Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
rmsnorm.py 3.15 KB
Newer Older
yanjl1's avatar
Initial  
yanjl1 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
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 
import hipdnn
import torch


def build_rmsnorm_graph(
    hipdnn_handle,
    torch_tensor_x,
    torch_tensor_scale,
    torch_tensor_bias,
    torch_tensor_epsilon,
    norm_forward_phase,
    eps,
    hipdnn_data_type,
):
    # Create graph
    graph = hipdnn.pygraph(
        handle=hipdnn_handle,
        io_data_type=hipdnn_data_type,
        intermediate_data_type=hipdnn.data_type.FLOAT,
        compute_data_type=hipdnn.data_type.FLOAT,
        name="rmsnorm_inference",
    )

    # Create hipdnn tensors
    hipdnn_tensor_x = graph.tensor_like(torch_tensor_x)
    hipdnn_tensor_scale = graph.tensor_like(torch_tensor_scale)
    hipdnn_tensor_bias = graph.tensor_like(torch_tensor_bias)
    hipdnn_tensor_epsilon = graph.tensor_like(torch_tensor_epsilon)
    hipdnn_tensor_epsilon.set_value(eps)

    # Create op
    hipdnn_tensor_y, hipdnn_tensor_inv_var = graph.rmsnorm(
        norm_forward_phase=norm_forward_phase,
        input=hipdnn_tensor_x,
        scale=hipdnn_tensor_scale,
        bias=hipdnn_tensor_bias,
        epsilon=hipdnn_tensor_epsilon,
        compute_data_type=hipdnn.data_type.FLOAT,
        name="rmsnorm",
    )
    hipdnn_tensor_y.set_output(True)
    hipdnn_tensor_inv_var.set_output(True)
    graph.build(hipdnn_handle)

    return (graph, hipdnn_tensor_x, hipdnn_tensor_scale, hipdnn_tensor_y, hipdnn_tensor_inv_var)


if __name__ == "__main__":
    # Input dimensions
    batch = 2  # Batch size
    seq_len = 1024  # Number of input channels
    embedding_dim = 768  # Number of feature
    norm_forward_phase = hipdnn.norm_forward_phase.TRAINING  # Norm forward phase
    eps = 1e-5

    hipdnn_data_type = hipdnn.data_type.FLOAT
    torch_data_type = torch.float32

    torch_tensor_x = torch.rand(batch, seq_len, embedding_dim, dtype=torch_data_type, device="cuda")
    torch_tensor_scale = torch.rand(1, 1, embedding_dim, dtype=torch_data_type, device="cuda")
    torch_tensor_bias = torch.rand(1, 1, embedding_dim, dtype=torch_data_type, device="cuda")
    torch_tensor_epsilon = torch.full(
        (1, 1, 1, 1), eps, dtype=torch.float32, requires_grad=False, device="cpu"
    )

    hipdnn_handle = hipdnn.create_handle()

    graph, hipdnn_tensor_x, hipdnn_tensor_scale, hipdnn_tensor_y, hipdnn_tensor_inv_var = (
        build_rmsnorm_graph(
            hipdnn_handle,
            torch_tensor_x,
            torch_tensor_scale,
            torch_tensor_bias,
            torch_tensor_epsilon,
            norm_forward_phase,
            eps,
            hipdnn_data_type,
        )
    )

    torch_tensor_y = torch.empty(hipdnn_tensor_y.get_dim(), dtype=torch_data_type, device="cuda")
    torch_tensor_inv_var = torch.empty(
        hipdnn_tensor_inv_var.get_dim(), dtype=torch_data_type, device="cuda"
    )
    variant_pack = {
        hipdnn_tensor_x: torch_tensor_x.data_ptr(),
        hipdnn_tensor_scale: torch_tensor_scale.data_ptr(),
        hipdnn_tensor_inv_var: torch_tensor_inv_var.data_ptr(),
        hipdnn_tensor_y: torch_tensor_y.data_ptr(),
    }
    workspace = torch.empty(graph.get_workspace_size(), dtype=torch.uint8, device="cuda")

    graph.exec(variant_pack=variant_pack, workspace=workspace.data_ptr())
    print("rmsnorm graph execution complete.")