[Feature]: RX 7600 XT support

[jonasgf]

Suggestion Description

I recently bought an RX 7600 XT for primarily GPGPU purposes. I did unfortunately however not check the compatibility list before purchase and was sad to learn that my graphics card is currently not working with ROCm.

---

I have the latest ROCm packages installed (6.0.0).

rocminfo gives me the following:

$ /opt/rocm/bin/rocminfo --support
ROCk module is loaded
hsa api call failure at: /usr/src/debug/rocminfo/rocminfo-rocm-6.0.0/rocminfo.cc:1219
Call returned HSA_STATUS_ERROR_OUT_OF_RESOURCES: The runtime failed to allocate the necessary resources. This error may also occur when the core runtime library needs to spawn threads or create internal OS-specific events.

rocm-smi lists the GPU as such:

0       [0x0518 : 0xc0]       N/A     N/A    N/A, N/A        None  None  0%   unknown  Unsupported    0%   0%    
        Navi 33 [Radeon RX 7                                                                                     

---

I first found the page listing supported GPUs on Linux. There is very few GPUs listed here.

The supported GPUs on Windows page does list quite a few more GPUs, also the ones that AFAIK work well on Linux. I do not however see the RX 7600 XT listed here at all, so this gives me some hope that the GPU might be supported at some point.

I hope you can consider adding support for the RX 7600 XT, since it seems like a fantastic choice for ML and GPGPU in this budget range.

Operating System

Arch Linux (Linux 6.7.4-arch1-1 x86_64)

GPU

AMD Radeon RX 7600 XT

ROCm Component

No response

[jonasgf]

Or am I mistaken, and this is supported as gfx1102? There might just be something wrong with my setup if others have gotten this to work. I am trying to get this to work on a headless server I just set up today.

[jonasgf]

FWIW, I get the following in the kernel log when attempting to run rocminfo:

amdgpu 0000:09:00.0: amdgpu: Timeout waiting for VM flush ACK!
amdgpu 0000:09:00.0: amdgpu: Timeout waiting for VM flush ACK!
[drm] PCIE GART of 512M enabled (table at 0x0000008000000000).
[drm] PSP is resuming...
[drm:psp_hw_start [amdgpu]] *ERROR* PSP load kdb failed!
[drm:psp_resume [amdgpu]] *ERROR* PSP resume failed
[drm:amdgpu_device_fw_loading [amdgpu]] *ERROR* resume of IP block <psp> failed -62
amdgpu 0000:09:00.0: amdgpu: amdgpu_device_ip_resume failed (-62).
[drm:mes_v11_0_submit_pkt_and_poll_completion.constprop.0 [amdgpu]] *ERROR* MES failed to response msg=14
[drm:amdgpu_mes_flush_shader_debugger [amdgpu]] *ERROR* failed to set_shader_debugger

[jonasgf]

So it seems like the problem was something with my setup. It works fine if I have a monitor attached to the graphics card.

I do however have to run my pytorch programs with HSA_OVERRIDE_GFX_VERSION=11.0.0.

And when it comes to the supported GPUs list, if the support is implied with the RX 7600 (non-XT), then my bad for misunderstanding. If there is a relevant difference then I think it should be added to the list if there is support for it.

[das-auge]

I am not sure if I understood right: Can you confirm that the RX 7600XT is provididing hardware acceleration for ROCm and Pytorch?

[jonasgf]

I was able to use it, yes. But I guess it is not officially supported.

[kozuch]

AMD really sucks failing to provide proper support info. How they want to attack nVidia's market share like this? Is there maybe a community site where people would list supported cards themselves?

[kozuch]

@jonasgf Do you have token generation speed from LLM (Llama 3 etc.) on this card?

[bdutta]

@jonasgf would you mind sharing some information about the Linux OS/release you are running and what is the CPU/motherboard that you have ? I am looking at buying a RX 7600 XT for local LLM/inference (ollama/lmstudio), so would be needing ROCm support, but clearly AMD remains laggard with their documentation. They do mention need of PCI atomics in the processor, but reading some more posts, I gather that it might be a combination of processor and motherboard that ensure availability of PCI atomics for this GPU to work fine. I was hoping to reuse my existing i5-8400 it possible with this GPU.

[TeaCult]

I made it work on archlinux with rocm 6.02 and torch 5.7
gemma-2-27b-Q4 10.5 toks/sec
SDXL %85 of RTX 3060 performance (I dont remember it/sec) , force-fp16 fixes a lot in vae decoding

Generally consumes too much ram in generations and training in pytorch compared to RTX although I preform frequent torch.cuda_empty_cache().

If bitsandbytes work (6.2 + torch nightly) , controlnets(comfy) does not , if controlnets work bitsandbytes does not. Is it a deal breaker ? Not sure.

Bottom line:
It works, You can generate images and use LLM inference on ggml gguf quantized files without any problem performantly. QLora tuning wont be working because of bitsandbytes and peft.

[Deng-Xian-Sheng]

run yolo12 by Ubuntu24 7600XT 8gb fail

[henfri]

Hello,

what about wsl? Is the 7600 XT supported?

The official Matrix does not show it as supported, but elsewhere I read that it works.

Greetings,
Hendrik