If you are not seeing almost maximum watt usage of your GPU on nvitop like this, especially if you a

F

Furkan Gözükara SECoursesOP•8/4/25, 10:16 PM

 my prompt <refiner> <lora:Phantom_Wan_14B_FusionX_LoRA:1.0> <base> <lora:Phantom_Wan_14B_FusionX_LoRA:0.9>

 my prompt <refiner> <lora:Phantom_Wan_14B_FusionX_LoRA:1.0> <base> <lora:Phantom_Wan_14B_FusionX_LoRA:0.9>

 my prompt <refiner> <lora:Phantom_Wan_14B_FusionX_LoRA:1.0> <base> <lora:Phantom_Wan_14B_FusionX_LoRA:0.9>

 my prompt <refiner> <lora:Phantom_Wan_14B_FusionX_LoRA:1.0> <base> <lora:Phantom_Wan_14B_FusionX_LoRA:0.9>

F

Furkan Gözükara SECoursesOP•8/9/25, 10:54 AM

 pip install flash-attn-triton and it should work on RTX 20xx (Turing), see https://github.com/rationalism/flash-attn-triton

 pip install flash-attn-triton and it should work on RTX 20xx (Turing), see https://github.com/rationalism/flash-attn-triton

 pip install flash-attn-triton and it should work on RTX 20xx (Turing), see https://github.com/rationalism/flash-attn-triton

 pip install flash-attn-triton and it should work on RTX 20xx (Turing), see https://github.com/rationalism/flash-attn-triton

F

Furkan Gözükara SECoursesOP•9/1/25, 11:21 AM

@FurkanGozukara If you just run pip install sageattention, then it's SageAttention 1. It's hosted on pypi.org and you can see the wheel at https://pypi.org/project/sageattention/#files . It only uses Triton kernels, not CUDA kernels, so it's easy to install.

For FlashAttention that only uses Triton kernels, you can run pip install flash-attn-triton.

@FurkanGozukara If you just run pip install sageattention, then it's SageAttention 1. It's hosted on pypi.org and you can see the wheel at https://pypi.org/project/sageattention/#files . It only uses Triton kernels, not CUDA kernels, so it's easy to install.

For FlashAttention that only uses Triton kernels, you can run pip install flash-attn-triton.

@FurkanGozukara If you just run pip install sageattention, then it's SageAttention 1. It's hosted on pypi.org and you can see the wheel at https://pypi.org/project/sageattention/#files . It only uses Triton kernels, not CUDA kernels, so it's easy to install.

For FlashAttention that only uses Triton kernels, you can run pip install flash-attn-triton.

@FurkanGozukara If you just run pip install sageattention, then it's SageAttention 1. It's hosted on pypi.org and you can see the wheel at https://pypi.org/project/sageattention/#files . It only uses Triton kernels, not CUDA kernels, so it's easy to install.

For FlashAttention that only uses Triton kernels, you can run pip install flash-attn-triton.

F

Furkan Gözükara SECoursesOP•9/28/25, 12:46 PM

F

Furkan Gözükara SECoursesOP•9/28/25, 12:46 PM

https://github.com/kohya-ss/musubi-tuner/blob/main/src/musubi_tuner/dataset/dataset_config.md#sample-for-image-dataset-with-control-images

F

Furkan Gözükara SECoursesOP•10/15/25, 1:21 PM

triton 3.5 : https://github.com/woct0rdho/triton-windows/releases/tag/v3.5.0-windows.post21

F

Furkan Gözükara SECoursesOP•10/19/25, 10:28 AM

Use --ddp_gradient_as_bucket_view param of DDP, that will solve this issue and yes the double memory consumption is expected without this param

Use --ddp_gradient_as_bucket_view param of DDP, that will solve this issue and yes the double memory consumption is expected without this param

Use --ddp_gradient_as_bucket_view param of DDP, that will solve this issue and yes the double memory consumption is expected without this param

Use --ddp_gradient_as_bucket_view param of DDP, that will solve this issue and yes the double memory consumption is expected without this param

F

Furkan Gözükara SECoursesOP•11/6/25, 10:49 AM

this tokenizer is useful : https://sd-tokenizer.rocker.boo/

Diffusion Tokenizer

Informs you about how your prompt/words gets turned into tokens, privately in your browser. For Diffusion models like Stable Diffusion, Flux, Pixart

F

Furkan Gözükara SECoursesOP•11/12/25, 12:34 AM

RIFE is included with Windows_Install_SwarmUI.bat file

also it is named as : Video Frame Interpolation Multiplier

and : Text2Video Frame Interpolation Multiplier

F

Furkan Gözükara SECoursesOP•11/14/25, 11:27 AM

onetrainer dreambooth setting

F

Furkan Gözükara SECoursesOP•11/14/25, 11:27 AM

set class images are prior prediction

F

Furkan Gözükara SECoursesOP•11/14/25, 11:29 AM

only works with lora

F

Furkan Gözükara SECoursesOP•11/18/25, 7:18 PM

If you are not seeing almost maximum watt usage of your GPU on nvitop like this, especially if you a

Similar Threads