backward performance optimization for MI350#4925

Open

liligwu wants to merge 51 commits intopytorch:mainfrom

Commits on Nov 6, 2025

Add gfx950 build support + fp16 fix + index type fix
avbokovoy
committed
Change int64_t to index_t as template parameters in load_raw_per_warp
avbokovoy
committed
Implement llvm fp16 buffer load for gfx950
avbokovoy
committed
Fix c-style half to float cast
avbokovoy
committed
Patch 256 half stores
avbokovoy
committed
cta_per_row workgroup optim

shbiswas834
authored and
avbokovoy
committed
Added mi350 guards

shbiswas834
authored and
avbokovoy
committed
Fix index overflow in row load

shbiswas834
authored and
avbokovoy
committed
cta_per_row workgroup reduce by 4 optim

shbiswas834
authored and
avbokovoy
committed
Fix mixed_D frontend to backend connection
avbokovoy
committed
changed max_segment_length_per_cta to 4096

kudomcho
authored and
avbokovoy
committed
added rocm guards and removed comment

shbiswas834
authored and
avbokovoy
committed
clean debug statements in Hip.cmake

liligwu
authored and
avbokovoy
committed
Merge pull request #121

shbiswas834
authored and
avbokovoy
committed
Guard f16 llvm intrinsics with ROCm >=7.0
avbokovoy
committed
fix the bug in dimention 160 in ROCm optimization

liligwu
authored and
avbokovoy
committed
Cleanup optimized warp_per_raw kernel
avbokovoy
committed
Add 320 embedding dim support for optimized warp_per_row kernel
avbokovoy
committed
changed the max length per warp and cta per row WG size

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

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

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

kudomcho
authored and
avbokovoy
committed
opt for grad_indice_weights kernel

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

kudomcho
authored and
avbokovoy
committed
workgroup tuning and loop unrolled

shbiswas834
authored and
avbokovoy
committed
specialize

Hardcode84
authored and
avbokovoy
committed
explicitly link to tbb

liligwu
authored and
avbokovoy
committed
added warpReduceAllSum with rocm guards

shbiswas834
authored and
avbokovoy
committed
revert unroll and wg tuning

shbiswas834
authored and
avbokovoy
committed
Minor update embedding_forward_split_kernel_template.cu

liligwu
authored and
avbokovoy
committed
add tbb-devel to the install_build_tools ()

liligwu
authored and
avbokovoy
committed
fix lint issues

liligwu
authored and
avbokovoy
committed
solve lint issues

liligwu
authored and
avbokovoy
committed
applied jinja is_rocm onto optimizations for backward and forward parameters

kudomcho
authored and
avbokovoy
committed
Guard supported grad_t for optimized warp_per_row dispatch
avbokovoy
committed
Forward index_t to the optimizer
avbokovoy
committed
Guard f16 llvm intrinsics with ROCm >=7.0
avbokovoy
committed
Fix buffer offset for emb_dim == 160
avbokovoy
committed
Remove sanity check
avbokovoy
committed
address the potential lint issues and revert the change in indices_generator.cpp

liligwu
authored and
avbokovoy
committed
addresss code style issue

liligwu
authored and
avbokovoy
committed
removed guard rocm on mixed_D and refactored mixed_D var assignment

kudomcho
authored and
avbokovoy
committed
Remove general load/store methods
avbokovoy
committed
Move weight type check to compile-time
avbokovoy
committed
Switch to 256B stores for float type
avbokovoy
committed
removed jinj is_rocm on total_L as USE_ROCM is already applied

kudomcho
authored and
avbokovoy
committed
Change mixed_D default value to false
avbokovoy
committed
Make const work_group_size for CUDA
avbokovoy
committed
Add jinja comments to grad_indice_weights kernel
avbokovoy
committed
Remove redundand comment
avbokovoy
committed
Unify cuda and rocm loops
avbokovoy
committed