+- Inference (https://www.datasculpting.ai/inference)
+-- Forum: Drivers / Frameworks / Toolkits (https://www.datasculpting.ai/inference/forumdisplay.php?fid=8)
+--- Forum: AMD / ROCm (https://www.datasculpting.ai/inference/forumdisplay.php?fid=32)
+--- Thread: ROCm version that finally behaves for Qwen (/showthread.php?tid=500)
ROCm version that finally behaves for Qwen - routecraft - 09-26-2026
I figured I'd ask here before changing everything.
I'm mostly working with FLUX in MLX on 5090. The part I'm trying to understand is tokens/sec.
If you've tested something similar, what did you keep and what did you stop bothering with?
Re: ROCm version that finally behaves for Qwen - rocmriley - 09-27-2026
That matches what I was seeing. I don't think your conclusion is far off. The biggest change for me was simplifying the stack instead of adding one more optimization layer.
Re: ROCm version that finally behaves for Qwen - tensorworks - 09-27-2026
A couple things I'd check first. On Threadripper box, the easiest win is often to back off the aggressive settings and see whether the problem survives. Once you know the plain path works, then add the faster path back in piece by piece.
Re: ROCm version that finally behaves for Qwen - macropoint - 09-27-2026
What worked for me was pretty simple. On 7900 XTX, the easiest win is often to back off the aggressive settings and see whether the problem survives.
Re: ROCm version that finally behaves for Qwen - dockworks23 - 09-29-2026
Ran a quick check on my box. On A6000 with PyTorch, I saw basically the same behavior. I expected the model choice to matter more than it did.
Re: ROCm version that finally behaves for Qwen - batchyard - 09-29-2026
A couple questions before I guess too hard. Is this local-only or are you routing anything externally?
Re: ROCm version that finally behaves for Qwen - quantshed - 09-29-2026
A couple things I'd check first. On dual 3090, the easiest win is often to back off the aggressive settings and see whether the problem survives.