[V1] Scatter and gather placeholders in the model runner (#15712)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Roger Wang <ywang@roblox.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Roger Wang <ywang@roblox.com>
This commit is contained in:
Cyrus Leung
2025-04-04 21:26:44 +00:00
committed by GitHub
co-authored by mgoin Roger Wang
parent 651cf0fec1
commit f5722a5052
42 changed files with 496 additions and 942 deletions
@@ -55,7 +55,10 @@ def server(request, audio_assets):
for key, value in request.param.items()
]
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
with RemoteOpenAIServer(MODEL_NAME,
args,
env_dict={"VLLM_AUDIO_FETCH_TIMEOUT":
"30"}) as remote_server:
yield remote_server
@@ -167,7 +167,7 @@ VLM_TEST_SETTINGS = {
"cherry_blossom": "<image>What is the season?", # noqa: E501
}),
multi_image_prompt="<image><image>Describe the two images in detail.", # noqa: E501
max_model_len=8192,
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
vllm_runner_kwargs={"mm_processor_kwargs": {"crop_to_patches": True}}
@@ -176,6 +176,8 @@ def test_chat(
model,
dtype=dtype,
tokenizer_mode="mistral",
load_format="mistral",
config_format="mistral",
max_model_len=max_model_len,
limit_mm_per_prompt=LIMIT_MM_PER_PROMPT,
) as vllm_model:
@@ -198,22 +200,14 @@ def test_chat(
@large_gpu_test(min_gb=48)
@pytest.mark.parametrize(
"prompt,expected_ranges",
[(_create_engine_inputs_hf(IMG_URLS[:1]), [{
"offset": 11,
"length": 494
}]),
(_create_engine_inputs_hf(IMG_URLS[1:4]), [{
"offset": 11,
"length": 266
}, {
"offset": 277,
"length": 1056
}, {
"offset": 1333,
"length": 418
}])])
@pytest.mark.parametrize("prompt,expected_ranges",
[(_create_engine_inputs_hf(IMG_URLS[:1]),
[PlaceholderRange(offset=11, length=494)]),
(_create_engine_inputs_hf(IMG_URLS[1:4]), [
PlaceholderRange(offset=11, length=266),
PlaceholderRange(offset=277, length=1056),
PlaceholderRange(offset=1333, length=418)
])])
def test_multi_modal_placeholders(vllm_runner, prompt,
expected_ranges: list[PlaceholderRange],
monkeypatch) -> None: