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

Fix the gfx950 numerical errors (#2911)

Browse Source 
* Update grouped_gemm example and pipeline

* find the root cause error in did not enable the transpose in gfx950 correctly

* Fix v3 pipeline, row and col major

* Disable f8 datatype tests, it fails on gfx950

* fix the abd test by clear the runtime argument unsupported

---------

Co-authored-by: AviralGoelAMD <aviral.goel@amd.com>
Co-authored-by: Mateusz Ozga <mateusz.ozga@amd.com>
This commit is contained in:
Thomas Ning
2025-09-23 22:54:52 -07:00
committed by GitHub
parent f161b5b738
commit b159841a06
9 changed files with 63 additions and 123 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
example/ck_tile/17_grouped_gemm/quant_run_grouped_gemm_example.inc
View File

@@ -183,12 +183,24 @@ int run_grouped_gemm_example_with_layouts(int argc,
if(!valid_input_data(group_count, Ms, Ns, Ks, stride_As, stride_Bs, stride_Cs))
{
std::cout << "Please check the input data. Default values will be used." << std::endl;
// Clear existing (invalid) data before adding defaults
Ms.clear();
Ns.clear();
Ks.clear();
stride_As.clear();
stride_Bs.clear();
stride_Cs.clear();
stride_AQs.clear();
stride_BQs.clear();
for(int i = 0; i < group_count; i++)
{
Ms.push_back(256 + 256 * i);
Ns.push_back(256 + 512 * i);
Ks.push_back(512 + 128 * i);
// Let get_default_stride calculate based on layout
stride_As.push_back(0);
stride_Bs.push_back(0);
stride_Cs.push_back(0);

16
example/ck_tile/17_grouped_gemm/run_grouped_gemm_example.inc
View File

@@ -172,15 +172,25 @@ int run_grouped_gemm_example_with_layouts(int argc,
std::cout << "Default values: Ms (256, 512, 768, 1024..), Ns (256, 768, 1280..), Ks (512, "
"896, 1280..)"
<< std::endl;
// Clear existing (invalid) data before adding defaults
Ms.clear();
Ns.clear();
Ks.clear();
stride_As.clear();
stride_Bs.clear();
stride_Cs.clear();
for(int i = 0; i < group_count; i++)
{
Ms.push_back(256 + 256 * i);
Ns.push_back(256 + 512 * i);
Ks.push_back(512 + 384 * i);
stride_As.push_back(Ks[i]);
stride_Bs.push_back(Ks[i]);
stride_Cs.push_back(Ns[i]);
// Set default strides based on layout later using get_default_stride
stride_As.push_back(0);
stride_Bs.push_back(0);
stride_Cs.push_back(0);
}
}