leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 25 Packages Projects Releases Wiki Activity

v4.2 release. (#2587)

Browse Source 
* Fix default cluster callback values to 1 to avoid profiler failure when these values are not set in command line.

* v4.2 release.
This commit is contained in:
Junkai-Wu
2025-08-22 18:11:24 -04:00
committed by GitHub
parent 11cad1f67b
commit a49a78ffef
351 changed files with 28182 additions and 2032 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/profiler/src/block_scaled_gemm_operation_profiler.cu
+11 -8
View File
@@ -176,7 +176,7 @@ Status BlockScaledGemmOperationProfiler::GemmProblem::parse(
if (!arg_as_int(this->cluster_m, "cluster_m", problem_space, problem)) {
// default value
this->cluster_m = 1;
this->cluster_m = std::string(operation_desc.name).find("_2sm") != std::string::npos ? 2 : 1;
}
if (!arg_as_int(this->cluster_n, "cluster_n", problem_space, problem)) {
@@ -191,17 +191,17 @@ Status BlockScaledGemmOperationProfiler::GemmProblem::parse(
if (!arg_as_int(this->cluster_m_fallback, "cluster_m_fallback", problem_space, problem)) {
// default value
this->cluster_m_fallback = 0;
this->cluster_m_fallback = (this->cluster_m % 2 == 0) ? 2 : 1;
}
if (!arg_as_int(this->cluster_n_fallback, "cluster_n_fallback", problem_space, problem)) {
// default value
this->cluster_n_fallback = 0;
this->cluster_n_fallback = 1;
}
if (!arg_as_int(this->cluster_k_fallback, "cluster_k_fallback", problem_space, problem)) {
// default value
this->cluster_k_fallback = 0;
this->cluster_k_fallback = 1;
}
if (!arg_as_SplitKModeID(this->split_k_mode, "split_k_mode", problem_space, problem)) {
@@ -540,6 +540,8 @@ Status BlockScaledGemmOperationProfiler::initialize_configuration(
gemm_workspace_.arguments.use_pdl = problem_.use_pdl;
cudaStreamCreateWithFlags(&gemm_workspace_.stream, cudaStreamNonBlocking);
// initialize reduction operation for parallel splitKMode
if (problem_.split_k_mode == library::SplitKMode::kParallel) {
if (!initialize_reduction_configuration_(operation, problem)) {
@@ -643,6 +645,7 @@ void BlockScaledGemmOperationProfiler::initialize_result_(
result.bytes = problem_.bytes(operation_desc);
result.flops = problem_.flops(operation_desc);
result.runtime = 0;
result.runtime_vector.resize(options.device.devices.size(), 0);
}
@@ -1578,7 +1581,7 @@ Status BlockScaledGemmOperationProfiler::profile_cutlass_(
}
}
auto func = [&](cudaStream_t, int iteration) {
auto func = [&](cudaStream_t stream, int iteration) {
// Iterate over copies of the problem in memory
int problem_idx = (iteration % gemm_workspace_.problem_count) * problem_.batch_count;
@@ -1599,7 +1602,7 @@ Status BlockScaledGemmOperationProfiler::profile_cutlass_(
arguments,
host_workspace,
device_workspace,
nullptr);
stream);
if (status != Status::kSuccess) {
return status;
@@ -1611,7 +1614,7 @@ Status BlockScaledGemmOperationProfiler::profile_cutlass_(
&gemm_workspace_.reduction_arguments,
gemm_workspace_.reduction_host_workspace.data(),
nullptr,
nullptr);
stream);
if (status != Status::kSuccess) {
return status;
@@ -1621,7 +1624,7 @@ Status BlockScaledGemmOperationProfiler::profile_cutlass_(
return status;
};
return profile_kernel_(result, options, func);
return profile_kernel_(result, options, func, gemm_workspace_.stream);
}
/////////////////////////////////////////////////////////////////////////////////////////////////