vllm/tests/models/language/pooling/test_token_classification.py
Cyrus Leung 0f29dca988
[CI/Build] Fix model nightly tests (#26466)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
2025-10-08 23:44:16 -07:00

71 lines
2.4 KiB
Python

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
import torch
from transformers import AutoModelForTokenClassification
from tests.models.utils import softmax
@pytest.mark.parametrize("model", ["boltuix/NeuroBERT-NER"])
# The float32 is required for this tiny model to pass the test.
@pytest.mark.parametrize("dtype", ["float"])
@torch.inference_mode
def test_bert_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
) -> None:
with vllm_runner(model, max_model_len=None, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.encode(example_prompts)
with hf_runner(
model, dtype=dtype, auto_cls=AutoModelForTokenClassification
) as hf_model:
tokenizer = hf_model.tokenizer
hf_outputs = []
for prompt in example_prompts:
inputs = tokenizer([prompt], return_tensors="pt")
inputs = hf_model.wrap_device(inputs)
output = hf_model.model(**inputs)
hf_outputs.append(softmax(output.logits[0]))
# check logits difference
for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
hf_output = torch.tensor(hf_output).cpu().float()
vllm_output = torch.tensor(vllm_output).cpu().float()
assert torch.allclose(hf_output, vllm_output, 1e-2)
@pytest.mark.parametrize("model", ["disham993/electrical-ner-ModernBERT-base"])
@pytest.mark.parametrize("dtype", ["float"])
@torch.inference_mode
def test_modernbert_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
) -> None:
with vllm_runner(model, max_model_len=None, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.encode(example_prompts)
with hf_runner(
model, dtype=dtype, auto_cls=AutoModelForTokenClassification
) as hf_model:
tokenizer = hf_model.tokenizer
hf_outputs = []
for prompt in example_prompts:
inputs = tokenizer([prompt], return_tensors="pt")
inputs = hf_model.wrap_device(inputs)
output = hf_model.model(**inputs)
hf_outputs.append(softmax(output.logits[0]))
# check logits difference
for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
hf_output = torch.tensor(hf_output).cpu().float()
vllm_output = torch.tensor(vllm_output).cpu().float()
assert torch.allclose(hf_output, vllm_output, atol=1e-2)