Bring the v1 card up to main

Bring the v1 card up to main

Bring the v1 sources and card up to main

Uploaded using `kernel-builder`.

Bring the v1 sources and card up to main

Uploaded using `kernel-builder`.

Uploaded using `kernel-builder`.

Uploaded using `kernel-builder`.

Uploaded using `kernel-builder`.

Uploaded using `kernel-builder`.

Keep the float32 state contraction on tf32 and the scan on fma

Add the three-term tf32 fragment primitives

Rewrite the float32 measurement notes

Restore the float32 scan's layout and record the measurement

Record the sm90 comparison after the float32 row-tile change

Cut the float32 state kernel's row tile to the d axis

Add the CuTe decode comparison and share the backward shared-memory sizing

Tie the streamed state kernel's register tile to the tile constants

Size the backward state recompute from sm_state_bytes

Add the sm90 comparison, both dtypes

Record the measured k-prefetch result in the remaining-work notes

Check forward agreement at sequence lengths to 32768

Range-reduced rotation and contiguous backward operands

Uploaded using `kernel-builder`.

Update the card on the release branch to the current measurements

Uploaded using `kernel-builder`.

Uploaded using `kernel-builder`.

Sync v1 sources and card to main

Uploaded using `kernel-builder`.

Sync v1 sources and card to main

Uploaded using `kernel-builder`.

Uploaded using `kernel-builder`.

Uploaded using `kernel-builder`.

Uploaded using `kernel-builder`.

Mamba-3 MIMO chunked forward and recurrent decode, native CUDA

initial commit