fix(llama): respect NODE_LLAMA_CPP_GPU override

[Kaylebor]

Problem

On some systems with AMD GPUs (and potentially others), getLlamaGpuTypes() returns an empty array even when Vulkan is available. This causes QMD to fall back to CPU-only mode, which is significantly slower for embedding operations.

Root Cause

While the detection logic looks correct (see src/bindings/utils/getLlamaGpuTypes.ts), it appears to fail on some Linux/Vulkan configurations (tested on AMD 7900 XTX). The exact root cause is unclear - could be path resolution issues or async handling.

Solution

Check for NODE_LLAMA_CPP_GPU environment variable before calling the potentially broken auto-detection. This allows users to explicitly specify their preferred GPU backend (cuda, metal, or vulkan).

Why this variable? Because it is what node-llama-cpp already uses: src/config.ts#L56

Why this approach:

• Non-breaking change: If the env var is not set, behavior remains identical to before
• Follows conventions: Uses the same naming pattern as other NODE_LLAMA_CPP_* variables used by the library
• User control: Allows overriding broken auto-detection without waiting for upstream fixes
• Clean workaround: Doesn't require modifying node-llama-cpp internals

Usage:

Force Vulkan on AMD GPUs

export NODE_LLAMA_CPP_GPU=vulkan
qmd status  # Now shows GPU: vulkan with offloading

Testing:

• Tested on AMD Radeon RX 7900 XTX (24GB VRAM) with CachyOS (Arch-based)
• Before: GPU: none (running on CPU)
• After: GPU: vulkan (offloading: yes) with 37GB VRAM detected

References:

• node-llama-cpp/src/config.ts:56 - env var definition
• node-llama-cpp/src/bindings/utils/getLlamaGpuTypes.ts - detection issue

[Kaylebor]

Maybe non-critical, but on my particular setup qmd completely fails to detect my fully-functional AMD GPU (yes, I do have Vulkan installed properly; in fact with the fix above it works well)