CUBLAS_STATUS_INVALID_VALUE (1) cuda runtime_error ai_generated true

RuntimeError: CUBLAS_STATUS_INVALID_VALUE when calling cublasSgemmStridedBatched with lda=0

ID: cuda/cublas-gemm-invalid-lda

Also available as: JSON · Markdown · 中文
90%Fix Rate
85%Confidence
1Evidence
2023-06-15First Seen

Version Compatibility

VersionStatusIntroducedDeprecatedNotes
CUDA 11.8 active — — —
CUDA 12.1 active — — —
cuBLAS 11.11 active — — —
cuBLAS 12.0 active — — —

Root Cause

cuBLAS GEMM or batched GEMM routines require leading dimensions (lda, ldb, ldc) to be non-zero and at least as large as the matrix width or height; a zero leading dimension causes an invalid value error.

generic

中文

cuBLAS GEMM 或批处理 GEMM 例程要求前导维度（lda、ldb、ldc）非零且至少与矩阵宽度或高度相同；零前导维度会导致无效值错误。

Official Documentation

https://docs.nvidia.com/cuda/cublas/index.html#cublas-t-routine

Workarounds

  1. 95% success Ensure lda, ldb, and ldc are at least 1 and >= number of rows (for column-major) or columns (for row-major). Example fix: if (lda == 0) lda = matrix_width; 
    Ensure lda, ldb, and ldc are at least 1 and >= number of rows (for column-major) or columns (for row-major). Example fix: if (lda == 0) lda = matrix_width;
  2. 85% success Add a debug assertion before the cuBLAS call to check that leading dimensions are positive and valid. Example: assert(lda > 0 && lda >= rows); 
    Add a debug assertion before the cuBLAS call to check that leading dimensions are positive and valid. Example: assert(lda > 0 && lda >= rows);

中文步骤

  1. Ensure lda, ldb, and ldc are at least 1 and >= number of rows (for column-major) or columns (for row-major). Example fix: if (lda == 0) lda = matrix_width;
  2. Add a debug assertion before the cuBLAS call to check that leading dimensions are positive and valid. Example: assert(lda > 0 && lda >= rows);

Dead Ends

Common approaches that don't work:

  1. 95% fail

    Increasing batch size or matrix dimensions does not fix a zero leading dimension; the error persists because the root cause is a programming mistake, not resource limits.

  2. 90% fail

    Switching to a different cuBLAS routine (e.g., from gemmStridedBatched to gemm) only masks the symptom; the caller still passes lda=0, which may crash later or produce wrong results.

  3. 85% fail

    Setting lda to a negative value (e.g., -1) causes a different error (CUBLAS_STATUS_NOT_INITIALIZED or segfault) but does not resolve the zero-layout issue.