File size: 3,991 Bytes
cf2f35c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.amp as amp

try:
    import xfuser
    from xfuser.core.distributed import (get_sequence_parallel_rank,
                                         get_sequence_parallel_world_size,
                                         get_sp_group,
                                         init_distributed_environment,
                                         initialize_model_parallel)
    from xfuser.core.long_ctx_attention import xFuserLongContextAttention
except Exception as ex:
    get_sequence_parallel_world_size = None
    get_sequence_parallel_rank = None
    xFuserLongContextAttention = None
    get_sp_group = None
    init_distributed_environment = None
    initialize_model_parallel = None

def pad_freqs(original_tensor, target_len):
    seq_len, s1, s2 = original_tensor.shape
    pad_size = target_len - seq_len
    padding_tensor = torch.ones(
        pad_size,
        s1,
        s2,
        dtype=original_tensor.dtype,
        device=original_tensor.device)
    padded_tensor = torch.cat([original_tensor, padding_tensor], dim=0)
    return padded_tensor

@amp.autocast('cuda', enabled=False)
def rope_apply(x, grid_sizes, freqs):
    """

    x:          [B, L, N, C].

    grid_sizes: [B, 3].

    freqs:      [M, C // 2].

    """
    s, n, c = x.size(1), x.size(2), x.size(3) // 2
    # split freqs
    freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1)

    # loop over samples
    output = []
    for i, (f, h, w) in enumerate(grid_sizes.tolist()):
        seq_len = f * h * w

        # precompute multipliers
        x_i = torch.view_as_complex(x[i, :s].to(torch.float32).reshape(
            s, n, -1, 2))
        freqs_i = torch.cat([
            freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
            freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
            freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
        ],
        dim=-1).reshape(seq_len, 1, -1)

        # apply rotary embedding
        sp_size = get_sequence_parallel_world_size()
        sp_rank = get_sequence_parallel_rank()
        freqs_i = pad_freqs(freqs_i, s * sp_size)
        s_per_rank = s
        freqs_i_rank = freqs_i[(sp_rank * s_per_rank):((sp_rank + 1) *
                                                       s_per_rank), :, :]
        x_i = torch.view_as_real(x_i * freqs_i_rank).flatten(2)
        x_i = torch.cat([x_i, x[i, s:]])

        # append to collection
        output.append(x_i)
    return torch.stack(output)

def usp_attn_forward(self,

                     x,

                     seq_lens,

                     grid_sizes,

                     freqs,

                     dtype=torch.bfloat16):
    b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
    half_dtypes = (torch.float16, torch.bfloat16)

    def half(x):
        return x if x.dtype in half_dtypes else x.to(dtype)

    # query, key, value function
    def qkv_fn(x):
        q = self.norm_q(self.q(x)).view(b, s, n, d)
        k = self.norm_k(self.k(x)).view(b, s, n, d)
        v = self.v(x).view(b, s, n, d)
        return q, k, v

    q, k, v = qkv_fn(x)
    q = rope_apply(q, grid_sizes, freqs)
    k = rope_apply(k, grid_sizes, freqs)

    # TODO: We should use unpaded q,k,v for attention.
    # k_lens = seq_lens // get_sequence_parallel_world_size()
    # if k_lens is not None:
    #     q = torch.cat([u[:l] for u, l in zip(q, k_lens)]).unsqueeze(0)
    #     k = torch.cat([u[:l] for u, l in zip(k, k_lens)]).unsqueeze(0)
    #     v = torch.cat([u[:l] for u, l in zip(v, k_lens)]).unsqueeze(0)

    x = xFuserLongContextAttention()(
        None,
        query=half(q),
        key=half(k),
        value=half(v),
        window_size=self.window_size)

    # TODO: padding after attention.
    # x = torch.cat([x, x.new_zeros(b, s - x.size(1), n, d)], dim=1)

    # output
    x = x.flatten(2)
    x = self.o(x)
    return x