v2.0.4

Release Notes

Features

• Added support for operators including Conv1D, RotaryEmbedding, GridSample, and ArgMax.
• Introduced the custom MultiHeadMatmul operator.
• Added the SPACEMIT_EP_DISABLE_PASSES_FILTER option.

Optimizations

• Improved the implementation of the Transpose operator.
• Optimized LightGlue FP16 performance from 300 ms to 80 ms.

Bug Fixes

• Fixed TCM usage issues in multi-session scenarios.
• Fixed FP16 accuracy issues in several Model Zoo models.
• Fixed correctness issues in several Transformer networks.