feat: new ltxv2 nodes for multi gpu and cpu support

[dreamfast]

This PR adds some new nodes to support the recent LTXV2. It is working okay for me although I've just tested the simple nodes only.

Checkpoint node combines the vae and audio_vae into one.

Had worked well to offload the text encoder to CPU which enabled me to generate videos without OOM errors.

Can pick and choose what you like, I had made some changes to logging output but understand if you want to preserve some stuff too.

[IL3DCR]

i need this!

[dreamfast]

i need this!

@IL3DCR https://github.com/dreamfast/ComfyUI-LTX2-MultiGPU - Can try this for now. It's a separate package but maintains compatibility with the current MultiGPU. So both should work together okay. Has the same nodes.

I might have to add GGUF too as that seems to be emerging.

[an80sPWNstar]

i need this!

@IL3DCR https://github.com/dreamfast/ComfyUI-LTX2-MultiGPU - Can try this for now. It's a separate package but maintains compatibility with the current MultiGPU. So both should work together okay. Has the same nodes.

I might have to add GGUF too as that seems to be emerging.

Are you able to integrate the distorch2 part? I am using 3 gpus and am still getting OOM's on LTX2. I have to do two workarounds to get it to even work. With the distorch2 options on other workflows, I'm totally fine.

[fazeaction]

@dreamfast Thank you for this PR. Do you know why using DualClipLoaderMultiGPU I'm getting always this error:

av.error.ValueError: [Errno 22] Invalid argument: 'avcodec_send_frame()'; last error log: [aac] Input contains (near) NaN/+-Inf

It looks like is happening also to other people too: Comfy-Org/ComfyUI#11664 (comment)

I'm trying to use this setup, replacing DualClipLoader with DualClipLoaderMultiGPU (gpu:1). Using x2 T4:

[dreamfast]

@dreamfast Thank you for this PR. Do you know why using DualClipLoaderMultiGPU I'm getting always this error:

av.error.ValueError: [Errno 22] Invalid argument: 'avcodec_send_frame()'; last error log: [aac] Input contains (near) NaN/+-Inf

It looks like is happening also to other people too: Comfy-Org/ComfyUI#11664 (comment)

I'm trying to use this setup, replacing DualClipLoader with DualClipLoaderMultiGPU (gpu:1). Using x2 T4:

I had that exact same issue and was able to get around it by using the 'Video Combine' node, so avoided using the 'Save Video' one.

I also had updated my ffmpeg version to the latest, but unsure if that also was required as I did that first thing.

[MeiYi-dev]

This PR adds some new nodes to support the recent LTXV2. It is working okay for me although I've just tested the simple nodes only.

Checkpoint node combines the vae and audio_vae into one.

Had worked well to offload the text encoder to CPU which enabled me to generate videos without OOM errors.

Can pick and choose what you like, I had made some changes to logging output but understand if you want to preserve some stuff too.

I really don't think we should bloat up this custom node since an official way of doing inference using classic nodes is possible https://huggingface.co/Kijai/LTXV2_comfy

[dreamfast]

This PR adds some new nodes to support the recent LTXV2. It is working okay for me although I've just tested the simple nodes only.
Checkpoint node combines the vae and audio_vae into one.
Had worked well to offload the text encoder to CPU which enabled me to generate videos without OOM errors.
Can pick and choose what you like, I had made some changes to logging output but understand if you want to preserve some stuff too.

I really don't think we should bloat up this custom node since an official way of doing inference using classic nodes is possible https://huggingface.co/Kijai/LTXV2_comfy

Sure. There are alternatives available now. I remember the day LTX2 came out I searched high and low for a way to offload to CPU or GPUs, however there was no options available at the time which led me to create this for my own use.

There's also the distorch2 features from the MultiGPU nodes which people may find useful for LTX2. I'll have some time over the next couple of days to add those to my other package. They are however included in this PR.

[MeiYi-dev]

This PR adds some new nodes to support the recent LTXV2. It is working okay for me although I've just tested the simple nodes only.
Checkpoint node combines the vae and audio_vae into one.
Had worked well to offload the text encoder to CPU which enabled me to generate videos without OOM errors.
Can pick and choose what you like, I had made some changes to logging output but understand if you want to preserve some stuff too.

I really don't think we should bloat up this custom node since an official way of doing inference using classic nodes is possible https://huggingface.co/Kijai/LTXV2_comfy

Sure. There are alternatives available now. I remember the day LTX2 came out I searched high and low for a way to offload to CPU or GPUs, however there was no options available at the time which led me to create this for my own use.

There's also the distorch2 features from the MultiGPU nodes which people may find useful for LTX2. I'll have some time over the next couple of days to add those to my other package. They are however included in this PR.

Yeah the ComfyUI MultiGPU nodes don't work for me with LTX 2, getting some errors. It would be nice to have that work with LTX 2 as well.

[90s0]

This PR adds some new nodes to support the recent LTXV2. It is working okay for me although I've just tested the simple nodes only.
Checkpoint node combines the vae and audio_vae into one.
Had worked well to offload the text encoder to CPU which enabled me to generate videos without OOM errors.
Can pick and choose what you like, I had made some changes to logging output but understand if you want to preserve some stuff too.

I really don't think we should bloat up this custom node since an official way of doing inference using classic nodes is possible https://huggingface.co/Kijai/LTXV2_comfy

Sure. There are alternatives available now. I remember the day LTX2 came out I searched high and low for a way to offload to CPU or GPUs, however there was no options available at the time which led me to create this for my own use.

There's also the distorch2 features from the MultiGPU nodes which people may find useful for LTX2. I'll have some time over the next couple of days to add those to my other package. They are however included in this PR.

Hello,
It seems the pollockjj still hasn’t updated to support LTX2.
Has the distorch2 functionality not yet been implemented in your ComfyUI-LTX2-MultiGPU ?
I hope you could consider adding this feature. Thank you.