Stuck on Model Loading

[koopa990]

From console

Loading Model 'ckpts\ltx-2-19b-distilled_Q6_K.gguf'

[koopa990]

[vadhorvad]

This issue usually happens if the backend gets stuck loading a heavy GGUF model due to a missing or misconfigured kernel, or if the system runs out of RAM during the process. When the backend locks up, the Gradio interface stops responding, which is why the "Abort" button hangs indefinitely.

Try these steps to get it running:
Force restart: Completely close your terminal/console to kill the frozen Python process.
Update your kernels: Before launching again, run the update script to make sure all optimized kernels (like GGUF, Triton, or Flash Attention) are properly built and up to date:

• Windows: Run scripts\update.bat and choose Upgrade.

• Linux/macOS: Run scripts/update.sh.

Clear VRAM/RAM: Once the app restarts, use the Unload All button on the WanGP toolbar to free up memory before loading the LTX-2 model again.

That being said, there is too little information here to pinpoint the exact cause. To give you a more specific answer, we need more context. If the issue persists, please share your hardware specs (GPU, VRAM, RAM) and a screenshot or log dump of your actual terminal/console output where the error occurs.

[koopa990]

This issue usually happens if the backend gets stuck loading a heavy GGUF model due to a missing or misconfigured kernel, or if the system runs out of RAM during the process. When the backend locks up, the Gradio interface stops responding, which is why the "Abort" button hangs indefinitely.

Try these steps to get it running:
Force restart: Completely close your terminal/console to kill the frozen Python process.
Update your kernels: Before launching again, run the update script to make sure all optimized kernels (like GGUF, Triton, or Flash Attention) are properly built and up to date:

• Windows: Run scripts\update.bat and choose Upgrade.

• Linux/macOS: Run scripts/update.sh.

Clear VRAM/RAM: Once the app restarts, use the Unload All button on the WanGP toolbar to free up memory before loading the LTX-2 model again.

That being said, there is too little information here to pinpoint the exact cause. To give you a more specific answer, we need more context. If the issue persists, please share your hardware specs (GPU, VRAM, RAM) and a screenshot or log dump of your actual terminal/console output where the error occurs.

Quick question. Isn't GGUF model the least intensive resource version of ltx?

I'm also running wan2gp on stability matrix. So I don't see update script in the folder

[koopa990]

[vadhorvad]

Hi @koopa990, thanks for sharing the console output! That actually points exactly to the problem.

Just a heads-up: I am not a deep expert on this, just a beginner trying to help out based on what I have seen!

You are absolutely right that GGUF models are designed to be much less resource-intensive. However, if you look closely at your log, you will see this line:

[GGUF][llama.cpp CUDA] kernels unavailable, using fallback

Because the required hardware-accelerated CUDA kernels for GGUF are missing from your environment, WanGP is forced to use the CPU fallback. This fallback is extremely slow and resource-heavy, which is why the loading process hangs indefinitely and the UI stops responding.

Regarding Stability Matrix: you are correct. Stability Matrix manages its own isolated virtual environments, so it does not use the standard scripts\update.bat file that comes with the standalone WanGP installation.

How to fix this:

• Option 1: Switch to an INT8 model (Highly Recommended)
  Your logs also show [Quanto][INT8] Injected int8 kernels ACTIVE (backend=triton). This means your environment is already perfectly configured for INT8 quantization! The easiest fix is to simply download the INT8 version of the LTX-2 model instead of the GGUF version. It should load right up.

• Option 2: Manually install the GGUF kernels
  If you specifically need to use the GGUF format, you will have to open the dedicated terminal for the Wan2GP package inside Stability Matrix and manually install the missing llama.cpp CUDA kernels.

[koopa990]

Hi @koopa990, thanks for sharing the console output! That actually points exactly to the problem.

Just a heads-up: I am not a deep expert on this, just a beginner trying to help out based on what I have seen!

You are absolutely right that GGUF models are designed to be much less resource-intensive. However, if you look closely at your log, you will see this line:

[GGUF][llama.cpp CUDA] kernels unavailable, using fallback

Because the required hardware-accelerated CUDA kernels for GGUF are missing from your environment, WanGP is forced to use the CPU fallback. This fallback is extremely slow and resource-heavy, which is why the loading process hangs indefinitely and the UI stops responding.

Regarding Stability Matrix: you are correct. Stability Matrix manages its own isolated virtual environments, so it does not use the standard scripts\update.bat file that comes with the standalone WanGP installation.

How to fix this:

• Option 1: Switch to an INT8 model (Highly Recommended)
  Your logs also show [Quanto][INT8] Injected int8 kernels ACTIVE (backend=triton). This means your environment is already perfectly configured for INT8 quantization! The easiest fix is to simply download the INT8 version of the LTX-2 model instead of the GGUF version. It should load right up.

• Option 2: Manually install the GGUF kernels
  If you specifically need to use the GGUF format, you will have to open the dedicated terminal for the Wan2GP package inside Stability Matrix and manually install the missing llama.cpp CUDA kernels.

Do you recommend me to use wan2gp on its own instead of SM?

in other case, is my RTX 3090 good for LTX 2? I once tried out 2.3 on comfy, but got nowhere in progress. I suppose GGUF doesn't do anything for my GPU

[vadhorvad]

Hi @koopa990, thanks for sharing the console output! That actually points exactly to the problem.
Just a heads-up: I am not a deep expert on this, just a beginner trying to help out based on what I have seen!
You are absolutely right that GGUF models are designed to be much less resource-intensive. However, if you look closely at your log, you will see this line:
[GGUF][llama.cpp CUDA] kernels unavailable, using fallback
Because the required hardware-accelerated CUDA kernels for GGUF are missing from your environment, WanGP is forced to use the CPU fallback. This fallback is extremely slow and resource-heavy, which is why the loading process hangs indefinitely and the UI stops responding.
Regarding Stability Matrix: you are correct. Stability Matrix manages its own isolated virtual environments, so it does not use the standard scripts\update.bat file that comes with the standalone WanGP installation.

How to fix this:

• Option 1: Switch to an INT8 model (Highly Recommended)
  Your logs also show [Quanto][INT8] Injected int8 kernels ACTIVE (backend=triton). This means your environment is already perfectly configured for INT8 quantization! The easiest fix is to simply download the INT8 version of the LTX-2 model instead of the GGUF version. It should load right up.
• Option 2: Manually install the GGUF kernels
  If you specifically need to use the GGUF format, you will have to open the dedicated terminal for the Wan2GP package inside Stability Matrix and manually install the missing llama.cpp CUDA kernels.

Do you recommend me to use wan2gp on its own instead of SM?

in other case, is my RTX 3090 good for LTX 2? I once tried out 2.3 on comfy, but got nowhere in progress. I suppose GGUF doesn't do anything for my GPU

Glad I could help point you in the right direction.

To answer your questions:

1. Standalone Wan2GP vs. Stability Matrix

I definitely recommend using Wan2GP on its own using the standalone installation. The 1-click automated scripts for Windows (.bat) make installation, environment management, and updates as seamless as possible. These scripts automatically detect your hardware and will install the best acceleration kernels (like Triton, Flash Attention, and even the GGUF kernels) specifically for your config. It saves a ton of manual troubleshooting!

2. Is the RTX 3090 good for LTX-2?

Your RTX 3090 is an absolute beast for this! It has 24GB of VRAM. WanGP is specifically designed to be accessible to the "GPU Poor" and can run select models with as little as 6 GB of VRAM. Because of WanGP's heavy VRAM optimizations under the hood, your 3090 will easily handle LTX-2 and LTX-2.3. If you struggled in ComfyUI, you will likely have a much smoother experience here. For context, I run it without any problems on my RTX 3060 with 12GB VRAM and 64GB of system RAM, so your 3090 will absolutely crush it!

3. Does GGUF do anything for your GPU?

You are totally onto something here. GGUF is a lifesaver for people who need to squeeze large models into cards with low VRAM. But since your RTX 3090 has a massive 24GB of VRAM, you don't even need the extreme compression of GGUF to fit LTX-2. You have more than enough room to run the INT8 models (which your console showed you already have the kernels for), and they will run extremely fast on your hardware.

[koopa990]

Hi @koopa990, thanks for sharing the console output! That actually points exactly to the problem.
Just a heads-up: I am not a deep expert on this, just a beginner trying to help out based on what I have seen!
You are absolutely right that GGUF models are designed to be much less resource-intensive. However, if you look closely at your log, you will see this line:
[GGUF][llama.cpp CUDA] kernels unavailable, using fallback
Because the required hardware-accelerated CUDA kernels for GGUF are missing from your environment, WanGP is forced to use the CPU fallback. This fallback is extremely slow and resource-heavy, which is why the loading process hangs indefinitely and the UI stops responding.
Regarding Stability Matrix: you are correct. Stability Matrix manages its own isolated virtual environments, so it does not use the standard scripts\update.bat file that comes with the standalone WanGP installation.

How to fix this:

• Option 1: Switch to an INT8 model (Highly Recommended)
  Your logs also show [Quanto][INT8] Injected int8 kernels ACTIVE (backend=triton). This means your environment is already perfectly configured for INT8 quantization! The easiest fix is to simply download the INT8 version of the LTX-2 model instead of the GGUF version. It should load right up.
• Option 2: Manually install the GGUF kernels
  If you specifically need to use the GGUF format, you will have to open the dedicated terminal for the Wan2GP package inside Stability Matrix and manually install the missing llama.cpp CUDA kernels.

Do you recommend me to use wan2gp on its own instead of SM?
in other case, is my RTX 3090 good for LTX 2? I once tried out 2.3 on comfy, but got nowhere in progress. I suppose GGUF doesn't do anything for my GPU

Glad I could help point you in the right direction.

To answer your questions:

1. Standalone Wan2GP vs. Stability Matrix

I definitely recommend using Wan2GP on its own using the standalone installation. The 1-click automated scripts for Windows (.bat) make installation, environment management, and updates as seamless as possible. These scripts automatically detect your hardware and will install the best acceleration kernels (like Triton, Flash Attention, and even the GGUF kernels) specifically for your config. It saves a ton of manual troubleshooting!

2. Is the RTX 3090 good for LTX-2?

Your RTX 3090 is an absolute beast for this! It has 24GB of VRAM. WanGP is specifically designed to be accessible to the "GPU Poor" and can run select models with as little as 6 GB of VRAM. Because of WanGP's heavy VRAM optimizations under the hood, your 3090 will easily handle LTX-2 and LTX-2.3. If you struggled in ComfyUI, you will likely have a much smoother experience here. For context, I run it without any problems on my RTX 3060 with 12GB VRAM and 64GB of system RAM, so your 3090 will absolutely crush it!

3. Does GGUF do anything for your GPU?

You are totally onto something here. GGUF is a lifesaver for people who need to squeeze large models into cards with low VRAM. But since your RTX 3090 has a massive 24GB of VRAM, you don't even need the extreme compression of GGUF to fit LTX-2. You have more than enough room to run the INT8 models (which your console showed you already have the kernels for), and they will run extremely fast on your hardware.

Where do I find the INT8 model version in the UI?

[vadhorvad]

Where do I find the INT8 model version in the UI?

You can find and download them directly through the Models Manager tab that you have open. If you look closely at the small text just below the "Filtered Finetunes for 'LTX-2'" heading, it says:

Missing files assume config: transformer qtype=int8, text-encoder qtype=bf16

This means that any of the base models in that list that don't specifically have "GGUF" or "NVFP4" in their name are the INT8 versions by default.

Steps to get it:

1. Click the blue View button next to LTX-2 2.0 Distilled 19B or the LTX-2 2.3 version.
2. Download the required files for it.

[koopa990]

Where do I find the INT8 model version in the UI?

You can find and download them directly through the Models Manager tab that you have open. If you look closely at the small text just below the "Filtered Finetunes for 'LTX-2'" heading, it says:

Missing files assume config: transformer qtype=int8, text-encoder qtype=bf16

This means that any of the base models in that list that don't specifically have "GGUF" or "NVFP4" in their name are the INT8 versions by default.

Steps to get it:

1. Click the blue View button next to LTX-2 2.0 Distilled 19B or the LTX-2 2.3 version.
2. Download the required files for it.

I assume it's the Dev version?

[vadhorvad]

Where do I find the INT8 model version in the UI?

You can find and download them directly through the Models Manager tab that you have open. If you look closely at the small text just below the "Filtered Finetunes for 'LTX-2'" heading, it says:

Missing files assume config: transformer qtype=int8, text-encoder qtype=bf16

This means that any of the base models in that list that don't specifically have "GGUF" or "NVFP4" in their name are the INT8 versions by default.
Steps to get it:

1. Click the blue View button next to LTX-2 2.0 Distilled 19B or the LTX-2 2.3 version.
2. Download the required files for it.

I assume it's the Dev version?

Not)
You go to "Plugins" you will find the "Model Manager" there. Next, activate it, save the changes and restart WanGP.