dequantize_awq.h 1.61 KB
Newer Older
1
2
3
4
5
6
7
#ifndef __INFINIOP_DEQUANTIZE_AWQ_API_H__
#define __INFINIOP_DEQUANTIZE_AWQ_API_H__

#include "../operator_descriptor.h"

typedef struct InfiniopDescriptor *infiniopDequantizeAWQDescriptor_t;

8
__INFINI_C __export infiniStatus_t infiniopCreateDequantizeAWQDescriptor(infiniopHandle_t handle,
9
10
11
12
13
14
                                                                  infiniopDequantizeAWQDescriptor_t *desc_ptr,
                                                                  infiniopTensorDescriptor_t out_desc,
                                                                  infiniopTensorDescriptor_t qweight_desc,
                                                                  infiniopTensorDescriptor_t scales_desc,
                                                                  infiniopTensorDescriptor_t zeros_desc);

15
__INFINI_C __export infiniStatus_t infiniopGetDequantizeAWQWorkspaceSize(infiniopDequantizeAWQDescriptor_t desc, size_t *size);
16

17
__INFINI_C __export infiniStatus_t infiniopDequantizeAWQ(infiniopDequantizeAWQDescriptor_t desc,
18
19
20
21
22
23
24
25
                                                  void *workspace,
                                                  size_t workspace_size,
                                                  void *out,
                                                  const void *qweight,
                                                  const void *scales,
                                                  const void *zeros,
                                                  void *stream);

26
__INFINI_C __export infiniStatus_t infiniopDestroyDequantizeAWQDescriptor(infiniopDequantizeAWQDescriptor_t desc);
27
28

#endif