biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI/Build] Split up decoder-only LM tests (#9488)

Browse Source 
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
This commit is contained in:
Cyrus Leung
2024-10-21 12:27:50 +08:00
committed by GitHub
parent 855e0e6f97
commit 696b01af8f
3 changed files with 18 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/models/decoder_only/language/test_big_models.py
View File

@@ -21,10 +21,14 @@ MODELS = [
]
if not current_platform.is_cpu():
# MiniCPM requires fused_moe which is not supported by CPU
MODELS.append("openbmb/MiniCPM3-4B")
MODELS += [
# fused_moe which not supported on CPU
"openbmb/MiniCPM3-4B",
# Head size isn't supported on CPU
"h2oai/h2o-danube3-4b-base",
]
#TODO: remove this after CPU float16 support ready
# TODO: remove this after CPU float16 support ready
target_dtype = "float" if current_platform.is_cpu() else "half"

52
tests/models/decoder_only/language/test_danube3_4b.py
View File

@@ -1,52 +0,0 @@
"""Compare the outputs of HF and vLLM when using greedy sampling.
This tests danube3 separately because its head size isn't supported on CPU yet.
Run `pytest tests/models/test_danube3_4b.py`.
"""
import pytest
from ...utils import check_outputs_equal
MODELS = ["h2oai/h2o-danube3-4b-base"]
target_dtype = "half"
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [32])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
model: str,
dtype: str,
max_tokens: int,
) -> None:
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", [target_dtype])
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)