test_gritlm.py 7.53 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
from __future__ import annotations
4

5
import numpy as np
6
7
8
9
import openai
import pytest
from scipy.spatial.distance import cosine

10
11
from vllm import LLM, SamplingParams
from vllm.config import ModelConfig
12
13
14
15
16

from ....utils import RemoteOpenAIServer

MODEL_NAME = "parasail-ai/GritLM-7B-vllm"
MAX_MODEL_LEN = 4000
17
ATOL = 0.002
18
19
20
21
22
23


def _arr(arr):
    """
    Convert a list of integers to an array of integers.
    """
24
    return np.array(arr)
25
26


27
def test_find_array():
28
    from vllm.model_executor.models.gritlm import GritLMMeanPool
29

30
31
    model_config = ModelConfig(
        MODEL_NAME,
32
        runner="pooling",
33
34
35
        dtype="bfloat16",
        seed=0,
    )
36
    pooling = GritLMMeanPool(model_config=model_config)
37

38
    arr = _arr([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
39

40
41
42
43
44
45
    assert pooling._find_array(arr, _arr([3, 4, 5]), start_idx=0) == 3
    assert pooling._find_array(arr, _arr([3, 4, 5]), start_idx=1) == 3
    assert pooling._find_array(arr, _arr([3, 4, 5]), start_idx=5) == -1
    assert pooling._find_array(arr, _arr([3, 4, 5]), end_idx=3) == -1
    assert pooling._find_array(arr, _arr([3, 4, 5]), end_idx=4) == 3
    assert pooling._find_array(arr, _arr([3, 5]), start_idx=0) == -1
46

47
    with pytest.raises(ValueError):
48
        pooling._find_array(arr, _arr([3, 4, 5]), start_idx=-1)
49
50


51
def run_llm_encode(
52
    llm: LLM,
53
54
    queries: list[str],
    instruction: str,
55
56
) -> list[list[float]]:
    outputs = llm.embed([instruction + q for q in queries])
57
58
59
    return [output.outputs.embedding for output in outputs]


60
async def run_client_embeddings(
61
    client: openai.AsyncOpenAI,
62
63
    queries: list[str],
    instruction: str,
64
) -> list[list[float]]:
65
66
67
68
69
70
71
72
    outputs = await client.embeddings.create(
        model=MODEL_NAME,
        input=[instruction + q for q in queries],
    )
    return [data.embedding for data in outputs.data]


def gritlm_instruction(instruction):
73
74
75
    return (
        "<|user|>\n" + instruction + "\n<|embed|>\n" if instruction else "<|embed|>\n"
    )
76
77
78
79
80
81
82
83


def get_test_data():
    """
    Grabbed this test data and the expected values from
    README.md in https://github.com/ContextualAI/gritlm
    """
    q_instruction = gritlm_instruction(
84
85
        "Given a scientific paper title, retrieve the paper's abstract",
    )
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
    queries = [
        "Bitcoin: A Peer-to-Peer Electronic Cash System",
        "Generative Representational Instruction Tuning",
    ]

    d_instruction = gritlm_instruction("")
    documents = [
        # ruff: noqa: E501
        "A purely peer-to-peer version of electronic cash would allow online payments to be sent directly from one party to another without going through a financial institution. Digital signatures provide part of the solution, but the main benefits are lost if a trusted third party is still required to prevent double-spending. We propose a solution to the double-spending problem using a peer-to-peer network. The network timestamps transactions by hashing them into an ongoing chain of hash-based proof-of-work, forming a record that cannot be changed without redoing the proof-of-work. The longest chain not only serves as proof of the sequence of events witnessed, but proof that it came from the largest pool of CPU power. As long as a majority of CPU power is controlled by nodes that are not cooperating to attack the network, they'll generate the longest chain and outpace attackers. The network itself requires minimal structure. Messages are broadcast on a best effort basis, and nodes can leave and rejoin the network at will, accepting the longest proof-of-work chain as proof of what happened while they were gone.",
        "All text-based language problems can be reduced to either generation or embedding. Current models only perform well at one or the other. We introduce generative representational instruction tuning (GRIT) whereby a large language model is trained to handle both generative and embedding tasks by distinguishing between them through instructions. Compared to other open models, our resulting GritLM 7B sets a new state of the art on the Massive Text Embedding Benchmark (MTEB) and outperforms all models up to its size on a range of generative tasks. By scaling up further, GritLM 8X7B outperforms all open generative language models that we tried while still being among the best embedding models. Notably, we find that GRIT matches training on only generative or embedding data, thus we can unify both at no performance loss. Among other benefits, the unification via GRIT speeds up Retrieval-Augmented Generation (RAG) by > 60% for long documents, by no longer requiring separate retrieval and generation models. Models, code, etc. are freely available at https://github.com/ContextualAI/gritlm.",
    ]

    return queries, q_instruction, documents, d_instruction


101
def validate_embed_output(q_rep: list[list[float]], d_rep: list[list[float]]):
102
    cosine_sim_q0_d0 = 1 - cosine(q_rep[0], d_rep[0])
103
    assert cosine_sim_q0_d0 == pytest.approx(0.609, abs=ATOL)
104
105

    cosine_sim_q0_d1 = 1 - cosine(q_rep[0], d_rep[1])
106
    assert cosine_sim_q0_d1 == pytest.approx(0.101, abs=ATOL)
107
108

    cosine_sim_q1_d0 = 1 - cosine(q_rep[1], d_rep[0])
109
    assert cosine_sim_q1_d0 == pytest.approx(0.120, abs=ATOL)
110
111

    cosine_sim_q1_d1 = 1 - cosine(q_rep[1], d_rep[1])
112
    assert cosine_sim_q1_d1 == pytest.approx(0.534, abs=ATOL)
113
114


115
116
def test_gritlm_offline_embedding(vllm_runner):
    queries, q_instruction, documents, d_instruction = get_test_data()
117

118
    with vllm_runner(
119
120
121
        MODEL_NAME,
        runner="pooling",
        max_model_len=MAX_MODEL_LEN,
122
    ) as vllm_model:
123
        llm = vllm_model.llm
124

125
126
127
128
129
130
131
132
133
134
        d_rep = run_llm_encode(
            llm,
            documents,
            d_instruction,
        )
        q_rep = run_llm_encode(
            llm,
            queries,
            q_instruction,
        )
135

136
    validate_embed_output(q_rep, d_rep)
137
138
139
140
141
142


@pytest.mark.asyncio
async def test_gritlm_api_server_embedding():
    queries, q_instruction, documents, d_instruction = get_test_data()

143
    args = ["--runner", "pooling", "--max_model_len", str(MAX_MODEL_LEN)]
144

145
    with RemoteOpenAIServer(MODEL_NAME, args) as server:
146
        client_embedding = server.get_async_client()
147

148
149
        d_rep = await run_client_embeddings(
            client_embedding,
150
151
152
            documents,
            d_instruction,
        )
153
154
        q_rep = await run_client_embeddings(
            client_embedding,
155
156
157
            queries,
            q_instruction,
        )
158

159
    validate_embed_output(q_rep, d_rep)
160
161


162
def test_gritlm_offline_generate(monkeypatch: pytest.MonkeyPatch, vllm_runner):
163
    input = "<|user|>\nWhat is the capital of France?\n<|assistant|>\n"
164

165
    with vllm_runner(
166
167
168
        MODEL_NAME,
        runner="generate",
        max_model_len=MAX_MODEL_LEN,
169
    ) as vllm_model:
170
        llm = vllm_model.llm
171

172
173
        sampling_params = SamplingParams(temperature=0.0, max_tokens=256)
        outputs = llm.generate(input, sampling_params=sampling_params)
174

175
    assert outputs[0].outputs[0].text == "The capital of France is Paris."
176
177
178


@pytest.mark.asyncio
179
async def test_gritlm_api_server_generate():
180
181
    input = "<|user|>\nWhat is the capital of France?\n<|assistant|>\n"

182
    args = ["--runner", "generate", "--max_model_len", str(MAX_MODEL_LEN)]
183

184
    with RemoteOpenAIServer(MODEL_NAME, args) as server:
185
186
187
188
189
190
191
192
        client_generate = server.get_async_client()

        outputs = await client_generate.completions.create(
            model=MODEL_NAME,
            prompt=input,
            max_tokens=256,
            temperature=0.0,
        )
193
194

    assert outputs.choices[0].text == "The capital of France is Paris."