Mila::Dnn::QwenAttentionBlockWorkspace< TDeviceType, TPrecision > Struct Template Reference

Transformer-owned shared activation workspace for QwenAttentionBlock (pooling). More...

Public Types
using	MR = typename DeviceTypeTraits<TDeviceType>::memory_resource
using	TensorType = Tensor<TPrecision, MR>

Public Member Functions
std::size_t	deviceStorageBytes () const
	Total device bytes held, for memory accounting.

Public Attributes
std::shared_ptr< TensorType >	attn
std::shared_ptr< TensorType >	ffn_act
std::shared_ptr< TensorType >	ffn_down
std::shared_ptr< TensorType >	ffn_in
std::shared_ptr< TensorType >	gate
std::shared_ptr< TensorType >	gate_up
std::shared_ptr< TensorType >	gated
std::shared_ptr< TensorType >	k
std::shared_ptr< TensorType >	k_normed
std::shared_ptr< TensorType >	normed
std::shared_ptr< TensorType >	o
std::shared_ptr< TensorType >	q
std::shared_ptr< TensorType >	q_normed
std::shared_ptr< TensorType >	qkv
std::shared_ptr< TensorType >	query_gate
std::shared_ptr< TensorType >	res1
std::shared_ptr< TensorType >	stream
std::shared_ptr< TensorType >	v

Detailed Description

template<DeviceType TDeviceType, TensorDataType TPrecision>
requires PrecisionSupportedOnDevice<TPrecision, TDeviceType>
struct Mila::Dnn::QwenAttentionBlockWorkspace< TDeviceType, TPrecision >

Transformer-owned shared activation workspace for QwenAttentionBlock (pooling).

One slot set serves every layer: the inference path is strictly sequential, so exactly one block is live at a time. Slots are sized [B, chunk, width]; components view prefixes. The single stream slot is alias-safe – a block's input is last read at res_1, mid-block, and only overwritten by its own res_2 at block end.

The DeltaNet block will need a DIFFERENT workspace (chunk-parallel delta-rule buffers, convolution staging), which is exactly the mismatch section 7 records against ITransformerBlock's concrete GqaState. Keeping this struct per-block-kind rather than per-family is what leaves room for that.

---

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

• Mila/Src/Dnn/Components/Transformers/Qwen/Qwen.AttentionBlock.ixx