ryomennagi/HDC-AM-Master-Weights

HDC-AM Master Fast-Weights Overlay

Hyperdimensional Computing Address Mapping master knowledge overlay, generated by the 444.4 MiB master fold (d=4096, strict bipolar {-1,+1}^D).

field	value
artifact	fast_weights_master.safetensors
size	465,985,624 B (444.4 MiB budget)
layout	single slots int8 tensor [113,766, 4096]
validations	mmap zero-copy open · TR-1 strict ±1 · associative recall top-1

Load (zero-copy mmap)

from huggingface_hub import hf_hub_download
from hdc_am_pipeline import MemoryMapped_OverlayLoader, HDC_Recall_Layer

path = hf_hub_download("ryomennagi/HDC-AM-Master-Weights",
                       "fast_weights_master.safetensors")
with MemoryMapped_OverlayLoader(path, writable=False) as mm:
    names = [k for k in mm.header_keys if k.startswith("slot_")]
    layer = HDC_Recall_Layer(mm, names, d=4096)
    print(layer.recall("Taj Mahal mausoleum Yamuna"))  # associative recall

Reproduce

python master_fill.py          # folds exact 113,766 rows -> 444.4 MiB
python verify_master.py        # gate checks (mmap / TR-1 / recall)