Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
common.cu 4.51 KB
Newer Older
Przemek Tredak's avatar
Inital code drop  
Przemek Tredak committed
1 
/*************************************************************************
Przemyslaw Tredak's avatar
Update copyright year (#48)  
Przemyslaw Tredak committed
2 
 * Copyright (c) 2022-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
Przemek Tredak's avatar
Inital code drop  
Przemek Tredak committed
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 
 *
 * See LICENSE for license information.
 ************************************************************************/

#include "common.h"
#include "transformer_engine/transformer_engine.h"


transformer_engine::DType getTransformerEngineFP8Type(bool e4m3_if_hybrid,
                                                      const std::string &fp8_recipe) {
    // if e4m3 or hybrid + forward
    if ( (fp8_recipe == "E4M3") || ( (fp8_recipe == "HYBRID") && e4m3_if_hybrid ) ) {
        return transformer_engine::DType::kFloat8E4M3;
    }
    return transformer_engine::DType::kFloat8E5M2;
}

transformer_engine::TensorWrapper makeTransformerEngineTensor(
    void* data_ptr,
    const NVTEShape& shape,
    const transformer_engine::DType type) {
  return transformer_engine::TensorWrapper(data_ptr, shape, type);
}


transformer_engine::TensorWrapper makeTransformerEngineTensor(
    void* data_ptr,
    const std::vector<size_t>& shape,
    const transformer_engine::DType type) {
  return transformer_engine::TensorWrapper(data_ptr, shape, type);
}


transformer_engine::TensorWrapper makeTransformerEngineTensor(at::Tensor tensor) {
    transformer_engine::DType dtype = GetTransformerEngineDType(tensor.scalar_type());
    std::vector<size_t> shape;

    for (auto s : tensor.sizes()) {
        shape.push_back(s);
    }
    return makeTransformerEngineTensor(tensor.data_ptr(), shape, dtype);
}
Przemyslaw Tredak's avatar
Move the amax/scale/scale_inv into the TE Tensor struct. (#33)  
Przemyslaw Tredak committed
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81 
transformer_engine::TensorWrapper makeTransformerEngineTensor(
    void* data_ptr,
    const std::vector<size_t>& shape,
    const transformer_engine::DType type,
    void* amax_ptr,
    void* scale_ptr,
    void* scale_inv_ptr) {
  return transformer_engine::TensorWrapper(data_ptr, shape, type,
                                           reinterpret_cast<float*>(amax_ptr),
                                           reinterpret_cast<float*>(scale_ptr),
                                           reinterpret_cast<float*>(scale_inv_ptr));
}


transformer_engine::TensorWrapper makeTransformerEngineTensor(at::Tensor tensor,
                                                              at::Tensor amax,
                                                              const at::Tensor scale,
                                                              at::Tensor scale_inv) {
    transformer_engine::DType dtype = GetTransformerEngineDType(tensor.scalar_type());
    std::vector<size_t> shape;

    for (auto s : tensor.sizes()) {
        shape.push_back(s);
    }
    NVTE_CHECK(amax.scalar_type() == at::kFloat);
    NVTE_CHECK(scale.scalar_type() == at::kFloat);
    NVTE_CHECK(scale_inv.scalar_type() == at::kFloat);

    return makeTransformerEngineTensor(tensor.data_ptr(), shape, dtype,
                                       amax.data_ptr(),
                                       scale.data_ptr(),
                                       scale_inv.data_ptr());
}
Przemek Tredak's avatar
Inital code drop  
Przemek Tredak committed
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126 
size_t product(const std::vector<size_t> &shape) {
    size_t ret = 1;
    for (auto s : shape) {
        ret *= s;
    }
    return ret;
}


at::Tensor allocateSpace(const NVTEShape &shape,
                         const transformer_engine::DType type,
                         bool init_to_zeros) {
    auto size = shape.ndim;
    if (size == 2 && init_to_zeros) {
        return at::zeros({static_cast<int64_t>(shape.data[0]),
                          static_cast<int64_t>(shape.data[1])},
                          at::CUDA(GetATenDType(type)));
    } else if (size == 2) {
        return at::empty({static_cast<int64_t>(shape.data[0]),
                          static_cast<int64_t>(shape.data[1])},
                          at::CUDA(GetATenDType(type)));
    } else if (size == 1 && init_to_zeros) {
        return at::zeros({static_cast<int64_t>(shape.data[0])}, at::CUDA(GetATenDType(type)));
    } else if (size == 1) {
        return at::empty({static_cast<int64_t>(shape.data[0])}, at::CUDA(GetATenDType(type)));
    }
    NVTE_CHECK(false, "Should never reach here! func: allocateSpace");
}


at::Tensor allocateTorchTensor(int M,
                               int N,
                               transformer_engine::DType dtype
) {
    return at::empty({static_cast<int64_t>(M), static_cast<int64_t>(N)},
                     at::CUDA(GetATenDType(dtype)));
}


at::Tensor allocateTorchTensor(int M,
                               transformer_engine::DType dtype
) {
    return at::empty({static_cast<int64_t>(M)},
                     at::CUDA(GetATenDType(dtype)));
}