deeplearning4j/deeplearning4j · error · BlasException

cusolverDnSgeqrf_bufferSize failed

Error message

cusolverDnSgeqrf_bufferSize failed

What it means

sgeqrf first calls cusolverDnSgeqrf_bufferSize to query the workspace size needed for the float QR factorization. When the returned status is not CUSOLVER_STATUS_SUCCESS, JcublasLapack throws this BlasException. It means the cuSOLVER query itself failed before any factorization ran - typically bad dimensions (M<=0/N<=0), a null/invalid device pointer, or an uninitialized/failed cuSOLVER handle/stream.

Solutions

  1. Validate input: ensure the INDArray is non-empty and rows/columns are > 0 before calling sgeqrf.
  2. Check GPU health and memory (nvidia-smi); an OOM or driver failure corrupts subsequent calls - restart the JVM/app.
  3. Verify nd4j-cuda backend version matches the installed CUDA toolkit (e.g. nd4j-cuda-11.6 with CUDA 11.6).
  4. Confirm cusolverDnSetStream succeeded just before the call (see 'solverSetStream failed' for the setup path).
  5. Fall back to CPU backend for this computation to isolate whether the failure is CUDA-specific.

Example fix

// before
INDArray a = Nd4j.create(new float[0]); // empty
INDArrayInfo info = Nd4j.getBlasWrapper().lapack().geqrf(a, tau);

// after
if (a.isEmpty() || a.rows() < 1 || a.columns() < 1) {
    throw new IllegalArgumentException("geqrf requires a non-empty matrix, got " + a.shape());
}
INDArrayInfo info = Nd4j.getBlasWrapper().lapack().geqrf(a, tau);
Defensive patterns

Strategy: validation

Validate before calling

// Java/ND4j
if (a == null || a.isEmpty() || a.rows() < 1 || a.columns() < 1) {
    throw new IllegalArgumentException("sgeqrf requires a non-empty matrix, got shape " + (a == null ? "null" : java.util.Arrays.toString(a.shape())));
}

Try / catch

// Java
try {
    lapack.geqrf(a, tau);
} catch (BlasException e) {
    // fall back to CPU backend or abort with a clear message
    throw new IllegalStateException("QR failed on GPU: " + e.getMessage(), e);
}

Prevention

When it happens

Trigger: Calling Nd4j's sgeqrf on a CUDA backend when the input INDArray is empty, has non-positive rows/columns, is not in a layout cuSOLVER accepts, or when the preceding cusolverDnSetStream/handle setup failed; also when the NativeBlas binding could not map the device pointer (uninitialized buffer).

Common situations: Passing zero-length or broadcast/empty arrays to QR decomposition on GPUs; using an NDArray with an unsupported stride/offset; mixing CUDA backend jars with mismatched cuSOLVER versions (CUDA toolkit upgrade without updating nd4j-cuda-*-platform); GPU out of memory corrupting pointer allocation.

Related errors


AI-assisted analysis of deeplearning4j/deeplearning4j@4c22ac5fe4 (2026-09-07). Data as JSON: /api/errors/4734150188a622c1. Report an issue: GitHub.

Appendix: source

Thrown at nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/blas/JcublasLapack.java:246

            if (result != 0)
                throw new IllegalStateException("solverSetStream failed");

            // transfer the INDArray into GPU memory
            CublasPointer xAPointer = new CublasPointer(a, ctx);
            CublasPointer xTauPointer = new CublasPointer(tau, ctx);

            // this output - indicates how much memory we'll need for the real operation
            val worksizeBuffer = (BaseCudaDataBuffer) Nd4j.getDataBufferFactory().createInt(1);
            worksizeBuffer.lazyAllocateHostPointer();

            int stat = cusolverDnSgeqrf_bufferSize(solverDn, M, N,
                    (FloatPointer) xAPointer.getDevicePointer(), M,
                    (IntPointer) worksizeBuffer.addressPointer() // we intentionally use host pointer here
            );


            if (stat != CUSOLVER_STATUS_SUCCESS) {
                throw new BlasException("cusolverDnSgeqrf_bufferSize failed", stat);
            }
            int worksize = worksizeBuffer.getInt(0);
            // Now allocate memory for the workspace, the permutation matrix and a return code
            Pointer workspace = new Workspace(worksize * Nd4j.sizeOfDataType());

            // Do the actual QR decomp
            stat = cusolverDnSgeqrf(solverDn, M, N,
                    (FloatPointer) xAPointer.getDevicePointer(), M,
                    (FloatPointer) xTauPointer.getDevicePointer(),
                    new CudaPointer(workspace).asFloatPointer(),
                    worksize,
                    new CudaPointer(allocator.getPointer(INFO, ctx)).asIntPointer()
            );
            if (stat != CUSOLVER_STATUS_SUCCESS) {
                throw new BlasException("cusolverDnSgeqrf failed", stat);
            }

            allocator.registerAction(ctx, a);

View on GitHub (pinned to 4c22ac5fe4)