petermg
[Info] Running on Linux using an AMD GPU
@petermg You dont have Discussions enabled so Im just posting it here.
[My system: R9 5950X, 64GB RAM, 7900XTX+6800XT]
I just wanted to quickly write that this projects works with ROCM when using "pure" Linux. Works with both my cards separately. Didnt try to combine as it takes barely any VRAM.
I just did two additional steps.
-
Comment out the torch lines in requirements.txt
-
Installed the pytorch-rocm package before the requirements.
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.3
Thats it.
Works really well besides a bunch of spam in the console.
MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 40026112, provided ptr: 0 size: 0
MIOpen(HIP): Warning [IsEnoughWorkspace] [EvaluateInvokers] Solver <GemmFwdRest>, workspace required: 40026112, provided ptr: 0 size: 0
This isnt really problematic for the operation of the program, its more about the rocm drivers and what level of debug info that should be displayed. It can actually be fixed entirely by the original creators, not this fork here though. I think, been a while since I looked into it.
To suppress these warning you can just paste this in the terminal before running to set it to error only:
export MIOPEN_LOG_LEVEL=1
AMD ROCm™ is an open software stack including drivers, development tools, and APIs that enable GPU programming from low-level kernel to end-user applications. ROCm is optimized for Generative AI and HPC applications, and is easy to migrate existing code into.
https://www.amd.com/en/products/software/rocm.html
Its the thing that makes it possible to run CUDA applications on AMD hardware.
I created this issue here mostly for other users that are unsure if it will work on their AMD GPU.
Not sure how well it works on windows though, dont use it.
As this is a torch only project, without any xformers for attention or bitsandbytes quantization, it works without issue on AMD GPUs. Those two packages are the only pain points for AMD nowadays, as they often need to be compiled manually and depending on the required torch version that can be difficult.
Semi-related to this, I have a Pascal (10-series) GPU and I was able to get it to run just changing to cu124 and torch/torchaudio 2.6.0
Very cool! Thanks for the update. It will help people with your series GPU.
On Mon, Jun 2, 2025, 6:22 PM efhosci @.***> wrote:
efhosci left a comment (petermg/Chatterbox-TTS-Extended#2) https://github.com/petermg/Chatterbox-TTS-Extended/issues/2#issuecomment-2933045531
Semi-related to this, I have a Pascal (10-series) GPU and I was able to get it to run just changing to cu124 and torch/torchaudio 2.6.0
— Reply to this email directly, view it on GitHub https://github.com/petermg/Chatterbox-TTS-Extended/issues/2#issuecomment-2933045531, or unsubscribe https://github.com/notifications/unsubscribe-auth/AE6NOTIXVKDSPTS5BAAIKGL3BT2GZAVCNFSM6AAAAAB6LLKLZCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDSMZTGA2DKNJTGE . You are receiving this because you were mentioned.Message ID: @.***>