Expand description
Weight loader: CMF tensor directory → Pipeline.
Storage rule: models WITH task masks are dequantized to f32 (masked
execution needs f32 row access; skill files are small by design).
Models without masks keep quantized matrices zero-copy from the mmap
(QTensor::Mapped) — this is what lets a 15B file run in a few GB
of RSS instead of 60 GB of f32.
Layer kinds come from arch.layer_types: FullAttention loads
self_attn.* (with auto-detected Qwen3.5 extras: per-head qk-norm by
tensor presence, output gate by q_proj row count); LinearAttention
loads the canonical core vmf_attn.* (folded at convert time).
Enums§
- Overlay
- Tensor source selector (spec §9): backbone, one skill’s overlay, or a soft superposition of top-m skills (claim 14 working tensors).