Add cuTile Rust backend#48

Open

robobryce wants to merge 29 commits into

mni-ml:mainfrom

robobryce:cutile-backend

Commits on Apr 21, 2026

add cuTile Rust backend as src/cutile

brycelelbach
and
claude
committed
replace host-fallback reductions with N-pass cuTile kernel

brycelelbach
and
claude
committed
make cuTile reductions a strict two-pass via zero-padded partitions
brycelelbach
committed
split cuTile kernels into per-op files

brycelelbach
and
claude
committed
port tier-1 fused cuTile kernels one-for-one from CUDA

brycelelbach
and
claude
committed
port softmax/cross_entropy/layernorm/fused_ops/reduce-along-dim/embedding/data

brycelelbach
and
claude
committed
add pooling and KV-quant cuTile kernels

brycelelbach
and
claude
committed

Commits on Apr 22, 2026

add Conv1D/Conv2D cuTile kernels

brycelelbach
and
claude
committed
add FlashAttention-2 forward + backward cuTile kernels

brycelelbach
and
claude
committed
add check_inf_nan_f32 cuTile kernel

brycelelbach
and
claude
committed
add generic permute cuTile kernel
brycelelbach
committed
add ops wrappers for elementwise/activation/reduce extension + softmax/norm/fused/dropout/grad_util/optimizer/mixed_precision
brycelelbach
committed
add cutile ops wrappers for avgpool2d and maxpool2d

brycelelbach
and
claude
committed
add cutile ops wrappers for conv1d and conv2d

brycelelbach
and
claude
committed
add cutile flash_attention forward + backward ops

brycelelbach
and
claude
committed
add cutile ops wrappers for embedding, data, loss, and kv_quant

brycelelbach
and
claude
committed
add cutile cargo feature on native crate

brycelelbach
and
claude
committed
flatten cutile pooling kernels to rank 3

brycelelbach
and
claude
committed
pass layernorm inv_c as runtime f32 parameter

brycelelbach
and
claude
committed
rename cutile runtime permute kernel to avoid intrinsic collision

brycelelbach
and
claude
committed
add cuTile correctness tests for newly-ported fused kernels

brycelelbach
and
claude
committed
fix cuTile shape issues in conv + flash_attention kernels

brycelelbach
and
claude
committed
add conv/flash-attention backward tests and fused-kernel benches

brycelelbach
and
claude
committed
Revert "add cutile cargo feature on native crate"
brycelelbach
committed
refresh src/native Cargo.lock after cutile feature revert

brycelelbach
and
claude
committed
bump cutile clang include path to llvm 15

brycelelbach
and
claude
committed
switch cutile global reductions to single-pass atomic

brycelelbach
and
claude
committed
document why sample_batch can't use TMA-lowerable view loads

brycelelbach
and
claude
committed
rewrite cutile benches to cover the training critical path

brycelelbach
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add cuTile Rust backend#48

Add cuTile Rust backend#48
robobryce wants to merge 29 commits into
mni-ml:mainfrom
robobryce:cutile-backend

Commits on Apr 21, 2026

add cuTile Rust backend as src/cutile

replace host-fallback reductions with N-pass cuTile kernel

make cuTile reductions a strict two-pass via zero-padded partitions

split cuTile kernels into per-op files

port tier-1 fused cuTile kernels one-for-one from CUDA

port softmax/cross_entropy/layernorm/fused_ops/reduce-along-dim/embedding/data

add pooling and KV-quant cuTile kernels

Commits on Apr 22, 2026

add Conv1D/Conv2D cuTile kernels

add FlashAttention-2 forward + backward cuTile kernels

add check_inf_nan_f32 cuTile kernel

add generic permute cuTile kernel

add ops wrappers for elementwise/activation/reduce extension + softmax/norm/fused/dropout/grad_util/optimizer/mixed_precision

add cutile ops wrappers for avgpool2d and maxpool2d

add cutile ops wrappers for conv1d and conv2d

add cutile flash_attention forward + backward ops

add cutile ops wrappers for embedding, data, loss, and kv_quant

add cutile cargo feature on native crate

flatten cutile pooling kernels to rank 3

pass layernorm inv_c as runtime f32 parameter

rename cutile runtime permute kernel to avoid intrinsic collision

add cuTile correctness tests for newly-ported fused kernels

fix cuTile shape issues in conv + flash_attention kernels

add conv/flash-attention backward tests and fused-kernel benches

Revert "add cutile cargo feature on native crate"

refresh src/native Cargo.lock after cutile feature revert

bump cutile clang include path to llvm 15

switch cutile global reductions to single-pass atomic

document why sample_batch can't use TMA-lowerable view loads

rewrite cutile benches to cover the training critical path