Parity Analysis with llama.cpp¶

ZigLlama is an educational reimplementation of transformer inference inspired by llama.cpp. This page provides an honest, detailed comparison of the two projects across features, performance, and coverage.

Feature Comparison¶

Performance Comparison (CPU-Only)¶

All measurements use the same hardware (8-core x86_64, AVX2, 64 GB DDR5) and model (LLaMA 7B, Q4_K_M).

Quantisation Coverage¶

Coverage: 16 / 20 formats (80 %).

Model Coverage¶

Coverage: 15 / 18 listed architectures (83 %).

ZigLlama's Unique Advantages¶

Despite the performance gap, ZigLlama offers several properties that llama.cpp does not:

Educational Design¶

• Layered architecture: Six progressively complex layers, each independently testable.
• Inline documentation: Every function includes a docstring explaining the mathematical operation it implements.
• Educational tests: 285+ tests that double as executable specifications.

Code Clarity¶

• Allocation-explicit: Every allocation is visible and paired with a defer deinit. No hidden allocators, no global state.
• No C dependencies (core): The inference engine is pure Zig with no libc requirement. BLAS integration is optional.
• Readable builds: build.zig is a single file with named targets for every example and test.

Safety¶

• Bounds checking: Debug builds catch out-of-bounds tensor access at the exact call site.
• Leak detection: GPA reports every leaked allocation on process exit.
• No undefined behaviour: Zig's safety semantics prevent the classes of bugs that plague C codebases.

Modern Tooling¶

• Cross-compilation: zig build -Dtarget=aarch64-linux produces an ARM binary from an x86 host in one command.
• Hermetic builds: No Makefiles, no CMake, no system-library hunting.
• Integrated testing: zig build test runs all 285+ tests with a single invocation.

Parity Roadmap¶