gpu_blas_lt_gemm_runner

[ScXfjiang]

No description provided.

[i-chaochen]

we also have this TF_USE_CUBLASLT for enable hipblaslt
https://github.com/ROCm/tensorflow-upstream/blob/r2.13-rocm-enhanced-hipblaslt/tensorflow/core/kernels/gpu_utils.cc#L98-L107

and it seems for CUDA-only https://github.com/ROCm/tensorflow-upstream/blob/r2.13-rocm-enhanced-hipblaslt_gpu_blas_lt_runner/tensorflow/core/kernels/matmul_op_impl.h#L411-L413

I'm thinking maybe we should unify these flags as one for non-xla @ScXfjiang @pemeliya

[ScXfjiang]

we also have this TF_USE_CUBLASLT for enable hipblaslt https://github.com/ROCm/tensorflow-upstream/blob/r2.13-rocm-enhanced-hipblaslt/tensorflow/core/kernels/gpu_utils.cc#L98-L107

and it seems for CUDA-only https://github.com/ROCm/tensorflow-upstream/blob/r2.13-rocm-enhanced-hipblaslt_gpu_blas_lt_runner/tensorflow/core/kernels/matmul_op_impl.h#L411-L413

I'm thinking maybe we should unify these flags as one for non-xla @ScXfjiang @pemeliya

I prefer flags in debug_options_flags.cc to env vars in general. The latter is hard to manage. How do you think about it? @pemeliya

[i-chaochen]

debug_options_flags.cc is for XLA specific stuff.

If this is TF and non-xla, these flags are more messy.