int m = params.m, n = params.n, k = params.k; float h_alpha = params.alpha; void* workspace = nullptr; size_t workspaceSize = 32 * GitHub Tag:"gpu" | Microsoft Community Hub The set of legal kernel and algorithm choices changes with them. And that is the point most people miss. The runtime is not just r... Microsoft Community Hub 6 sites Accelerating MoE's with a Triton Persistent Cache-Aware Grouped ... Aug 18, 2025 —
// 4. Algorithm Heuristic Search cublasLtMatmulPreference_t preference; cublasLtMatmulPreferenceInit(&preference); size_t workspaceSize = 1024 * 1024; // 1MB workspace cublasLtMatmulPreferenceSetAttribute(preference, CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES, &workspaceSize, sizeof(workspaceSize)); cublaslt grouped gemm
cublasLtMatrixLayoutInit(&Adesc, CUDA_R_16F, M, K, lda); cublasLtMatrixLayoutSetAttribute(Adesc, CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT, &batchCount, sizeof(batchCount)); cublasLtMatrixLayoutSetAttribute(Adesc, CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET, &strideA, sizeof(strideA)); int m = params
// d_A, d_B, d_C are device pointers to the start of the batched data cublasLtMatmul(ltHandle, matmulDesc, &alpha, d_A, Adesc, d_B, Bdesc, &beta, d_C, Cdesc, d_C, Cdesc, &heuristicResult.algo, workspace, workspaceSize, stream); Microsoft Community Hub 6 sites Accelerating MoE's with
Enables fusing operations like activation functions (ReLU, GELU) and bias addition directly into the GEMM kernel.