text-generation-inference/server/exllama_kernels/exllama_kernels/cuda_func/q4_matmul.cuh
OlivierDehaene 0c207f71ed feat: experimental support for cuda graphs (#1428)
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>
2024-04-24 13:15:45 +03:00

38 lines
675 B
Plaintext

// Adapted from turboderp exllama: https://github.com/turboderp/exllama
#ifndef _q4_matmul_cuh
#define _q4_matmul_cuh
#include <cuda_runtime.h>
#include <cuda_fp16.h>
#include <cstdint>
#include <cstdio>
#include <ATen/cuda/CUDAContext.h>
#include "q4_matrix.cuh"
#include "../tuning.h"
void q4_matmul_cuda
(
ExLlamaTuning* tuningParams,
const half* x,
const int x_height,
const Q4Matrix* w,
half* out,
bool no_zero,
cudaStream_t alt_stream
);
void q4_matmul_recons_cuda
(
ExLlamaTuning* tuningParams,
const half* x,
const int x_height,
Q4Matrix* w,
half* out,
bool no_zero,
const cublasHandle_t handle
);
#endif