STEEL

• S reamlined
• T ensor
• E xecution
• E ngine
• L ibrary

STEEL is a fun experiment writing essentially pure C++ to try to get to the most elegant tensor/ML library I can get while understanding how these are implemented, from the math, to cache aware efficient algorithims, to good ml practices.

Build

cmake -S . -B build
cmake --build build -j$(nproc)

Binaries

• build/matrix_bench — low-level matrix multiplication benchmark
• build/qwen_infer — interactive Qwen2 inference
• build/steel_bench — inference benchmark (prefill / decode / end-to-end)
• build/steel_tests — unit tests

Benchmark

./build/steel_bench --model qwen2.5-0.5b-instruct-fp16.gguf --threads 8

Options:

Flag	Default	Description
--model	qwen2.5-0.5b-instruct-fp16.gguf	Path to GGUF model
--threads	auto	CPU threads
--decode-tokens	64	Tokens to generate per decode test
--warmup	1	Warmup iterations
--iters	3	Benchmark iterations (use 5 to match llama-bench)

Reports mean ± stddev and best tok/s for prefill, decode, and end-to-end generation.