Fast RMS Norm (#862)

* fast rmsnorm

* no rms gpu

* kernel

* fix shared mem

* looped rms and donation in softmax

* Make the squaring in float32 to avoid underflow

* Fix the default StreamOrDevice for rope and rms_norm in fast

* nits

---------

Co-authored-by: Angelos Katharopoulos <a_katharopoulos@apple.com>
This commit is contained in:
Awni Hannun
2024-03-21 07:20:54 -07:00
committed by GitHub
parent 4650d94d98
commit a54f06b16f
17 changed files with 493 additions and 41 deletions

View File

@@ -44,7 +44,6 @@ target_sources(
${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
${CMAKE_CURRENT_SOURCE_DIR}/quantized.cpp
${CMAKE_CURRENT_SOURCE_DIR}/reduce.cpp
${CMAKE_CURRENT_SOURCE_DIR}/rope.cpp
${CMAKE_CURRENT_SOURCE_DIR}/scan.cpp
${CMAKE_CURRENT_SOURCE_DIR}/select.cpp
${CMAKE_CURRENT_SOURCE_DIR}/softmax.cpp