Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
compat.cuh 1.59 KB
Newer Older
Peng Zhang's avatar
[2/n]decouple quantization implementation from vLLM dependency (#8112)  
Peng Zhang committed
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
59
60
61
62 
/*
Copied from https://github.com/turboderp/exllamav2
*/

#ifndef _compat_cuh
#define _compat_cuh

namespace sglang {
namespace gptq {
// atomicAdd for half types, to support CC < 7.x

__device__ __forceinline__ void atomicAdd_half(half* address, half val) {
  unsigned int* address_as_ui = (unsigned int*)((char*)address - ((size_t)address & 2));
  unsigned int old = *address_as_ui;
  unsigned int assumed;

  do {
    assumed = old;
    __half_raw hsum;
    hsum.x = (size_t)address & 2 ? (old >> 16) : (old & 0xffff);
    half tmpres = __hadd(hsum, val);
    hsum = __half_raw(tmpres);
    old = (size_t)address & 2 ? (old & 0xffff) | (hsum.x << 16) : (old & 0xffff0000) | hsum.x;
    old = atomicCAS(address_as_ui, assumed, old);
  } while (assumed != old);
}

// atomicAdd for half2 types

__device__ __forceinline__ void atomicAdd_half2(half2* address, half2 val) {
  unsigned int* address_as_ui = (unsigned int*)address;
  unsigned int old = *address_as_ui;
  unsigned int assumed;
  do {
    assumed = old;
    half2 old_val = *((half2*)&old);
    half2 new_val = __hadd2(old_val, val);
    old = atomicCAS(address_as_ui, assumed, *((unsigned int*)&new_val));
  } while (assumed != old);
}

//

#if defined(__CUDA_ARCH__) || defined(USE_ROCM)
#if __CUDA_ARCH__ < 700 || defined(USE_ROCM)

__device__ __forceinline__ void atomicAdd(half* address, half val) {
  atomicAdd_half(address, val);
}

#if __CUDA_ARCH__ < 600 || defined(USE_ROCM)
__device__ __forceinline__ void atomicAdd(half2* address, half2 val) {
  atomicAdd_half2(address, val);
}
#endif

#endif
#endif

}  // namespace gptq
}  // namespace sglang
#endif