ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Dev/a8w4 and a8w8splitk (#3447)

Browse Source 
* Ck moe bs splitk pr (#3440)

* splitk kick-off. Compilation fail

* splitk hack pass

* fix scale offset calc.

* clang-format for a8w8_moe_blk_gemm1 splitk change

* fix testcase error

---------

Co-authored-by: oscar <huaiguxu@amd.com>
Co-authored-by: huaiguxu <145733371+huaiguxu@users.noreply.github.com>

* Zan/moe a8w4 (#3441)

* update

* update

* update ck moe a8w4

* update

* update

* update

* compile pass

* update

* update

* python3 op_tests/test_moe_2stage.py -t 16 -e 1 -k 1 -dim 256,256 ready

* support new a8w4 kernel

* update

* update ck_tile

* re format

* update

* update

* fix conflict

* fix build

* update ck_tile moe

* fix clang format

* fix the problem

* fix accruacy issue

* fix

---------

Co-authored-by: oscar <huaiguxu@amd.com>
Co-authored-by: huaiguxu <145733371+huaiguxu@users.noreply.github.com>
Co-authored-by: Zzz9990 <zanzhang@amd.com>
Co-authored-by: felix <felix.li@amd.com>
This commit is contained in:
yadaish
2025-12-19 09:26:52 +08:00
committed by GitHub
parent ba897f8435
commit c0ee71d735
13 changed files with 2911 additions and 139 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_moe_blockscale_b_preshuffle_v1.hpp
View File

@@ -360,6 +360,7 @@ struct BlockwiseGemmXdlops_pipeline_moe_blockscale_bpreshuffle_v1<
});
});
__builtin_amdgcn_sched_barrier(0);
// Local prefill A1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
@@ -550,6 +551,7 @@ struct BlockwiseGemmXdlops_pipeline_moe_blockscale_bpreshuffle_v1<
});
});
__builtin_amdgcn_sched_barrier(0);
a_scale_thread_copy.Run(a_scale_grid_desc,
a_scale_grid_buf,
a_scale_thread_desc,
@@ -677,6 +679,7 @@ struct BlockwiseGemmXdlops_pipeline_moe_blockscale_bpreshuffle_v1<
});
block_sync_lds();
__builtin_amdgcn_sched_barrier(0);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {