: resolution, steps, then swap the model. This is the fastest way to learn what each node actually contributes.
5. Then generalize: once you can read one workflow, you can read almost all of them — loaders change, but the loader → encode → sample → decode skeleton is shared by image, video, and audio workflows.

If you hit any specific node you can't make sense of while doing this, feel free to open a discussion under the relevant tutorial page and we'll take a look.