| --- |
| license: mit |
| library_name: pytorch |
| tags: |
| - modular-arithmetic |
| - neural-arithmetic |
| - bit-serial |
| - gru |
| --- |
| |
| # NeuralHorner v8 compact inference experiment |
|
|
| This local derivative tests storage and inference optimizations against the |
| published MIT-licensed `TrickyRex/bitserial-modmul-v8` checkpoint. It is not a |
| newly trained model. Original model and weights by Robert Sneiderman: |
| <https://huggingface.co/TrickyRex/bitserial-modmul-v8>. |
|
|
| Changes: |
|
|
| - checkpoint tensors stored as bfloat16; |
| - direct `logit > 0` decisions instead of `sigmoid(logit) > 0.5`; |
| - one operand is reduced and the other is streamed directly through the |
| learned multiplication transition, removing one modulus-width pass; |
| - static multiplicand/modulus feature channels are allocated once per scan |
| instead of rebuilt at every recurrent step. |
|
|
| The last change is mathematically equivalent for an exact transition, but the |
| learned cell is not proven exact. Benchmark equivalence therefore must be |
| measured before this is treated as an accuracy-preserving release. |
|
|
| ## Local validation |
|
|
| Hardware: Apple GPU through PyTorch MPS. Dataset: the official 100-case public |
| benchmark for each tier. |
|
|
| - Tiers 1-7: 700/700 exact. |
| - Tier 6 runtime: 19.07 seconds versus 27.02 seconds for the published wrapper |
| (29.4% lower wall-clock in this comparison). The earlier short-schedule |
| version took 20.45-21.31 seconds before static-channel reuse. |
| - Tier 6 outputs: 100/100 byte-for-byte identical to the published wrapper. |
| - The bfloat16 checkpoint with the original three-pass schedule also produced |
| 100/100 outputs identical to the fp32 checkpoint on Tier 6. |
| - Official static analysis: clean. |
|
|
| The checkpoint is 948,196 bytes versus 1,887,610 bytes (49.8% smaller). The |
| shorter schedule removes exactly one `Leff` recurrent pass: 20% of recurrent |
| step calls on public Tiers 3 and 5-10, 14.3% on Tier 4, 25% on Tier 2, and |
| 33.3% on Tier 1. |
|
|
| ## Release gate |
|
|
| Do not claim unchanged Tier 1-10 accuracy yet. The current machine is too slow |
| for a practical full-width Tier 8-10 MPS run. Before release, run the official |
| 1,100-case scorer on CUDA for the three published seeds and require 100% on |
| every scored tier, then repeat the held-out 768-case adversarial battery. |
|
|
