v0.1.7-a0

SpacemiT MTMD Refactor Release Notes

This is a validation build for the MTMD refactor branch, not a production-ready release, we created a new branch named mtmd-backend.
This pre-release includes two major updates:

1. Refactored SMT MTMD Media Backend

• Reworked the multimodal integration to better align with upstream llama-server.
• Added a dedicated SMT MTMD media backend layer under tools/smt-mtmd.
• Introduced server-side media abstraction for vision/audio backend selection.
• Moved SMT vision/audio ONNX Runtime execution behind media backend wrappers.
• Added dedicated media worker thread execution so backend init/inference errors can be isolated and reported clearly.
• Renamed the build option to LLAMA_SERVER_SMT_MTMD to reflect full multimodal support instead of vision-only usage.
• Added SMT MTMD CI/build support, including RISC-V glibc and OHOS cross-compilation flow.

2. SpacemiT GGML/Gemma Optimization Updates

• Ported SpacemiT GGML backend optimizations from the previous branch.
• Added RVV optimized paths for unary operations such as tanh / gelu.
• Added optimized GEGLU execution paths used by Gemma-family models.
• Integrated IME/TCM thread coordination improvements so TCM resource management is handled consistently by the main scheduling flow.
• Improved RISC-V SpacemiT backend behavior for GGUF decode workloads while keeping ONNX EP affinity and llama.cpp GGUF worker affinity independent.