Mila::Dnn::LlamaBlock< TDeviceType, TPrecision, TWeightQuant, TKvPolicy >::BlockBuildContexts Struct Reference

The per-child build contexts this block implies. More...

Public Attributes
dim_t	batch { 0 }
BuildContext	gate_up
dim_t	head_dim { 0 }
BuildContext	hidden
BuildContext	main
dim_t	model_dim { 0 }
dim_t	num_heads { 0 }
dim_t	num_kv_heads { 0 }
BuildContext	qkv
dim_t	sequence { 0 }

Detailed Description

template<DeviceType TDeviceType, TensorDataType TPrecision, WeightQuantPolicy TWeightQuant = NoWeightQuant, KvCachePolicy TKvPolicy = NoKvCompression>
struct Mila::Dnn::LlamaBlock< TDeviceType, TPrecision, TWeightQuant, TKvPolicy >::BlockBuildContexts

The per-child build contexts this block implies.

Shared with onBuilding() so the two cannot disagree. The inference and training paths differ only in the sequence extent – one prefill chunk against the full context – so both fall out of the same derivation.

---

The documentation for this struct was generated from the following file:

• Mila/Src/Dnn/Components/Transformers/LlaMa/Llama.Block.ixx