The hexagon backend introduces support for splitting models across multiple devices using a row-split strategy. This change enables work distribution in fused kernels and updates device configuration to handle device groups.
- Adds multi-device row splitting capabilities to the hexagon architecture.
- Implements work splitting logic for fused kernels and various operations like matmul and argsort.
- Updates session initialization, fence handling, and error recovery for robustness in multi-device scenarios.
- Refactors partitioning logic and synchronizes state across device groups.
This enhancement allows models to be distributed across multiple Hexagon cores or devices, potentially improving throughput and enabling larger model inference.