Wan2GP chooses RAM instead of VRAM

[DUFUSxD]

Hello, I'm kinda new in local video generation. Today i've got a problem:

Loading Model 'ckpts/ltxv_0.9.8_13B_distilled_quanto_bf16_int8.safetensors' ...
************ Memory Management for the GPU Poor (mmgp 3.5.3) by DeepBeepMeep ************
Pytorch compilation of 'transformer' is scheduled for these modules : ['transformer_blocks.*'].
Pinning data of 'transformer' to reserved RAM
The whole model was pinned to reserved RAM: 56 large blocks spread across 12625.60 MB
Hooked to model 'transformer' (Transformer3DModel)
Pinning data of 'vae' to reserved RAM
The whole model was pinned to reserved RAM: 11 large blocks spread across 2378.30 MB
Hooked to model 'vae' (CausalVideoAutoencoder)
Pinning data of 'text_encoder' to reserved RAM
The whole model was pinned to reserved RAM: 41 large blocks spread across 9083.39 MB
Hooked to model 'text_encoder' (T5EncoderModel)
Pinning data of 'latent_upsampler' to reserved RAM
The whole model was pinned to reserved RAM: 1 large blocks spread across 240.80 MB
Hooked to model 'latent_upsampler' (LatentUpsampler)
Pinning data of 'prompt_enhancer_image_caption_model' to reserved RAM
Found 2 tied weights for a total of 400.70 MB, last : language_model.model.shared.weight <-> language_model.model.decoder.embed_tokens.weight
The whole model was pinned to reserved RAM: 13 large blocks spread across 3139.31 MB
Hooked to model 'prompt_enhancer_image_caption_model' (Florence2ForConditionalGeneration)
Pinning data of 'prompt_enhancer_llm_model' to reserved RAM
Tied weights of 751.50 MB detected: model.embed_tokens.weight <-> lm_head.weight
The whole model was pinned to reserved RAM: 13 large blocks spread across 3481.29 MB
Hooked to model 'prompt_enhancer_llm_model' (LlamaForCausalLM)
Initializing BaseSegmenter to cuda
Loading Model 'ckpts/ltxv_0.9.8_13B_distilled_quanto_bf16_int8.safetensors' ...
Pytorch compilation of 'transformer' is scheduled for these modules : ['transformer_blocks.*'].
Pinning data of 'transformer' to reserved RAM
The whole model was pinned to reserved RAM: 56 large blocks spread across 12625.60 MB
Hooked to model 'transformer' (Transformer3DModel)
Hooked to model 'vae' (CausalVideoAutoencoder)
Hooked to model 'text_encoder' (T5EncoderModel)
Hooked to model 'latent_upsampler' (LatentUpsampler)
Hooked to model 'prompt_enhancer_image_caption_model' (Florence2ForConditionalGeneration)
Hooked to model 'prompt_enhancer_llm_model' (LlamaForCausalLM)
Initializing BaseSegmenter to cuda

GPU utilization is about 1-5%, and that's not about PY.
What can cause it? And what else should I provide/check to get this situation more understandable?

Specs:
CPU: Intel(R) Core(TM) i7-14700KF
GPU: 5090
RAM: 128gb

Wan2GP Settings:
Profile: 1
Model: LTXV 0.9.8 Distilled 13B
Using venv virtual for Wan2GP

[wodeshijie1234]

Your computing power device can run Comfyui directly instead of wan2gp

[DUFUSxD]

I know, but I decided to use wan2gp because its API is more understandable to me. I use it mainly as an intermediate AI in a chain with SD

[Mark-sys1]

I use a 5090 and 64 gb of ram with profile 2 and 8 gb preloading Vram (you can increase this value if you have a lot of Vram free)
and I use this parameter --perc-reserved-mem-max 0.2

[palimbro]

I use a 5090 and 64 gb of ram with profile 2 and 8 gb preloading Vram (you can increase this value if you have a lot of Vram free) and I use this parameter --perc-reserved-mem-max 0.2

Hi, i have the same configuration, can you explain me where putting this parameter "--perc-reserved-mem-max 0.2" ? Is it on the web interface ? A screenshot would be nice :) thanks

[Mark-sys1]

Use notepad++ and modify the line 16 of the file start.js in the directory wan.git