One IR from H100 to your phone

2026-06-30 · Caldrix Engineering

FlashAttention 3 is a brilliant algorithm expressed as a CUDA artifact. Port it to Metal and you are not porting an algorithm — you are re-deriving it against different memory hierarchies, different warp semantics, different everything.

Caldrix IR will encode FlashAttention as what it actually is: a tiling and recomputation strategy over the attention computation, parameterized by the memory hierarchy it runs against. On sm_90 the lowering will reproduce the kernel you know. On Apple Silicon it will re-tile for the unified-memory profile. On a Hexagon NPU it may decide the strategy doesn't pay at all and lower to something humbler — but it will decide that from the same IR, with the same cost model that placed your KV cache and picked your quantization.

That is the whole bet behind Caldrix: stop porting artifacts between vendor stacks, and start compiling concepts down to whatever silicon shows up.