Running invoke on Linux on AMD GPU (6900xt)

[merry hearth]

Hi, I have a 6900xt and thinking about dual booting ubuntu to get invoke. Rocm says it supports 20.04.2 but there isn't a download for that version on Ubuntu's website. Is there anyone else that's also using amd gpu that can guide me?

[merry hearth]

bump

[merry hearth]

bump

[vital cargo]

I'm on ubuntu 22.10 with a 6750XT and InvokeAI runs fine.

Ping me if you need help when you're installing it

[vagrant spire]

@vital cargo I have the same problem, I tried auto and manual installation, the script detect the GPU, but then when I run it, it uses CPU
I'm on ubuntu 22.04, last version of ROCm and invoke source from github

[vital cargo]

I'm busy for the next 2 days but I'd be available to check for a solution with you @vagrant spire Thursday(?)

[vagrant spire]

I think I found something, I cannot download pytorch rocm 5.3, so I'm trying to use 5.1.1, anyway it's ok for me (:

[vagrant spire]

the installer already uses it, I removed everything, purged the cache of pip, I'm installing again 😅

ok, I reinstalled, it downloaded rocm 5.1.1, but when it start seems that is loading cuda and I don't understand why...

UserWarning: HIP initialization: Unexpected error from hipGetDeviceCount(). Did you run some cuda functions before calling NumHipDevices() that might have already set an error? Error 101: hipErrorInvalidDevice (Triggered internally at ../c10/hip/HIPFunctions.cpp:110.)
return torch._C._cuda_getDeviceCount() > 0

I checked also if ROCm are installed and from the smi seems all ok...

======================= ROCm System Management Interface =======================
================================= Concise Info =================================
GPU Temp (DieEdge) AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU%
0 60.0c 32.0W 500Mhz 1000Mhz 0% auto 264.0W 7% 0%

============================= End of ROCm SMI Log ==============================

[red gulch]

Issues with AMD GPUs on Linux with ROCm installed not showing a recognized GPU when starting Invoke.sh giving the error "hipErrorNoBinaryForGpu: Unable to find code object for all current devices!"

If you have ROCm installed and ROCminfo reports a GPU, you can add
"HSA_OVERRIDE_GFX_VERSION=10.3.0"
Before running invoke.sh
Example is
$ HSA_OVERRIDE_GFX_VERSION=10.3.0 /home/invokeai/invoke.sh

This will allow the script to run and utilize your GPU.

This is on my set up using Ubuntu 22.04 and ROCm 5.3.3 install with a RX 5700XT GPU

This is what I had to do to get running Ubuntu with a 5700xt

[vagrant spire]

I already tried, but nothing so far...
I think I need to check the rocm I installed, maybe downgrade it 🙄

[vagrant spire]

I have ROCm 5.4.1, maybe is better roll back? @red gulch

I run again rocminfo:

---

Agent 1

---

Name: AMD Ryzen 7 2700X Eight-Core Processor

---

Agent 2

---

Name: gfx1030

[vagrant spire]

So far...
I compiled last pytorch, but invoke remove it and place 1.13, it's fine,
I roll back to 5.3.3 and even python don't detect the GPU into a simple script:

import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)

x = torch.rand(5, 3)
print(x)

this is sad

[vagrant spire]

I found something,
I installed with sudo pip3 torch [...] and run the script with sudo and now I can see the GPU with it,
so I think that is something related to permissions that I missed or conf badly (:

[red gulch]

On the rocm site I used 5.3.3 after you run the first script on the installation page for Ubuntu, if you scroll down there is a section for single install there is an option for for multiple use cases
sudo amdgpu-install --usecase=hiplibsdk,rocm

[vagrant spire]

I removed everything, installed 5.3.3 and run also that, I removed and regroupped the user (from video group) and now I can use the pytorch from python.
I tried the manual installation and no luck also with the string you shared, now I'm trying to install with auto script will see

[red gulch]

I used auto myself

It might give you the same error at the end of the auto script when it attempts to set up invoke

[vagrant spire]

seems ok, I'm downloading the dataset again, but this time it wrote cuda instead of CPU

[red gulch]

I'm sending you good linux and amd vibes, the dozens of us need all the help we can get

[vagrant spire]

AMD an hard platform, but I wanted to test this AI, I'm so curious, I didn't expect to fight to get it run xD
I installed after a very long time linux and seems run right now.

Now I need to investigate how to configure this https://github.com/ROCmSoftwarePlatform/MIOpen#installing-miopen-kernels-package 🙂

now I'm using the GPU (:
idk why I don't need your string, maybe because with 6900XT isn't required, who knows?

a thing I noticed is that I must recompile opencv for patchmatch.patch_match so I can fix the warnings/errors, I did before, but I removed everything to start again, so I'll do later.

if anyone need, to fix the MIOpen warning it's pretty easy:

sudo apt list miopenkernels*
from the list select the version that you like to use, I installed miopenkernels-gfx1030-36kdb because my GPU is the 6900XT
(
not fixed yet, I need to work on it because sometimes it rebuild the SQLite DB
found something here: https://github.com/ROCmSoftwarePlatform/MIOpen/issues/1606
)

[vagrant spire]

I fixed Patchmatch warnings in easier way than compile opencv:

sudo apt install libopencv-dev
I got into pkgconfig folder opencv4.pc instead of opencv.pc
so I decided to link it:
cd /usr/lib/x86_64-linux-gnu/pkgconfig/
sudo ln -s opencv4.pc opencv.pc
then I relaunched invoke and it compiled and loaded correctly

[vagrant spire]

@red gulch I fixed also the manual installation.
I removed the cmd for torch from requirements.txt (I could also edit it, but I checked it before)

-r environments-and-requirements/requirements-base.txt

Get hardware-appropriate torch/torchvision

#--extra-index-url https://download.pytorch.org/whl/rocm5.2 --trusted-host https://download.pytorch.org
#torch>=1.13.1
#torchvision>=0.14.1
-e .
then I run:
pip --python invokeai install --use-pep517 --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm5.3
and after:
pip --python invokeai install --use-pep517 .
So I can use 5.3 version 👍

[red gulch]

Great to hear! Glad everyone is working to get AMD working since they can't seem to get their mess together and fix it.

[vagrant spire]

Still sad that AMD hasn't a proper support on windows and in general it's a pain use it also on Linux xD

I'm pretty happy with the 6900xt, but software side, Nvidia is still better imho

[red gulch]

100% I have a 6800xt in my main rig and it runs great but always feels like it's just on the verge of a crash or driver reset.

[river sonnet]

Hey!
Can you please write what steps helped you get rid of the error 'HIP initialization: Unexpected error from hipGetDeviceCount()' ?
I'm using Ubuntu 20.04 and ROCm 5.4.3 install with a 6600XT GPU. Already tried manual and automatic install but got the same result with Unexpected error from hipGetDeviceCount() 😦
I will be very grateful for your help!

[vagrant spire]

Hey I removed ROCm 5.4.3 and installed 5.3 using pytorch command like I wrote in my message
THIS IS FOR MANUAL INSTALL.

I removed the cmd for torch from requirements.txt (I could also edit it, but I checked it before)
-r environments-and-requirements/requirements-base.txt

Get hardware-appropriate torch/torchvision

#--extra-index-url https://download.pytorch.org/whl/rocm5.2 --trusted-host https://download.pytorch.org/
#torch>=1.13.1
#torchvision>=0.14.1
-e .
then I run:
pip --python invokeai install --use-pep517 --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm5.3
and after:
pip --python invokeai install --use-pep517 .
So I can use 5.3 version

Did you add your user to video group?

[river sonnet]

Yes

[vagrant spire]

rocm-smi output? @river sonnet

[river sonnet]

@vagrant spire Still trying to install ROCm 5.3.3..
AMD repo.radeon.com has horrible download speed and often even Connection fails 😦 So amdgpu-install -y --accept-eula --usecase=rocm command can't finish downloading process

[vagrant spire]

can you show me nano /etc/apt/sources.list.d/rocm.list?

deb [arch=amd64] https://repo.radeon.com/rocm/apt/5.3.3 jammy main

I have 22.04, you need to change the jammy with 20.04 name 🙂

[river sonnet]

deb [arch=amd64] http://repo.radeon.com/rocm/apt/5.3.3 focal main

[vagrant spire]

that is correct also the usercase is better use the complete one like @red gulch said

sudo amdgpu-install --usecase=hiplibsdk,rocm

if I'm not wrong

[vagrant spire]

anyway from italy I have not so much problem with AMD repo

[river sonnet]

perhaps this problem appears only in Ukraine😔

[vagrant spire]

It's not a good time right now in Ukraine unfortunately 😟

[river sonnet]

Agree😔

[river sonnet]

i'll try it tomorrow - maybe repo.radeon.com will work better

Thank you for your help!

[vagrant spire]

np, I'm here, now I'm try to play with GPT2 but ROCm is hard, mainly on older code 😩

[river sonnet]

Hey, @vagrant spire !
I finally installed ROCm 5.3.3 🙂 (with usecases as @red gulch said - sudo amdgpu-install --usecase=hiplibsdk,rocm )

[vagrant spire]

hey 🙂

[river sonnet]

can you please provide more info about this operation "removed the cmd for torch from requirements.txt (I could also edit it, but I checked it before)"
How did you do that?

I'm following manual installation instruction https://invoke-ai.github.io/InvokeAI/installation/020_INSTALL_MANUAL/#linux-install-with-a-runtime-container
And now on step where i should run command pip install InvokeAI --use-pep517 --extra-index-url https://download.pytorch.org/whl/rocm5.2. Am I correct that at this stage I should make these changes?

P.S. I apologize in advance some stupid questions, I don't have much technical knowledge in this area

[vagrant spire]

edit your requirements.txt

-r environments-and-requirements/requirements-base.txt
-e .
I left only those two

pip --python invokeai install --use-pep517 --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm5.3

pip --python invokeai install --use-pep517 .

it's better use 5.3 with 5.3

[vagrant spire]

oh they changed manual, nice 🙂
if you need something else, let me know 😉

maybe it's better use this string:
pip install InvokeAI --use-pep517 --extra-index-url https://download.pytorch.org/whl/nightly/rocm5.3

should be the same, but they changed a lot in the manual installation, so I need to test it (:

I tested the command and it install "normal" torch instead of rocm version 😩
so...
first you must run

pip install --use-pep517 --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm5.3
then
pip install InvokeAI --use-pep517

I try to install again invoke so I'm sure that is working 🙂

I'm downloading only SD1.5 model

[river sonnet]

So i should manually create requirements.txt and add this
-r environments-and-requirements/requirements-base.txt -e .
into it
But there not such file environments-and-requirements/requirements-base.txt 😟

https://github.com/invoke-ai/InvokeAI/issues/2499#issuecomment-1416415056
"As of 2.3.0 you no longer need to use requirements.txt, and support has been removed for requirements.txt and environment.yaml."

[river sonnet]

so i can just run this two commands, without any prerequisites and changes, right?

[vagrant spire]

no with the new manual isn't requested anymore for what I saw

should be enough:

pip install --use-pep517 --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm5.3
then
pip install InvokeAI --use-pep517

I'm trying the first generation right now

ok it's working 🙂

@river sonnet delete the invoke forlder, start again and use only those 2 commands 😉

<@&1020322764311253052> someone can update the manual with:
pip install --use-pep517 --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm5.3
and after that command
pip install InvokeAI --use-pep517

because AMD hate a bit their users and with normal command you get still cuda version and not rocm 😉

[red gulch]

Yeah amd are just about useless with this.

[vagrant spire]

mainly it's unfair, they merged cuda into main package and not rocm :/

also tensorflow is the same, they merged tf-gpu into main tf, but rocm is another package and only for 2.x so I'm rewriting some code for test gpt-2 finetuning 😄

[red gulch]

Godspeed to you!

[vagrant spire]

@river sonnet any update?

[river sonnet]

In downloading process 🙂

[red gulch]

We're here for you! I struggled through multiple OS installs and week of installing and reinstalling rocm distros,

[true shard]

There No longer is a requirements.txt. We’ve changed the build system.

[vagrant spire]

yes, I know, I reinstalled and tested, check the topic, when I mentioned the maintainer group 🙂

[river sonnet]

tried to generate image, but it still using CPU instead of GPU 😔

[vagrant spire]

can you show me the rocm-smi command?

and try to create test.py:

import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
and run it with python test.py

you should be inside the InvokeAI env 😉

[vagrant spire]

which model do you have?

[river sonnet]

cpu

[river sonnet]

6600 XT

[vagrant spire]

mobile?

[river sonnet]

no

[vagrant spire]

can you show me the groups command?

[vagrant spire]

do you have 20.04? If I remember correctly you should have 2 groups for ROCm
the second should be render, I check the doc and I let you know

[river sonnet]

yes, it's 20.04

[vagrant spire]

sudo usermod -a -G render $LOGNAME
add also this

and run again the little script

I think you will need to use also and env variable like shown by @red gulch

HSA_OVERRIDE_GFX_VERSION=10.3.0 invokeai --web
or something like that

[red gulch]

After adding your user to the render group be sure to reboot as well.

[river sonnet]

Finally, it works!!!

https://tenor.com/view/frankenstein-frankensteins-monster-its-alive-alive-not-dead-yet-gif-23425340

@vagrant spire Thank you sooo much!!! You are my saviour!

[vagrant spire]

🙂

[red gulch]

Grats! Good work @vagrant spire

[true shard]

@vagrant spire - It looks like there was a lot that went into getting this to finally work. Thanks for this! I wonder if you could write a brief summary of the necessary steps for the next version of the installation documentation?

[vagrant spire]

ok, I'll do again and write down the single step soon as possible 👍

[vagrant spire]

@true shard those are the steps I did to launch it, about ROCm it's also a pain because AMD manual isn't so clear 😅

[grim sierra]

Thanks for the steps, i'll try this too on a fresh install! As it has been a real pain until now. The GPU is working but the image resulting is a grey square, so not that much of a Win.
I also managed to run Stable Diffusion on windows using nod-ai/Shark and a special version of the v22 AMD Drivers (note: updated v23 drivers supports now IREE compiler using MLIR interface on Vulkan, just saw this, maybe there is a way there too to use InvokeAI through that on windows in a next future?)

Just a question, which version of the AMDGPU drivers did you install to make it work ?

[vagrant spire]

ROCm 5.3, I tried 5.4 but last time I didn't get them working, so I stay with 5.3 until pytorch will support 5.4 (:
about windows yes, MS tried to implement a layer to let AMD to run also on windows, but you will lose performance (there is another webUI that run smoothly on windows both native and bash)
about the grey square I think that is related to @red gulch command to add before launch invoke, which one GPU do you have?

[grim sierra]

Sapphire Pulse RX 5600XT 6gb
You mean HSA_OVERRIDE ? Yeah, i used this

[vagrant spire]

6GB is very low memory for run AI models

but like I said, you need to use:

HSA_OVERRIDE_GFX_VERSION=10.3.0 invokeai --web

[vagrant spire]

did you get grey img on windows?

[grim sierra]

It works "fine" on windows with the way I found, there are issues with the software. But I got Images, yes (I mean actual images, not grey square) I suspect a wrong install of ROCm or the driver or kernel module though, on linux.

[vagrant spire]

I only said that is very low memory 😄

sometimes I have problems with my 6900XT infact I'm using google colab to train audioLM 🙈

[grim sierra]

😄

[grim sierra]

I saw ppl online that used RX580 just fine, so the 5600 xt shouldn't be a problem xD

on windows, it took 30sec, for a 512x512, Euler scheduler, AnythingV3 model, CFG 6, 50 steps. Compared to 20 min, I say it is progress xD. Well it tooks 15 min for a first compile of I don't know what, but that's once. Not too bad I'd say for a 6gb :p

[vagrant spire]

I think that first run load the model on VRAM 😉

[grim sierra]

I see 🙂

[grim sierra]

Doesn't explain why grey square on linux (for now) and not on windows with the alternative software and drivers :p

[vagrant spire]

uhm tbh I got few when I didn't set it properly, so I think that there is something not correct in your conf

[grim sierra]

I'll have to try your way 🙂 and I hope that'll solve for the author of this post too 😄

[vagrant spire]

mainly with 5.4 I got grey squares

[grim sierra]

Hmm, know what? I installed 5.4.1 now that you mention it ^^;

[vagrant spire]

I started too with last version, but then I rollback to 5.3 because I saw that pytorch support it and 5.2

[grim sierra]

The InvokeAI installer installs 5.2 no ?

[vagrant spire]

nop

at least to me torch install normal torch without rocm 🤷‍♂️

[grim sierra]

Hm, because, at setup, i choose AMD GPU, and during the setup process I saw ROCm 5.2 😕

[vagrant spire]

[grim sierra]

me too 😄

[vagrant spire]

I used manual installation, but I got normal torch and not torch-rocm, because of that I install manually torch-rocm before invoke

and the pain is also for other type of models...I trained GPT-2 as chatbot, but I edited the code for rocm, because all code was for normal torch

[grim sierra]

Seems like a lot of fun there xD

[vagrant spire]

yes, I like the AI stuff tbh, I'm not so good to write a model from ground up, but I'm starting to see how they working, I'd like to test musicLM, but it is based on audioLM and it require a lot of train steps and I'm studying how to get enough CU to train it without spending too much xD

[grim sierra]

Especially now that the pricing is still quite high 😒 well, it's better, but still...

[vagrant spire]

cloud isn't so expensive, but still is for hobby project

[grim sierra]

Agree ^^

[vagrant spire]

@grim sierra did you test?

[grim sierra]

I didn't, I'm at work. I might be able to do that this weekend or on Friday :)

[red gulch]

There is a launch argument for Float32 that cleared up the grey block issue on linux

[grim sierra]

Really?? What is it ? 🤩

[red gulch]

First render was a solid grey picture with my card, I have to go into the invoke.init and add "--precision float32" argument

No quotes of course

This fixed the issue on my 5700xt

[grim sierra]

I'll test this evening!!

U might just be saving me from reinstalling for the 4th times a Linux distro 😅

[red gulch]

Best of luck, once I'm home I can get the exact line i added it at.

[red gulch]

I've gone from Ubuntu to manjaro, back to Ubuntu multiple times, amd has made this such a mess.

[grim sierra]

True!

[red gulch]

I have to reinstall Automatic1111 for the deforum support. I had issues with it on Ubuntu, hopefully it will go a bit smoother this time.

[grim sierra]

May Tux be with you!

[vagrant spire]

AMD ROCm is a little bit messy, but for what I read on their github there isn't so much budget for it so 🤷‍♂️

[grim sierra]

Meh, might retry that after a fresh install, looks like I broke something, I'm trying to reconfigure invokeai before but already have issues lol. Hope that the author of the main post did have more luck than me 😅

Update: yup, I broke something, UI starts, model loads, but AttributeError: 'Combine' ... Anyhow, have to get a new install 😦

Is it better on arch or ubuntu? 😁

[grim sierra]

I did a new install, following your steps. However I got an issue with dkms during the building process of the kernel modules. Don't understand...

I've installed all the rest, AMD drivers, rocm 5.3.2 (may be I should have get 5.3 or 5.3.3, will uninstall tomorrow) etc until invokeai, then at configure: "hipErrorNoBinaryForGpu: Unable to find code object for all current devices!". Then the installer Aborted. I'm desperate now 😅 Ubuntu is 22.4.1, kernel is 5.15 something.

[red gulch]

i got that error installing mine. go ahead and start up the invokeai with the HSA command in front of it. it will finish installing and have the command going for the installer

i did use 5.3.3 in my install seemed to work the best

[vagrant spire]

I need to check my version tbh

But for what I saw, 6800/6900xt are the only two GPU that has less problem with ROCm, I hope that pro version still will be better than those

[vagrant spire]

Package: rocm-libs
Version: 5.3.3.50303-99~22.04
and I'm using source
https://repo.radeon.com/rocm/apt/5.3.3 jammy/main amd64 Packages

[grim sierra]

will uninstall later today and install 5.3.3 🙂

Did you guys had the DKMS issue ?

Paramétrage de amdgpu-dkms (1:5.18.2.22.40.50303-1504718.22.04) ... Loading new amdgpu-5.18.2.22.40-1504718.22.04 DKMS files... Building for 5.15.0-60-generic Building for architecture x86_64 Building initial module for 5.15.0-60-generic ERROR: Cannot create report: [Errno 17] File exists: '/var/crash/amdgpu-dkms.0.crash' Error! Bad return status for module build on kernel: 5.15.0-60-generic (x86_64) Consult /var/lib/dkms/amdgpu/5.18.2.22.40-1504718.22.04/build/make.log for more information. dpkg: erreur de traitement du paquet amdgpu-dkms (--configure) : installed amdgpu-dkms package post-installation script subprocess returned error exit status 10

[vagrant spire]

this afternoon I'll go back on linux and I'll help better if you will can stay on 😉

[red gulch]

Once I'm home ill take a look and see if I get the same issue

[vagrant spire]

I haven't the command amdgpu-dkms installed 🤔

I used amdgpu-install

[daring egret]

When I tried to install in 12/2022 (Ubuntu (non LTS) 20.10 + RX 6900) I had to use the "--no-dkms" option. Otherwise it threw me a bunch of errors which I didn't know what to make of / how to fix.
So I installed with: "sudo amdgpu-install --usecase=rocm,hip,mllib --no-dkms"

[grim sierra]

same 😕 but it tries to build kernel module 😕

[grim sierra]

and it worked ? I mean SD on GPU ? Anyhow, i'll try that too if there is still issue. I'm gonna try like "with the error"

[vagrant spire]

which one ubuntu version are you using?

[daring egret]

with the rx6900, yes. (just updated to invoke 2.3.0 yesterday. I think i'm on rocm 5.4.0 by now, but need to check. On windows atm.)

Here are my notes how I installed the rocm driver: #1053043950262427648 message

[grim sierra]

Jammy

[vagrant spire]

22.04 ok

[grim sierra]

Thanks you for that 🙂 I'll check that 🙂

[vagrant spire]

I did the same too, but I haven't the problem with --no-dkms, I check amdgpu-install version

amdgpu-install/jammy,jammy,now 22.20.50205-1511376~22.04 all [installed]

to check yours, apt list amdgpu-install

[grim sierra]

amdgpu-install/jammy,jammy,now 5.3.50303-1504718.22.04 all [installé]

[vagrant spire]

😅

try to remove amdgpu-install and run again the list command

[grim sierra]

I don't understand, what should I do ? I've downloaded the rocm package from the rocm website 😕

(well with the command line they put for download)

[vagrant spire]

ok

sudo apt remove amdgpu-install

then

apt list amdgpu-install

[grim sierra]

I leave rocm installed ?

It does something, but no progress on the invokeai bash window ^^;

Did you download the amdgpu-install from AMD ?

[vagrant spire]

yes, but I have also their reposity

one minute that I'm checking the docs

they upgraded the docs for 5.4 finally with 22.04 on it

can you run this command?

cd /etc/apt/sources.list.d && ls

[grim sierra]

amdgpu.list amdgpu.list.save amdgpu-proprietary.list amdgpu-proprietary.list.save rocm.list rocm.list.save

[vagrant spire]

my rocm.list:

deb [arch=amd64] https://repo.radeon.com/rocm/apt/5.3.3 jammy main
my amdgpu.list:
deb https://repo.radeon.com/amdgpu/latest/ubuntu jammy main

before I was using instead of latest 22.20.5, because of that my amdgpu-install is different than yours 🙂
now I updated and upgraded after adding latest and I'll test invoke after a reboot

[grim sierra]

My rom.list:
deb [arch=amd64] https://repo.radeon.com/rocm/apt/5.3.3 jammy main

My amdgpu.list
deb https://repo.radeon.com/amdgpu/5.3.3/ubuntu jammy main

[vagrant spire]

ok, is still working

[grim sierra]

^^

[vagrant spire]

change amdgpu that is better

[grim sierra]

so the .deb from amd website I guess ?

[vagrant spire]

about rocm idk, like I said, I stay with 5.3 until pytorch will support it

[vagrant spire]

no, you can edit directly with nano

sudo nano amdgpu.list

and change 5.3.3 with latest

then run:

sudo apt update
sudo apt upgrade
sudo apt install amdgpu-install

in this way I didn't get any issues, only few warnings about missing firmware 👀

[grim sierra]

damn, forgot to uninstall amdgpu-install

[vagrant spire]

should be upgraded with upgrade command 😉

[grim sierra]

Ok, it tries to build kernel module

[vagrant spire]

yes, also on my side, if you see W: those are warnings

[grim sierra]

Woooooooooooooh!! I only got this
Can't load /var/lib/shim-signed/mok/.rnd into RNG 40E7961E1E7F0000:error:12000079:random number generator:RAND_load_file:Cannot open file:../crypto/rand/randfile.c:106:Filename=/var/lib/shim-signed/mok/.rnd The rest just compiled

[vagrant spire]

strange thing

[grim sierra]

indeed

[vagrant spire]

did you already install invoke?

[vagrant spire]

something related to openssl

[grim sierra]

yes, i did and it loads, but still "Preparing" to render

maybe it compiles something I don't know of ^^

[vagrant spire]

first run is always slow

[grim sierra]

Oh Ok, i didn't wait last time ^^;

f*** it is "disconnected" on the UI

[vagrant spire]

check the console

[grim sierra]

i'm gonna wait a bit

[vagrant spire]

maybe is downloading something

[grim sierra]

nope 😕
2023-02-16T16:42:58+0100: Disconnected from server

and in the console, it is

[vagrant spire]

terminal, sorry, not the console on the UI 😄

[grim sierra]

^^

The GPU is used, that's for sure ^^;

99% of it

let's wait ^^

[vagrant spire]

😄

[grim sierra]

^^; TBH, on windows, with the other solution I got, it takes 20 min for the first compile. However, I got nothing in the terminal ^^;

Are u kidding me xD
just updated xD

[vagrant spire]

😅

[grim sierra]

i'm not sure I should ^^;

[vagrant spire]

idk I do everything with terminal xD

[grim sierra]

it just popped-up ^^

[vagrant spire]

was 5 years that I wasn't using linux other than in the bash on windows 10, I have to admit that I must say ty to AMD to forced me to go back to linux XD

I'd like to test the 7900 XTX in the future 😄

[grim sierra]

me too, but the budget is not there xD

and i'd have too much of a CPU bottleneck

[vagrant spire]

yeh I have already some problems with my 2700x and 6900XT 😄
I'm waiting for the all new x3d, but change all the platform is expensive right now

[grim sierra]

lol, last time I checked I took a new AMD AM5 sochet CPU (7600 I think), not the most expensive motherboard, 32gb of DDR5, and a powersupply ~700W and I was around 800€ xD

I was litteraly on my ass

[vagrant spire]

I need 64gb of ram and I'd like to buy a mobo with at least 3 PCI5 M2 🙂
that is expensive

[grim sierra]

hu hu, very ^^;

I start to be discouraged, there is nothing happening, I tried without --precision float32 and it crashed, then used again float32, and not it is doing something, don't know what, and nothing appears in the terminal, server is disconnected from the webui xD

AND... ubuntu it upgraded to ROCm 5.4

was not working before that anyway

It's been half-an-hour, i think it is not doing anything actually 😕

It is soooooooooo annoying to be that close and nothing works

[vagrant spire]

uhm

I don't understand why you have ROCm 5.4

run this

apt show rocm-libs -a
mine is:
Package: rocm-libs
Version: 5.3.3.50303-99~22.04
Priority: optional
Section: devel
Maintainer: ROCm Libs Support [email protected]
Installed-Size: 13,3 kB
Depends: hipblas (= 0.52.0.50303-99~22.04), hipfft (= 1.0.9.50303-99~22.04), hipsolver (= 1.5.0.50303-99~22.04), hipsparse (= 2.3.2.50303-99~22.04), miopen-hip (= 2.18.0.50303-99~22.04), rccl (= 2.12.12.50303-99~22.04), rocalution (= 2.1.2.50303-99~22.04), rocblas (= 2.45.0.50303-99~22.04), rocfft (= 1.0.18.50303-99~22.04), rocrand (= 2.10.9.50303-99~22.04), rocsolver (= 3.19.0.50303-99~22.04), rocsparse (= 2.3.3.50303-99~22.04), rocm-core (= 5.3.3.50303-99~22.04), hipblas-dev (= 0.52.0.50303-99~22.04), hipcub-dev (= 2.10.12.50303-99~22.04), hipfft-dev (= 1.0.9.50303-99~22.04), hipsolver-dev (= 1.5.0.50303-99~22.04), hipsparse-dev (= 2.3.2.50303-99~22.04), miopen-hip-dev (= 2.18.0.50303-99~22.04), rccl-dev (= 2.12.12.50303-99~22.04), rocalution-dev (= 2.1.2.50303-99~22.04), rocblas-dev (= 2.45.0.50303-99~22.04), rocfft-dev (= 1.0.18.50303-99~22.04), rocprim-dev (= 2.10.9.50303-99~22.04), rocrand-dev (= 2.10.9.50303-99~22.04), rocsolver-dev (= 3.19.0.50303-99~22.04), rocsparse-dev (= 2.3.3.50303-99~22.04), rocthrust-dev (= 2.10.9.50303-99~22.04), rocwmma-dev (= 0.7.0.50303-99~22.04)
Homepage: https://github.com/RadeonOpenCompute/ROCm
Download-Size: 1.000 B
APT-Sources: https://repo.radeon.com/rocm/apt/5.3.3 jammy/main amd64 Packages
Description: Radeon Open Compute (ROCm) Runtime software stack

[grim sierra]

I wish i knew 😦

I'll get back to it but I need a break from it 😄

I think It updated the source list with 5.4

[vagrant spire]

mhm strange, I still have 5.3.3

[grim sierra]

I think I'm doomed to not use this xD
@vagrant spire I've tried your way and @daring egret' (no-dkms), while setting ROCm to 5.3.3 but still having the latest amdgpu-install.
InvokeAI starts, but when I try to use any model (here stable-diffusion 1.5), the UI disconnects from the server after like 2-3 minutes and I'm left with a Terminal loading and showing nothing and a GPU running at 100% for a very long time without anything outputted.

[daring egret]

• Did you add your user to the video/render groups?
• If you haven't yet since the rocm installation, try rebooting (I know, I know...).
• After reboot, what does "rocminfo" show? Does it list your gpu?

[grim sierra]

• Yes 😄
• Yes, can retry though
• Attached 🙂 (Yup, it does)

[daring egret]

Ok, so you have a RX 5600 XT ?

[grim sierra]

indeed

[daring egret]

Did you try this stuff with "export HSA_OVERRIDE_GFX_VERSION=10.3.0"

[grim sierra]

I did 🙂

[daring egret]

ok

when you start invoke.sh and enter the development console

[grim sierra]

I can retry by putting it in front before the call to invoke ai

[daring egret]

instead of the webui (i.e. don't press 2, but press the number for development console

Can you do that? Enter the developer console?

i.e. start the invoke.sh script, don't press 2 for the web-ui, instead press the number for the developer mode/console

[grim sierra]

hm, i did the install manually, I don't have invokeai.sh =/

[daring egret]

ok

i'm not entirely familiar with the current instructions for manual install

you probably need to activate a virtual environment?

to start the invoke web ui

[grim sierra]

indeed

[daring egret]

ok, please activate the virtual environment 😉

[grim sierra]

It is activated 🙂

[daring egret]

then start a "python" shell/console inside that virtual environment. i.e. type "python"

then "import torch"

inside that python shell

and finally type: "torch.cuda.is_available()"

it should return "True"

if it doesn't, then we have a problem with the rocm installation. And if it does return True, then your GPU should work properly and we need to find the error with the invoke AI installation.

[grim sierra]

Hm, grumpf
"hipErrorNoBinaryForGpu: Unable to find code object for all current devices!"
Aborted (core dumped)

[daring egret]

ok. try the following:

activate the virtual environment. type: "export HSA_OVERRIDE_GFX_VERSION=10.3.0"

and then everything above inside that same virtual environment: i.e. "python", "import torch", and "torch.cuda.is_available()"

[grim sierra]

"True"

[daring egret]

ok!

[grim sierra]

that's already something ^^;

[daring egret]

So everytime before you start the invokeai webui, make sure that after you activate the virtual enviroment, you enter that command "export HSA_OVERRIDE_GFX_VERSION=10.3.0"

then start the webui

and lets see if we have more success

[grim sierra]

what I do is : "HSA_OVERRIDE_GFX_VERSION=10.3.0 invokeai --web"
should it be "export HSA_OVERRIDE_GFX_VERSION=10.3.0 invokeai --web" ?

[daring egret]

I'm afraid I'm not a linux guru. I assume that "HSA_OVERRIDE_GFX_VERSION=10.3.0 invokeai --web" does the same as entering the export command before calling the web ui

so if you did that already that way, we probably only confirmed that that command really is necessary

[grim sierra]

lemme try again :p

[daring egret]

So when you start the webui. It lists some stuff in the terminal window. It should print something like ">> Using device_type cuda"

[grim sierra]

I get to the UI, i set a prompt, choose 11 steps to be quick, and after
">> ESRGAN Parameters: False

Facetool Parameters: False
Setting Sampler to k_euler (EulerDiscreteScheduler)"
it seems to do nothing but keeps my GPU busy

[grim sierra]

it is

[daring egret]

the first time after start up, it can take some time (i.e. 1 or 2 minutes maybe more)

if it doesn't crash, I would suggest to wait quite a bit just to be sure

[grim sierra]

The thing is that after 1-2 minutes, the UI shows disconnected and the console shows "Disconnected from server"
while terminal shows still the same message about setting sampler and nothing after it

[daring egret]

i'm rebooting to my ubuntu and check how long it takes for me

[vagrant spire]

If you have amdgpu-install 5.4 is fine, you need to check the other command I sent you about rocm-libs

[grim sierra]

Package: rocm-libs
Version: 5.3.3.50303-99~22.04
Priority: optional
Section: devel
Maintainer: ROCm Libs Support [email protected]
Installed-Size: 13,3 kB
Depends: hipblas (= 0.52.0.50303-99~22.04), hipfft (= 1.0.9.50303-99~22.04), hipsolver (= 1.5.0.50303-99~22.04), hipsparse (= 2.3.2.50303-99~22.04), miopen-hip (= 2.18.0.50303-99~22.04), rccl (= 2.12.12.50303-99~22.04), rocalution (= 2.1.2.50303-99~22.04), rocblas (= 2.45.0.50303-99~22.04), rocfft (= 1.0.18.50303-99~22.04), rocrand (= 2.10.9.50303-99~22.04), rocsolver (= 3.19.0.50303-99~22.04), rocsparse (= 2.3.3.50303-99~22.04), rocm-core (= 5.3.3.50303-99~22.04), hipblas-dev (= 0.52.0.50303-99~22.04), hipcub-dev (= 2.10.12.50303-99~22.04), hipfft-dev (= 1.0.9.50303-99~22.04), hipsolver-dev (= 1.5.0.50303-99~22.04), hipsparse-dev (= 2.3.2.50303-99~22.04), miopen-hip-dev (= 2.18.0.50303-99~22.04), rccl-dev (= 2.12.12.50303-99~22.04), rocalution-dev (= 2.1.2.50303-99~22.04), rocblas-dev (= 2.45.0.50303-99~22.04), rocfft-dev (= 1.0.18.50303-99~22.04), rocprim-dev (= 2.10.9.50303-99~22.04), rocrand-dev (= 2.10.9.50303-99~22.04), rocsolver-dev (= 3.19.0.50303-99~22.04), rocsparse-dev (= 2.3.3.50303-99~22.04), rocthrust-dev (= 2.10.9.50303-99~22.04), rocwmma-dev (= 0.7.0.50303-99~22.04)
Homepage: https://github.com/RadeonOpenCompute/ROCm
Download-Size: 1.000 B
APT-Sources: https://repo.radeon.com/rocm/apt/5.3.3 jammy/main amd64 Packages
Description: Radeon Open Compute (ROCm) Runtime software stack

[daring egret]

i'm on rocm 5.4.2 btw. and before that on 5.3.0. So I don't think it has do be some specific version.

[vagrant spire]

I had grey images with 5.4, so I roll back to 5.3

[grim sierra]

maybe a combination of CPU/MB/GPU, which for some reasons don't work for this altogether

[vagrant spire]

No, I don't think so, its only config problems

[grim sierra]

Which graphic drivers do you have, the open or proprietary one ?
(invokeai is sucking 25% of my CPU and 2.5GB of Ram xD)

[vagrant spire]

Uhm I need to check, because I installed a minimal Ubuntu with check on install drivers

[grim sierra]

same here

kubuntu for me, but same version as ubuntu 22.04

[vagrant spire]

If you see true with python code it's not a problem for how you installed rocm

Because it is working

[grim sierra]

That's what I think, maybe it is doing something and I just have to wait 😕

[daring egret]

If invokeai still doesn't create/show your image after several minutes, what are the latest entries in your terminal window?

[vagrant spire]

Mhm strange

Can you show us rocm-smi when you are using InvokeAI?

[vagrant spire]

GPU is working and has loaded the model on the memory

[daring egret]

well it clearly tried to use your gpu to create the image

[vagrant spire]

Can you show us also the terminal where you launch InvokeAI?

So I can check if there are problems with installation

[grim sierra]

The three last line since the message from the console of the web UI (2023-02-16T20:05:54+0100:Disconnected from server)

[vagrant spire]

And the other console?

[grim sierra]

at launch of invokeai ?

[daring egret]

does the rx5600 have 6GB Memory?

[grim sierra]

yes
works fine on windows with nod-ai/Shark and latest amd drivers (well, until it doesn't lol)

Everything looks just fine 😕

[vagrant spire]

When I tested 1 month ago, it worked on a specific version not lasted, nice that they updated

[grim sierra]

(Well, unfortunately, it is not working on windows with InvokeAI)

[vagrant spire]

Because it's require to write a specific code to use direcfml from Microsoft

And cut down performance xD

[grim sierra]

GPU is working his ass out :

[vagrant spire]

Anyway, you need to follow that GitHub link and install what is missing

[grim sierra]

Patchmatch ?

weird that it is not part of the invoke install

[vagrant spire]

Yes, it's better install also it

Because it's a thing required for in painting feature if I'm not wrong

[grim sierra]

for inpainting and outpainting

but is it when you already use an image, right ?

Will install patchmatch tomorrow I think, for now I'll have a pizza. I'm frustrated that this is not working. Thanks you for your time today 🙂 !

[vagrant spire]

Yes should be not a big deal, but it's better try everything at this point

[grim sierra]

so true ^^

[vagrant spire]

I didn't follow the docs on InvokeAI, but I installed it and you find in this topic how, follow the guide that I believe that is the same I already did

[daring egret]

My only remaining suggestions would be to try a fresh (automated) installation for InvokeAI.

[grim sierra]

I can do that if I just delete the manual one ?

[vagrant spire]

Can be a solution, but I edited the requirement txt for make it work, now they changed how they install it, so should be checked if install torch or torch-rocm

Yep

[grim sierra]

good to know

[vagrant spire]

You can also keep the manual installation

And use another folder if needed

[daring egret]

if you have the diskspace and don't want to delete the old installation completely, because you want to try more with it later or something, you can just rename the old folder

while you try the automated installation

or you install the automated installation to a different folder

you don't have to delete the manual installation if you don't want to

[grim sierra]

Cool 😉 Will save me some time !

Thanks for the advice !

[vagrant spire]

Before launch the auto install, do a pip cache purge

[grim sierra]

noted !

[vagrant spire]

And keep on eye on torch, if it download torch 1.13 not rocm you will not use GPU

[grim sierra]

I think that's why i took the manual way 😕

But I think the autoinstaller is asking about AMD/NVidia or CPU

[vagrant spire]

I'll check the install script and I'll edit it then

[grim sierra]

i can't remember

but is installing pytorch rocm 5.2 not 5.3

[daring egret]

It did for me to, so I guess that shouldn't be a problem.

[vagrant spire]

Should be the same, important is that it install rocm, then 5.2 is the stable version

[daring egret]

I found this article, where they benchmark different GPUs for stable diffusion: https://www.tomshardware.com/news/stable-diffusion-gpu-benchmarks
And it says that the 5600 "failed", so they didn't benchmark it. Whatever "failed" means. Mabye they couldn't get it to run either.

[vagrant spire]

I checked the install script on GitHub and seems that they run first torch rocm and then install invoke, maybe it can work

4090 is a monster, the 7900 xtx is impressive, mine is very bad xD

6gb of VRAM isn't so much plus 5xxx isn't very good on ML stuff

[daring egret]

You can find some reddit posts or other sites on google claiming that all sorts of AMD GPUs work, for example this one: https://www.reddit.com/r/StableDiffusion/comments/ww436j/howto_stable_diffusion_on_an_amd_gpu/

But its ~6 months old, and the version of some libraries, etc. changed in the meantime. For example it also states the the RX480 8GB should work (which I have in an old pc), and it absolutely did not work for me in december 2022, no matter what I tried.
One specific reason seems to be that AMD stated that they no longer support the latest features for rocm on older gpus, and then some other developers dropped support for these gpus entirely in their libraries.

So even If you find older articles stating that the 5600 should work, that might no longer be the case.

[grim sierra]

Thanks (btw, I installed patchmatch and... nope, not at all, still same beahviour)

Thanks for those links 🙂
Well, strange thing is that I can produce stuff in 30 sec on windows with shark and amd drivers 😕

[vagrant spire]

Time to buy a 7900xtx then xD

[grim sierra]

soooo, i guess that's all for nvidia supremacy 😕

[vagrant spire]

Because it uses vulkan and emulate in some sort the process

[grim sierra]

loool, I wish, my wallet is trying to kill itself xD

[grim sierra]

Vulkan is great 😄

[vagrant spire]

Thanks to Microsoft there is a way to use AMD GPUs on windows, but then it should be implemented into the code and I don't know the difficulty tbh

I need to check how is working (:

Sorry it's not vulkan but dx12 🤣

[grim sierra]

Right, vulkan is on Linux (thanks valve btw 🤩)

(they managed to use vulkan and other tricks to enable us to play DirectX games 😁)

[vagrant spire]

Meh vulkan is everywhere, thanks to AMD xD
Then everyone join the project also Microsoft

[grim sierra]

I can't wait to install the official desktop version of steam os 3 (when valve offers it to the world 😁)

[grim sierra]

Ha ha, yeah, thanks AMD!

[vagrant spire]

Don't know, I need to watch the steamos because I'm using windows for gaming 🤣

[grim sierra]

I'm using windows too for gaming I've a steam deck and tbh, it is very good

[vagrant spire]

Anyway the main problem is that direct ml uses a custom init for the devide and didn't share CUDA calls like rocm, that I don't understand why they don't use GPU() instead of cuda() 🤣🤣🤣

[grim sierra]

SteamOS 3, because the original one was based on debian, now it is arch.

[vagrant spire]

Arch isn't based on Ubuntu? 👀

Don't kill me, 5 years away from Linux xD

[grim sierra]

Really, I don't know, been awhile since I put my head in all of this stuff ^^;

[grim sierra]

Don't think so 😕 But been awhile too away from Linux. I had dual boot etc, then I just kept windows 😕

Arch is... Well, Arch 😅

And Ubuntu is Debian based if I'm not mistaken

Been a long time 😅

[vagrant spire]

No ok I'm sorry arch is arch 😝

[grim sierra]

😁

[vagrant spire]

Now if I can I'll try directml on windows, then I'll looking into invokeai how much is hard to implement 👀

[grim sierra]

I think, really, you'll become a god for a lot of windows users 🤩

[vagrant spire]

I'm not so good, depend on codebase of invokeai 😄

[grim sierra]

Trying to install after the graphic driver from amdgpu-install --use-case=graphics, and auto install of invokeai, wish me luckw after that, I give up 😅

[vagrant spire]

never giveup 😄
I investigated 3 days in it before find a way to make it run xD

I just discovered that you download from pip invokeai instead to use the one you download with sources 😄
so it's not a proper manual installation

[vagrant spire]

I'm following the code, I stuck on encoding the prompt, because I need to load the device in another way 🤣 but then someone decided to turn off the electricity on all block

[grim sierra]

😖

I followed the guide on their wiki 😦 Indeed, I didn't clone the repository 😕

Now, I did the auto install, after the AMD graphics installation, and now I'm back to (grey) square 1. I'm adding the argument --precision float32.

It Works... But... HIP out of memory, cannot generate image 😣

I'm sooooo frustrated, It works on windows with some magic trickery, but here... it just don't. I even see the image progress in the UI xD But... nope

i'll see tomorrow about the HIP stuff, weird thing is, If I try to make 256x256, the system tells me that there is 0 image created, but the image is present in the folder xD

I'll try the argument "--free_gpu_mem" that could help I think, found it here : https://invoke-ai.github.io/InvokeAI/features/CLI/#list-of-arguments-recognized-at-the-command-line

[grim sierra]

Funny thing is, the requirements for invokeai is a 4gb vram GPU 😅

[vagrant spire]

I think that depend also by the model

[grim sierra]

makes sense 🙂

🤣 Warning: '--free_gpu_mem' is not yet supported when generating image using model based on HuggingFace Diffuser.
Ok, I tried an embedding and i didn't got the message about hip, and it generated and image in 256x256 in 6 seconds. Trying with 512x512, and... worked xD Ok so I should not use the provided stable diffusion model, that's ok 😄

Today is a good day 😄

(so free_gpu_mem is needed for me 🙂 )
However, I might disconnect and reconnect between each model change.
For now a 512x512, 50 iterations, k_euler_a, CFG 9, takes 30sec and generates fine, with free_gpu_mem and precision float32.

I call that a win 😄

[vagrant spire]

yep

[grim sierra]

Thanks a lot for your support @vagrant spire and @daring egret !! Your help was very helpful ! 😄
I might also try with ROCm 5.4 if there is any improvements 🙂

[vagrant spire]

🙂

[vagrant spire]

I need ONNX format for models, so I hope that in the future invoke will support those type of models 🙂

[grim sierra]

fingers crossed 😄 ! never used ONNX.
I know that one option to use Stable diffusion on windows was to convert to ONNX, but all I know 🙂

[true shard]

This is fixed in upcoming 2.3.1

[river sonnet]

Hi @vagrant spire
Sorry to bother you again..
A few months ago you helped me install InvokeAI on Ubuntu, for an XT6600 video card. Back then we were installing version 2.3.0. I recently tried to update to the latest version and again I don't have the image generation using the video card, only the CPU.

Perhaps you have already tried to install the latest versions of InvokeAI and already know some new nuances that arise during installation?

[vagrant spire]

I can do 🙂

tomorrow I'll try to install the new version and I'll let you know if for you is ok 😉

[river sonnet]

Thank you for helping me again! 🥹

[vagrant spire]

I'm updating ubuntu...it's sad that with kernel 6 I must use repo for 22.10 instead of LTS 22.04 😕

[vagrant spire]

@river sonnet I need more time because I was testing kernel 6 for other stuff and ROCm ofc doesn't support it yet 🥹
tomorrow I'll reinstall ubuntu 22.04 so I can check the install script

[vagrant spire]

@river sonnet I'm installing with automated script from 2.3.5, so far I see that it installed pytorch in the correct version and I'm downloading models (again xD) I let you know in a while

[vagrant spire]

I finished the setup and downloads, but it start with CPU -.-

[river sonnet]

Of course! I'll wait as long as it takes 🙂

[river sonnet]

Yeah, same story 😔

[vagrant spire]

I don't think that is invokeAI issue, but still ROCm, because it doesn't use GPU even with a simple python code, so I need to investigate

[vagrant spire]

a little update...seems that last update of 22.04 broken dkms drivers, because rocm 5.3.3 uses kernel 5.18 and ubuntu uses 5.19 🙈
so I need to understand how to fix the issue, because rocm 5.4 seems doesn't work with pytorch, 5.4 uses torch 2.x instread of 1.13 xD

next time Nvidia for this purposes

[tiny thunder]

please write down the guide how you installed everything. because the video card is of the same generation, but I don't understand how to make ai run through the video card

[vagrant spire]

you find all guides and test in this topic, but seems broken rn
https://www.tomshardware.com/news/amd-rocm-comes-to-windows-on-consumer-gpus <--- I hope it will be a game changer, when? idk

for now with rocm 5.5 no windows support nor 7900 XTX O_O https://github.com/RadeonOpenCompute/ROCm/issues/2128

[vagrant spire]

@tiny thunder you find a method here #1049702853490511872 message

problems I have now it's with kernel drivers that didn't compile with last 22.04 version I downloaded from ubuntu.com so I need to check how to fix it 😄
rocm 5.4 never worked with torch 1.13 on my conf 😦

without kernel dkms I cannot use torch, so I need to work on it 😄

[tiny thunder]

I read in general what they write. The only useful thing I found in the messages. I put it on 22.04.2 lts and Romc 5.4.2 from amd's off-site. Maybe I'm using the wrong software versions?... I just honestly have already reread so much and am completely confused

[vagrant spire]

when I tested it before 5.4 doesn't work with torch 1.13 😦 @tiny thunder
about 22.04, the last version 22.04.2 with updated kernel on my conf doesn't work, I need to find the old 22.04 with kernel 5.18 or rollback to it

AMD is pretty bad with ROCm unfortunately 😦

[tiny thunder]

maybe use 20.04 LTS and rmc 5.3.3?

[vagrant spire]

yes, I used 20.04 LTS and rocm 5.3.3 and everything worked

[tiny thunder]

-_- i have no mouth and i must scream... these are about the emotions I have)))) I hope it works. you use manual Installation?

[vagrant spire]

rn I didn't installed invoke

because I have problems with dkms, but I noticed that is improved a lot the automatically script

in the past I used the manual mode because was the only way to get rocm torch xD

[tiny thunder]

ok) I'll try it tomorrow

[daring egret]

I think you are trying to make it work with a 6600XT in the discussion here atm, right? So dunno how helpful this is. But with a 6900XT it worked for me without a problem when I installed everything a few days ago with ubuntu 22.04.2 lts and rocm 5.5

[vagrant spire]

I cannot run it with 22.04.2 and rocm 5.3.3, I'll try with rocm 5.5 @daring egret

[vagrant spire]

@daring egret I just installed on fresh ubuntu 22.04.2 ROCm 5.5 and with simple python script:

import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
and venv with pytorch 1.13 and rocm5.2 (that is lasted that is using 1.13.x) I get:
UserWarning: HIP initialization: Unexpected error from hipGetDeviceCount(). Did you run some cuda functions before calling NumHipDevices() that might have already set an error? Error 101: hipErrorInvalidDevice (Triggered internally at ../c10/hip/HIPFunctions.cpp:110.)
return torch._C._cuda_getDeviceCount() > 0
cpu

so doesn't work anymore with last 22.04.2, at least on my machine

ok, I fixed, now also 22.04 require render group 😄

@tiny thunder and @river sonnet follow this guide https://docs.amd.com/bundle/ROCm-Installation-Guide-v5.5/page/Prerequisites.html
with 5.3.3 you don't need the render group, but now it's mandatory also for 22.04 😄

maybe you can fix, remove all amdgpu/rocm stuff and try to reinstall following the guide for ROCm 5.5 😉

[vagrant spire]

I confirm that also automated script work after following 5.5 installation guide

[tiny thunder]

In general, I tried to install it all in different ways and honestly - I was desperate for today. Maybe you can tell me what I'm doing wrong. Let's take the new Ubuntu 22.04.02 LTS as a basis. It is not a problem to reinstall because it is on the m2 ssd. After that, I install the updates that are offered and reboot. Then I take from this site инструкцию(https://docs.amd.com/bundle/ROCm-Installation-Guide-v5.5/page/How_to_Install_Rom.html#_How_to_Install ) and do the following -

sudo apt-get update
wget https://repo.radeon.com/amdgpu-install/5.5/ubuntu/jammy/amdgpu-install_5.5.50500-1_all.deb
sudo apt-get install ./amdgpu-install_5.5.50500-1_all.deb

Next is
sudo amdgpu-install --usecase=hiplibsdk,rocm

Then I do the following from this site -
https://docs-amd-com.translate.goog/bundle/ROCm-Installation-Guide-v5.5/page/Prerequisites.html?_x_tr_sl=auto&_x_tr_tl=ru&_x_tr_hl=ru

sudo apt list --installed | grep 'wget|gnupg2|gawk|curl'

Next
sudo usermod -a -G render $LOGNAME
sudo usermod -a -G video $LOGNAME

Further
echo 'ADD_EXTRA_GROUPS=1' | sudo tee -a /etc/adduser.conf
echo 'EXTRA_GROUPS=video' | sudo tee -a /etc/adduser.conf
echo 'EXTRA_GROUPS=render' | sudo tee -a /etc/adduser.conf

Then I follow this instruction and install everything from this instruction automatically -
https://invoke-ai.github.io/InvokeAI/installation/010_INSTALL_AUTOMATED/#installation

sudo apt update
sudo apt install -y python3 python3-pip python3-venv
sudo update-alternatives --install /usr/local/bin/python python /usr/bin/python3.10 3

Next
sudo apt update && sudo apt install -y libglib2.0-0 libgl1-mesa-glx

And then I just have to install the automatic installation by selecting the 2nd item. Where did I make a mistake or what didn't I do?

[vagrant spire]

It's all right, only after installing rocm reboot

[tiny thunder]

not working:C I'll try again tomorrow from the beginning

[vagrant spire]

which one GPU have you got? @tiny thunder

[tiny thunder]

yes. there is also a video card in the processor (rysen 5 5600 g), but I disabled it in the bios and it does not appear in the system.

[vagrant spire]

ok, but what have you got? 6900xt or less?

[tiny thunder]

6750xt

[tiny thunder]

Maybe it doesn't work on my graphics card? Because I have already reinstalled it 2 times with your correction for a reboot

I want to cry:C Decided to reboot 2 times just like that. And it worked after 2 times. Eh?

YAY. I WIN IT!!!!!

[vagrant spire]

Congr

[tiny thunder]

When generating, it just crashes without errors... I'm going to drink)

[tiny thunder]

do you know what these mistakes are

[vagrant spire]

Uhm nope never saw

But when I'll back to my PC I'll checking out

[vagrant spire]

@tiny thunder try to add:

sudo amdgpu-install --usecase=opencl

[tiny thunder]

the same mistakes

[vagrant spire]

🤔

try this command:

export HSA_OVERRIDE_GFX_VERSION=10.3.0
before run invoke

[tiny thunder]

RROR_COMPILATION (6)
MIOpen(HIP): Error [BuildHip] HIPRTC status = HIPRTC_ERROR_COMPILATION (6), source file: naive_conv.cpp
MIOpen(HIP): Warning [BuildHip] /tmp/comgr-ad3b08/input/naive_conv.cpp:39:10: fatal error: 'limits' file not found
#include <limits> // std::numeric_limits
^~~~~~~~
1 error generated when compiling for gfx1030.
terminate called after throwing an instance of 'miopen::Exception'
what(): /long_pathname_so_that_rpms_can_package_the_debug_info/data/driver/MLOpen/src/hipoc/hipoc_program.cpp:304: Code object build failed. Source: naive_conv.cpp
/home/puma/invokeai/invoke.sh: line 51: 5618 Aborted (core dumped) invokeai --web $PARAMS

maybe something went wrong?

[vagrant spire]

it's something related to HIP

[tiny thunder]

https://github.com/RadeonOpenCompute/ROCm/issues/1889

[vagrant spire]

https://github.com/RadeonOpenCompute/ROCm/issues/1889#issuecomment-1382501319

[tiny thunder]

Uh-huh. now it doesn't fly out. now it just freezes.... AHAHAHAHAHA

[vagrant spire]

LUL

first time need time

[tiny thunder]

okay. it generates black pictures

[vagrant spire]

did you use: export HSA_OVERRIDE_GFX_VERSION=10.3.0 ?

[tiny thunder]

yep

[tiny thunder]

without this, it crashes

[vagrant spire]

🤔

[vagrant spire]

close XD

I think there is a way to get a debug log

[tiny thunder]

come on))) see how cool it generates black squares

[tiny thunder]

how?

[vagrant spire]

I dont remember 😦

[tiny thunder]

This?

[vagrant spire]

nope it's only a list of generated images

try with:

--precision fp16
or
--precision fp32

how many model do you keep loaded?
try also this command:

--free_gpu_mem
it's slower but maybe help

but still I don't think is related to those settings

[tiny thunder]

When generating, other lines appear. Maybe some driver got up badly?

[vagrant spire]

for me can be precision at this point tbh

[tiny thunder]

Rephrase. I don't quite understand you

[vagrant spire]

try to use the precision flag when you start the program, I read few months ago that with 6700 precision causes some troubles

[tiny thunder]

Where is it?...

[vagrant spire]

when you start invoke.sh, add --precision fp32 or --precision fp16

it should change precision 😉

[tiny thunder]

I just didn't quite understand where to add it, because if I just write to the command line, he doesn't know this command

[vagrant spire]

uhm

can you show me the content of invoke.sh?

because I'm not in linux rn and I try to remember

[vagrant spire]

can you show me what it print out with 10?

"Command-line help"

[tiny thunder]

This?

[vagrant spire]

we need to find where is the setting file 😄

can you screen the content of invokeai pls?

[vagrant spire]

ok, give me the invoke.init

[vagrant spire]

ok

try this

[tiny thunder]

I did it

[tiny thunder]

it didn't help. changed this setting already

[vagrant spire]

🤔

[tiny thunder]

I think this driver got up somewhere wrong

[vagrant spire]

have you got miopen-hip?

[tiny thunder]

????

[vagrant spire]

I don't know, I'm trying to understand why it doesn't generate

[tiny thunder]

And what is it?

[vagrant spire]

can be something related to the HIP problem you have, but not sure 🤔
when you generate an image, what did you get in the terminal?

[vagrant spire]

uhm strange

[tiny thunder]

++

[vagrant spire]

maybe try to download or use another model?

smaller one

to understand if at least it can work

[tiny thunder]

I already did. Maybe the problem is that if I don't use the code in developer console - pip install torch torch vision --force-reinstall --extra-index-url https://download.pytorch.org/whl/rocm5 .4.2
Then my video card does not work. Maybe that's the problem?

[vagrant spire]

🤔

[tiny thunder]

I'll try to just reinstall. I will make a backup of what is already there

[vagrant spire]

invoke use venv

are you in a virtual machine?

or native?

[tiny thunder]

native

[vagrant spire]

ok, let's reinstall then

install only ROCm then for invoke we will check the settings 😉

[tiny thunder]

reinstall working...

[vagrant spire]

👍