fix: respect NODE_LLAMA_CPP_GPU and GGML_CUDA env vars for CPU-only mode

[MichalSy]

Problem

On systems without a physical GPU, qmd still attempts to use CUDA if the CUDA toolkit or libraries are installed. This happens because:

1. getLlamaGpuTypes() from node-llama-cpp detects CUDA libraries on the system
2. qmd then calls getLlama({ gpu: "cuda", ... }) to use CUDA
3. This triggers a CUDA build of llama.cpp (which can take significant time and CPU)
4. On systems without a GPU, this build ultimately fails or runs very slowly in CPU fallback mode

Even when setting environment variables like:

• NODE_LLAMA_CPP_GPU=false
• GGML_CUDA=OFF
• NODE_LLAMA_CPP_CMAKE_OPTION_GGML_CUDA=OFF (related issue macOS Intel: unnecessary CUDA build attempt causes ~30s delay on every startup #185)

qmd would ignore these variables and still try to use CUDA if detected.

Impact

• High CPU usage: Users without GPUs see 100-400% CPU usage from CUDA build attempts
• Slow startup: qmd takes much longer to initialize while building CUDA binaries
• Wasted disk space: CUDA binaries (~50MB) are built even though they can't be used
• Confusing behavior: Setting NODE_LLAMA_CPP_GPU=false has no effect

Solution

This PR modifies the ensureLlama() function in src/llm.ts to:

1. Check environment variables first before calling getLlamaGpuTypes()
2. If NODE_LLAMA_CPP_GPU=false or GGML_CUDA=OFF is set, skip GPU detection entirely
3. Force CPU-only mode with getLlama({ gpu: false, ... })

Environment Variables Supported

Variable	Value	Effect
NODE_LLAMA_CPP_GPU	"false"	Force CPU-only mode
GGML_CUDA	"OFF"	Force CPU-only mode

Usage

Users without GPUs can now reliably force CPU-only mode:

# Option 1: Set in shell
export NODE_LLAMA_CPP_GPU=false
qmd search "query"

# Option 2: Set inline
NODE_LLAMA_CPP_GPU=false qmd search "query"

# Option 3: In wrapper script
#!/bin/bash
export NODE_LLAMA_CPP_GPU=false
export GGML_CUDA=OFF
qmd "$@"

Testing

Tested on a system without GPU:

• ✅ No CUDA builds triggered when NODE_LLAMA_CPP_GPU=false is set
• ✅ qmd uses pre-built CPU binaries from @node-llama-cpp/linux-x64
• ✅ CPU usage remains normal during embedding operations

Related

• Issue macOS Intel: unnecessary CUDA build attempt causes ~30s delay on every startup #185 - Discusses NODE_LLAMA_CPP_CMAKE_OPTION_GGML_CUDA=OFF but this alone is insufficient because qmd explicitly requests CUDA mode regardless of cmake options

Notes

• This is a non-breaking change: existing behavior is preserved for users who don't set the environment variables
• Users with working GPU setups are unaffected
• This aligns qmd's behavior with user expectations from node-llama-cpp environment variables