Bug fixes

v0.2.2

bump version

v0.2.1

fix act checkpointing and add local model config file support

Bug fixes and MoE training analysis support

This release fixes a few bugs when calculating memory usage (e.g. activation, optimizer states), and adds support to analysis MoE training.

Bug fixes and Llama 2 inference support

This release:

• adds group query attention (GQA) support
• changes the activation memory calculation in inference to assume maximum tensor buffer
• fixes the kv cache size calculation
• adds a gpu cost analysis in the inference
• adds llama2 inference case study

v0.1.0

release