Identity policy - no compression. More...

Static Public Attributes
static constexpr bool	kIsActive = false

Detailed Description

Identity policy - no compression.

Zero storage overhead and zero runtime cost; all if constexpr branches on kKvCompressed compile away entirely. Default for all GroupedQueryAttention instantiations.

When this policy is active, kCacheDtype on GroupedQueryAttention falls back to TComputePrecision, so no kStorageDtype field is required or present.

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

Mila/Src/Dnn/Quantization/KvCache/Policy.ixx

Static Public Attributes

Detailed Description