MmapLlamaLoader

Memory-mapped GGUF loader that provides zero-copy tensor access.

This loader memory-maps the GGUF file and creates tensor views that reference the mapped memory directly, avoiding data copies. This is particularly efficient for large models as:

• Memory is loaded on-demand by the OS (lazy loading)

• Multiple processes can share the same mapped pages

• No explicit memory allocation for tensor data

Limitations:

• JVM only (uses java.nio.MappedByteBuffer)

• Currently only supports F32 tensors (no quantized tensor support yet)

• Quantized models require dequantization which defeats the zero-copy benefit

Usage:

val loader = MmapLlamaLoader(path)
val weights = loader.loadToMap<FP32, Float>(ctx)
// Use weights...
loader.close() // Release mmap when done

Parameters