masahi commented on code in PR #16413: URL: https://github.com/apache/tvm/pull/16413#discussion_r1466127313
########## src/runtime/contrib/cublas/cublas_utils.h: ########## @@ -80,7 +80,14 @@ struct CuBlasThreadEntry { struct CuBlasLtThreadEntry { CuBlasLtThreadEntry(); ~CuBlasLtThreadEntry(); + cublasLtHandle_t handle{nullptr}; + cublasLtMatmulPreference_t matmul_pref_desc{nullptr}; + void* workspace_ptr{nullptr}; + // 32MB workspace as suggested by NVIDIA + // https://docs.nvidia.com/cuda/cublas/index.html#cublassetworkspace. + static constexpr const size_t workspace_size = 33554432; Review Comment: I'm assuming that 32MB is also good for pre-Hopper since it is bigger than the recommended size, 4MB. @vinx13 -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: commits-unsubscr...@tvm.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org