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

66
tests/v1/tpu/test_perf.py
View File

@@ -4,6 +4,7 @@
Run `pytest tests/v1/tpu/test_perf.py`.
"""
from __future__ import annotations
import time
@@ -37,7 +38,6 @@ TEST_PARAMS = [
# open(/dev/vfio/0): Device or resource busy: Device or resource busy;
# Couldn't open iommu group /dev/vfio/0
# => Investigate
# TestParams(
# model="Qwen/Qwen2.5-1.5B-Instruct",
# num_prompts=1,
@@ -59,16 +59,14 @@ TEST_PARAMS = [
num_prompts=64,
prefix_len=500,
decode_len=50,
# commit id: ccb246776d93ef105904a8ec015b3587240a1183
# tpu: v5lite (old vllm CI/CD)
# expected_avg_time=1.4,
# err_tol=0.30,
# (This is the active CI/CD instance)
# commit id: ccb246776d93ef105904a8ec015b3587240a1183
# tpu: v6e (current vllm CI/CD)
expected_avg_time=1.7, # measured with VLLM_XLA_CACHE_PATH=
expected_avg_time=1.7, # measured with VLLM_XLA_CACHE_PATH=
err_tol=0.20,
),
]
@@ -81,44 +79,50 @@ MAX_NUM_SEQS = 32
GPU_UTIL = 0.9
@pytest.mark.skipif(not current_platform.is_tpu(),
reason="This is a basic performance test for TPU only")
@pytest.mark.skipif(
not current_platform.is_tpu(),
reason="This is a basic performance test for TPU only",
)
@pytest.mark.parametrize("params", TEST_PARAMS)
def test_perf(
vllm_runner: type[VllmRunner],
monkeypatch: pytest.MonkeyPatch,
params: TestParams,
) -> None:
tokenizer = get_tokenizer(params.model,
tokenizer_mode="auto",
trust_remote_code=True)
tokenizer = get_tokenizer(
params.model, tokenizer_mode="auto", trust_remote_code=True
)
prompts = []
for i in range(params.num_prompts):
prefix_token_ids = np.random.randint(0,
tokenizer.vocab_size,
size=params.prefix_len).tolist()
prefix_token_ids = np.random.randint(
0, tokenizer.vocab_size, size=params.prefix_len
).tolist()
prompt = tokenizer.decode(prefix_token_ids)
prompts.append(prompt)
print(
"-- Running: num_prompts = {} prefix_len = {} decode_len = {}".format(
len(prompts), params.prefix_len, params.decode_len))
len(prompts), params.prefix_len, params.decode_len
)
)
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1")
sampling_params = SamplingParams(max_tokens=params.decode_len,
temperature=1.0,
min_p=0.0)
sampling_params = SamplingParams(
max_tokens=params.decode_len, temperature=1.0, min_p=0.0
)
with vllm_runner(params.model,
max_num_batched_tokens=MAX_MODEL_LEN,
max_model_len=MAX_MODEL_LEN,
max_num_seqs=MAX_NUM_SEQS,
gpu_memory_utilization=GPU_UTIL,
enforce_eager=False,
tensor_parallel_size=1) as vllm_model:
with vllm_runner(
params.model,
max_num_batched_tokens=MAX_MODEL_LEN,
max_model_len=MAX_MODEL_LEN,
max_num_seqs=MAX_NUM_SEQS,
gpu_memory_utilization=GPU_UTIL,
enforce_eager=False,
tensor_parallel_size=1,
) as vllm_model:
print(" -- Warmup / Compile")
for i in range(NUM_WARMUPS):
_ = vllm_model.generate(prompts, sampling_params)
@@ -133,14 +137,18 @@ def test_perf(
avg_time = sum(times) / len(times)
print(" -- avg_time = {}".format(avg_time))
print(" -- expected_avg_time = {} with err_tol = {}".format(
params.expected_avg_time, params.err_tol))
print(
" -- expected_avg_time = {} with err_tol = {}".format(
params.expected_avg_time, params.err_tol
)
)
diff = avg_time - params.expected_avg_time
ok = diff < params.err_tol
if diff < -params.err_tol:
print(" !! WARNING !! Performance has improved by {}, "
"it may be necessary to fine-tune the "
"expected_avg_time = {}".format(
-diff, params.expected_avg_time))
print(
" !! WARNING !! Performance has improved by {}, "
"it may be necessary to fine-tune the "
"expected_avg_time = {}".format(-diff, params.expected_avg_time)
)
assert ok, " !! ERROR !! Regression detected"