mod-arithmetic / manifest.json
mikelou's picture
Add faster compact NeuralHorner v8 prototype
efe8193 verified
Raw
History Blame Contribute Delete
784 Bytes
{
"entry_class": "model.BitSerialReducer",
"output_base": 2,
"framework": "pytorch",
"model_description": "Compute-optimized NeuralHorner v8 inference wrapper using the original ~471K-parameter bidirectional two-layer GRU transition. Stores weights in bfloat16, thresholds logits directly, reuses static feature channels, and removes one redundant modular-reduction pass by streaming one original operand directly through the learned Horner multiply transition.",
"training_description": "Uses the published TrickyRex/bitserial-modmul-v8 weights (MIT), warm-started and fine-tuned by its author on one-step modular transitions. This derivative changes checkpoint precision and the mathematically equivalent inference schedule; it does not retrain the learned transition."
}