Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_activation.py 2.27 KB
Newer Older
huangwb's avatar
add custom vllm source code  
huangwb 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 
import pytest
import torch
import torch.nn.functional as F
from transformers.activations import get_activation

from vllm import activation_ops

DTYPES = [torch.half, torch.bfloat16, torch.float]
NUM_TOKENS = [7, 83, 2048]  # Arbitrary values for testing
D = [512, 4096, 5120, 13824]  # Arbitrary values for testing
SEEDS = [0]


def ref_silu_and_mul(x: torch.Tensor) -> torch.Tensor:
    x1, x2 = x.chunk(chunks=2, dim=1)
    return F.silu(x1) * x2


@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@pytest.mark.parametrize("d", D)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("seed", SEEDS)
@torch.inference_mode()
def test_silu_and_mul(
    num_tokens: int,
    d: int,
    dtype: torch.dtype,
    seed: int,
) -> None:
    torch.random.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    x = torch.randn(num_tokens, 2 * d, dtype=dtype, device="cuda")
    out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
    activation_ops.silu_and_mul(out, x)
    ref_out = ref_silu_and_mul(x)
    assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)


@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@pytest.mark.parametrize("d", D)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("seed", SEEDS)
@torch.inference_mode()
def test_gelu_new(
    num_tokens: int,
    d: int,
    dtype: torch.dtype,
    seed: int,
) -> None:
    torch.random.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    x = torch.randn(num_tokens, d, dtype=dtype, device="cuda")
    out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
    activation_ops.gelu_new(out, x)
    ref_out = get_activation("gelu_new")(x)
    assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)


@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@pytest.mark.parametrize("d", D)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("seed", SEEDS)
def test_gelu_fast(
    num_tokens: int,
    d: int,
    dtype: torch.dtype,
    seed: int,
) -> None:
    torch.random.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    x = torch.randn(num_tokens, d, dtype=dtype, device="cuda")
    out = torch.empty(num_tokens, d, dtype=dtype, device="cuda")
    activation_ops.gelu_fast(out, x)
    ref_out = get_activation("gelu_fast")(x)
    assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)