|
Mila
Deep Neural Network Library
|
Roles a Gated DeltaNet block instantiates, on top of the feed-forward pair. More...
Roles a Gated DeltaNet block instantiates, on top of the feed-forward pair.
The DeltaNet block is Phase 3 and does not exist yet; this concept is here because the plan below already names its roles, and a role no concept ever checks is a comment. When the block lands it requires this alongside FeedForwardPrecisionRoles, exactly as QwenAttentionBlock requires AttentionPrecisionRoles.
The query/key pair separates from the value/gate/output group because section 5 ranks them a half-step apart: the recurrence reads q and k to address the state matrix, while v, the output gate and o carry the magnitude that enters it.