Write a good bug report

Collect versions

Downloadcollect-rocm-report.shand run it on the box that fails:

curl -fsSL https://amdai.net/rocm-problems/collect-rocm-report.sh -o collect-rocm-report.sh
chmod +x collect-rocm-report.sh
./collect-rocm-report.sh | tee rocm-report.txt

The script prints OS (without hostname), amdgpu/kfd, rocminfo lines that are not UUIDs, hipconfig, amd-smi or rocm-smi, optional clinfo, and the env vars people actually set (HSA_OVERRIDE_GFX_VERSION, device isolation). Read the file before you paste. It does not upload anything.

Template

## What I expected
Tokens on the discrete GPU (gfx ID: ____).

## What happened
Paste the first error, not a paraphrase.

## Hardware
- GPU / gfx:
- OS / distro / kernel:
- ROCm / HIP prefix:
- Backend + version (llama.cpp / Strata / vLLM / Ollama):
- Model / quant / context (Q4 floor):

## Command
```
# exact argv
```

## Already tried
- [ ] llama.cpp Vulkan on the same GGUF
- [ ] Unset HSA_OVERRIDE_GFX_VERSION
- [ ] One GPU visible (HIP_VISIBLE_DEVICES / GGML_VK_VISIBLE_DEVICES)
- [ ] Matching thread: (amdai tracker URL or GitHub issue)

## Logs
Attach `rocm-report.txt` and `dmesg | grep -iE 'amdgpu|MODE1|MES'` if it hung.

Where to file

• Driver / runtime / HIP: ROCm/ROCm issues
• llama.cpp: ggml-org/llama.cpp
• vLLM: vllm-project/vllm
• Ollama: ollama/ollama
• PyTorch: pytorch/pytorch
• Home-lab discussion with a rig badge: Ask the forum

Search the tracker and the upstream issue first. If it matches, add repro data there. See how you can help.