Research/rosa torch fusion 20260811 - #1

Merged

aabbdev merged 29 commits into

research/rosa-torch-fusion-20260811

Aug 11, 2026

Commits on Aug 10, 2026

Use dense Numba arrays for exact ROSA
aabbdev
committed
8918661
Use sparse Numba transitions
aabbdev
committed
1748a0f
Avoid Numba parallel overhead on small inputs
aabbdev
committed
ce812fe
Integrate optional Numba inference backend
aabbdev
committed
0b57a07
Use a Link-Cut Tree for online ROSA
aabbdev
committed
bcd61ae
Fuse stateful ROSA prefill in Numba
aabbdev
committed
f5c2628
Expose production stateful ROSA inference
aabbdev
committed
d6ddc69
Use open-addressed SAM transitions
aabbdev
committed
792bc0f
Fuse stateful prefill into one kernel
aabbdev
committed
420363f
Use pinned buffers for CUDA steps
aabbdev
committed
ae48b0d
Revert "Use pinned buffers for CUDA steps"
aabbdev
committed
6faf6e3
Parallelize large stateful batches
aabbdev
committed
a795dcf
Revert "Parallelize large stateful batches"
aabbdev
committed
f23eea0
Use optional native stateful stepping
aabbdev
committed
133a6f8
Build stateful ROSA prefill in bulk
aabbdev
committed
cf8f8e3
Package ROSA native acceleration for 0.2
aabbdev
committed
3c20281
Build stateful prefill in native code
aabbdev
committed
a8bad50
Track rich stateful ROSA candidates
aabbdev
committed
2555552
Support recyclable ragged inference slots
aabbdev
committed
a2c75ad
Add structured differentiable candidates
aabbdev
committed
9d6b62a
Use native rich candidates in ROSA forward
aabbdev
committed
528c6d9
Add native rich prefill and ragged stepping
aabbdev
committed
cd09e7e

Commits on Aug 11, 2026

Unify top-one rich and ragged inference state
aabbdev
committed
a71a3bd
Document unified rich inference
aabbdev
committed
9a7b9b7
Use fused rich prefill in ROSA forward
aabbdev
committed
2cc75cc
Parallelize native batch prefill
aabbdev
committed
d0e93b6
Add reusable native candidate buffers
aabbdev
committed
a708243
Reuse candidate tensor views
aabbdev
committed
3b43225
Release rosa-torch 0.2.0
aabbdev
committed
7864fb0