[CI/Build] Reorganize models tests (#7820)

This commit is contained in:
Cyrus Leung
2024-09-13 10:20:06 -07:00
committed by GitHub
parent 0a4806f0a9
commit a84e598e21
55 changed files with 415 additions and 498 deletions
@@ -15,12 +15,15 @@ from vllm.utils import is_hip
from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
from ..models.utils import check_outputs_equal
from ..utils import multi_gpu_test
MODELS = [
"facebook/opt-125m",
"meta-llama/Llama-2-7b-hf",
]
TARGET_TEST_SUITE = os.environ.get("TARGET_TEST_SUITE", "L4")
def test_vllm_gc_ed():
"""Verify vllm instance is GC'ed when it is deleted"""
@@ -70,6 +73,65 @@ def test_models(
)
@multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize(
"model, distributed_executor_backend, attention_backend, "
"test_suite", [
("facebook/opt-125m", "ray", "", "L4"),
("facebook/opt-125m", "mp", "", "L4"),
("meta-llama/Llama-2-7b-hf", "ray", "", "L4"),
("meta-llama/Llama-2-7b-hf", "mp", "", "L4"),
("facebook/opt-125m", "ray", "", "A100"),
("facebook/opt-125m", "mp", "", "A100"),
("facebook/opt-125m", "mp", "FLASHINFER", "A100"),
("meta-llama/Meta-Llama-3-8B", "ray", "FLASHINFER", "A100"),
])
def test_models_distributed(
hf_runner,
vllm_runner,
example_prompts,
model: str,
distributed_executor_backend: str,
attention_backend: str,
test_suite: str,
) -> None:
if test_suite != TARGET_TEST_SUITE:
pytest.skip(f"Skip test for {test_suite}")
if model == "meta-llama/Llama-2-7b-hf" and distributed_executor_backend == "ray" and attention_backend == "" and test_suite == "L4": # noqa
# test ray adag
os.environ['VLLM_USE_RAY_SPMD_WORKER'] = "1"
os.environ['VLLM_USE_RAY_COMPILED_DAG'] = "1"
if attention_backend:
os.environ["VLLM_ATTENTION_BACKEND"] = attention_backend
dtype = "half"
max_tokens = 5
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
with vllm_runner(model,
dtype=dtype,
tensor_parallel_size=2,
distributed_executor_backend=distributed_executor_backend
) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_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",
)
def test_model_with_failure(vllm_runner) -> None:
try:
with patch("vllm.model_executor.models.opt.OPTForCausalLM.forward",
@@ -6,11 +6,13 @@ prefill requests are chunked.
Run `pytest tests/models/test_chunked_prefill.py`.
"""
import os
from contextlib import nullcontext
import pytest
from ..models.utils import check_logprobs_close, check_outputs_equal
from ..utils import multi_gpu_test
MODELS = [
"facebook/opt-125m",
@@ -66,6 +68,59 @@ def test_models(
)
@multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize("distributed_executor_backend", ["ray", "mp"])
@pytest.mark.parametrize("model", MODELS)
def test_models_distributed(
hf_runner,
vllm_runner,
example_prompts,
model: str,
distributed_executor_backend: str,
) -> None:
if (model == "meta-llama/Llama-2-7b-hf"
and distributed_executor_backend == "ray"):
# test ray adag
os.environ['VLLM_USE_RAY_SPMD_WORKER'] = "1"
os.environ['VLLM_USE_RAY_COMPILED_DAG'] = "1"
dtype = "half"
max_tokens = 5
chunked_prefill_token_size = 16
# Add a chunked prefill config.
max_num_seqs = min(chunked_prefill_token_size, 256)
assert chunked_prefill_token_size != -1
enable_chunked_prefill = True
max_num_batched_tokens = chunked_prefill_token_size
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
with vllm_runner(
model,
dtype=dtype,
tensor_parallel_size=2,
max_num_seqs=max_num_seqs,
enable_chunked_prefill=enable_chunked_prefill,
max_num_batched_tokens=max_num_batched_tokens,
distributed_executor_backend=distributed_executor_backend,
) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_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(
"kv_cache_dtype,model",
[("fp8_e4m3",
+7 -4
View File
@@ -19,10 +19,13 @@ MODELS = [
"facebook/opt-125m",
]
assert ENABLE_ARTIFICIAL_PREEMPT is True, (
"Use an env var VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1. "
"`VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest "
"tests/basic_correctness/test_preemption.py`")
@pytest.fixture(scope="module", autouse=True)
def check_settings():
assert ENABLE_ARTIFICIAL_PREEMPT is True, (
"Use an env var VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1. "
"`VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest "
"tests/basic_correctness/test_preemption.py`")
@pytest.fixture