cusparseLt questions

[Septend-fun]

Hi, experts. I met a question when I use cusparseLt for model inference. In my case, I set weights as sparse matmul. I hope the initialization ,prune, and compression,Init run only once. Now I have a question,if there are different weights and inputs which have same shape, could I run these code only once?

 // matrix descriptor initialization
    CHECK_CUSPARSE( cusparseLtStructuredDescriptorInit(
                                            &handle, &matA, num_A_rows,
                                            num_A_cols, lda, alignment,
                                            type_AB, order,
                                            CUSPARSELT_SPARSITY_50_PERCENT) )

    CHECK_CUSPARSE( cusparseLtDenseDescriptorInit(
                                            &handle, &matB, num_B_rows,
                                            num_B_cols, ldb, alignment,
                                            type_AB, order) )
    CHECK_CUSPARSE( cusparseLtDenseDescriptorInit(
                                            &handle, &matC, num_C_rows,
                                            num_C_cols, ldc, alignment,
                                            type_C, order) )

    // matmul, algorithm selection, and plan initialization
    CHECK_CUSPARSE( cusparseLtMatmulDescriptorInit(
                                            &handle, &matmul, opA, opB,
                                            &matA, &matB, &matC, &matC,
                                            compute_type) )

    CHECK_CUSPARSE( cusparseLtMatmulAlgSelectionInit(
                                            &handle, &alg_sel, &matmul,
                                            CUSPARSELT_MATMUL_ALG_DEFAULT) )

    CHECK_CUSPARSE( cusparseLtMatmulPlanInit(&handle, &plan, &matmul, &alg_sel))

    CHECK_CUSPARSE(cusparseLtMatmulDescSetAttribute(&handle,
                                                    &matmul,
                                                    CUSPARSELT_MATMUL_SPARSE_MAT_POINTER,
                                                    &dA,
                                                    sizeof(dA)));
 if (matmul_search) {
        CHECK_CUSPARSE( cusparseLtMatmulSearch(&handle, &plan, &alpha,
                                               dA_compressed, dB, &beta,
                                               dC, dD, nullptr,
                                               streams, num_streams) )
        // dC accumulates so reset dC for correctness check
        CHECK_CUDA( cudaMemcpy(dC, hC, C_size, cudaMemcpyHostToDevice) )
    } else {
    // otherwise, it is possible to set it directly:
        int alg = 0;
        CHECK_CUSPARSE( cusparseLtMatmulAlgSetAttribute(
                                               &handle, &alg_sel,
                                               CUSPARSELT_MATMUL_ALG_CONFIG_ID,
                                               &alg, sizeof(alg)))
    }

    //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    size_t workspace_size;
    CHECK_CUSPARSE( cusparseLtMatmulPlanInit(&handle, &plan, &matmul, &alg_sel))

    CHECK_CUSPARSE( cusparseLtMatmulGetWorkspace(&handle, &plan,
                                                 &workspace_size))
    void* d_workspace;
    CHECK_CUDA( cudaMalloc((void**) &d_workspace, workspace_size) )

In fact, I tried it.But I met the error:

CUSPARSE API failed at line 330 with error : internal error (7)

The line 330 is

    CHECK_CUSPARSE( cusparseLtMatmul(&handle, &plan, &alpha, dA_compressed, dB,
                                     &beta, dC, dD, d_workspace, streams,
                                     num_streams) )

[j4yan]

@Septend-fun Is it possible to make a reproducer?