|
#pragma once |
|
|
|
#include <vector> |
|
|
|
#include <torch/torch.h> |
|
|
|
std::vector<torch::Tensor> |
|
symmetric_quantize_last_axis_of_tensor(torch::Tensor const &weight, |
|
at::ScalarType quant_type, |
|
bool return_unprocessed_quantized_tensor); |
|
|
|
torch::Tensor preprocess_weights_cuda(torch::Tensor const &ori_weight, |
|
bool is_int4); |
|
|
|
torch::Tensor w8_a16_gemm_forward_cuda(torch::Tensor const &input, |
|
torch::Tensor const&weight, |
|
torch::Tensor const &scale); |
|
|
|
torch::Tensor w8_a16_gemm_forward_cuda_(torch::Tensor const &input, |
|
torch::Tensor const &weight, |
|
torch::Tensor const &scale, |
|
torch::Tensor &output, |
|
const int64_t m, |
|
const int64_t n, |
|
const int64_t k); |
|
|
