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

[1/n][CI] Load models in CI from S3 instead of HF (#13205)

Browse Source 
Signed-off-by: <>
Co-authored-by: EC2 Default User <ec2-user@ip-172-31-20-117.us-west-2.compute.internal>
This commit is contained in:
Kevin H. Luu
2025-02-18 23:34:59 -08:00
committed by GitHub
parent fd84857f64
commit d5d214ac7f
43 changed files with 225 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/engine/test_computed_prefix_blocks.py
View File

@ -2,12 +2,15 @@
import pytest
from vllm.config import LoadFormat
from vllm.engine.arg_utils import EngineArgs
from vllm.engine.llm_engine import LLMEngine
from vllm.sampling_params import SamplingParams
from ..conftest import MODEL_WEIGHTS_S3_BUCKET
@pytest.mark.parametrize("model", ["facebook/opt-125m"])
@pytest.mark.parametrize("model", [f"{MODEL_WEIGHTS_S3_BUCKET}/distilgpt2"])
@pytest.mark.parametrize("block_size", [16])
def test_computed_prefix_blocks(model: str, block_size: int):
# This test checks if we are able to run the engine to completion
@ -24,6 +27,7 @@ def test_computed_prefix_blocks(model: str, block_size: int):
"decoration.")
engine_args = EngineArgs(model=model,
load_format=LoadFormat.RUNAI_STREAMER,
block_size=block_size,
enable_prefix_caching=True)

7
tests/engine/test_detokenization.py
View File

@ -2,11 +2,14 @@
import pytest
from vllm.config import LoadFormat
from vllm.entrypoints.llm import LLM
from vllm.sampling_params import SamplingParams
from ..conftest import MODEL_WEIGHTS_S3_BUCKET
@pytest.mark.parametrize("model", ["facebook/opt-125m"])
@pytest.mark.parametrize("model", [f"{MODEL_WEIGHTS_S3_BUCKET}/distilgpt2"])
def test_computed_prefix_blocks(model: str):
# This test checks if the engine generates completions both with and
# without optional detokenization, that detokenization includes text
@ -17,7 +20,7 @@ def test_computed_prefix_blocks(model: str):
"paper clips? Is there an easy to follow video tutorial available "
"online for free?")
llm = LLM(model=model)
llm = LLM(model=model, load_format=LoadFormat.RUNAI_STREAMER)
sampling_params = SamplingParams(max_tokens=10,
temperature=0.0,
detokenize=False)

17
tests/engine/test_executor.py
View File

@ -6,12 +6,17 @@ from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import pytest
from vllm.config import LoadFormat
from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
from vllm.engine.async_llm_engine import AsyncLLMEngine
from vllm.engine.llm_engine import LLMEngine
from vllm.executor.uniproc_executor import UniProcExecutor
from vllm.sampling_params import SamplingParams
from ..conftest import MODEL_WEIGHTS_S3_BUCKET
RUNAI_STREAMER_LOAD_FORMAT = LoadFormat.RUNAI_STREAMER
class Mock:
...
@ -33,10 +38,11 @@ class CustomUniExecutor(UniProcExecutor):
CustomUniExecutorAsync = CustomUniExecutor
@pytest.mark.parametrize("model", ["facebook/opt-125m"])
@pytest.mark.parametrize("model", [f"{MODEL_WEIGHTS_S3_BUCKET}/distilgpt2"])
def test_custom_executor_type_checking(model):
with pytest.raises(ValueError):
engine_args = EngineArgs(model=model,
load_format=RUNAI_STREAMER_LOAD_FORMAT,
distributed_executor_backend=Mock)
LLMEngine.from_engine_args(engine_args)
with pytest.raises(ValueError):
@ -45,7 +51,7 @@ def test_custom_executor_type_checking(model):
AsyncLLMEngine.from_engine_args(engine_args)
@pytest.mark.parametrize("model", ["facebook/opt-125m"])
@pytest.mark.parametrize("model", [f"{MODEL_WEIGHTS_S3_BUCKET}/distilgpt2"])
def test_custom_executor(model, tmp_path):
cwd = os.path.abspath(".")
os.chdir(tmp_path)
@ -54,6 +60,7 @@ def test_custom_executor(model, tmp_path):
engine_args = EngineArgs(
model=model,
load_format=RUNAI_STREAMER_LOAD_FORMAT,
distributed_executor_backend=CustomUniExecutor,
enforce_eager=True, # reduce test time
)
@ -68,7 +75,7 @@ def test_custom_executor(model, tmp_path):
os.chdir(cwd)
@pytest.mark.parametrize("model", ["facebook/opt-125m"])
@pytest.mark.parametrize("model", [f"{MODEL_WEIGHTS_S3_BUCKET}/distilgpt2"])
def test_custom_executor_async(model, tmp_path):
cwd = os.path.abspath(".")
os.chdir(tmp_path)
@ -77,6 +84,7 @@ def test_custom_executor_async(model, tmp_path):
engine_args = AsyncEngineArgs(
model=model,
load_format=RUNAI_STREAMER_LOAD_FORMAT,
distributed_executor_backend=CustomUniExecutorAsync,
enforce_eager=True, # reduce test time
)
@ -95,7 +103,7 @@ def test_custom_executor_async(model, tmp_path):
os.chdir(cwd)
@pytest.mark.parametrize("model", ["facebook/opt-125m"])
@pytest.mark.parametrize("model", [f"{MODEL_WEIGHTS_S3_BUCKET}/distilgpt2"])
def test_respect_ray(model):
# even for TP=1 and PP=1,
# if users specify ray, we should use ray.
@ -104,6 +112,7 @@ def test_respect_ray(model):
engine_args = EngineArgs(
model=model,
distributed_executor_backend="ray",
load_format=RUNAI_STREAMER_LOAD_FORMAT,
enforce_eager=True, # reduce test time
)
engine = LLMEngine.from_engine_args(engine_args)

9
tests/engine/test_skip_tokenizer_init.py
View File

@ -2,16 +2,21 @@
import pytest
from vllm.config import LoadFormat
from vllm.entrypoints.llm import LLM
from vllm.sampling_params import SamplingParams
from ..conftest import MODEL_WEIGHTS_S3_BUCKET
@pytest.mark.parametrize("model", ["facebook/opt-125m"])
@pytest.mark.parametrize("model", [f"{MODEL_WEIGHTS_S3_BUCKET}/distilgpt2"])
def test_skip_tokenizer_initialization(model: str):
# This test checks if the flag skip_tokenizer_init skips the initialization
# of tokenizer and detokenizer. The generated output is expected to contain
# token ids.
llm = LLM(model=model, skip_tokenizer_init=True)
llm = LLM(model=model,
skip_tokenizer_init=True,
load_format=LoadFormat.RUNAI_STREAMER)
sampling_params = SamplingParams(prompt_logprobs=True, detokenize=True)
with pytest.raises(ValueError, match="cannot pass text prompts when"):

2
tests/engine/test_stop_reason.py
View File

@ -12,7 +12,7 @@ import transformers
from vllm import SamplingParams
MODEL = "facebook/opt-350m"
MODEL = "distilbert/distilgpt2"
STOP_STR = "."
SEED = 42
MAX_TOKENS = 1024