AMD Strix Halo Local LLM Troubleshooting

September 19 functional update: the scoped runtime qualification records text (Qwen3.6), image (Qwen2.5-VL), executed tools (Devstral) and pinned Open WebUI on the existing 0.32.15 service after restart. Isolated 0.34.2 is useful but not default; official Qwen3.8 and full-reboot candidate acceptance remain open. The tested host's Ollama listener was LAN-reachable, not local-only. Released llama.cpp v0.4.1 passed bounded direct/server/HIP controls; this does not qualify every model, long-context shape or maximum-memory allocation.

Evidence reviewed: September 19, 2026.

This page extracts the most useful checks from the canonical README troubleshooting, known-issues section, and concise setup guide. Those files remain the source of truth for full commands, dated evidence, and caveats.

Ollama Runs, But Falls Back To CPU

Symptom: generation completes but is far slower than the guide's measured GPU routes. The measured Ollama 0.31.x builds could detect Radeon 8060S and then drop the integrated-GPU path when OLLAMA_IGPU_ENABLE=1 was missing.

Check: run ollama ps, then inspect journalctl -u ollama for Vulkan, iGPU, and GPU messages.

Fix: make sure the service has OLLAMA_VULKAN=1, OLLAMA_IGPU_ENABLE=1, and HIP_VISIBLE_DEVICES=-1; reload systemd and restart Ollama. Follow the full README troubleshooting entry.

Prompt Processing Is Slow Because -ub Exceeds -b

Symptom: prompt processing is below the relevant guide row while generation looks normal.

Check: compare the -b and -ub values in the exact command. llama.cpp silently clamps -ub to min(n_batch, n_ubatch) when -ub is larger.

Fix: use a micro-batch no larger than the batch and rerun the same command before changing the model or backend. See the canonical batch-clamp troubleshooting note.

AMDVLK Silently Wins ICD Selection Instead Of RADV

Symptom: Vulkan inference works, but prompt processing is unexpectedly slow or output identifies the AMD open-source driver instead of RADV STRIX_HALO.

Check: inspect vulkaninfo --summary and the device/driver lines printed by llama-bench.

Fix: remove AMDVLK so its ICD file cannot override RADV. An explicit RADV ICD selection can be used as a diagnostic, but the guide's normal path keeps AMDVLK uninstalled. Read the full Vulkan driver correction.

linux-firmware-20251125 Breaks The ROCm Path

Symptom: instability, crashes, or ROCm containers failing to start after a firmware-package change.

Check: inspect the installed linux-firmware package version.

Fix: do not use linux-firmware-20251125 on the documented Strix Halo ROCm path; inspect available distro packages, restore an unaffected version, reboot and repeat the GPU/model checks. Holding a package does not downgrade it. See Step 4.4 of the README.

Large Models Cannot See The Expected GTT Pool

Symptom: the OS or runtime exposes too little GPU-accessible shared memory, or a model that should fit fails during allocation.

Check: inspect /sys/module/amdgpu/parameters/gttsize, /sys/module/ttm/parameters/pages_limit, the active kernel command line, and free -h.

Fix: compare the selected RAM/kernel profile with actual allocation needs. The guide's amdgpu.gttsize=131072 ttm.pages_limit=31457280 values belong to the 128GB Beelink profile, not a 96GB preset or a universal OOM fix. Preserve unrelated settings, resolve conflicts, and verify live values after any reboot. Use the complete kernel-parameter procedure.

A Stale ROCm Override Reports gfx1100 Or Segfaults

Symptom: a current container reports gfx1100 instead of native gfx1151, or crashes during model load.

Check: run printenv HSA_OVERRIDE_GFX_VERSION and compare device detection with the variable unset.

Fix: remove a stale global HSA_OVERRIDE_GFX_VERSION from host shell or service startup files and retry. Keep the older 11.5.1 value only in commands that deliberately reproduce the dated b8460/kernel 6.19.4 evidence. See the current ROCm migration check.

HIP Loads, But Does The Exact Workload Remain Correct?

Symptom: long-context, vision, or multi-slot HIP output repeats or becomes garbled on the integrated-host compute path described by upstream reports.

Check (September 19 status): compare exact outputs on historical stock b10687 and released v0.4.1, recording the model, prompt, context, backend commit, usable memory and buffer path. Issue #26209 remains open. PR #25863 closed unmerged; #28604 shipped the revert mitigation. Broader scheduler PR #27311 remains open. The September 19 Coder controls passed on the release at 4,579 and 14,029 prompt tokens and with two concurrent distinct markers; historical b10687 failed retrieval and both slots. Both builds passed the Gemma image fixture. These are scoped controls, not general Qwen3.8, maximum-memory or all-model HIP qualification.

Fix: pin a known-good or patched HIP build and run exact-output controls before making a practical-model recommendation; use the documented Vulkan route as a comparator when appropriate. The guide's b10046 result is only a small-model allocation/setup smoke, not long-context, multimodal, or multi-slot correctness. Read the scoped upstream compatibility alert.

Independence And Affiliate Disclosure

This guide contains no affiliate links as of September 19, 2026. Future affiliate, loaned, gifted, sponsored, or early-access relationships must be disclosed near the relevant links or results and do not buy positive conclusions.