ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-04 21:51:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] layernorm have more accurate residual (#1623)

Browse Source 
* more accurate residual

* modify comment

* Fix literal case in README.md

---------

Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
This commit is contained in:
carlushuang
2024-11-02 05:30:16 +00:00
committed by GitHub
parent 03c6448ba3
commit cb6c5d39dc
6 changed files with 97 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
example/ck_tile/02_layernorm2d/README.md
View File

@@ -69,7 +69,7 @@ args:
```
## limitations
Note that `fquant=2`, `fadd=2`, `prec_sx/prec_sy` other than `fp32` are not by default generated. though our kernel template suppor this. (TBD: add some flag in generate.py) to generate those instance on demand. Beside, N>8192 case will by default using two-pass pipeline, and `-fquant=1/2` are not supported yet.
Note that `fquant=2`, `fadd=2`, `prec_sx/prec_sy` other than `fp32` are not by default generated. Though our kernel template suppor this. (TBD: add some flag in generate.py) to generate those instance on demand. Beside, `N>8192` case will by default using two-pass pipeline, and `-fquant=1/2` are not supported yet. If need suport `N>8192` and `fused+residual+store`, you can use this example together with `12_smoothquant`, to construct layernorm+residual, and smoothquant, 2 kernels for this purpose.
```
# some case
@@ -82,4 +82,4 @@ Note that `fquant=2`, `fadd=2`, `prec_sx/prec_sy` other than `fp32` are not by d
# standard fp16 layernorm 2d, m=10. n=1024, fused-smooth-quant+fused-add-store, output in int8
./build/bin/tile_example_layernorm2d_fwd -m=10 -n=1024 -prec_o=int8 -fquant=1 -fadd=1
```
```

5
example/ck_tile/02_layernorm2d/generate.py
View File

@@ -202,8 +202,9 @@ float layernorm2d_fwd_(const S& s, A a)
using Default2DEpilogueProblem = ck_tile::Default2DEpilogueProblem<ComputeDataType, YDataType, false, Traits_::kPadN, false>;
using Default2DEpilogue = ck_tile::Default2DEpilogue<Default2DEpilogueProblem>;
using DynamicQuantEpilogueProblem = ck_tile::DynamicQuantEpilogueProblem<ComputeDataType, YScaleDataType, YDataType, typename Traits_::Shape,
ck_tile::DynamicQuantEpilogueTraits<false, Traits_::kPadN, false, true/*max3*/>>;
static constexpr bool UseSmoothInputScale = Traits_::kFusedQuant == 1;
using DynamicQuantEpilogueProblem = ck_tile::DynamicQuantEpilogueProblem<ComputeDataType, XScaleDataType, YScaleDataType, YDataType, typename Traits_::Shape,
ck_tile::DynamicQuantEpilogueTraits<false, Traits_::kPadN, UseSmoothInputScale, false, true/*max3*/>>;
using DynamicQuantEpilogue = ck_tile::DynamicQuantEpilogue<DynamicQuantEpilogueProblem>;