custom
code
sovereign-compute
File size: 4,221 Bytes
fd6abd3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
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
// ============================================================================
// flash_attention.cpp - PyTorch C++ Extension
// ============================================================================

#include <torch/extension.h>
#include <cuda_runtime.h>
#include <cuda_fp16.h>
#include <vector>
#include <cassert>

void launch_flash_attention_fwd(
    const half* Q, const half* K, const half* V, half* O,
    int batch_size, int num_heads, int N, int d, float scale
);

void launch_flash_attention_bwd(
    const half* Q, const half* K, const half* V,
    const half* O, const half* dO,
    const float* L, const float* M,
    half* dQ, half* dK, half* dV,
    int batch_size, int num_heads, int N, int d, float scale
);

std::vector<at::Tensor> flash_attention_fwd(
    at::Tensor Q, at::Tensor K, at::Tensor V, double scale_ = -1.0
) {
    TORCH_CHECK(Q.is_cuda() && K.is_cuda() && V.is_cuda(), "Inputs must be CUDA tensors");
    TORCH_CHECK(Q.dtype() == at::kHalf, "Inputs must be fp16");

    int batch_size = Q.size(0);
    int num_heads = Q.size(1);
    int N = Q.size(2);
    int d = Q.size(3);
    float scale = (scale_ > 0) ? scale_ : 1.0f / std::sqrt(d);

    auto O = at::empty_like(Q);
    auto L = at::empty({batch_size, num_heads, N}, Q.options().dtype(at::kFloat));
    auto M = at::empty({batch_size, num_heads, N}, Q.options().dtype(at::kFloat));

    launch_flash_attention_fwd(
        static_cast<const half*>(Q.data_ptr()),
        static_cast<const half*>(K.data_ptr()),
        static_cast<const half*>(V.data_ptr()),
        static_cast<half*>(O.data_ptr()),
        batch_size, num_heads, N, d, scale
    );

    return {O, L, M};
}

std::vector<at::Tensor> flash_attention_bwd(
    at::Tensor Q, at::Tensor K, at::Tensor V,
    at::Tensor O, at::Tensor dO,
    at::Tensor L, at::Tensor M, double scale_ = -1.0
) {
    TORCH_CHECK(Q.is_cuda() && dO.is_cuda(), "Inputs must be CUDA tensors");

    int batch_size = Q.size(0);
    int num_heads = Q.size(1);
    int N = Q.size(2);
    int d = Q.size(3);
    float scale = (scale_ > 0) ? scale_ : 1.0f / std::sqrt(d);

    auto dQ = at::empty_like(Q);
    auto dK = at::empty_like(K);
    auto dV = at::empty_like(V);

    launch_flash_attention_bwd(
        static_cast<const half*>(Q.data_ptr()),
        static_cast<const half*>(K.data_ptr()),
        static_cast<const half*>(V.data_ptr()),
        static_cast<const half*>(O.data_ptr()),
        static_cast<const half*>(dO.data_ptr()),
        static_cast<const float*>(L.data_ptr()),
        static_cast<const float*>(M.data_ptr()),
        static_cast<half*>(dQ.data_ptr()),
        static_cast<half*>(dK.data_ptr()),
        static_cast<half*>(dV.data_ptr()),
        batch_size, num_heads, N, d, scale
    );

    return {dQ, dK, dV};
}

class FlashAttentionFunction : public torch::autograd::Function<FlashAttentionFunction> {
public:
    static torch::autograd::variable_list forward(
        torch::autograd::AutogradContext* ctx,
        torch::Tensor Q, torch::Tensor K, torch::Tensor V, double scale = -1.0
    ) {
        auto outputs = flash_attention_fwd(Q, K, V, scale);
        ctx->save_for_backward({Q, K, V, outputs[0], outputs[1], outputs[2]});
        ctx->saved_data["scale"] = scale;
        return {outputs[0]};
    }

    static torch::autograd::variable_list backward(
        torch::autograd::AutogradContext* ctx,
        torch::autograd::variable_list grad_outputs
    ) {
        auto saved = ctx->get_saved_variables();
        auto grads = flash_attention_bwd(
            saved[0], saved[1], saved[2], saved[3],
            grad_outputs[0], saved[4], saved[5],
            ctx->saved_data["scale"].toDouble()
        );
        return {grads[0], grads[1], grads[2], torch::Tensor()};
    }
};

torch::Tensor flash_attention(torch::Tensor Q, torch::Tensor K, torch::Tensor V, double scale = -1.0) {
    return FlashAttentionFunction::apply({Q, K, V, scale})[0];
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("flash_attention_fwd", &flash_attention_fwd, "FlashAttention Forward");
    m.def("flash_attention_bwd", &flash_attention_bwd, "FlashAttention Backward");
    m.def("flash_attention", &flash_attention, "FlashAttention (autograd)");
}