v0.6.11.2 — AMD: fix the in-process engine landing on an integrated GPU (issue #18)
On AMD rigs with motherboard/integrated graphics, the in-process Vulkan inference engine could load the model onto the iGPU — into shared system RAM instead of a discrete GPU's VRAM — which made the miner exit or restart-loop.
Fix
The engine now selects the discrete GPU against its own device list, so the choice is always valid: integrated GPUs are never used and a discrete AMD card is preferred. The model loads into real VRAM on a discrete card. (The earlier attempt selected a discrete GPU by an index taken from a separate enumeration, which on mixed iGPU + discrete boxes could still resolve to the iGPU or abort — that path is gone.)
Controls
- Default: auto-selects a discrete GPU. No configuration needed.
KERYX_LLAMA_VK_DEVICE=N— pin the engine to a specific GPU (the engine's device index: 0 = first, 1 = second, …). Safe on any index now.- The zero-dup VRAM saving still applies on RDNA cards; on other/GCN cards the engine hosts inference while each card keeps its full-speed OpenCL walk (unchanged, never slower).
Validation
Live-tested on an Intel-iGPU + 3× AMD box (the reproduction): the engine now loads on a discrete AMD card in real VRAM (system RAM normal, not UMA), with no crash; both auto-pick and the manual override work; 29.4 Mh/s, 0 rejects; consensus/PoM math unchanged and byte-exact.
AMD assets
| file | sha256 |
|---|---|
keryx-miner-supr-amd-0.6.11.2.tar.gz (HiveOS)
| edc062d9052c84d59550363e50f708ea32724b0c40cb88fbb74f429a8ecfccfc
|
keryx-miner-supr-amd-mmpos_0.6.11.2.tar.gz (mmpOS)
| 5f38c3769e075509bbf40ce9e7540c9be53667c669cb093b362b75c02525780f
|
keryx-miner-supr-amd-0.6.11.2.zip (Linux runnable)
| 5d9a8bc49ac08c69ff047d28690ee7447999a8620886335f4cec7cea0fb28bb4
|
Windows AMD build follows via CI (Windows has no in-process engine; it keeps the OpenCL walk + candle inference, so this change does not affect it).
NVIDIA Linux assets
12 tarballs (modern sm_75+ / legacy sm_70+ / pascal sm_60-61) + per-line SHA256SUMS. NVIDIA binaries are functionally identical to v0.6.11.1 (the issue-#18 fix is AMD/Vulkan-only); rebuilt so every platform installs as 0.6.11.2. Windows zips + macOS attached by CI.