test_embedding_models.py 2.27 KB
Newer Older
1
2
3
4
5
"""
Copyright 2023-2024 SGLang Team
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
6

7
    http://www.apache.org/licenses/LICENSE-2.0
8

9
10
11
12
13
14
15
16
17
18
19
20
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""

import unittest

import torch

from sglang.test.runners import DEFAULT_PROMPTS, HFRunner, SRTRunner
21
from sglang.test.test_utils import get_similarities
22

23
24
25
26
MODELS = [
    ("Alibaba-NLP/gte-Qwen2-1.5B-instruct", 1, 1e-5),
    ("intfloat/e5-mistral-7b-instruct", 1, 1e-5),
]
27
28
29
TORCH_DTYPES = [torch.float16]


30
class TestEmbeddingModels(unittest.TestCase):
31
32
33
34
35
36
37

    def assert_close_prefill_logits(
        self,
        prompts,
        model_path,
        tp_size,
        torch_dtype,
38
        prefill_tolerance,
39
40
    ) -> None:
        with HFRunner(
41
            model_path, torch_dtype=torch_dtype, is_generation=False
42
43
44
45
46
47
48
        ) as hf_runner:
            hf_outputs = hf_runner.forward(prompts)

        with SRTRunner(
            model_path,
            tp_size=tp_size,
            torch_dtype=torch_dtype,
49
            is_generation=False,
50
        ) as srt_runner:
51
            srt_outputs = srt_runner.forward(prompts)
52
53

        for i in range(len(prompts)):
54
55
56
            hf_logits = torch.Tensor(hf_outputs.embed_logits[i])
            srt_logits = torch.Tensor(srt_outputs.embed_logits[i])

57
58
            similarity = torch.tensor(get_similarities(hf_logits, srt_logits))
            print("similarity diff", abs(similarity - 1))
59

60
            if len(prompts[i]) <= 1000:
61
62
63
                assert torch.all(
                    abs(similarity - 1) < prefill_tolerance
                ), "embeddings are not all close"
64
65

    def test_prefill_logits(self):
66
        for model, tp_size, prefill_tolerance in MODELS:
67
68
            for torch_dtype in TORCH_DTYPES:
                self.assert_close_prefill_logits(
69
                    DEFAULT_PROMPTS, model, tp_size, torch_dtype, prefill_tolerance
70
71
72
73
74
                )


if __name__ == "__main__":
    unittest.main(warnings="ignore")