llamastash / llamastash / 26778936539

87%

Build Type

push

github

Committed by web-flow

Commit Message

fix: Windows model-launch backend + Host-panel RAM/UMA + daemon self-heal (#11)

* fix(gpu): unify hardware-info codepath; fix Host-panel RAM + UMA star

The Host panel showed a wrong RAM value and a spurious unified `*` on
Windows AMD while `init` was correct. Two root causes: the DXGI probe
inferred UMA from `SharedSystemMemory` (~half system RAM on every
adapter, discrete included), and the host pane then subtracted those
bytes from the RAM gauge.

- Add `GpuInfo::is_unified()` as the single source of truth for the
  unified-memory decision; both init and the host pane derive from it.
- Host pane no longer subtracts UMA bytes: RAM always shows the true
  sysinfo total (matching init); a derived `unified` flag drives `RAM*`.
- DXGI: detect genuine UMA via the authoritative D3D12 `UMA` architecture
  flag instead of the shared-memory heuristic; fold the shared pool into
  VRAM total only for true APUs.

Linux / macOS / NVIDIA paths unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(cli): self-heal stale runtime.json so the daemon auto-respawns

After an unclean daemon exit (crash, `daemon stop --force`, power loss)
the leftover runtime.json pointed at a dead control plane. `Client::connect`
doesn't probe reachability, so `connect_or_spawn` trusted the file and
every call failed instead of auto-spawning — the TUI/CLI could never
restart the daemon until the file was hand-deleted.

- connect_or_spawn now gates the "runtime.json present" path on the PID
  lock (OS-released the instant the daemon dies, on both flock and
  LockFileEx). If no live daemon holds it, clear the stale runtime.json
  and auto-spawn. Spawn logic factored into `spawn_and_attach`.
- `daemon stop --force` removes the runtime.json it orphans (the hard
  kill skips the daemon's own shutdown cleanup).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(init): route Windows AMD GPUs to the Vulkan llama.cpp build

`win-hip-rad... (continued)

Coverage Stats

104 of 121 new or added lines in 27 files covered. (85.95%)

20 existing lines in 8 files now uncovered.

42229 of 48914 relevant lines covered (86.33%)

1630.37 hits per line