blob: 20d6605f4fd0b67e0142321ce68c0bd5a17d8250 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
|
#ifndef _GPU_OPS_KERNELS_H_
#define _GPU_OPS_KERNELS_H_
#include <cuda_runtime_api.h>
#include <cstddef>
#include <cstdint>
#include<stdlib.h>
#include<stdint.h>
namespace gpu_ops {
struct MHAParams {
uint32_t q_batch_stride;
uint32_t k_batch_stride;
uint32_t v_batch_stride;
uint32_t o_batch_stride;
uint32_t q_row_stride;
uint32_t k_row_stride;
uint32_t v_row_stride;
uint32_t o_row_stride;
uint32_t q_head_stride;
uint32_t k_head_stride;
uint32_t v_head_stride;
uint32_t o_head_stride;
uint32_t b;
uint32_t h;
uint32_t h_k;
uint32_t d;
uint32_t d_rounded;
float softmax_scale;
float softcap;
uint32_t seqlen_q;
uint32_t seqlen_k;
uint32_t seqlen_q_rounded;
uint32_t seqlen_k_rounded;
int window_size_left;
int window_size_right;
int is_causal;
int is_bf16;
};
void run_mha_fwd_j(cudaStream_t stream, void **buffers,
const char *opaque,
std::size_t opaque_len);
void run_mha_bwd_j(cudaStream_t stream, void **buffers,
const char *opaque,
std::size_t opaque_len);
}
#endif
|
