google/jax
main
test_983470154
[Pallas:SC] Lower swaps to tpu.compress_store_vreg or tpu.store without layout passes
On the path that does not run the layout passes, masked (compressed) swaps now lower to tpu.compress_store_vreg instead of a masked tpu.vector_store.
tpu.compress_store_vreg
tpu.vector_store