Mila
Deep Neural Network Library
Loading...
Searching...
No Matches
Mila::Dnn::Quant::Weight::DecoderPrecisionPlan Concept Referenceexport
module Dnn.Quantization.Weight.PrecisionPlan

What an attention-plus-feed-forward decoder block requires. More...

Concept definition

template<typename TPlan>
concept DecoderPrecisionPlan =
AttentionPrecisionRoles<TPlan> && FeedForwardPrecisionRoles<TPlan>
Mila::Dnn::Quant::Weight::AttentionPrecisionRoles
Fused QKV projection and the attention output projection.
Definition PrecisionPlan.ixx:57
Mila::Dnn::Quant::Weight::DecoderPrecisionPlan
What an attention-plus-feed-forward decoder block requires.
Definition PrecisionPlan.ixx:78
Mila::Dnn::Quant::Weight::FeedForwardPrecisionRoles
Fused gate+up projection and the down projection of a gated feed-forward network.
Definition PrecisionPlan.ixx:64

Detailed Description

What an attention-plus-feed-forward decoder block requires.

Every block kind in the tree today instantiates exactly these four projections. A block with a different mixer – a recurrent one, say – composes its own concept from its own roles rather than widening this one, so that adding a block kind cannot invalidate the plans of the blocks that already work.