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

144
tests/models/multimodal/generation/test_ultravox.py
View File

@@ -15,12 +15,12 @@ from ...registry import HF_EXAMPLE_MODELS
MODEL_NAME = "fixie-ai/ultravox-v0_5-llama-3_2-1b"
AUDIO_PROMPTS = AUDIO_ASSETS.prompts({
"mary_had_lamb":
"Transcribe this into English.",
"winning_call":
"What is happening in this audio clip?",
})
AUDIO_PROMPTS = AUDIO_ASSETS.prompts(
{
"mary_had_lamb": "Transcribe this into English.",
"winning_call": "What is happening in this audio clip?",
}
)
MULTI_AUDIO_PROMPT = "Describe each of the audios above."
@@ -33,7 +33,7 @@ CHUNKED_PREFILL_KWARGS = {
"enable_chunked_prefill": True,
"max_num_seqs": 2,
# Use a very small limit to exercise chunked prefill.
"max_num_batched_tokens": 16
"max_num_batched_tokens": 16,
}
@@ -43,27 +43,33 @@ def params_kwargs_to_cli_args(params_kwargs: dict[str, Any]) -> list[str]:
for key, value in params_kwargs.items():
if isinstance(value, bool):
if value:
args.append(f"--{key.replace('_','-')}")
args.append(f"--{key.replace('_', '-')}")
else:
args.append(f"--{key.replace('_','-')}={value}")
args.append(f"--{key.replace('_', '-')}={value}")
return args
@pytest.fixture(params=[
pytest.param({}, marks=pytest.mark.cpu_model),
pytest.param(CHUNKED_PREFILL_KWARGS),
])
@pytest.fixture(
params=[
pytest.param({}, marks=pytest.mark.cpu_model),
pytest.param(CHUNKED_PREFILL_KWARGS),
]
)
def server(request, audio_assets: AudioTestAssets):
args = [
"--dtype", "bfloat16", "--max-model-len", "4096", "--enforce-eager",
"--dtype",
"bfloat16",
"--max-model-len",
"4096",
"--enforce-eager",
"--limit-mm-per-prompt",
json.dumps({"audio": len(audio_assets)}), "--trust-remote-code"
json.dumps({"audio": len(audio_assets)}),
"--trust-remote-code",
] + params_kwargs_to_cli_args(request.param)
with RemoteOpenAIServer(MODEL_NAME,
args,
env_dict={"VLLM_AUDIO_FETCH_TIMEOUT":
"30"}) as remote_server:
with RemoteOpenAIServer(
MODEL_NAME, args, env_dict={"VLLM_AUDIO_FETCH_TIMEOUT": "30"}
) as remote_server:
yield remote_server
@@ -77,12 +83,11 @@ def _get_prompt(audio_count, question, placeholder):
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
placeholder = f"{placeholder}\n" * audio_count
return tokenizer.apply_chat_template([{
'role': 'user',
'content': f"{placeholder}{question}"
}],
tokenize=False,
add_generation_prompt=True)
return tokenizer.apply_chat_template(
[{"role": "user", "content": f"{placeholder}{question}"}],
tokenize=False,
add_generation_prompt=True,
)
def run_multi_audio_test(
@@ -99,19 +104,21 @@ def run_multi_audio_test(
model_info.check_available_online(on_fail="skip")
model_info.check_transformers_version(on_fail="skip")
with vllm_runner(model,
dtype=dtype,
enforce_eager=True,
limit_mm_per_prompt={
"audio":
max((len(audio) for _, audio in prompts_and_audios))
},
**kwargs) as vllm_model:
with vllm_runner(
model,
dtype=dtype,
enforce_eager=True,
limit_mm_per_prompt={
"audio": max((len(audio) for _, audio in prompts_and_audios))
},
**kwargs,
) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
[prompt for prompt, _ in prompts_and_audios],
max_tokens,
num_logprobs=num_logprobs,
audios=[audios for _, audios in prompts_and_audios])
audios=[audios for _, audios in prompts_and_audios],
)
# The HuggingFace model doesn't support multiple audios yet, so
# just assert that some tokens were generated.
@@ -122,21 +129,25 @@ def run_multi_audio_test(
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@pytest.mark.parametrize("vllm_kwargs", [
pytest.param({}, marks=pytest.mark.cpu_model),
pytest.param(CHUNKED_PREFILL_KWARGS),
])
def test_models_with_multiple_audios(vllm_runner,
audio_assets: AudioTestAssets, dtype: str,
max_tokens: int, num_logprobs: int,
vllm_kwargs: dict) -> None:
vllm_prompt = _get_prompt(len(audio_assets), MULTI_AUDIO_PROMPT,
VLLM_PLACEHOLDER)
@pytest.mark.parametrize(
"vllm_kwargs",
[
pytest.param({}, marks=pytest.mark.cpu_model),
pytest.param(CHUNKED_PREFILL_KWARGS),
],
)
def test_models_with_multiple_audios(
vllm_runner,
audio_assets: AudioTestAssets,
dtype: str,
max_tokens: int,
num_logprobs: int,
vllm_kwargs: dict,
) -> None:
vllm_prompt = _get_prompt(len(audio_assets), MULTI_AUDIO_PROMPT, VLLM_PLACEHOLDER)
run_multi_audio_test(
vllm_runner,
[(vllm_prompt, [audio.audio_and_sample_rate
for audio in audio_assets])],
[(vllm_prompt, [audio.audio_and_sample_rate for audio in audio_assets])],
MODEL_NAME,
dtype=dtype,
max_tokens=max_tokens,
@@ -149,28 +160,25 @@ def test_models_with_multiple_audios(vllm_runner,
async def test_online_serving(client, audio_assets: AudioTestAssets):
"""Exercises online serving with/without chunked prefill enabled."""
messages = [{
"role":
"user",
"content": [
*[{
"type": "audio_url",
"audio_url": {
"url": audio.url
}
} for audio in audio_assets],
{
"type":
"text",
"text":
f"What's happening in these {len(audio_assets)} audio clips?"
},
],
}]
messages = [
{
"role": "user",
"content": [
*[
{"type": "audio_url", "audio_url": {"url": audio.url}}
for audio in audio_assets
],
{
"type": "text",
"text": f"What's happening in these {len(audio_assets)} audio clips?",
},
],
}
]
chat_completion = await client.chat.completions.create(model=MODEL_NAME,
messages=messages,
max_tokens=10)
chat_completion = await client.chat.completions.create(
model=MODEL_NAME, messages=messages, max_tokens=10
)
assert len(chat_completion.choices) == 1
choice = chat_completion.choices[0]