May 2026 (2026-05) focused on correctness and test coverage in the TT-MLIR pipeline. Delivered a dtype propagation fix for TTNNKVCacheDtypeConversion to ensure correct dtype wiring through TP model paths, along with targeted test coverage to prevent regression. These changes reduce runtime dtype errors and improve reliability when mesh_shard sits between operations, enabling safer experimentation with TP model configurations.

April 2026 monthly summary for tenstorrent/tt-mlir focused on enabling memory-efficient KV cache handling via BFP8. Delivered a new conversion pass and data-type support to reduce runtime memory footprint while preserving accuracy and performance. Key changes: - Implemented experimental KV cache dtype conversion pass (TTNNKVCacheDtypeConversion) to convert KV cache tensors to BFP8 and updated related operations (fill_cache, update_cache) to operate with the new types. Commit: aeb247375459f8a0accc6e886c8e3d1025aef66d. - Extended data-type support by adding BFP type handling to TensorDesc and generalizing WeightDtype to BFPDtype to be shared across conversion passes. - Strengthened runtime validation by constraining UpdateKVCacheOperation::validate_on_program_cache_miss to allow only FLOAT32, BFLOAT16, and BFLOAT8_B for both input and cache tensors, preventing unsupported BFLOAT4_B usage at runtime. - Resulting in clear business value: reduced KV cache memory usage, enabling larger effective batch sizes and models within the same hardware constraints, while maintaining correctness and integration with existing TTNN paths.

Concise monthly summary for March 2026 focused on the tt-xla repository contributions and impact.