Current GPU compute stacks route integer inference through hardware designed for floating-point arithmetic, then compensate with conversion layers, precision negotiation, NaN handling, loss scaling, and rounding mode management. VDR-LLM-Prolog eliminates floating-point computation entirely — the model's forward pass, attention mechanism, softmax, training loop, and all infrastructure operations run on exact integer arithmetic with fixed denominators. TensorProlog is the GPU compute layer built for this architecture. It defines: a type system with three fixed-denominator Q-bases (Q16, Q32, Q335) and no float types; an instruction set where multiply-accumulate is a widening integer multiply plus bit shift; a memory model organized around fixed-size knowledge base structs at integer addresses; session-scoped streams with integer credential enforcement; a Prolog engine parallelized as batched cross-multiply unification across warps; grammar-directed structural token generation that bypasses the LLM forward pass; a runner system that provides autonomous background execution with snapshot-based recycling; and a server layer that clones sessions per connection with time-bounded integer credentials. The spec covers 23 modules, approximately 580 API functions (versus ~4,000+ in CUDA), approximately 30,000 lines of implementation across 168 files, and builds in six phases from pure-Zig CPU arithmetic through GCP GPU deployment. The entire float software stack — cuBLAS precision variants, cuDNN mixed-precision management, TensorRT quantization calibration, NCCL non-deterministic allreduce, loss scaling, gradient clipping, NaN recovery, and the Transformer Engine — is replaced, not optimized.
Geoffrey Howland (Fri,) studied this question.