this works with AMD?

[KillyTheNetTerminal]

No description provided.

[ArkhamInsanity]

Currently using it on my ubuntu pc with AMD 6800XT graphic card, about 45 seconds for 45 seconds songs with defaults parameters (100 steps)

[KillyTheNetTerminal]

oh, cool, but using rocm? I use directml on windows (Rx580) how did you install this? i get errors and my comfy installation corrupted

[ArkhamInsanity]

oh, cool, but using rocm? I use directml on windows (Rx580) how did you install this? i get errors and my comfy installation corrupted

Yes with rocm, never tried to use directml on windows myself, sorry.

[ghost]

hmm, requires flash_attn. How did you get it working with rocm?

TypeError: expected string or bytes-like object
      [end of output]
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for flash_attn
  Running setup.py clean for flash_attn
Failed to build flash_attn

[ArkhamInsanity]

hmm, requires flash_attn which requires nvidia no? How did you get it working with rocm?

TypeError: expected string or bytes-like object
      [end of output]
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for flash_attn
  Running setup.py clean for flash_attn
Failed to build flash_attn

I don't know how to help you, I used comfyui manager to install like everything I install in comfyui and it worked right away

[ghost]

I think it is since StableAudioSampler was updated today/yesterday? I installed it fine from the comfy manager last week. But today, no go. Anyway, I found there is a ROCm flash_attn from our friends @ AMD.com

https://rocm.docs.amd.com/en/latest/how-to/llm-fine-tuning-optimization/model-acceleration-libraries.html

This installs Flash Attention 2 ROCm and the node is available. In process of testing if it actually works but I'm confident.
:-)

NO!! installed but complained I needed a MI200 or MI300 GPU when I ran a workflow with StableAudioSampler. GaAAhHhH!!

[KillyTheNetTerminal]

I'm a little confused about all this haha... flash attention is something that is not available in Directml? Rocm, what is it about? I installed a comfy Zluda version and it also disables that option

[ArkhamInsanity]

I think it is since StableAudioSampler was updated today/yesterday? I installed it fine from the comfy manager last week. But today, no go. Anyway, I found there is a ROCm flash_attn from our friends @ AMD.com

https://rocm.docs.amd.com/en/latest/how-to/llm-fine-tuning-optimization/model-acceleration-libraries.html

This installs Flash Attention 2 ROCm and the node is available. In process of testing if it actually works but I'm confident. :-)

I have to update^^ Tomorrow I may have more time to try it

I'm a little confused about all this haha... flash attention is something that is not available in Directml? Rocm, what is it about? I installed a comfy Zluda version and it also disables that option

I am not really good with the tech, but Rocm works on linux and soon on windows.

[ghost]

Well I tried,
option one installed but told me it only works on MI 200/300 GPU's (it tells me that when I try to run a workflow) ,
the other option 'Triton' wont build, it fails getting the pytorch version.

No issues with ROCm, pytorch and onnxruntime work fine using ROCm 6.0.2 (stable)

As flash_attn errors out of building the wheel at the pytorch version check I am assuming it wants an older ROCm like 5.7 maybe? as this changes the version number of pytorch (adds rocm-6.0 to the version number)

Arch Linux
Pythons 3.10 through 3.12 tested
Pytorch ROCm 6.0.2 from official site (https://pytorch.org/get-started/locally/)
Radeon RX 6900XT (gfx1030)
Clean venv and just install pytorch then try to install flash_attn
I tried just installing it from comfyui manager and there is no difference in the errors. it wants this flash_attn and I can't get it to work.

[ghost]

I'm a little confused about all this haha... flash attention is something that is not available in Directml? Rocm, what is it about? I installed a comfy Zluda version and it also disables that option.

seems on linux one of the dependencies for one of the requirements for StableAudioSampler is flash_attn, I have never come accross it before but there it is. Don't know what changed as I did install StableAudioSampler only a week ago without issues somwhere in between this flash_attn appeared as a dependency.

windows uses DirectML/Zluda. On Linux, for AMD, we use ROCm/OpenCL.

[ghost]

I'm a little confused about all this haha... flash attention is something that is not available in Directml? Rocm, what is it about? I installed a comfy Zluda version and it also disables that option

oops, sorry, then Iv hijacked your thread, just assumed linux, , err yes, As far as I am aware this works for AMD , just not today on linux LOL

[ghost]

err, hmm, ok, well, remove the flash_attn line from your requirements.txt, then do
pip install -r requirements.txt everything goes fine and dandy without flash_attn, ComfyUI-StableAudioSampler workflow working fine and outputting music. Solved, pheeewww, tested on python 3.10

[ArkhamInsanity]

You should definitely post your own issue topic if you want the developper to take care of it