From 67dde781f51178800c8f63b29e33d659d1a71f86 Mon Sep 17 00:00:00 2001 From: Jatin K Malik Date: Wed, 5 Aug 2026 21:56:02 -0700 Subject: [PATCH] docs(user-guide): note CUDA device 0 limit for dual NVIDIA (#644) Document that CUDA-backed pywhispercpp always uses device 0, and point to CUDA_VISIBLE_DEVICES or a Vulkan build as workarounds. Co-authored-by: Cursor Agent Co-authored-by: Jatin K Malik --- docs/USER_GUIDE.md | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/docs/USER_GUIDE.md b/docs/USER_GUIDE.md index 9987f70..3d17d19 100644 --- a/docs/USER_GUIDE.md +++ b/docs/USER_GUIDE.md @@ -158,6 +158,10 @@ English-only whisper.cpp specializations limit the language selector to English. On multi-GPU machines, Vocalinux prefers a **discrete** Vulkan device when one is present. You can override the device under **Advanced** settings (`whispercpp_gpu_device`). +Pip wheels of pywhispercpp are often CUDA builds. In that case Vocalinux always uses CUDA device 0 (the first NVIDIA GPU), because Vulkan GPU indices do not match CUDA ordinals. On a hybrid laptop the NVIDIA GPU is usually Vulkan GPU 1; feeding that index into CUDA would land on the CPU fallback instead. + +If you need a second NVIDIA GPU, either set `CUDA_VISIBLE_DEVICES` before starting Vocalinux or install a Vulkan-built pywhispercpp so the Advanced GPU picker applies. + To check which backend is being used, look for these log messages when starting Vocalinux: ``` [INFO] whisper.cpp using Vulkan GPU backend: AMD Radeon RX 6800