LlamaDenseWeight.h 1.99 KB
Newer Older
Li Zhang's avatar
Li Zhang committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
/*
 * Copyright (c) OpenMMLab. All rights reserved.
 * Copyright (c) 2019-2023, NVIDIA CORPORATION.  All rights reserved.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

AllentDan's avatar
AllentDan committed
18
// Modified from https://github.com/NVIDIA/FasterTransformer/blob/main/src/fastertransformer/layers/DenseWeight.h
Li Zhang's avatar
Li Zhang committed
19
20
21
22
23
24
25
26
27

#pragma once

#include "src/fastertransformer/layers/FfnWeight.h"
#include "src/fastertransformer/layers/attention_layers/AttentionWeight.h"
#include "src/fastertransformer/utils/cuda_utils.h"

namespace fastertransformer {

AllentDan's avatar
AllentDan committed
28
enum class WeightType : int {
Li Zhang's avatar
Li Zhang committed
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
63
64
65
66
67
    kFP32,
    kFP16,
    kFP8,  // not supported yet
    kINT8,
    kINT4
};

inline size_t getBitSize(WeightType type)
{
    switch (type) {
        case WeightType::kFP32:
            return 32;
        case WeightType::kFP16:
            return 16;
        case WeightType::kFP8:
            return 8;
        case WeightType::kINT8:
            return 8;
        case WeightType::kINT4:
            return 4;
    }
}

template<typename T>
struct LlamaDenseWeight {

    size_t     input_dims;
    size_t     output_dims;
    void*      kernel;
    WeightType type;
    T*         bias;
    T*         scales;
    T*         zeros;
};

template<typename T>
struct LlamaAttentionWeight {
    LlamaDenseWeight<T> qkv;
    LlamaDenseWeight<T> output;
AllentDan's avatar
AllentDan committed
68
    std::vector<float>  past_kv_scale;
Li Zhang's avatar
Li Zhang committed
69
70
71
72
73
74
75
76
77
};

template<typename T>
struct LlamaFfnWeight {
    LlamaDenseWeight<T> gating;
    LlamaDenseWeight<T> intermediate;
    LlamaDenseWeight<T> output;
};

AllentDan's avatar
AllentDan committed
78
}  // namespace fastertransformer