fix(installer): route RDNA1 and 6GB AMD GPUs to Vulkan and a fitting model

[Dario-PC-and-Gaming]

Summary

• Stop treating unknown discrete AMD GPUs as Strix Halo gfx1151 / ROCm 7. Navi 10 (RX 5600 XT, gfx1010, PCI 0x731f) is not ROCm 7-capable, so Linux now selects Lemonade Vulkan and skips the HIP llama-server build.
• Run AMD topology on a single GPU (not only 2+), add Navi 10 entries to the GPU database, and for known VRAM under 8 GB pick a 6 GB-safe GGUF (Qwen 4B static / catalog fit such as Phi-4 Mini) instead of Qwen 9B.
• Windows installer: pass VRAM into tier mapping, and stop PowerShell 5.1 from treating curl/npm stderr as a fatal NativeCommandError so AMD installs can finish.

AI Assistance

Cursor Agent drafted the installer/GPU routing changes, contract tests, and this PR text. I reviewed the diff, ran the focused tests, and validated inference on local hardware.

Release Lane

• Stable hotfix targeting release/2.6.x
• Mainline change targeting main
• Next-minor work targeting the next feature/minor release
• Not sure; reviewer should help classify

Stable hotfix reason:

N/A — mainline hardware-reach fix.

Changed Surface

• Docs only
• Tests only
• Dashboard UI
• Dashboard API / host agent
• Installer / bootstrap / lifecycle
• Docker Compose / service manifests
• Model routing / Hermes / capabilities
• Network exposure / auth / proxy
• Dependencies / runtime wiring

Risk And Validation

• Risk level: High (installer phases, Windows installer, AMD compose overlay, model selector)
• Validation run:
  • git diff --check
  • Markdown/link sanity for docs
  • Focused tests listed below
  • Dashboard lint/test/build
  • Extension audit / compose validation
  • Release-grade fleet or scoped hardware validation
  • Stable-lane patch validation, if targeting release/2.6.x
  • Not required because:

Commands/results:

bash tests/test-rx5600xt-compat.sh
  21/21 passed

Live Windows (AMD Radeon RX 5600 XT, ~6128 MB, discrete):
  classify: rx_5600_xt, T1, 336 GB/s
  installer: Tier 1 Entry Level (6GB), catalog pick phi-4-mini
  Lemonade 10.0.0: --llamacpp vulkan, device=gpu
  chat completions: HTTP 200; first load ~90s; follow-up ~216 ms / ~48 tok/s
  Full Windows AMD stack later completed (dashboard :3001, Open WebUI :3000, Lemonade :8080).

Deferred:
  Linux ROCm-vs-Vulkan compose path not exercised on this host (Windows uses native Lemonade).
  Fleet / make gate not run.

Operational Change Check

• This is not an operational change.
• This is an operational change and validation is recorded above.
• This is an operational change and validation is intentionally deferred for:

Notes For Reviewers

• Rollback: revert this PR. Existing installs keep their .env until re-run; new AMD gfx10 installs will stop defaulting to ROCm 7 / gfx1151.
• Known pre-existing: Windows PnP can report HasNpu=True on this desktop (false positive). Not changed here.
• Windows-only NativeCommandError guards in ui.ps1 / 07-devtools.ps1 were required to get past OpenCode/npm downloads on PowerShell 5.1 during the live install. They are bundled because that path blocked the hardware validation; happy to split if review prefers.
• Out of scope: Docker image disk vs INSTALL_DIR preflight (follow-up PR); SearXNG not started with -NoRecommended while Open WebUI still enables web search.