Show HN: TRiP – a complete transformer engine in C built from scratch just by me

TRiP is a solo-built, from-scratch transformer engine in pure C, released on HackerNews after 18 months of development. It supports Llama2, Gemma 1.0/1.1, PaliGemma, and GPT-2 with full forward and backward passes, AdamW optimizer, cosine annealing, BPE tokenization, and multimodal inference, requiring only gcc 13+, libjpeg, and libx11. Unlike llama.cpp, TRiP prioritizes pedagogical transparency over performance, making it uniquely valuable for engineers wanting ground-level understanding of attention, backprop, and tokenization without abstraction layers. CPU practitioners should note the confirmed float32 performance advantage over bf16/fp16 on non-accelerated hardware.