test_client.py 5.99 KB
Newer Older
1
2
3
4
import pytest

from text_generation import Client, AsyncClient
from text_generation.errors import NotFoundError, ValidationError
5
from text_generation.types import FinishReason, InputToken
6
7


8
9
def test_generate(flan_t5_xxl_url, hf_headers):
    client = Client(flan_t5_xxl_url, hf_headers)
10
    response = client.generate("test", max_new_tokens=1, decoder_input_details=True)
11

12
    assert response.generated_text == ""
13
14
15
16
    assert response.details.finish_reason == FinishReason.Length
    assert response.details.generated_tokens == 1
    assert response.details.seed is None
    assert len(response.details.prefill) == 1
17
    assert response.details.prefill[0] == InputToken(id=0, text="<pad>", logprob=None)
18
    assert len(response.details.tokens) == 1
19
    assert response.details.tokens[0].id == 3
20
    assert response.details.tokens[0].text == " "
21
    assert not response.details.tokens[0].special
22
23


24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
def test_generate_max_new_tokens_not_set(flan_t5_xxl_url, hf_headers):
    client = Client(flan_t5_xxl_url, hf_headers)
    response = client.generate("test", decoder_input_details=True)

    assert response.generated_text != ""
    assert response.details.finish_reason == FinishReason.EndOfSequenceToken
    assert response.details.generated_tokens > 1
    assert response.details.seed is None
    assert len(response.details.prefill) == 1
    assert response.details.prefill[0] == InputToken(id=0, text="<pad>", logprob=None)
    assert len(response.details.tokens) > 1
    assert response.details.tokens[0].id == 3
    assert response.details.tokens[0].text == " "
    assert not response.details.tokens[0].special


40
41
def test_generate_best_of(flan_t5_xxl_url, hf_headers):
    client = Client(flan_t5_xxl_url, hf_headers)
42
43
44
    response = client.generate(
        "test", max_new_tokens=1, best_of=2, do_sample=True, decoder_input_details=True
    )
45
46
47
48
49
50
51

    assert response.details.seed is not None
    assert response.details.best_of_sequences is not None
    assert len(response.details.best_of_sequences) == 1
    assert response.details.best_of_sequences[0].seed is not None


52
53
54
55
56
57
58
59
60
61
62
63
def test_generate_not_found(fake_url, hf_headers):
    client = Client(fake_url, hf_headers)
    with pytest.raises(NotFoundError):
        client.generate("test")


def test_generate_validation_error(flan_t5_xxl_url, hf_headers):
    client = Client(flan_t5_xxl_url, hf_headers)
    with pytest.raises(ValidationError):
        client.generate("test", max_new_tokens=10_000)


64
65
def test_generate_stream(flan_t5_xxl_url, hf_headers):
    client = Client(flan_t5_xxl_url, hf_headers)
66
67
68
69
70
71
72
    responses = [
        response for response in client.generate_stream("test", max_new_tokens=1)
    ]

    assert len(responses) == 1
    response = responses[0]

73
    assert response.generated_text == ""
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
    assert response.details.finish_reason == FinishReason.Length
    assert response.details.generated_tokens == 1
    assert response.details.seed is None


def test_generate_stream_not_found(fake_url, hf_headers):
    client = Client(fake_url, hf_headers)
    with pytest.raises(NotFoundError):
        list(client.generate_stream("test"))


def test_generate_stream_validation_error(flan_t5_xxl_url, hf_headers):
    client = Client(flan_t5_xxl_url, hf_headers)
    with pytest.raises(ValidationError):
        list(client.generate_stream("test", max_new_tokens=10_000))


@pytest.mark.asyncio
92
93
async def test_generate_async(flan_t5_xxl_url, hf_headers):
    client = AsyncClient(flan_t5_xxl_url, hf_headers)
94
95
96
    response = await client.generate(
        "test", max_new_tokens=1, decoder_input_details=True
    )
97

98
    assert response.generated_text == ""
99
100
101
102
    assert response.details.finish_reason == FinishReason.Length
    assert response.details.generated_tokens == 1
    assert response.details.seed is None
    assert len(response.details.prefill) == 1
103
    assert response.details.prefill[0] == InputToken(id=0, text="<pad>", logprob=None)
104
    assert len(response.details.tokens) == 1
105
    assert response.details.tokens[0].id == 3
106
    assert response.details.tokens[0].text == " "
107
    assert not response.details.tokens[0].special
108
109


110
111
112
113
@pytest.mark.asyncio
async def test_generate_async_best_of(flan_t5_xxl_url, hf_headers):
    client = AsyncClient(flan_t5_xxl_url, hf_headers)
    response = await client.generate(
114
        "test", max_new_tokens=1, best_of=2, do_sample=True, decoder_input_details=True
115
116
117
118
119
120
121
122
    )

    assert response.details.seed is not None
    assert response.details.best_of_sequences is not None
    assert len(response.details.best_of_sequences) == 1
    assert response.details.best_of_sequences[0].seed is not None


123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
@pytest.mark.asyncio
async def test_generate_async_not_found(fake_url, hf_headers):
    client = AsyncClient(fake_url, hf_headers)
    with pytest.raises(NotFoundError):
        await client.generate("test")


@pytest.mark.asyncio
async def test_generate_async_validation_error(flan_t5_xxl_url, hf_headers):
    client = AsyncClient(flan_t5_xxl_url, hf_headers)
    with pytest.raises(ValidationError):
        await client.generate("test", max_new_tokens=10_000)


@pytest.mark.asyncio
138
139
async def test_generate_stream_async(flan_t5_xxl_url, hf_headers):
    client = AsyncClient(flan_t5_xxl_url, hf_headers)
140
141
142
143
144
145
146
    responses = [
        response async for response in client.generate_stream("test", max_new_tokens=1)
    ]

    assert len(responses) == 1
    response = responses[0]

147
    assert response.generated_text == ""
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
    assert response.details.finish_reason == FinishReason.Length
    assert response.details.generated_tokens == 1
    assert response.details.seed is None


@pytest.mark.asyncio
async def test_generate_stream_async_not_found(fake_url, hf_headers):
    client = AsyncClient(fake_url, hf_headers)
    with pytest.raises(NotFoundError):
        async for _ in client.generate_stream("test"):
            pass


@pytest.mark.asyncio
async def test_generate_stream_async_validation_error(flan_t5_xxl_url, hf_headers):
    client = AsyncClient(flan_t5_xxl_url, hf_headers)
    with pytest.raises(ValidationError):
        async for _ in client.generate_stream("test", max_new_tokens=10_000):
            pass