Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_punica.py 5.82 KB
Newer Older
Antoni Baum's avatar
[Experimental] Add multi-LoRA support (#1804)  
Antoni Baum 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
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
162
163
164
165
166
167
168
169
170
171
172
173
174
175 
# Based on code from https://github.com/punica-ai/punica

import pytest
import torch

import vllm.lora.punica as punica


def assert_close(a, b):
    rtol, atol = {
        torch.float16: (5e-3, 5e-3),
        torch.bfloat16: (3e-2, 2e-2),
        torch.float32: (None, None),
    }[a.dtype]
    torch.testing.assert_close(a, b, rtol=rtol, atol=atol)


def _lora_ref_impl(
    y_final: torch.Tensor,
    x: torch.Tensor,
    wa_T_all: torch.Tensor,
    wb_T_all: torch.Tensor,
    indicies: torch.LongTensor,
    layer_idx: int,
    scale: float,
):
    y_stage_1 = torch.empty(
        (x.size(0), wa_T_all.size(-2)),
        dtype=torch.float32,
        device=x.device,
    )
    bs = x.shape[0]
    s = torch.tensor(scale, dtype=torch.float32, device=x.device)
    for i, lora_idx in zip(range(bs), indicies.cpu().tolist()):
        xi = x[i].unsqueeze(0).to(torch.float32)
        wa = wa_T_all[lora_idx, layer_idx].transpose(-1, -2).to(torch.float32)
        wb = wb_T_all[lora_idx, layer_idx].transpose(-1, -2).to(torch.float32)

        tmp = xi @ wa
        y_stage_1[i] = tmp.squeeze(0)
        y_final[i] += (tmp @ wb).squeeze(0) * s
    return y_final, y_stage_1


H1 = H2 = [
    128, 256, 512, 1024, 1280, 2048, 2560, 2752, 3072, 3456, 3584, 4096, 5120,
    5504, 5632, 6912, 7168, 8192, 9216, 10240, 11008, 13824, 14336, 32000,
    32256, 32512, 32768, 33024
]
SEED = [0xabcdabcd987]


@pytest.mark.parametrize("dtype_str", ["float16", "bfloat16"])
@pytest.mark.parametrize("h1", H1)
@pytest.mark.parametrize("h2", H2)
@pytest.mark.parametrize("seed", SEED)
@torch.inference_mode()
def test_lora_correctness(dtype_str, h1, h2, seed):
    torch.manual_seed(seed)
    num_loras = 4
    num_layers = 1
    r = 8
    bs = 32
    scale = 0.123
    dtype = getattr(torch, dtype_str)
    device = torch.device("cuda")

    wa_T_all = torch.randn(num_loras,
                           num_layers,
                           r,
                           h1,
                           dtype=dtype,
                           device=device)
    wb_T_all = torch.randn(num_loras,
                           num_layers,
                           h2,
                           r,
                           dtype=dtype,
                           device=device)
    indices = torch.randint(num_loras, (bs, ), dtype=torch.long, device=device)

    for layer_idx in range(num_layers):
        x = torch.randn(bs, h1, dtype=dtype, device=device)
        y = torch.randn(bs, h2, dtype=dtype, device=device)

        y_ref = y.clone()
        _lora_ref_impl(y_ref, x, wa_T_all, wb_T_all, indices, layer_idx, scale)

        y_our = y.clone()
        punica.add_lora(y_our, x, wa_T_all, wb_T_all, indices, layer_idx,
                        scale)

        assert_close(y_ref, y_our)


@pytest.mark.parametrize("dtype_str", ["float16", "bfloat16"])
@pytest.mark.parametrize("h1", H1)
@pytest.mark.parametrize("h2", H2)
@pytest.mark.parametrize("seed", SEED)
@torch.inference_mode()
def test_lora_correctness_slice(dtype_str, h1, h2, seed):
    if h2 % 3 != 0 or h2 // 3 not in H1:
        pytest.skip("h2 must be divisible by 3 and in supported shapes")
    torch.manual_seed(seed)
    num_loras = 4
    num_layers = 1
    r = 8
    bs = 32
    scale = 0.123
    dtype = getattr(torch, dtype_str)
    device = torch.device("cuda")

    wa_T_all_0 = torch.randn(num_loras,
                             num_layers,
                             r,
                             h1,
                             dtype=dtype,
                             device=device)
    wa_T_all_1 = torch.randn(num_loras,
                             num_layers,
                             r,
                             h1,
                             dtype=dtype,
                             device=device)
    wa_T_all_2 = torch.randn(num_loras,
                             num_layers,
                             r,
                             h1,
                             dtype=dtype,
                             device=device)
    wb_T_all_0 = torch.randn(num_loras,
                             num_layers,
                             h2 // 3,
                             r,
                             dtype=dtype,
                             device=device)
    wb_T_all_1 = torch.randn(num_loras,
                             num_layers,
                             h2 // 3,
                             r,
                             dtype=dtype,
                             device=device)
    wb_T_all_2 = torch.randn(num_loras,
                             num_layers,
                             h2 // 3,
                             r,
                             dtype=dtype,
                             device=device)

    indices = torch.randint(num_loras, (bs, ), dtype=torch.long, device=device)

    for layer_idx in range(num_layers):
        x = torch.randn(bs, h1, dtype=dtype, device=device)
        y = torch.randn(bs, h2, dtype=dtype, device=device)
        s = h2 // 3

        y_ref = y.clone()
        _lora_ref_impl(y_ref[:, :s], x, wa_T_all_0, wb_T_all_0, indices,
                       layer_idx, scale)
        _lora_ref_impl(y_ref[:, s:s * 2], x, wa_T_all_1, wb_T_all_1, indices,
                       layer_idx, scale)
        _lora_ref_impl(y_ref[:, s * 2:], x, wa_T_all_2, wb_T_all_2, indices,
                       layer_idx, scale)

        y_our = y.clone()
        punica.add_lora_slice(y_our, x, wa_T_all_0, wb_T_all_0, indices,
                              layer_idx, scale, 0, s)
        punica.add_lora_slice(y_our, x, wa_T_all_1, wb_T_all_1, indices,
                              layer_idx, scale, s, s)
        punica.add_lora_slice(y_our, x, wa_T_all_2, wb_T_all_2, indices,
                              layer_idx, scale, s * 2, s)

        assert_close(y_ref[:, :s], y_our[:, :s])
        assert_close(y_ref[:, s:s * 2], y_our[:, s:s * 2])
        assert_close(y_ref[:, s * 2:], y_our[:, s * 2:])