File size: 5,835 Bytes
e45d058
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
# Run test with:
# torchrun --no_python --nproc_per_node=8 pytest -q -s tests/modules/test_mha_parallel.py

import math

import pytest
import torch
import torch.nn.functional as F
from apex.transformer import parallel_state, tensor_parallel
from einops import rearrange
from flash_attn.modules.mha import MHA, ParallelMHA

is_sm8x = torch.cuda.get_device_capability("cuda")[0] >= 8


@pytest.mark.parametrize("dtype", [torch.float16] + ([torch.bfloat16] if is_sm8x else []))
# @pytest.mark.parametrize('dtype', [torch.float16])
@pytest.mark.parametrize("world_size", [1, 2, 4, 8])
# @pytest.mark.parametrize('world_size', [2])
@pytest.mark.parametrize("sequence_parallel", [True, False])
# @pytest.mark.parametrize('sequence_parallel', [False])
@pytest.mark.parametrize("head_dim", [64, 128])
# @pytest.mark.parametrize('head_dim', [64])
@pytest.mark.parametrize("embed_dim", [1024, 4096])
# @pytest.mark.parametrize('embed_dim', [1024])
def test_mha_parallel(embed_dim, head_dim, sequence_parallel, world_size, dtype):
    assert embed_dim % head_dim == 0
    num_heads = embed_dim // head_dim
    assert num_heads % world_size == 0
    rtol, atol = (3e-3, 1e-2) if dtype == torch.bfloat16 else (3e-3, 1e-3)
    if not torch.distributed.is_initialized():
        torch.distributed.init_process_group(backend="nccl", init_method="env://")
    device = f"cuda:{torch.distributed.get_rank()}"
    assert world_size <= torch.distributed.get_world_size()
    parallel_state.initialize_model_parallel(tensor_model_parallel_size_=world_size)
    rank = parallel_state.get_tensor_model_parallel_rank()
    # set seed
    torch.random.manual_seed(0)
    batch_size = 2
    seqlen = 1024
    assert (batch_size * seqlen) % world_size == 0
    x_pt = torch.randn(
        batch_size * seqlen, embed_dim, device=device, dtype=dtype, requires_grad=True
    )
    # We need to generate g here so that all processes get the same gradient,
    # as rank 0 will have an extra bias that changes the RNG.
    # If we don't divide by batch_size, the gradient gets a bit too large.
    g = torch.randn_like(x_pt) / 32
    if sequence_parallel:
        x = (
            tensor_parallel.scatter_to_sequence_parallel_region(x_pt)
            .detach()
            .clone()
            .requires_grad_()
        )
    else:
        x = x_pt.detach().clone().requires_grad_()

    model_pt = MHA(
        embed_dim,
        num_heads,
        rotary_emb_dim=int(head_dim // 2),
        use_flash_attn=True,
        device=device,
        dtype=dtype,
    )
    partition_dim = embed_dim // world_size
    model = ParallelMHA(
        embed_dim,
        num_heads,
        parallel_state.get_tensor_model_parallel_group(),
        rotary_emb_dim=int(head_dim // 2),
        use_flash_attn=True,
        sequence_parallel=sequence_parallel,
        device=device,
        dtype=dtype,
    )

    with torch.no_grad():
        model.Wqkv.weight.copy_(
            rearrange(
                rearrange(model_pt.Wqkv.weight, "(three o) i -> three o i", three=3)[
                    :, rank * partition_dim : (rank + 1) * partition_dim
                ],
                "three o i -> (three o) i",
            )
        )
        model.Wqkv.bias.copy_(
            rearrange(
                rearrange(model_pt.Wqkv.bias, "(three o) -> three o", three=3)[
                    :, rank * partition_dim : (rank + 1) * partition_dim
                ],
                "three o -> (three o)",
            )
        )
        model.out_proj.weight.copy_(
            model_pt.out_proj.weight[:, rank * partition_dim : (rank + 1) * partition_dim]
        )
        if rank == 0:
            model.out_proj.bias.copy_(model_pt.out_proj.bias)

    out = model(x, seqlen=seqlen)
    out_pt = rearrange(model_pt(rearrange(x_pt, "(b s) d -> b s d", s=seqlen)), "b s d -> (b s) d")
    partition_batch_dim = batch_size * seqlen // world_size
    assert torch.allclose(
        out,
        out_pt[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
        if sequence_parallel
        else out_pt,
        rtol=rtol,
        atol=atol,
    )

    out_pt.backward(g)
    out.backward(
        g[rank * partition_batch_dim : (rank + 1) * partition_batch_dim] if sequence_parallel else g
    )
    parallel_state.destroy_model_parallel()

    assert torch.allclose(
        x.grad,
        x_pt.grad[rank * partition_batch_dim : (rank + 1) * partition_batch_dim]
        if sequence_parallel
        else x_pt.grad,
        rtol=rtol,
        atol=atol / 100,  # magnitude of x.grad is quite small
    )
    # The error for d_weight and d_bias is quite a bit higher
    assert torch.allclose(
        model.Wqkv.weight.grad,
        rearrange(
            rearrange(model_pt.Wqkv.weight.grad, "(three o) i -> three o i", three=3)[
                :, rank * partition_dim : (rank + 1) * partition_dim
            ],
            "three o i -> (three o) i",
        ),
        rtol=rtol,
        atol=atol * 10,
    )
    assert torch.allclose(
        model.Wqkv.bias.grad,
        rearrange(
            rearrange(model_pt.Wqkv.bias.grad, "(three o) -> three o", three=3)[
                :, rank * partition_dim : (rank + 1) * partition_dim
            ],
            "three o -> (three o)",
        ),
        rtol=rtol,
        atol=atol * 5,
    )
    assert torch.allclose(
        model.out_proj.weight.grad,
        model_pt.out_proj.weight.grad[:, rank * partition_dim : (rank + 1) * partition_dim],
        rtol=rtol,
        atol=atol * 10,
    )
    if rank == 0:
        assert torch.allclose(
            model.out_proj.bias.grad, model_pt.out_proj.bias.grad, rtol=rtol, atol=atol * 5
        )