[bug]: Add support for ROCm version for gfx1151 (Rocm >= 6.4)

[drag-c]

Is there an existing issue for this problem?

• I have searched the existing issues

Install method

Manual

Operating system

Linux

GPU vendor

AMD (ROCm)

GPU model

8060s gfx1151

GPU VRAM

96GB

Version number

6.9.0

Browser

Firefox

System Information

AMD RYZEN AI MAX+ 395 w/ Radeon 8060S
Linux Fedora 43: 6.17.7-300.fc43.x86_64

What happened

I'm trying to manually install invokeai so that I can choose to install ROCm version 6.4 (compatible with gfx1151), since the installer installs version 6.3.

uv pip install invokeai==6.9.0 --python 3.12 --python-preference only-managed --torch-backend=rocm6.4 --force-reinstall
but

error: invalid value 'rocm6.4' for '--torch-backend <TORCH_BACKEND>'
  [possible values: auto, cpu, cu130, cu129, cu128, cu126, cu125, cu124, cu123, cu122, cu121, cu120, cu118, cu117, cu116, cu115, cu114, cu113, cu112, cu111, cu110, cu102, cu101, cu100, cu92, cu91, cu90, cu80, rocm6.3, rocm6.2.4, rocm6.2, rocm6.1, rocm6.0, rocm5.7, rocm5.6, rocm5.5, rocm5.4.2, rocm5.4, rocm5.3, rocm5.2, rocm5.1.1, rocm4.2, rocm4.1, rocm4.0.1, xpu]

  tip: a similar value exists: 'rocm6.2.4'

For more information, try '--help'.

I also tried the installation with AppImage, but at Launch, I get this error:

bitsandbytes library load error: Configured ROCm binary not found at /home/fed/invokeaiapp/.venv/lib/python3.12/site-packages/bitsandbytes/libbitsandbytes_rocm63.so Traceback (most recent call last): File "/home/fed/invokeaiapp/.venv/lib/python3.12/site-packages/bitsandbytes/cextension.py", line 313, in <module> lib = get_native_library() ^^^^^^^^^^^^^^^^^^^^ File "/home/fed/invokeaiapp/.venv/lib/python3.12/site-packages/bitsandbytes/cextension.py", line 282, in get_native_library raise RuntimeError(f"Configured {BNB_BACKEND} binary not found at {cuda_binary_path}") RuntimeError: Configured ROCm binary not found at /home/fed/invokeaiapp/.venv/lib/python3.12/site-packages/bitsandbytes/libbitsandbytes_rocm63.so

What you expected to happen

I want it to run successfully, like Comfyui

How to reproduce the problem

I'm trying to manually install invokeai so that I can choose to install ROCm version 6.4 (compatible with gfx1151), since the installer installs version 6.3.

uv pip install invokeai==6.9.0 --python 3.12 --python-preference only-managed --torch-backend=rocm6.4 --force-reinstall
but

error: invalid value 'rocm6.4' for '--torch-backend <TORCH_BACKEND>'
  [possible values: auto, cpu, cu130, cu129, cu128, cu126, cu125, cu124, cu123, cu122, cu121, cu120, cu118, cu117, cu116, cu115, cu114, cu113, cu112, cu111, cu110, cu102, cu101, cu100, cu92, cu91, cu90, cu80, rocm6.3, rocm6.2.4, rocm6.2, rocm6.1, rocm6.0, rocm5.7, rocm5.6, rocm5.5, rocm5.4.2, rocm5.4, rocm5.3, rocm5.2, rocm5.1.1, rocm4.2, rocm4.1, rocm4.0.1, xpu]

  tip: a similar value exists: 'rocm6.2.4'

For more information, try '--help'.

Additional context

No response

Discord username

No response

[heathen711]

@drag-c --torch-backend is a feature of uv not invoke: https://docs.astral.sh/uv/reference/settings/#pip_torch-backend

If your system does have ROCM 6.4 installed, the auto mode should pick it up.

I was the one who updated the instructions due to the mix-matching difference between CUDA and ROCM; CUDA has support for mixing within the same family (e.g. 12.6 and 12.8 can work together) while ROCM does not support that, if you want 6.4 you must have 6.4 installed (either within the docker container or bare metal).

if not:

1. Open a bug with uv via https://github.com/astral-sh/uv/issues to inform them that rocm 6.4 isn't being detected correctly
2. fall back to using the --index-url https://download.pytorch.org/whl/rocm6.4 instead of --torch-backend

I should tell you, that InvokeAI has only been tested with 6.2.4 and is what is currently used in the docker images from https://github.com/invoke-ai/InvokeAI/pkgs/container/invokeai (I wrote the docker container for ROCM) so 6.4 may have crashes that others have no run into.

[ArKam]

@drag-c for information, I've done a documentation update (still to be merged) about AMD Rocm support on invoke-ai:

#8546

@heathen711 for information:
I've tested it with up to 6.9.0 and ROCm 7.1 it just works from gfx1030 up to gfx1201 so far (tested both).

Everything works perfectly fine.
We just need to finally validate and push this documentation update.