mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 18:42:06 +00:00

Files

ruanjm ce1d20c2c6 [CK_TILE] Improve RMS/Layer Normalization 2 Pass Pipeline Performance (#1861 )

* 50ms -> 28ms

* Fix bug in non fuse_add_store cases

* Fine tuned setting for 2 pass pipeline

* adjust workload

* remove unnecessary change

* add layernorm

* Adding output quant and unquant results at the same time.

* fix test

* fix format

* tune for cases 128x640 and 128x1024

* bug ifx

[ROCm/composable_kernel commit: d49abdaa87]

2025-03-25 20:09:45 +08:00

script

[CK_TILE] Improve RMS/Layer Normalization 2 Pass Pipeline Performance (#1861 )

2025-03-25 20:09:45 +08:00

CMakeLists.txt

Implement fp8 quant for layernorm and rmsnorm (#1814 )

2025-01-24 16:40:43 +08:00

example_rmsnorm2d_fwd.cpp

[CK_TILE] Improve RMS/Layer Normalization 2 Pass Pipeline Performance (#1861 )

2025-03-25 20:09:45 +08:00

generate.py

[CK_TILE] Improve RMS/Layer Normalization 2 Pass Pipeline Performance (#1861 )

2025-03-25 20:09:45 +08:00

README.md

[Ck tile] support rmsnorm and related fusion (#1605 )

2024-10-30 15:22:56 +08:00

rmsnorm2d_fwd.cpp

[CK_TILE] Improve RMS/Layer Normalization 2 Pass Pipeline Performance (#1861 )

2025-03-25 20:09:45 +08:00

rmsnorm2d_fwd.hpp

[CK_TILE] Improve RMS/Layer Normalization 2 Pass Pipeline Performance (#1861 )

2025-03-25 20:09:45 +08:00

README.md

Rmsnorm2D forward

This folder contains example for Rmsnorm2D forward using ck_tile tile-programming implementation.

build

# in the root of ck_tile
mkdir build && cd build
sh ../script/cmake-ck-dev.sh  ../ <arch>  # you can replace this <arch> to gfx90a, gfx942...
make tile_rmsnorm2d_fwd -j

This will result in an executable build/bin/tile_rmsnorm2d_fwd

cmdline

args:
          -m    m dimension (default:3328)
          -n    m dimension (default:4096)
          -e    epsilon (default:1e-5)
          -v    cpu validation or not (default:1)
       -prec    precision (default:fp16)