Expand description
Architecture-owned, explicitly admitted FreeToken Qwen3.5 text execution. The generic GGUF/Llama admission remains unchanged.
Modules§
- admission
- Cold admission with an explicit exporter contract and a reviewed artifact receipt.
- attention
- Architecture-owned attention admission and allocation-free execution. Packing, norm storage and rotary geometry arrive from the admitted exporter profile; neither tensor names nor a successful matrix multiply establish them.
- dense_
host - Exact-order FP8 projection over admitted immutable bytes. No allocation or activation quantization; scalar norm reads retain the canonical oracle.
- session
- Caller-budgeted cold state and allocation-free mixed-attention token execution. A token commits only after every layer, final norm, and output projection.
- weights
- Exact borrowed FTW projection views; all execution output is caller-owned.