Mach-1-Additive-35B + bf16 base vision tower (byte-identical clone, adds vision.safetensors)

a0dc6f0 verified 4 days ago

preview code

Raw

History Blame Contribute Delete

1.09 kB

Release notes

Current payload:

Experts: integer L1-ball trellis codes (K=1.5, V=8; book: packed/experts/codebook.safetensors) + wave_gamma per-wavefront scales + continuous fp16 su/sv side-streams. Expert files carry NO Wscale — sv is the complete scale (decode with wscale=None); wave_gamma applies in rotated space before the su/sv Hadamards.
Spine: 64-level integer lattice in the K=4 trellis. Head: int5-g64. Embedding: int4 + exact-overwrite exception list.
Ship gate: decode.py-primitive reconstruction == the served checkpoint (bf16 rounding, layers 0/20/39 x 4 experts x 3 projs).
Full measured board and retention table: see BOARD.md. Expert payload 6.207 GB.

Multimodal variant

This repo (Mach-1-Additive-35B-Multimodal) is a byte-identical clone of SyzygyResearch/Mach-1-Additive-35B plus vision.safetensors — the base model's bf16 vision tower (333 model.visual.* tensors, 0.8931 GB, byte-identical to Qwen/Qwen3.6-35B-A3B), as already accounted for in MANIFEST.json. The quantized language payload is unchanged.