fix: skip ggml_backend_load_all when use_gpu=0 in whisper filter

[dymoo]

Problem

The FFmpeg whisper filter (af_whisper.c in the jellyfin-ffmpeg submodule) unconditionally calls ggml_backend_load_all() during initialization, which triggers Vulkan and OpenCL backend registration. On CPU-only systems (e.g. Docker containers without GPU hardware or ICD drivers), this causes a segfault when:

1. ggml_backend_vk_reg() → ggml_vk_instance_init() → vk::createInstance() / vk::enumeratePhysicalDevices() crashes
2. The Vulkan C++ bindings may throw or the ICD loader may segfault before throwing

This means WhisperTranscriber.create({ useGpu: false }) still crashes on CPU-only containers, because ggml_backend_load_all() runs before the use_gpu parameter is consulted by whisper.cpp.

Fix

Update the jellyfin-ffmpeg submodule to include seydx/jellyfin-ffmpeg#1, which conditions ggml_backend_load_all() on use_gpu:

• use_gpu=0: Skip ggml_backend_load_all(). Only the CPU backend is needed (always available via GGML_USE_CPU). No GPU device enumeration occurs.
• use_gpu=1: Load all backends as before (preserving existing behavior).

Why this works

ggml_backend_registry (in ggml-backend-reg.cpp) uses C++11 function-local static initialization, making it lazy — it's only constructed on first call to get_reg(). Since ggml_backend_load_all() is the caller, skipping it means:

1. The lazy static constructor never runs
2. ggml_backend_vk_reg() never gets called
3. No Vulkan instance creation or device enumeration occurs

The CPU backend is independently available via static linking and doesn't need ggml_backend_load_all().

Impact

• Fixes segfault on CPU-only containers when use_gpu=0
• No behavior change when use_gpu=1 (default)
• Depends on: fix: skip ggml_backend_load_all when use_gpu=0 in whisper filter jellyfin-ffmpeg#1

Testing

• Verified on a CPU-only Docker container (Debian, no GPU/ICD): WhisperTranscriber.create({ useGpu: false }) no longer crashes
• Transcription works correctly using the CPU backend
• No Vulkan or OpenCL initialization occurs when use_gpu=0

[seydx]

Already merged upstream in seydx/jellyfin-ffmpeg and submodule updated. Closing.

[seydx]

https://github.com/seydx/node-av/releases/tag/v5.2.3