|
Mila
Deep Neural Network Library
|
One policy across every Qwen role, DeltaNet included. More...
Public Types | |
| using | DeltaNetGating = TPolicy |
| using | DeltaNetQueryKey = TPolicy |
| using | DeltaNetValueGateOutput = TPolicy |
| Public Types inherited from Mila::Dnn::Quant::Weight::UniformPrecisionPlan< TPolicy > | |
| using | EmbeddingTable = TPolicy |
| using | FeedForwardDown = TPolicy |
| using | FeedForwardGateUp = TPolicy |
| using | LanguageModelHead = TPolicy |
| using | OutputProjection = TPolicy |
| using | QkvProjection = TPolicy |
One policy across every Qwen role, DeltaNet included.
The generic UniformPrecisionPlan cannot name the DeltaNet roles – they are this family's, and a block kind must not be able to invalidate the plans of blocks that already work – so the family that adds roles adds its own uniform lift. Inheriting keeps the shared roles defined once: a role added to the generic plan arrives here.
Note this does NOT replace the generic lift for the attention block. That block requires only the shared roles, so QwenAttentionBlock<Cuda, BF16, PerGroupFp4<128>> still resolves through PrecisionPlanFor and stays a valid spelling.