ejkernel v0.0.50

Breaking Changes

• Renamed attention_sink to softmax_aux across ragged_page_attention_v3 and unified_attention (modules + kernels + tests).
• Renamed mla_attention to flash_mla in ejkernel.modules / ejkernel.modules.operations.

Added

• New paged-attention ops with Triton + XLA backends:
  • chunked_prefill_paged_decode (updates block-tabled KV cache + runs unified paged attention).
  • decode_attention (paged decode attention returning output + LSE).
• XLA backend support for flash_mla.

Changed

• make_dummy_rpa_inputs now supports padded page tables via total_num_pages for smaller physical caches.
• Attention-module call signatures now place cfg at the end (chunked_prefill_paged_decode, decode_attention, unified_attention).

Testing

• Added/expanded Triton↔XLA equivalence tests for chunked_prefill_paged_decode and decode_attention, including shard_map coverage.

Full Changelog: v0.0.47...v0.0.50