Work top down. Do not fine-tune a jinja template while rocminfo is empty.
Paste a log into the error troubleshooter (hipErrorNoBinaryForGpu, MODE1 reset, Ollama gfx1201, mixed apt repos, /dev/kfd permissions).
The GPU is not in rocminfo
lspcifirst. No AMD VGA/3D line → power cable, slot, or the card never initialized (dmesg | grep -i amdgpu).lsmod | grep amdgpuempty → kernel module / Secure Boot / DKMS. TTY or SSH recovery; do not install a third driver to “fix” it.- Userspace from Ubuntu archive
rocmon a box you thought was Core SDK 10.0 →gfx1201never existed in that tree. See ROCm and install. - You installed every
amdrocm-core-sdk10.0-*family package → TheRock wants one gfx family. Purge extras.
Permission denied on /dev/kfd
You are not in render/video, or you never logged out. Containers need /dev/kfd and /dev/dri and the right GIDs.
PyTorch or vLLM never sees the card
| Symptom | Likely cause |
|---|---|
torch.version.hip is None |
CUDA wheel from default PyPI |
pip install vllm then NVIDIA errors |
Same — you wanted --extra-index-url ROCm or vllm-openai-rocm |
Available false, rocminfo fine |
Groups, HIP_VISIBLE_DEVICES, or wheel built for the wrong gfx |
| Device name is the iGPU | Visibility env; pick the discrete index |
AMD’s 10.0 docs pin vLLM 0.27 + PyTorch 2.13. Community ROCm Docker is a different pin (vllm/vllm-openai-rocm:v0.30.0; previous v0.29.0). Do not mix them in one venv. Version notes.
Generation is glacial and VRAM is empty
The process is on CPU. Ollama/Windows ROCm for gfx1201 has been a frequent miss — use Linux + a backend that logs HIP/Vulkan, or llama.cpp you compiled.
llama.cpp HIP configure fails
HIP_PATH=/opt/rocm or /opt/rocm/core-10.0 (confirm). If CMake still sulks, Vulkan tonight, HIP tomorrow.
vLLM dies on warmup / RCCL
Single GPU first. Two home cards are an RCCL project. Read the image’s ROCm version. 0.21–0.25 warmup bugs were a documented class of pain — use a release the notes say is fixed, or AMD’s 0.27 image.
Qwen thinks forever or leaks tokens
Wrong chat template (--jinja), or an abliterated edit that broke termination. Compare official instruct. Disable thinking if the card says to.
Harness is “broken”
curl the inference URL. If curl is fine, the bug is Hermes/OpenClaw config (Ollama /v1 vs native API, missing VLLM_API_KEY, iGPU endpoint). Harnesses.
Firmware / MES / “it worked yesterday”
RDNA4 still sees firmware and MES surprises after a kernel upgrade. If a stack died after apt upgrade, bisect kernel vs ROCm userspace before you reinstall TheRock or rebuild llama.cpp. Check AMD’s release notes, not a Discord screenshot. If HIP is newly broken but the card still paints a desktop, try Vulkan llama.cpp before you nuke /opt/rocm/core-10.0.