Expand description
Virtual-layer → real-weight index scheduling shared by all families.
§Virtual-layer → real-weight scheduling
A {Model}Layers stack can run n_virtual_layers logical passes over only
n_real_layers weight sets (e.g. 48 logical from 12 real); each virtual
layer keeps its own cache but shares parameters. A Schedule maps a
virtual layer index to the real weight index to use.
For bidirectional stacks, BidiSchedule additionally interleaves the
two directions: even virtual indices run the straight (→) pass and odd
indices run the reverse (←) pass.
Each variant is documented with a worked virtual→real mapping example.
Enums§
- Bidi
Schedule - How a bidirectional layer stack maps virtual layer indices to real layer indices, interleaving the straight (→, even indices) and reverse (←, odd indices) directions.
- Schedule
- How a unidirectional layer stack maps virtual layer indices to real (weight-bearing) layer indices.