ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 05:01:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Reapply "[CK_TILE] support hdim=192/128 pair for deepseekv3 (#1961)" … (#1971)

Browse Source 
* Reapply "[CK_TILE] support hdim=192/128 pair for deepseekv3 (#1961)" (#1969)

This reverts commit 8cbcd3e0d0.

* fix codegen problem

* Update config.hpp

---------

Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
This commit is contained in:
carlushuang
2025-03-13 11:41:39 +08:00
committed by GitHub
parent d4a6d69643
commit 3e81279d26
10 changed files with 37 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp
View File

@@ -112,6 +112,13 @@ struct BlockFmhaPipelineQRKSVSAsync
else
return 2;
}
else if constexpr(kQKHeaddim <= 192)
{
if constexpr(kPadSeqLenK && BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 1;
else
return 2;
}
else if constexpr(kQKHeaddim <= 256)
{
return 1;

2
include/ck_tile/ops/fmha/pipeline/tile_fmha_shape.hpp
View File

@@ -13,6 +13,8 @@ static CK_TILE_HOST_DEVICE constexpr index_t ceil_to_qualified_tile_length(index
return 128;
if(len == 160)
return 256;
if(len == 192)
return 192;
// only length of 96, 160 and power-of-two is supported
if(!(len & (len - 1)))