[ci] Use env var to control whether to use S3 bucket in CI (#13634)

This commit is contained in:
Kevin H. Luu
2025-02-22 19:19:45 -08:00
committed by GitHub
parent 322d2a27d6
commit 2c5e637b57
30 changed files with 222 additions and 231 deletions
+27 -28
View File
@@ -8,21 +8,17 @@ import ray
from prometheus_client import REGISTRY
from vllm import EngineArgs, LLMEngine
from vllm.config import LoadFormat
from vllm.distributed import cleanup_dist_env_and_memory
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.engine.async_llm_engine import AsyncLLMEngine
from vllm.engine.metrics import RayPrometheusStatLogger
from vllm.sampling_params import SamplingParams
from ..conftest import MODEL_WEIGHTS_S3_BUCKET
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET
MODELS = [
"distilbert/distilgpt2",
]
RUNAI_STREAMER_LOAD_FORMAT = LoadFormat.RUNAI_STREAMER
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
@@ -146,9 +142,8 @@ def test_metric_set_tag_model_name(vllm_runner, model: str, dtype: str,
metrics_tag_content = stat_logger.labels["model_name"]
if served_model_name is None or served_model_name == []:
actual_model_name = f"{MODEL_WEIGHTS_S3_BUCKET}/{model}"
assert metrics_tag_content == actual_model_name, (
f"Metrics tag model_name is wrong! expect: {actual_model_name!r}\n"
assert metrics_tag_content == f"{MODEL_WEIGHTS_S3_BUCKET}/{model}", (
f"Metrics tag model_name is wrong! expect: {model!r}\n"
f"actual: {metrics_tag_content!r}")
else:
assert metrics_tag_content == served_model_name[0], (
@@ -174,10 +169,11 @@ async def test_async_engine_log_metrics_regression(
when disable_log_stats=False
(see: https://github.com/vllm-project/vllm/pull/4150#pullrequestreview-2008176678)
"""
engine_args = AsyncEngineArgs(model=model,
dtype=dtype,
disable_log_stats=disable_log_stats,
load_format=RUNAI_STREAMER_LOAD_FORMAT)
engine_args = AsyncEngineArgs(
model=model,
dtype=dtype,
disable_log_stats=disable_log_stats,
)
async_engine = AsyncLLMEngine.from_engine_args(engine_args)
for i, prompt in enumerate(example_prompts):
results = async_engine.generate(
@@ -189,7 +185,7 @@ async def test_async_engine_log_metrics_regression(
async for _ in results:
pass
assert_metrics(async_engine.engine, disable_log_stats,
assert_metrics(model, async_engine.engine, disable_log_stats,
len(example_prompts))
@@ -204,10 +200,11 @@ def test_engine_log_metrics_regression(
max_tokens: int,
disable_log_stats: bool,
) -> None:
engine_args = EngineArgs(model=model,
dtype=dtype,
disable_log_stats=disable_log_stats,
load_format=RUNAI_STREAMER_LOAD_FORMAT)
engine_args = EngineArgs(
model=model,
dtype=dtype,
disable_log_stats=disable_log_stats,
)
engine = LLMEngine.from_engine_args(engine_args)
for i, prompt in enumerate(example_prompts):
engine.add_request(
@@ -218,7 +215,8 @@ def test_engine_log_metrics_regression(
while engine.has_unfinished_requests():
engine.step()
assert_metrics(engine, disable_log_stats, len(example_prompts))
assert_metrics(f"{MODEL_WEIGHTS_S3_BUCKET}/{model}", engine,
disable_log_stats, len(example_prompts))
@pytest.mark.parametrize("model", MODELS)
@@ -285,14 +283,15 @@ def test_metric_spec_decode_interval(
) -> None:
k = 5
engine_args = EngineArgs(model=model,
dtype=dtype,
disable_log_stats=False,
gpu_memory_utilization=0.4,
speculative_model=model,
num_speculative_tokens=k,
enforce_eager=True,
load_format=RUNAI_STREAMER_LOAD_FORMAT)
engine_args = EngineArgs(
model=model,
dtype=dtype,
disable_log_stats=False,
gpu_memory_utilization=0.4,
speculative_model=model,
num_speculative_tokens=k,
enforce_eager=True,
)
engine = LLMEngine.from_engine_args(engine_args)
@@ -359,7 +358,7 @@ def test_metric_spec_decode_interval(
cleanup_dist_env_and_memory()
def assert_metrics(engine: LLMEngine, disable_log_stats: bool,
def assert_metrics(model: str, engine: LLMEngine, disable_log_stats: bool,
num_requests: int) -> None:
if disable_log_stats:
with pytest.raises(AttributeError):
@@ -370,7 +369,7 @@ def assert_metrics(engine: LLMEngine, disable_log_stats: bool,
# Ensure the count bucket of request-level histogram metrics matches
# the number of requests as a simple sanity check to ensure metrics are
# generated
labels = {'model_name': engine.model_config.model}
labels = {'model_name': model}
request_histogram_metrics = [
"vllm:e2e_request_latency_seconds",
"vllm:request_prompt_tokens",