#pragma once #include #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800 // No support for async #else torch::Tensor gptq_marlin_gemm(torch::Tensor &a, torch::Tensor &b_q_weight, torch::Tensor &b_scales, torch::Tensor &g_idx, torch::Tensor &perm, torch::Tensor &workspace, int64_t num_bits, int64_t size_m, int64_t size_n, int64_t size_k, bool is_k_full); torch::Tensor gptq_marlin_repack(torch::Tensor &b_q_weight, torch::Tensor &perm, int64_t size_k, int64_t size_n, int64_t num_bits); torch::Tensor marlin_gemm(torch::Tensor &a, torch::Tensor &b_q_weight, torch::Tensor &b_scales, torch::Tensor &workspace, int64_t size_m, int64_t size_n, int64_t size_k); #endif