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

[Misc] Replace os environ to monkeypatch in test suite (#14516)

Browse Source 
Signed-off-by: sibi <85477603+t-sibiraj@users.noreply.github.com>
Signed-off-by: Aaron Pham <contact@aarnphm.xyz>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: Aaron Pham <contact@aarnphm.xyz>
This commit is contained in:
Sibi
2025-03-17 11:35:57 +08:00
committed by GitHub
parent 1e799b7ec1
commit a73e183e36
43 changed files with 1900 additions and 1658 deletions
Download Patch File Download Diff File Expand all files Collapse all files

111
tests/prefix_caching/test_prefix_caching.py
View File

@@ -4,25 +4,29 @@
Run `pytest tests/prefix_caching/test_prefix_caching.py`.
"""
from __future__ import annotations
import pytest
from tests.conftest import VllmRunner
from tests.core.utils import SchedulerProxy, create_dummy_prompt
from tests.kernels.utils import override_backend_env_variable
from vllm import SamplingParams, TokensPrompt
from vllm.core.scheduler import Scheduler
from vllm.engine.llm_engine import LLMEngine
from vllm.platforms import current_platform
from vllm.utils import STR_BACKEND_ENV_VAR
from ..models.utils import check_outputs_equal
@pytest.fixture(scope="function", autouse=True)
def use_v0_only(monkeypatch):
def use_v0_only(monkeypatch: pytest.MonkeyPatch):
"""
This module relies on V0 internals, so set VLLM_USE_V1=0.
"""
monkeypatch.setenv('VLLM_USE_V1', '0')
with monkeypatch.context() as m:
m.setenv('VLLM_USE_V1', '0')
yield
MODELS = [
@@ -56,7 +60,7 @@ def test_mixed_requests(
cached_position: int,
enable_chunked_prefill: bool,
block_size: int,
monkeypatch,
monkeypatch: pytest.MonkeyPatch,
) -> None:
"""
Test the case when some sequences have the prefix cache hit
@@ -67,72 +71,77 @@ def test_mixed_requests(
pytest.skip("Flashinfer does not support ROCm/HIP.")
if backend == "XFORMERS" and current_platform.is_rocm():
pytest.skip("Xformers does not support ROCm/HIP.")
override_backend_env_variable(monkeypatch, backend)
with monkeypatch.context() as m:
m.setenv(STR_BACKEND_ENV_VAR, backend)
with hf_runner(model, dtype=dtype) as hf_model:
hf_outputs = hf_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)
cached_prompt = example_prompts[cached_position]
with vllm_runner(
model,
dtype=dtype,
enable_prefix_caching=True,
enable_chunked_prefill=enable_chunked_prefill,
block_size=block_size,
) as vllm_model:
# Run the first prompt so the cache is populated
vllm_outputs = vllm_model.generate_greedy([cached_prompt], max_tokens)
cached_prompt = example_prompts[cached_position]
with vllm_runner(
model,
dtype=dtype,
enable_prefix_caching=True,
enable_chunked_prefill=enable_chunked_prefill,
block_size=block_size,
) as vllm_model:
# Run the first prompt so the cache is populated
vllm_outputs = vllm_model.generate_greedy([cached_prompt],
max_tokens)
# Run all the promopts
greedy_params = SamplingParams(temperature=0.0, max_tokens=max_tokens)
req_outputs = vllm_model.model.generate(example_prompts, greedy_params)
# Run all the promopts
greedy_params = SamplingParams(temperature=0.0,
max_tokens=max_tokens)
req_outputs = vllm_model.model.generate(example_prompts,
greedy_params)
# Verify number of cached tokens
for i in range(len(req_outputs)):
if i == cached_position:
expected_num_cached_tokens = (
len(req_outputs[i].prompt_token_ids) //
block_size) * block_size
else:
expected_num_cached_tokens = 0
assert (
req_outputs[i].num_cached_tokens == expected_num_cached_tokens)
# Verify number of cached tokens
for i in range(len(req_outputs)):
if i == cached_position:
expected_num_cached_tokens = (
len(req_outputs[i].prompt_token_ids) //
block_size) * block_size
else:
expected_num_cached_tokens = 0
assert (req_outputs[i].num_cached_tokens ==
expected_num_cached_tokens)
vllm_outputs = [(
output.prompt_token_ids + list(output.outputs[0].token_ids),
output.prompt + output.outputs[0].text,
) for output in req_outputs]
vllm_outputs = [(
output.prompt_token_ids + list(output.outputs[0].token_ids),
output.prompt + output.outputs[0].text,
) for output in req_outputs]
check_outputs_equal(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
check_outputs_equal(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("backend", ["FLASH_ATTN", "FLASHINFER", "XFORMERS"])
def test_unstable_prompt_sequence(
vllm_runner,
backend: str,
monkeypatch,
monkeypatch: pytest.MonkeyPatch,
) -> None:
if backend == "FLASHINFER" and current_platform.is_rocm():
pytest.skip("Flashinfer does not support ROCm/HIP.")
if backend == "XFORMERS" and current_platform.is_rocm():
pytest.skip("Xformers does not support ROCm/HIP.")
override_backend_env_variable(monkeypatch, backend)
with monkeypatch.context() as m:
m.setenv(STR_BACKEND_ENV_VAR, backend)
with vllm_runner(
"Qwen/Qwen2.5-0.5B-Instruct",
enable_chunked_prefill=True,
enable_prefix_caching=True,
max_model_len=4096,
) as vllm_model:
for prompt in UNSTABLE_PROMPT_SEQUENCE:
vllm_model.generate(TokensPrompt(prompt_token_ids=prompt),
SamplingParams(max_tokens=1))
with vllm_runner(
"Qwen/Qwen2.5-0.5B-Instruct",
enable_chunked_prefill=True,
enable_prefix_caching=True,
max_model_len=4096,
) as vllm_model:
for prompt in UNSTABLE_PROMPT_SEQUENCE:
vllm_model.generate(TokensPrompt(prompt_token_ids=prompt),
SamplingParams(max_tokens=1))
@pytest.mark.parametrize("model", MODELS)