[Triton][gfx1250] gemm a16w16 cleanup - #3646

Merged

vgokhale merged 41 commits into

mainfrom

alizaidy/gemm_a16w16_cleanup

Jun 10, 2026

Commits on Jun 4, 2026

add torch compile guard we need to split into two callers internally cuz you cant use torch compile guard when the function doesnt have a fixed return signature
ahmed-bsod
committed
37a5948
tdm tensor descriptor needs last dim to be greater or equal to 16bytes
vgokhale
committed
b810d8e

Commits on Jun 5, 2026

Gluon kernel for a16w16 gemm

omuhamma
authored and
AlexAUT
committed
e30a521
Moving my config file into the triton location

omuhamma
authored and
AlexAUT
committed
09bc6bc
Update op_tests/op_benchmarks/triton/bench_gemm_a16w16.py

authored and
AlexAUT
committed
430df4b
Code style (black/ruff)

omuhamma
authored and
AlexAUT
committed
dad4b79
Make gemm kernel the only GPU program
AlexAUT
committed
863db86
WIP pipeline LDS
AlexAUT
committed
eeac730
WIP advance
AlexAUT
committed
b3ad0c6
Revert "WIP advance"
AlexAUT
committed
9af755c
Use loop load latency
AlexAUT
committed
eda04d7
Add testing for lds_pipeline kernel and fix bug in bias initialization

azaidy
authored and
AlexAUT
committed
47b8d31
Add TDM store

azaidy
authored and
AlexAUT
committed
7e0711f
Use TDM advance and fix bug in basic

azaidy
authored and
AlexAUT
committed
299ba34
Peel first loop iteration

Boss2002n
authored and
AlexAUT
committed
359faa6
Uncomment sched barrier since we no longer spill
AlexAUT
committed
83eead9
Adjust to new TDM
AlexAUT
committed
8ee055a
Adjusts to new TDM API
AlexAUT
committed
81a68cf

Commits on Jun 6, 2026

add configs and fix lds pipeline

lburzawa
authored and
Boss2002n
committed
a8b59d4

Commits on Jun 7, 2026

Commits on Jun 8, 2026

Tuned configs
vgokhale
committed
5b38439
More tuning
vgokhale
committed
f818e78

Commits on Jun 9, 2026

more tuning
vgokhale
committed
0e65cd1
Merge remote-tracking branch 'origin/main' into vinayak/gemm_bf16
vgokhale
committed
d070391
Revert non-bf16-gemm edits (MoE routing TDM guard, fused_rmsnorm_add)
azaidy
committed
1e89fda
Remove unnecessary kernel variants
azaidy
committed
451fec4
Cleanup unused prefetch functionality
azaidy
committed
61795b1
Cleanup unused init.py files
azaidy
committed
ab92b37
Cleanup wrapper and bench
azaidy
committed
e1f58e9
Rearrange helpers
azaidy
committed
5514af5
Move kernel to correct folder
azaidy
committed
e024d75

Commits on Jun 10, 2026

Move to standard layout notation
azaidy
committed
9d6e684
Remove padding and add temporary assert to exit on non padded shapes
azaidy
committed
c9e83ad
Change kernel names
azaidy
committed
0968a33
Minimize test and bench changes
azaidy
committed
693baa9
Skip triton on gfx1250
azaidy
committed
caddb3e
Add fake
azaidy
committed
095b8e7
Clean lint
azaidy
committed
0f791f3