v0.6.32

Allow `blend` op to run on very large matrices.

v0.6.31

Fix context leak.

v0.6.30

Fix WASM read back.

v0.6.29

WASM compatible.

v0.6.28

Bump version.

v0.6.27

Fix safetensors convertor for lora.

v0.6.26

Redesign lora blend pattern constructors.

v0.6.25

Small fix.

v0.6.24

Neo tensor (#20)

* Implement ref count cache.

* `Tensor`'s meta data is also `View`.

* When creating tensor, use ref count cache.

* Fully leverages ref count buffer.

* Do not use temporary cache when loading the model.

* Ref count cache correctly initializes resources.

* Unify cache types.

* Reformat code.

* Bump version to v0.6.25

v0.6.23

Allow loading tensor deltas in LoRA and for embedding.