Several kernel optimization from aiter team#5074

Draft

Bernard-Liu wants to merge 54 commits intopytorch:mainfrom

ROCm:aiter/mi350_kernel_opt

Commits on Oct 27, 2025

Add gfx950 build support + fp16 fix + index type fix

avbokovoy
authored and
liligwu
committed
Change int64_t to index_t as template parameters in load_raw_per_warp

avbokovoy
authored and
liligwu
committed
Implement llvm fp16 buffer load for gfx950

avbokovoy
authored and
liligwu
committed
Fix c-style half to float cast

avbokovoy
authored and
liligwu
committed
Patch 256 half stores

avbokovoy
authored and
liligwu
committed
cta_per_row workgroup optim

shbiswas834
authored and
liligwu
committed
Added mi350 guards

shbiswas834
authored and
liligwu
committed
Fix index overflow in row load

shbiswas834
authored and
liligwu
committed
cta_per_row workgroup reduce by 4 optim

shbiswas834
authored and
liligwu
committed
Fix mixed_D frontend to backend connection

avbokovoy
authored and
liligwu
committed
changed max_segment_length_per_cta to 4096

kudomcho
authored and
liligwu
committed
added rocm guards and removed comment

shbiswas834
authored and
liligwu
committed
clean debug statements in Hip.cmake
liligwu
committed
Merge pull request #121

shbiswas834
authored and
liligwu
committed
Guard f16 llvm intrinsics with ROCm >=7.0

avbokovoy
authored and
liligwu
committed
fix the bug in dimention 160 in ROCm optimization
liligwu
committed
Cleanup optimized warp_per_raw kernel

avbokovoy
authored and
liligwu
committed
Add 320 embedding dim support for optimized warp_per_row kernel

avbokovoy
authored and
liligwu
committed
changed the max length per warp and cta per row WG size

root
authored and
liligwu
committed
added DPP and changed max length per warp to 16k

kudomcho
authored and
liligwu
committed
guard max segment warp based on emb dim

kudomcho
authored and
liligwu
committed
added guarding opt of max segment for the case batch size list=1

kudomcho
authored and
liligwu
committed
opt for grad_indice_weights kernel

root
authored and
liligwu
committed
added store row per warp on emb 192 and added accuracy test functionality

kudomcho
authored and
liligwu
committed
workgroup tuning and loop unrolled

shbiswas834
authored and
liligwu
committed
specialize

Hardcode84
authored and
liligwu
committed
explicitly link to tbb
liligwu
committed
added warpReduceAllSum with rocm guards

shbiswas834
authored and
liligwu
committed
revert unroll and wg tuning

shbiswas834
authored and
liligwu
committed
Minor update embedding_forward_split_kernel_template.cu
liligwu
committed
add tbb-devel to the install_build_tools ()
liligwu
committed
fix lint issues
liligwu
committed
solve lint issues
liligwu
committed
applied jinja is_rocm onto optimizations for backward and forward parameters

kudomcho
authored and
liligwu
committed
Guard supported grad_t for optimized warp_per_row dispatch

avbokovoy
authored and
liligwu
committed
Forward index_t to the optimizer

avbokovoy
authored and
liligwu
committed
Guard f16 llvm intrinsics with ROCm >=7.0

avbokovoy
authored and
liligwu
committed
Fix buffer offset for emb_dim == 160

avbokovoy
authored and
liligwu
committed
Remove sanity check

avbokovoy
authored and
liligwu
committed
address the potential lint issues and revert the change in indices_generator.cpp
liligwu
committed
addresss code style issue
liligwu
committed

Commits on Nov 2, 2025

hack param
Bernard-Liu
committed
support opt code_gen
Bernard-Liu
committed
support subwarp

yadaish
authored and
Bernard-Liu
committed
update subwarp kernel
Bernard-Liu
committed
grad sum kernel unroll improvement

XingerZhu
authored and
Bernard-Liu
committed
fix performance issuse

yadaish
authored and
Bernard-Liu
committed
fix vbe opt not imply
Bernard-Liu
committed

Commits on Nov 3, 2025

fix smybol bug & rm comment
Bernard-Liu
committed