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

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

122
tests/basic_correctness/test_basic_correctness.py
View File

@@ -4,6 +4,7 @@
Run `pytest tests/basic_correctness/test_basic_correctness.py`.
"""
import os
import weakref
from unittest.mock import Mock
@@ -37,16 +38,21 @@ def test_vllm_gc_ed():
def _fix_prompt_embed_outputs(
vllm_outputs: list[tuple[list[int], str]], hf_model: HfRunner,
example_prompts: list[str]) -> list[tuple[list[int], str]]:
vllm_outputs: list[tuple[list[int], str]],
hf_model: HfRunner,
example_prompts: list[str],
) -> list[tuple[list[int], str]]:
fixed_vllm_outputs = []
for vllm_output, hf_input, prompt in zip(
vllm_outputs, hf_model.get_inputs(example_prompts),
example_prompts):
vllm_outputs, hf_model.get_inputs(example_prompts), example_prompts
):
hf_input_ids = hf_input["input_ids"].tolist()[0]
fixed_vllm_outputs.append(
(hf_input_ids + vllm_output[0][len(hf_input_ids):],
prompt + vllm_output[1]))
(
hf_input_ids + vllm_output[0][len(hf_input_ids) :],
prompt + vllm_output[1],
)
)
return fixed_vllm_outputs
@@ -69,8 +75,7 @@ def test_models(
enable_prompt_embeds: bool,
) -> None:
if backend == "XFORMERS" and model == "google/gemma-2-2b-it":
pytest.skip(
f"{backend} does not support gemma2 with full context length.")
pytest.skip(f"{backend} does not support gemma2 with full context length.")
with monkeypatch.context() as m:
m.setenv("VLLM_ATTENTION_BACKEND", backend)
@@ -78,34 +83,35 @@ def test_models(
# 5042 tokens for gemma2
# gemma2 has alternating sliding window size of 4096
# we need a prompt with more than 4096 tokens to test the sliding window
prompt = "The following numbers of the sequence " + ", ".join(
str(i) for i in range(1024)) + " are:"
prompt = (
"The following numbers of the sequence "
+ ", ".join(str(i) for i in range(1024))
+ " are:"
)
example_prompts = [prompt]
with hf_runner(model) as hf_model:
hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
if enable_prompt_embeds:
with torch.no_grad():
prompt_embeds = hf_model.get_prompt_embeddings(
example_prompts)
prompt_embeds = hf_model.get_prompt_embeddings(example_prompts)
with VllmRunner(
model,
max_model_len=8192,
enforce_eager=enforce_eager,
enable_prompt_embeds=enable_prompt_embeds,
gpu_memory_utilization=0.7,
async_scheduling=async_scheduling,
distributed_executor_backend=model_executor,
model,
max_model_len=8192,
enforce_eager=enforce_eager,
enable_prompt_embeds=enable_prompt_embeds,
gpu_memory_utilization=0.7,
async_scheduling=async_scheduling,
distributed_executor_backend=model_executor,
) as vllm_model:
if enable_prompt_embeds:
vllm_outputs = vllm_model.generate_greedy(
prompt_embeds, max_tokens)
vllm_outputs = vllm_model.generate_greedy(prompt_embeds, max_tokens)
vllm_outputs = _fix_prompt_embed_outputs(
vllm_outputs, hf_model, example_prompts)
vllm_outputs, hf_model, example_prompts
)
else:
vllm_outputs = vllm_model.generate_greedy(
example_prompts, max_tokens)
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=hf_outputs,
@@ -117,21 +123,18 @@ def test_models(
@multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize(
"model, distributed_executor_backend, attention_backend, "
"test_suite, extra_env", [
"model, distributed_executor_backend, attention_backend, test_suite, extra_env",
[
("distilbert/distilgpt2", "ray", "", "L4", {}),
("distilbert/distilgpt2", "mp", "", "L4", {}),
("distilbert/distilgpt2", "ray", "", "L4", {
"VLLM_SLEEP_WHEN_IDLE": "1"
}),
("distilbert/distilgpt2", "mp", "", "L4", {
"VLLM_SLEEP_WHEN_IDLE": "1"
}),
("distilbert/distilgpt2", "ray", "", "L4", {"VLLM_SLEEP_WHEN_IDLE": "1"}),
("distilbert/distilgpt2", "mp", "", "L4", {"VLLM_SLEEP_WHEN_IDLE": "1"}),
("meta-llama/Llama-3.2-1B-Instruct", "ray", "", "L4", {}),
("meta-llama/Llama-3.2-1B-Instruct", "mp", "", "L4", {}),
("distilbert/distilgpt2", "ray", "", "A100", {}),
("distilbert/distilgpt2", "mp", "", "A100", {}),
])
],
)
@pytest.mark.parametrize("enable_prompt_embeds", [True, False])
def test_models_distributed(
monkeypatch: pytest.MonkeyPatch,
@@ -149,11 +152,14 @@ def test_models_distributed(
pytest.skip(f"Skip test for {test_suite}")
with monkeypatch.context() as monkeypatch_context:
if model == "meta-llama/Llama-3.2-1B-Instruct" and distributed_executor_backend == "ray" and attention_backend == "" and test_suite == "L4": # noqa
if (
model == "meta-llama/Llama-3.2-1B-Instruct"
and distributed_executor_backend == "ray"
and attention_backend == ""
and test_suite == "L4"
): # noqa
if enable_prompt_embeds:
pytest.skip(
"enable_prompt_embeds does not work with ray compiled dag."
)
pytest.skip("enable_prompt_embeds does not work with ray compiled dag.")
monkeypatch_context.setenv("VLLM_USE_RAY_SPMD_WORKER", "1")
monkeypatch_context.setenv("VLLM_USE_RAY_COMPILED_DAG", "1")
@@ -175,30 +181,26 @@ def test_models_distributed(
# 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,
enable_prompt_embeds=enable_prompt_embeds,
gpu_memory_utilization=0.7,
model,
dtype=dtype,
tensor_parallel_size=2,
distributed_executor_backend=distributed_executor_backend,
enable_prompt_embeds=enable_prompt_embeds,
gpu_memory_utilization=0.7,
) as vllm_model:
if enable_prompt_embeds:
with hf_runner(model, dtype=dtype) as hf_model:
with torch.no_grad():
prompt_embeds = hf_model.get_prompt_embeddings(
example_prompts)
vllm_outputs = vllm_model.generate_greedy(
prompt_embeds, max_tokens)
prompt_embeds = hf_model.get_prompt_embeddings(example_prompts)
vllm_outputs = vllm_model.generate_greedy(prompt_embeds, max_tokens)
vllm_outputs = _fix_prompt_embed_outputs(
vllm_outputs, hf_model, example_prompts)
hf_outputs = hf_model.generate_greedy(
example_prompts, max_tokens)
vllm_outputs, hf_model, example_prompts
)
hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
else:
vllm_outputs = vllm_model.generate_greedy(
example_prompts, max_tokens)
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)
hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=hf_outputs,
@@ -209,27 +211,23 @@ def test_models_distributed(
def test_failed_model_execution(vllm_runner, monkeypatch) -> None:
from vllm.envs import VLLM_USE_V1
if not VLLM_USE_V1:
pytest.skip("Skipping V0 test, dump input not supported")
# Needed to mock an error in the same process
monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
with vllm_runner('facebook/opt-125m', enforce_eager=True) as vllm_model:
with vllm_runner("facebook/opt-125m", enforce_eager=True) as vllm_model:
if isinstance(vllm_model.llm.llm_engine, LLMEngineV1):
v1_test_failed_model_execution(vllm_model)
def v1_test_failed_model_execution(vllm_model):
engine = vllm_model.llm.llm_engine
mocked_execute_model = Mock(
side_effect=RuntimeError("Mocked Critical Error"))
engine.engine_core.engine_core.model_executor.execute_model =\
mocked_execute_model
mocked_execute_model = Mock(side_effect=RuntimeError("Mocked Critical Error"))
engine.engine_core.engine_core.model_executor.execute_model = mocked_execute_model
with pytest.raises(RuntimeError) as exc_info:
prompts = [

5
tests/basic_correctness/test_cpu_offload.py
View File

@@ -5,5 +5,6 @@ from ..utils import compare_two_settings
def test_cpu_offload():
compare_two_settings("meta-llama/Llama-3.2-1B-Instruct", [],
["--cpu-offload-gb", "1"])
compare_two_settings(
"meta-llama/Llama-3.2-1B-Instruct", [], ["--cpu-offload-gb", "1"]
)

23
tests/basic_correctness/test_cumem.py
View File

@@ -23,13 +23,13 @@ def test_python_error():
tensors = []
with allocator.use_memory_pool():
# allocate 70% of the total memory
x = torch.empty(alloc_bytes, dtype=torch.uint8, device='cuda')
x = torch.empty(alloc_bytes, dtype=torch.uint8, device="cuda")
tensors.append(x)
# release the memory
allocator.sleep()
# allocate more memory than the total memory
y = torch.empty(alloc_bytes, dtype=torch.uint8, device='cuda')
y = torch.empty(alloc_bytes, dtype=torch.uint8, device="cuda")
tensors.append(y)
with pytest.raises(RuntimeError):
# when the allocator is woken up, it should raise an error
@@ -41,17 +41,17 @@ def test_python_error():
def test_basic_cumem():
# some tensors from default memory pool
shape = (1024, 1024)
x = torch.empty(shape, device='cuda')
x = torch.empty(shape, device="cuda")
x.zero_()
# some tensors from custom memory pool
allocator = CuMemAllocator.get_instance()
with allocator.use_memory_pool():
# custom memory pool
y = torch.empty(shape, device='cuda')
y = torch.empty(shape, device="cuda")
y.zero_()
y += 1
z = torch.empty(shape, device='cuda')
z = torch.empty(shape, device="cuda")
z.zero_()
z += 2
@@ -74,16 +74,16 @@ def test_basic_cumem():
def test_cumem_with_cudagraph():
allocator = CuMemAllocator.get_instance()
with allocator.use_memory_pool():
weight = torch.eye(1024, device='cuda')
weight = torch.eye(1024, device="cuda")
with allocator.use_memory_pool(tag="discard"):
cache = torch.empty(1024, 1024, device='cuda')
cache = torch.empty(1024, 1024, device="cuda")
def model(x):
out = x @ weight
cache[:out.size(0)].copy_(out)
cache[: out.size(0)].copy_(out)
return out + 1
x = torch.empty(128, 1024, device='cuda')
x = torch.empty(128, 1024, device="cuda")
# warmup
model(x)
@@ -109,7 +109,7 @@ def test_cumem_with_cudagraph():
model_graph.replay()
# cache content is as expected
assert torch.allclose(x, cache[:x.size(0)])
assert torch.allclose(x, cache[: x.size(0)])
# output content is as expected
assert torch.allclose(y, x + 1)
@@ -123,7 +123,8 @@ def test_cumem_with_cudagraph():
("meta-llama/Llama-3.2-1B", True),
# sleep mode with pytorch checkpoint
("facebook/opt-125m", True),
])
],
)
def test_end_to_end(monkeypatch: pytest.MonkeyPatch, model: str, use_v1: bool):
with monkeypatch.context() as m:
assert use_v1