This is the D version of llama2.c by Andrej Karpathy. It runs inference for the llama2 model architecture recently published by Meta.

Initial code was generated by ctod tool and saved as ctod_initial.d

Some small manual adjustments:

• added cast(float*) to calloc and mmap
• because of lack clock_gettime on Darwin OS, it was changed with MonoTime from core.time
• commented out pragmas for OpenMP

To build inference:

dub build -b=release

To run example:

./llama2_d stories15M.bin -i "your_prompt"

Tested on:

• macOS (M1)
• Linux
• Windows

• Make code more iDiomatic
• Improve performance
• Add Windows support (port win.h/win.c files from original repo)
• Parallelize the code with std.parallel and SIMD

Any form of contribution is welcome. Feel free to open an issue or create a pull request. If you are contributing optimizations, please provide benchmarks and/or performance comparisons as well as the code to reproduce them.

• Andrej Karpathy for the original llama2.c implementation
• Dennis Korpel for great ctod tool
• cgbur for ideas for optimizations and readme structure