[BUG] SVR throwing cuBLAS error during inference

[aerdem4]

Describe the bug
SVR throws error during inference: RuntimeError: cuBLAS error encountered at: file=../src_prims/matrix/grammatrix.cuh line=134: call='raft::linalg::cublasgemm( cublas_handle, CUBLAS_OP_N, CUBLAS_OP_T, n1, n2, n_cols, &alpha, x1, ld1, x2, ld2, &beta, out, ld_out, stream)', Reason=7:CUBLAS_STATUS_INVALID_VALUE
Obtained 64 stack frames

Steps/Code to reproduce bug
You can reproduce the bug using this Kaggle kernel: https://www.kaggle.com/aerdem4/optiver-rapids-svr-wip

[dantegd]

@tfeher might be the best person to take a look at this

[tfeher]

Here is a simplified reproducer:

import cupy as cp
import cuml

shape = (10,1)
X = cp.random.uniform(size=shape, dtype=cp.float64)
y = cp.ones(shape)

model = cuml.svm.SVR(kernel="linear", degree=3, C=10, verbose=cuml.common.logger.level_debug)
model.fit(X, y)

print("Fit status:", model._fit_status_)
print("n_support:", model.n_support_)

model.predict(X)

Output:

SMO solver finished after 1 outer iterations, total inner iterations, and diff -0.200000
Fit status: 0
n_support: None

...
RuntimeError: cuBLAS error encountered at: file=../src_prims/matrix/grammatrix.cuh line=165: call='raft::linalg::cublasgemm(cublas_handle, CUBLAS_OP_N, CUBLAS_OP_T, n1, n2, n_cols, &alpha, x1, ld1, x2, ld2, &beta, out, ld_out, stream)', Reason=7:CUBLAS_STATUS_INVALID_VALUE

Note that the SVM is seems to be fitted correctly (fit_status == 0), but the parameters of the SVM is not stored (n_support=None). The example here can be fitted with a bias only. It seems that the classier does not handle the corner case where we have zero support vectors.

I will fix this.