[Model] Support Qwen2 embeddings and use tags to select model tests (#10184)
This commit is contained in:
@@ -33,6 +33,10 @@ def test_models(
|
||||
|
||||
with vllm_runner(model, dtype=dtype) as vllm_model:
|
||||
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
|
||||
# This test is for verifying whether the model's extra_repr
|
||||
# can be printed correctly.
|
||||
print(vllm_model.model.llm_engine.model_executor.driver_worker.
|
||||
model_runner.model)
|
||||
|
||||
for i in range(len(example_prompts)):
|
||||
hf_output_ids, hf_output_str = hf_outputs[i]
|
||||
@@ -293,17 +297,3 @@ def test_jamba_distributed_produces_identical_generation(
|
||||
name_0="vllm_tp_1",
|
||||
name_1="vllm_tp_2",
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize("dtype", ["float"])
|
||||
def test_model_print(
|
||||
vllm_runner,
|
||||
model: str,
|
||||
dtype: str,
|
||||
) -> None:
|
||||
with vllm_runner(model, dtype=dtype) as vllm_model:
|
||||
# This test is for verifying whether the model's extra_repr
|
||||
# can be printed correctly.
|
||||
print(vllm_model.model.llm_engine.model_executor.driver_worker.
|
||||
model_runner.model)
|
||||
|
||||
@@ -51,6 +51,10 @@ def test_models(
|
||||
|
||||
with vllm_runner(model, dtype=dtype) as vllm_model:
|
||||
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
|
||||
# This test is for verifying whether the model's extra_repr
|
||||
# can be printed correctly.
|
||||
print(vllm_model.model.llm_engine.model_executor.driver_worker.
|
||||
model_runner.model)
|
||||
|
||||
for i in range(len(example_prompts)):
|
||||
hf_output_ids, hf_output_str = hf_outputs[i]
|
||||
@@ -279,17 +283,3 @@ def test_state_cleanup(
|
||||
except ValueError:
|
||||
pytest.fail("Mamba inner state wasn't cleaned up between states, "
|
||||
"could be related to finished_requests_ids")
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize("dtype", ["float"])
|
||||
def test_model_print(
|
||||
vllm_runner,
|
||||
model: str,
|
||||
dtype: str,
|
||||
) -> None:
|
||||
with vllm_runner(model, dtype=dtype) as vllm_model:
|
||||
# This test is for verifying whether the model's extra_repr
|
||||
# can be printed correctly.
|
||||
print(vllm_model.model.llm_engine.model_executor.driver_worker.
|
||||
model_runner.model)
|
||||
|
||||
@@ -4,37 +4,52 @@ Run `pytest tests/models/test_models.py`.
|
||||
"""
|
||||
import pytest
|
||||
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
from ...utils import check_logprobs_close
|
||||
|
||||
MODELS = [
|
||||
"facebook/opt-125m", # opt
|
||||
"openai-community/gpt2", # gpt2
|
||||
# "Milos/slovak-gpt-j-405M", # gptj
|
||||
# "bigcode/tiny_starcoder_py", # gpt_bigcode
|
||||
# "EleutherAI/pythia-70m", # gpt_neox
|
||||
"bigscience/bloom-560m", # bloom - testing alibi slopes
|
||||
"microsoft/phi-2", # phi
|
||||
# "stabilityai/stablelm-3b-4e1t", # stablelm
|
||||
# "bigcode/starcoder2-3b", # starcoder2
|
||||
"google/gemma-1.1-2b-it", # gemma
|
||||
"Qwen/Qwen2.5-0.5B-Instruct", # qwen2
|
||||
"meta-llama/Llama-3.2-1B-Instruct", # llama
|
||||
]
|
||||
|
||||
if not current_platform.is_cpu():
|
||||
MODELS += [
|
||||
# fused_moe which not supported on CPU
|
||||
"openbmb/MiniCPM3-4B",
|
||||
]
|
||||
|
||||
target_dtype = "half"
|
||||
|
||||
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize("dtype", [target_dtype])
|
||||
@pytest.mark.parametrize(
|
||||
"model",
|
||||
[
|
||||
pytest.param(
|
||||
"bigscience/bloom-560m", # bloom - testing alibi slopes
|
||||
marks=[pytest.mark.core_model, pytest.mark.cpu_model],
|
||||
),
|
||||
pytest.param(
|
||||
"openai-community/gpt2", # gpt2
|
||||
marks=[pytest.mark.core_model, pytest.mark.cpu_model],
|
||||
),
|
||||
pytest.param("Milos/slovak-gpt-j-405M"), # gptj
|
||||
pytest.param("bigcode/tiny_starcoder_py"), # gpt_bigcode
|
||||
pytest.param("EleutherAI/pythia-70m"), # gpt_neox
|
||||
pytest.param(
|
||||
"google/gemma-1.1-2b-it", # gemma
|
||||
marks=[pytest.mark.core_model, pytest.mark.cpu_model],
|
||||
),
|
||||
pytest.param(
|
||||
"meta-llama/Llama-3.2-1B-Instruct", # llama
|
||||
marks=[pytest.mark.core_model, pytest.mark.cpu_model],
|
||||
),
|
||||
pytest.param(
|
||||
"openbmb/MiniCPM3-4B",
|
||||
# fused_moe not supported on CPU
|
||||
marks=[pytest.mark.core_model],
|
||||
),
|
||||
pytest.param(
|
||||
"facebook/opt-125m", # opt
|
||||
marks=[pytest.mark.core_model, pytest.mark.cpu_model],
|
||||
),
|
||||
pytest.param(
|
||||
"microsoft/phi-2", # phi
|
||||
marks=[pytest.mark.core_model],
|
||||
),
|
||||
pytest.param(
|
||||
"Qwen/Qwen2.5-0.5B-Instruct", # qwen2
|
||||
marks=[pytest.mark.core_model],
|
||||
),
|
||||
pytest.param("stabilityai/stablelm-3b-4e1t"), # stablelm
|
||||
pytest.param("bigcode/starcoder2-3b"), # starcoder2
|
||||
])
|
||||
@pytest.mark.parametrize("dtype", ["half"])
|
||||
@pytest.mark.parametrize("max_tokens", [32])
|
||||
@pytest.mark.parametrize("num_logprobs", [5])
|
||||
def test_models(
|
||||
|
||||
@@ -9,10 +9,14 @@ import pytest
|
||||
import torch
|
||||
from transformers import AutoModelForSequenceClassification
|
||||
|
||||
CLASSIFICATION_MODELS = ["jason9693/Qwen2.5-1.5B-apeach"]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model", CLASSIFICATION_MODELS)
|
||||
@pytest.mark.parametrize(
|
||||
"model",
|
||||
[
|
||||
pytest.param("jason9693/Qwen2.5-1.5B-apeach",
|
||||
marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
|
||||
],
|
||||
)
|
||||
@pytest.mark.parametrize("dtype", ["float"])
|
||||
def test_classification_models(
|
||||
hf_runner,
|
||||
@@ -23,31 +27,19 @@ def test_classification_models(
|
||||
) -> None:
|
||||
with vllm_runner(model, dtype=dtype) as vllm_model:
|
||||
vllm_outputs = vllm_model.classify(example_prompts)
|
||||
# This test is for verifying whether the model's extra_repr
|
||||
# can be printed correctly.
|
||||
print(vllm_model.model.llm_engine.model_executor.driver_worker.
|
||||
model_runner.model)
|
||||
|
||||
with hf_runner(model,
|
||||
dtype=dtype,
|
||||
auto_cls=AutoModelForSequenceClassification) as hf_model:
|
||||
hf_outputs = hf_model.classify(example_prompts)
|
||||
|
||||
print(hf_outputs, vllm_outputs)
|
||||
|
||||
# check logits difference
|
||||
for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
|
||||
hf_output = torch.tensor(hf_output)
|
||||
vllm_output = torch.tensor(vllm_output)
|
||||
|
||||
assert torch.allclose(hf_output, vllm_output, 1e-3)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model", CLASSIFICATION_MODELS)
|
||||
@pytest.mark.parametrize("dtype", ["float"])
|
||||
def test_classification_model_print(
|
||||
vllm_runner,
|
||||
model: str,
|
||||
dtype: str,
|
||||
) -> None:
|
||||
with vllm_runner(model, dtype=dtype) as vllm_model:
|
||||
# This test is for verifying whether the model's extra_repr
|
||||
# can be printed correctly.
|
||||
print(vllm_model.model.llm_engine.model_executor.driver_worker.
|
||||
model_runner.model)
|
||||
|
||||
@@ -4,25 +4,25 @@ Run `pytest tests/models/embedding/language/test_embedding.py`.
|
||||
"""
|
||||
import pytest
|
||||
|
||||
from vllm.utils import current_platform
|
||||
|
||||
from ..utils import check_embeddings_close
|
||||
|
||||
# Model, Guard
|
||||
MODELS = [
|
||||
"intfloat/e5-mistral-7b-instruct",
|
||||
"BAAI/bge-base-en-v1.5",
|
||||
"BAAI/bge-multilingual-gemma2",
|
||||
"intfloat/multilingual-e5-large",
|
||||
]
|
||||
|
||||
ENCODER_ONLY = [
|
||||
"BAAI/bge-base-en-v1.5",
|
||||
"intfloat/multilingual-e5-large",
|
||||
]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize(
|
||||
"model",
|
||||
[
|
||||
# [Encoder-only]
|
||||
pytest.param("BAAI/bge-base-en-v1.5",
|
||||
marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
|
||||
pytest.param("intfloat/multilingual-e5-large"),
|
||||
# [Encoder-decoder]
|
||||
pytest.param("intfloat/e5-mistral-7b-instruct",
|
||||
marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
|
||||
pytest.param("BAAI/bge-multilingual-gemma2",
|
||||
marks=[pytest.mark.core_model]),
|
||||
pytest.param("ssmits/Qwen2-7B-Instruct-embed-base"),
|
||||
pytest.param("Alibaba-NLP/gte-Qwen2-1.5B-instruct"),
|
||||
],
|
||||
)
|
||||
@pytest.mark.parametrize("dtype", ["half"])
|
||||
def test_models(
|
||||
hf_runner,
|
||||
@@ -31,9 +31,6 @@ def test_models(
|
||||
model,
|
||||
dtype: str,
|
||||
) -> None:
|
||||
if model not in ENCODER_ONLY and current_platform.is_cpu():
|
||||
pytest.skip("Skip large embedding models test on CPU.")
|
||||
|
||||
# The example_prompts has ending "\n", for example:
|
||||
# "Write a short story about a robot that dreams for the first time.\n"
|
||||
# sentence_transformers will strip the input texts, see:
|
||||
@@ -46,8 +43,13 @@ def test_models(
|
||||
is_sentence_transformer=True) as hf_model:
|
||||
hf_outputs = hf_model.encode(example_prompts)
|
||||
|
||||
with vllm_runner(model, dtype=dtype, max_model_len=None) as vllm_model:
|
||||
with vllm_runner(model, task="embedding", dtype=dtype,
|
||||
max_model_len=None) as vllm_model:
|
||||
vllm_outputs = vllm_model.encode(example_prompts)
|
||||
# This test is for verifying whether the model's extra_repr
|
||||
# can be printed correctly.
|
||||
print(vllm_model.model.llm_engine.model_executor.driver_worker.
|
||||
model_runner.model)
|
||||
|
||||
check_embeddings_close(
|
||||
embeddings_0_lst=hf_outputs,
|
||||
|
||||
@@ -88,6 +88,7 @@ def _run_test(
|
||||
|
||||
@pytest.mark.skipif(transformers.__version__.startswith("4.46"),
|
||||
reason="Model broken with changes in transformers 4.46")
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize("dtype", ["half"])
|
||||
def test_models_text(
|
||||
@@ -112,6 +113,7 @@ def test_models_text(
|
||||
|
||||
|
||||
@large_gpu_test(min_gb=48)
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize("dtype", ["half"])
|
||||
def test_models_image(
|
||||
|
||||
@@ -74,6 +74,7 @@ def _run_test(
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize("dtype", ["half"])
|
||||
def test_models_text(
|
||||
@@ -98,6 +99,7 @@ def test_models_text(
|
||||
|
||||
|
||||
@large_gpu_test(min_gb=48)
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize("dtype", ["half"])
|
||||
def test_models_image(
|
||||
|
||||
@@ -14,8 +14,6 @@ from ....conftest import (DecoderPromptType, ExplicitEncoderDecoderPrompt,
|
||||
from ....utils import multi_gpu_test
|
||||
from ...utils import check_logprobs_close
|
||||
|
||||
MODELS = ["facebook/bart-base", "facebook/bart-large-cnn"]
|
||||
|
||||
|
||||
def vllm_to_hf_output(
|
||||
vllm_output: Tuple[List[int], str, Optional[SampleLogprobs]],
|
||||
@@ -170,7 +168,14 @@ def run_test(
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize(
|
||||
"model",
|
||||
[
|
||||
pytest.param("facebook/bart-base",
|
||||
marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
|
||||
pytest.param("facebook/bart-large-cnn"),
|
||||
],
|
||||
)
|
||||
@pytest.mark.parametrize("dtype", ["float", "bfloat16"])
|
||||
@pytest.mark.parametrize("max_tokens", [64])
|
||||
@pytest.mark.parametrize("num_logprobs", [5])
|
||||
|
||||
@@ -233,6 +233,7 @@ def clear_cache():
|
||||
|
||||
|
||||
@large_gpu_test(min_gb=48)
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.parametrize("model", models)
|
||||
@pytest.mark.parametrize(
|
||||
"sizes",
|
||||
@@ -278,6 +279,7 @@ def test_models_single_leading_image(hf_runner, vllm_runner, image_assets,
|
||||
|
||||
|
||||
@large_gpu_test(min_gb=48)
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.parametrize("model", models)
|
||||
@pytest.mark.parametrize("dtype", ["bfloat16"])
|
||||
@pytest.mark.parametrize("max_tokens", [128])
|
||||
@@ -326,6 +328,7 @@ def test_models_multi_leading_images(hf_runner, vllm_runner, image_assets,
|
||||
|
||||
|
||||
@large_gpu_test(min_gb=48)
|
||||
@pytest.mark.core_model
|
||||
@pytest.mark.parametrize("model", models)
|
||||
@pytest.mark.parametrize("dtype", ["bfloat16"])
|
||||
@pytest.mark.parametrize("max_tokens", [128])
|
||||
|
||||
@@ -129,9 +129,13 @@ _EMBEDDING_EXAMPLE_MODELS = {
|
||||
# [Text-only]
|
||||
"BertModel": _HfExamplesInfo("BAAI/bge-base-en-v1.5"),
|
||||
"Gemma2Model": _HfExamplesInfo("BAAI/bge-multilingual-gemma2"),
|
||||
"LlamaModel": _HfExamplesInfo("llama", is_available_online=False),
|
||||
"MistralModel": _HfExamplesInfo("intfloat/e5-mistral-7b-instruct"),
|
||||
"Qwen2Model": _HfExamplesInfo("ssmits/Qwen2-7B-Instruct-embed-base"),
|
||||
"Qwen2ForRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-RM-72B"),
|
||||
"Qwen2ForSequenceClassification": _HfExamplesInfo("jason9693/Qwen2.5-1.5B-apeach"), # noqa: E501
|
||||
"RobertaModel": _HfExamplesInfo("sentence-transformers/stsb-roberta-base-v2"), # noqa: E501
|
||||
"XLMRobertaModel": _HfExamplesInfo("intfloat/multilingual-e5-large"),
|
||||
# [Multimodal]
|
||||
"LlavaNextForConditionalGeneration": _HfExamplesInfo("royokong/e5-v"),
|
||||
"Phi3VForCausalLM": _HfExamplesInfo("TIGER-Lab/VLM2Vec-Full",
|
||||
|
||||
@@ -77,8 +77,8 @@ def test_registry_is_pp(model_arch, is_pp, init_cuda):
|
||||
|
||||
|
||||
def test_hf_registry_coverage():
|
||||
untested_archs = (HF_EXAMPLE_MODELS.get_supported_archs() -
|
||||
set(ModelRegistry.get_supported_archs()))
|
||||
untested_archs = (ModelRegistry.get_supported_archs() -
|
||||
HF_EXAMPLE_MODELS.get_supported_archs())
|
||||
|
||||
assert not untested_archs, (
|
||||
"Please add the following architectures to "
|
||||
|
||||
Reference in New Issue
Block a user