@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 48e925fab5

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 15, 12:15 UTC

[Misc] Clean up test docstrings and names (#17521)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

Cyrus Leung committed May 1, 2025 at 20:19 UTC 48e925fab596f513b76e53e3061e897729ff0cb9

19 files changed +51 -115

.buildkite/test-pipeline.yaml

+8 -2

@@ -395,10 +395,8 @@ steps:

395 - csrc/

396 - vllm/model_executor/layers/quantization

397 - tests/quantization

398 - - tests/models/quantization

398 commands:

399 - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization

401 - - pytest -v -s models/quantization

400

401 - label: LM Eval Small Models # 53min

402 working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"

@@ -509,6 +507,14 @@ steps:

507 - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git

508 - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'

509

510 +- label: Quantized Models Test

511 + #mirror_hardwares: [amd]

512 + source_file_dependencies:

513 + - vllm/model_executor/layers/quantization

514 + - tests/models/quantization

515 + commands:

516 + - pytest -v -s models/quantization

517 +

518 # This test is used only in PR development phase to test individual models and should never run on main

519 - label: Custom Models Test

520 mirror_hardwares: [amd]

tests/models/language/generation/test_common.py renamed

-5

@@ -1,9 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the outputs of HF and vLLM when using greedy sampling.

3 -

4 -Run `pytest tests/models/test_models.py`.

5 -"""

6 -

2 import pytest

3 import torch

tests/models/language/generation/test_granite.py

-4

@@ -1,8 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the outputs of HF and vLLM for Granite models using greedy sampling.

3 -

4 -Run `pytest tests/models/test_granite.py`.

5 -"""

2 import pytest

4 from ...utils import check_logprobs_close

tests/models/language/generation/test_mistral.py

-4

@@ -1,8 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the outputs of HF and vLLM for Mistral models using greedy sampling.

3 -

4 -Run `pytest tests/models/test_mistral.py`.

5 -"""

2 import copy

3 import json

tests/models/language/generation/test_phimoe.py

-4

@@ -1,8 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the outputs of HF and vLLM for moe models using greedy sampling.

3 -

4 -Run `pytest tests/models/test_phimoe.py`.

5 -"""

2 import pytest

3 import torch

tests/models/language/pooling/test_classification.py renamed

+1 -5

@@ -1,8 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the classification outputs of HF and vLLM models.

3 -

4 -Run `pytest tests/models/test_cls_models.py`.

5 -"""

2 import pytest

3 import torch

4 from transformers import AutoModelForSequenceClassification

@@ -19,7 +15,7 @@ from vllm.platforms import current_platform

15 )

16 @pytest.mark.parametrize("dtype",

17 ["half"] if current_platform.is_rocm() else ["float"])

22 -def test_classification_models(

18 +def test_models(

19 hf_runner,

20 vllm_runner,

21 example_prompts,

tests/models/language/pooling/test_embedding.py

-4

@@ -1,8 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the embedding outputs of HF and vLLM models.

3 -

4 -Run `pytest tests/models/embedding/language/test_embedding.py`.

5 -"""

2 import pytest

4 from vllm.config import PoolerConfig

tests/models/language/pooling/test_jina.py

+3 -8

@@ -1,9 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -# ruff: noqa: E501

3 -"""Compare the scoring outputs of HF and vLLM models.

4 -

5 -Run `pytest tests/models/embedding/language/test_jina.py`.

6 -"""

2 import math

4 import pytest

@@ -22,9 +17,9 @@ TEXTS_2 = [

17 "Organic skincare for sensitive skin with aloe vera and chamomile.",

18 "New makeup trends focus on bold colors and innovative techniques",

19 "Bio-Hautpflege für empfindliche Haut mit Aloe Vera und Kamille",

25 - "Neue Make-up-Trends setzen auf kräftige Farben und innovative Techniken",

26 - "Cuidado de la piel orgánico para piel sensible con aloe vera y manzanilla",

27 - "Las nuevas tendencias de maquillaje se centran en colores vivos y técnicas innovadoras",

20 + "Neue Make-up-Trends setzen auf kräftige Farben und innovative Techniken", # noqa: E501

21 + "Cuidado de la piel orgánico para piel sensible con aloe vera y manzanilla", # noqa: E501

22 + "Las nuevas tendencias de maquillaje se centran en colores vivos y técnicas innovadoras", # noqa: E501

23 "针对敏感肌专门设计的天然有机护肤产品",

24 "新的化妆趋势注重鲜艳的颜色和创新的技巧",

25 "敏感肌のために特別に設計された天然有機スキンケア製品",

tests/models/language/pooling/test_scoring.py

+22 -39

@@ -1,15 +1,11 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the scoring outputs of HF and vLLM models.

3 -

4 -Run `pytest tests/models/embedding/language/test_scoring.py`.

5 -"""

2 import math

4 import pytest

5 import torch

6 import torch.nn.functional as F

12 -MODELS = [

8 +CROSS_ENCODER_MODELS = [

9 "cross-encoder/ms-marco-MiniLM-L-6-v2", # Bert

10 "BAAI/bge-reranker-v2-m3", # Roberta

11 ]

@@ -28,21 +24,21 @@ TEXTS_2 = [

24 "The capital of Germany is Berlin.",

25 ]

27 +DTYPE = "half"

28 +

32 -@pytest.fixture(scope="module", params=MODELS)

30 +@pytest.fixture(scope="module", params=CROSS_ENCODER_MODELS)

31 def model_name(request):

32 yield request.param

37 -@pytest.mark.parametrize("dtype", ["half"])

38 -def test_llm_1_to_1(vllm_runner, hf_runner, model_name, dtype: str):

39 -

35 +def test_cross_encoder_1_to_1(vllm_runner, hf_runner, model_name):

36 text_pair = [TEXTS_1[0], TEXTS_2[0]]

42 - with hf_runner(model_name, dtype=dtype, is_cross_encoder=True) as hf_model:

38 + with hf_runner(model_name, dtype=DTYPE, is_cross_encoder=True) as hf_model:

39 hf_outputs = hf_model.predict([text_pair]).tolist()

45 - with vllm_runner(model_name, task="score", dtype=dtype,

41 + with vllm_runner(model_name, task="score", dtype=DTYPE,

42 max_model_len=None) as vllm_model:

43 vllm_outputs = vllm_model.score(text_pair[0], text_pair[1])

@@ -52,18 +48,16 @@ def test_llm_1_to_1(vllm_runner, hf_runner, model_name, dtype: str):

48 assert math.isclose(hf_outputs[0], vllm_outputs[0], rel_tol=0.01)

55 -@pytest.mark.parametrize("dtype", ["half"])

56 -def test_llm_1_to_N(vllm_runner, hf_runner, model_name, dtype: str):

57 -

51 +def test_cross_encoder_1_to_N(vllm_runner, hf_runner, model_name):

52 text_pairs = [

53 [TEXTS_1[0], TEXTS_2[0]],

54 [TEXTS_1[0], TEXTS_2[1]],

55 ]

63 - with hf_runner(model_name, dtype=dtype, is_cross_encoder=True) as hf_model:

57 + with hf_runner(model_name, dtype=DTYPE, is_cross_encoder=True) as hf_model:

58 hf_outputs = hf_model.predict(text_pairs).tolist()

66 - with vllm_runner(model_name, task="score", dtype=dtype,

60 + with vllm_runner(model_name, task="score", dtype=DTYPE,

61 max_model_len=None) as vllm_model:

62 vllm_outputs = vllm_model.score(TEXTS_1[0], TEXTS_2)

@@ -74,18 +68,16 @@ def test_llm_1_to_N(vllm_runner, hf_runner, model_name, dtype: str):

68 assert math.isclose(hf_outputs[1], vllm_outputs[1], rel_tol=0.01)

77 -@pytest.mark.parametrize("dtype", ["half"])

78 -def test_llm_N_to_N(vllm_runner, hf_runner, model_name, dtype: str):

79 -

71 +def test_cross_encoder_N_to_N(vllm_runner, hf_runner, model_name):

72 text_pairs = [

73 [TEXTS_1[0], TEXTS_2[0]],

74 [TEXTS_1[1], TEXTS_2[1]],

75 ]

85 - with hf_runner(model_name, dtype=dtype, is_cross_encoder=True) as hf_model:

77 + with hf_runner(model_name, dtype=DTYPE, is_cross_encoder=True) as hf_model:

78 hf_outputs = hf_model.predict(text_pairs).tolist()

88 - with vllm_runner(model_name, task="score", dtype=dtype,

80 + with vllm_runner(model_name, task="score", dtype=DTYPE,

81 max_model_len=None) as vllm_model:

82 vllm_outputs = vllm_model.score(TEXTS_1, TEXTS_2)

@@ -101,13 +93,10 @@ def emb_model_name(request):

93 yield request.param

104 -@pytest.mark.parametrize("dtype", ["half"])

105 -def test_llm_1_to_1_embedding(vllm_runner, hf_runner, emb_model_name,

106 - dtype: str):

107 -

96 +def test_embedding_1_to_1(vllm_runner, hf_runner, emb_model_name):

97 text_pair = [TEXTS_1[0], TEXTS_2[0]]

110 - with hf_runner(emb_model_name, dtype=dtype,

99 + with hf_runner(emb_model_name, dtype=DTYPE,

100 is_sentence_transformer=True) as hf_model:

101 hf_embeddings = hf_model.encode(text_pair)

102 hf_outputs = [

@@ -116,7 +105,7 @@ def test_llm_1_to_1_embedding(vllm_runner, hf_runner, emb_model_name,

105

106 with vllm_runner(emb_model_name,

107 task="embed",

119 - dtype=dtype,

108 + dtype=DTYPE,

109 max_model_len=None) as vllm_model:

110 vllm_outputs = vllm_model.score(text_pair[0], text_pair[1])

111

@@ -126,16 +115,13 @@ def test_llm_1_to_1_embedding(vllm_runner, hf_runner, emb_model_name,

115 assert math.isclose(hf_outputs[0], vllm_outputs[0], rel_tol=0.01)

116

117

129 -@pytest.mark.parametrize("dtype", ["half"])

130 -def test_llm_1_to_N_embedding(vllm_runner, hf_runner, emb_model_name,

131 - dtype: str):

132 -

118 +def test_embedding_1_to_N(vllm_runner, hf_runner, emb_model_name):

119 text_pairs = [

120 [TEXTS_1[0], TEXTS_2[0]],

121 [TEXTS_1[0], TEXTS_2[1]],

122 ]

123

138 - with hf_runner(emb_model_name, dtype=dtype,

124 + with hf_runner(emb_model_name, dtype=DTYPE,

125 is_sentence_transformer=True) as hf_model:

126 hf_embeddings = [

127 hf_model.encode(text_pair) for text_pair in text_pairs

@@ -147,7 +133,7 @@ def test_llm_1_to_N_embedding(vllm_runner, hf_runner, emb_model_name,

133

134 with vllm_runner(emb_model_name,

135 task="embed",

150 - dtype=dtype,

136 + dtype=DTYPE,

137 max_model_len=None) as vllm_model:

138 vllm_outputs = vllm_model.score(TEXTS_1[0], TEXTS_2)

139

@@ -158,16 +144,13 @@ def test_llm_1_to_N_embedding(vllm_runner, hf_runner, emb_model_name,

144 assert math.isclose(hf_outputs[1], vllm_outputs[1], rel_tol=0.01)

145

146

161 -@pytest.mark.parametrize("dtype", ["half"])

162 -def test_llm_N_to_N_embedding(vllm_runner, hf_runner, emb_model_name,

163 - dtype: str):

164 -

147 +def test_embedding_N_to_N(vllm_runner, hf_runner, emb_model_name):

148 text_pairs = [

149 [TEXTS_1[0], TEXTS_2[0]],

150 [TEXTS_1[1], TEXTS_2[1]],

151 ]

152

170 - with hf_runner(emb_model_name, dtype=dtype,

153 + with hf_runner(emb_model_name, dtype=DTYPE,

154 is_sentence_transformer=True) as hf_model:

155 hf_embeddings = [

156 hf_model.encode(text_pair) for text_pair in text_pairs

@@ -179,7 +162,7 @@ def test_llm_N_to_N_embedding(vllm_runner, hf_runner, emb_model_name,

162

163 with vllm_runner(emb_model_name,

164 task="embed",

182 - dtype=dtype,

165 + dtype=DTYPE,

166 max_model_len=None) as vllm_model:

167 vllm_outputs = vllm_model.score(TEXTS_1, TEXTS_2)

168

tests/models/language/pooling/test_snowflake_arctic_embed.py

-4

@@ -1,8 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the embedding outputs of HF and vLLM models.

3 -

4 -Run `pytest tests/models/embedding/language/test_snowflake_arctic_embed.py`.

5 -"""

2 import pytest

4 from ...utils import EmbedModelInfo, check_embeddings_close

tests/models/language/pooling/test_truncation_control.py

+12 -12

@@ -5,18 +5,18 @@ MODEL_NAME = "sentence-transformers/all-MiniLM-L12-v2"

5 max_model_len = 128

7 input_str = """Immerse yourself in the enchanting chronicle of calculus, a

8 - mathematical domain that has radically transformed our comprehension of

9 - change and motion. Despite its roots in ancient civilizations, the

10 - formal birth of calculus predominantly occurred in the 17th century,

11 - primarily under the influential guidance of Sir Isaac Newton and Gottfried

12 - Wilhelm Leibniz. The earliest traces of calculus concepts are found in

13 - ancient Greek mathematics,most notably in the works of Eudoxus and

14 - Archimedes, around 300 BCE. They utilized the 'method of exhaustion'—a

15 - technique for computing areas and volumes through the use of finite sums.

16 - This methodology laid crucial foundational work for integral calculus.

17 - In the 17th century, both Newton and Leibniz independently pioneered

18 - calculus, each contributing unique perspectives that would shape this new

19 - field."""

8 +mathematical domain that has radically transformed our comprehension of

9 +change and motion. Despite its roots in ancient civilizations, the

10 +formal birth of calculus predominantly occurred in the 17th century,

11 +primarily under the influential guidance of Sir Isaac Newton and Gottfried

12 +Wilhelm Leibniz. The earliest traces of calculus concepts are found in

13 +ancient Greek mathematics,most notably in the works of Eudoxus and

14 +Archimedes, around 300 BCE. They utilized the 'method of exhaustion'—a

15 +technique for computing areas and volumes through the use of finite sums.

16 +This methodology laid crucial foundational work for integral calculus.

17 +In the 17th century, both Newton and Leibniz independently pioneered

18 +calculus, each contributing unique perspectives that would shape this new

19 +field."""

22 def test_smaller_truncation_size(vllm_runner,

tests/models/multimodal/generation/test_pixtral.py

-4

@@ -1,8 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the outputs of HF and vLLM for Mistral models using greedy sampling.

3 -

4 -Run `pytest tests/models/test_mistral.py`.

5 -"""

2 import json

3 from dataclasses import asdict

4 from typing import TYPE_CHECKING, Any, Optional

tests/models/multimodal/generation/test_whisper.py

+2 -2

@@ -119,10 +119,10 @@ def run_test(

119 assert output.outputs[0].text == expected

120

121

122 -@create_new_process_for_each_test("spawn")

122 @pytest.mark.core_model

123 @pytest.mark.parametrize(

124 "model", ["openai/whisper-small", "openai/whisper-large-v3-turbo"])

125 +@create_new_process_for_each_test()

126 def test_models(vllm_runner, model) -> None:

127 run_test(

128 vllm_runner,

@@ -131,11 +131,11 @@ def test_models(vllm_runner, model) -> None:

131 )

132

133

134 -@create_new_process_for_each_test("spawn")

134 @multi_gpu_test(num_gpus=2)

135 @pytest.mark.core_model

136 @pytest.mark.parametrize("model", ["openai/whisper-large-v3-turbo"])

137 @pytest.mark.parametrize("distributed_executor_backend", ["ray", "mp"])

138 +@create_new_process_for_each_test()

139 def test_models_distributed(

140 vllm_runner,

141 model,

tests/models/quantization/test_aqlm.py

-5

@@ -1,9 +1,4 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compare the outputs of a AQLM model between vLLM and HF Transformers

3 -

4 -Run `pytest tests/models/test_aqlm.py`.

5 -"""

6 -

2 import pytest

4 from tests.quantization.utils import is_quant_method_supported

tests/models/quantization/test_bitblas.py

-2

@@ -8,8 +8,6 @@ bitblas/GPTQ models are in the top 3 selections of each other.

8 Note: bitblas internally uses locks to synchronize the threads. This can

9 result in very slight nondeterminism for bitblas. As a result, we re-run the

10 test up to 3 times to see if we pass.

11 -

12 -Run `pytest tests/models/test_bitblas.py`.

11 """

12 from dataclasses import dataclass

tests/models/quantization/test_gptq_bitblas.py

-2

@@ -8,8 +8,6 @@ bitblas/GPTQ models are in the top 3 selections of each other.

8 Note: bitblas internally uses locks to synchronize the threads. This can

9 result in very slight nondeterminism for bitblas. As a result, we re-run the

10 test up to 3 times to see if we pass.

11 -

12 -Run `pytest tests/models/test_bitblas.py`.

11 """

12 from dataclasses import dataclass

tests/models/quantization/test_gptq_marlin.py

+2 -3

@@ -1,13 +1,12 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Compares the outputs of gptq vs gptq_marlin

2 +"""Compares the outputs of gptq vs gptq_marlin.

3 +

4 Note: GPTQ and Marlin do not have bitwise correctness.

5 As a result, in this test, we just confirm that the top selected tokens of the

6 Marlin/GPTQ models are in the top 5 selections of each other.

7 Note: Marlin internally uses locks to synchronize the threads. This can

8 result in very slight nondeterminism for Marlin. As a result, we re-run the test

9 up to 3 times to see if we pass.

9 -

10 -Run `pytest tests/models/test_gptq_marlin.py`.

10 """

11 import os

tests/models/quantization/test_gptq_marlin_24.py

-2

@@ -4,8 +4,6 @@

4 Note: GPTQ and Marlin_24 do not have bitwise correctness.

5 As a result, in this test, we just confirm that the top selected tokens of the

6 Marlin/GPTQ models are in the top 3 selections of each other.

7 -

8 -Run `pytest tests/models/test_marlin_24.py`.

7 """

8 from dataclasses import dataclass

tests/models/test_transformers.py

+1 -4

@@ -1,8 +1,5 @@

1 # SPDX-License-Identifier: Apache-2.0

2 -"""Test the functionality of the Transformers backend.

3 -

4 -Run `pytest tests/models/test_transformers.py`.

5 -"""

2 +"""Test the functionality of the Transformers backend."""

3 import pytest

5 from ..conftest import HfRunner, VllmRunner