R1

2026-09-11 14:55:34 -04:00
parent 660294d0aa
commit 25b2c4be2e
33 changed files with 1293 additions and 34 deletions
@@ -0,0 +1,20 @@
+# Hardware compatibility
+
+JARVIS-QVAC has one inference owner: `jarvisd` owns the QVAC master and every
+model call goes through `daemon/qvac-master.js`. The harness, media jobs, and
+computer-use tools do not start competing QVAC servers. The GPU gate is strict:
+when QVAC cannot report a usable GPU backend, inference fails with a truthful
+diagnostic instead of silently falling back to CPU.
+
+| Hardware path | Recommended profile | Expected backend | Notes |
+| --- | --- | --- | --- |
+| AMD Vulkan | laptop-16gb or desktop-gpu | Vulkan/llama.cpp | Verify the QVAC runtime sees the ICD, not only `vulkaninfo`. |
+| NVIDIA CUDA | desktop-gpu | CUDA/cuBLAS | Install a driver compatible with the QVAC build and verify VRAM. |
+| Intel Vulkan | laptop-16gb | Vulkan | Use a current Mesa stack; larger media jobs may remain queued. |
+| CPU only | unsupported for inference | none | Wake and diagnostics can run, but the master refuses model inference. |
+
+Run `npm run gpu-doctor` before downloading models. `assessModelFit()` should
+select the largest profile that fits reported VRAM; an unavailable model stays
+visible as unavailable and never causes the voice loop to crash. The current
+machine's GPU visibility is an installation concern when this document is
+updated; record the exact doctor output in the release checklist.