#702 [AMD HIP] RX 9060 XT (gfx1200) Repeating output

描述

I've had several cases of the model starting to loop indefinitely until the context is full on the same output sentence.
I am not sure why this is happening, I've looked at the troubleshooting section and other issues but haven't found a solution yet.

I am new at using Strata, I normally use llama.cpp on that same system without any problem.

I've seen this problem with both ISTA coder-iq1_m and iq3_xxs, running at kv q8 with 64K context on a gfx1200 (RX 9060 XT 16GB), system has enough ram (42 GB free, 64GB total), performances are great otherwise.
I'm using the HIP backend with the auto ROCM downloaded (v 7.10.0a20251120).
I'm on linux Ubuntu 24.04 (CT), 7.0.14-16-pve # 1 SMP PREEMPT_DYNAMIC PMX 7.0.14-16 (2026-08-28T17:31Z) x86_64 x86_64 x86_64 GNU/Linux
Strata "version": "0.1.38"

I have had more success and less problem with long coding sessions with pi agent (although this could be luck, I'm not sure), most of the issues have been happening in a normal chat.

I haven't tuned anything, this is the "out of the box" ./setup.sh experience.

Otherwise, performances are great in general (30-45 T/s), it would be an amazing solution if I could find the reason why this is happening.