Add an env variable to disable addmm_cuda_lt kernel (#91436)
addmm_cuda_lt failed for some corner cases, so far we can not reproduce the corner cases in the unit tests, seems that the failures do not only depend on matrices' shape and strides. For now, add an environment variable to allow users disable this kernel for such corner cases.
**See the case one with more error logs:**
RuntimeError: 0CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasLtMatmul with transpose_mat1 1 transpose_mat2 0 m 80 n 1024 k 160 mat1_ld 160 mat2_ld 160 result_ld 80 abcType 14 computeType 68 scaleType 0 result_shape 1024 80 result_stride 80 1 self_shape 80 self_stride 1 mat1_shape 1024 160 mat1_stride 160 1 mat2_shape 160 80 mat2_stride 1 160
Exception raised from gemm_and_bias at fbcode/caffe2/aten/src/ATen/cuda/CUDABlas.cpp:1071 (most recent call first):
**another case with more error logs:**
RuntimeError: 0CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasLtMatmul with transpose_mat1 1 transpose_mat2 0 m 16 n 16384 k 48 mat1_ld 48 mat2_ld 48 result_ld 16 abcType 14 computeType 68 scaleType 0 result_shape 16384 16 result_stride 16 1 self_shape 16 self_stride 1 mat1_shape 16384 48 mat1_stride 48 1 mat2_shape 48 16 mat2_stride 1 48
Exception raised from gemm_and_bias at fbcode/caffe2/aten/src/ATen/cuda/CUDABlas.cpp:1071 (most recent call first):
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91436
Approved by: https://github.com/ngimel
