[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:
@@ -92,8 +92,8 @@ def _validate_image_prompt_replacements_one(
first_placeholder = image_placeholders[0]
# NOTE: There is a BOS token
assert first_placeholder["offset"] == 1
assert first_placeholder["length"] == (
assert first_placeholder.offset == 1
assert first_placeholder.length == (
len(processed_inputs["prompt_token_ids"]) - 1) // num_imgs
except Exception as exc:
@@ -92,8 +92,8 @@ def _validate_image_prompt_replacements_one(
first_placeholder = image_placeholders[0]
assert first_placeholder["offset"] == 0
assert first_placeholder["length"] == len(
assert first_placeholder.offset == 0
assert first_placeholder.length == len(
processed_inputs["prompt_token_ids"]) // num_imgs
except Exception as exc:
failed_size_excs.append((image_size, exc))
+3 -1
View File
@@ -277,7 +277,9 @@ _MULTIMODAL_EXAMPLE_MODELS = {
trust_remote_code=True,
hf_overrides={"architectures": ["GLM4VForCausalLM"]}), # noqa: E501
"H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m",
extras={"2b": "h2oai/h2ovl-mississippi-2b"}), # noqa: E501
extras={"2b": "h2oai/h2ovl-mississippi-2b"}, # noqa: E501
max_transformers_version="4.48", # noqa: E501
transformers_version_reason="HF model is not compatible."), # noqa: E501
"InternVLChatModel": _HfExamplesInfo("OpenGVLab/InternVL2-1B",
extras={"2B": "OpenGVLab/InternVL2-2B"}, # noqa: E501
trust_remote_code=True),
+9
View File
@@ -785,6 +785,7 @@ def test_find_update_tokens(
item_idx=0,
start_idx=6,
tokens=[32000, 32000],
is_embed=None,
),
],
"pattern_4": [
@@ -793,6 +794,7 @@ def test_find_update_tokens(
item_idx=0,
start_idx=3,
tokens=[32000],
is_embed=None,
),
],
}
@@ -807,12 +809,14 @@ def test_find_update_tokens(
item_idx=0,
start_idx=1,
tokens=[32000, 32000],
is_embed=None,
),
PlaceholderFeaturesInfo(
modality="pattern_1",
item_idx=1,
start_idx=5,
tokens=[32000, 32000],
is_embed=None,
),
],
"pattern_3": [
@@ -821,6 +825,7 @@ def test_find_update_tokens(
item_idx=0,
start_idx=7,
tokens=[1550, 918, 1550],
is_embed=None,
),
],
# No match for pattern_4 as it has lower priority than pattern_1
@@ -835,12 +840,14 @@ def test_find_update_tokens(
item_idx=0,
start_idx=1,
tokens=[32000, 32000],
is_embed=None,
),
PlaceholderFeaturesInfo(
modality="pattern_1",
item_idx=1,
start_idx=3,
tokens=[32000, 32000],
is_embed=None,
),
],
"pattern_4": [
@@ -849,6 +856,7 @@ def test_find_update_tokens(
item_idx=0,
start_idx=5,
tokens=[32000],
is_embed=None,
),
],
"pattern_3": [
@@ -857,6 +865,7 @@ def test_find_update_tokens(
item_idx=0,
start_idx=6,
tokens=[1550, 918, 1550],
is_embed=None,
),
],
}
+17 -29
View File
@@ -3,7 +3,7 @@
import pytest
import torch
from vllm.multimodal.inputs import MultiModalKwargs
from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
from vllm.sampling_params import SamplingParams
from vllm.utils import sha256
# disable yapf here as it formats differently than isort such that both fail
@@ -158,13 +158,10 @@ def test_generate_block_hash_extra_keys():
request = make_request(
request_id=0,
prompt_token_ids=[_ for _ in range(20)],
mm_positions=[{
"offset": 0,
"length": 5
}, {
"offset": 10,
"length": 5
}],
mm_positions=[
PlaceholderRange(offset=0, length=5),
PlaceholderRange(offset=10, length=5),
],
mm_hashes=["hash1", "hash2"],
)
@@ -222,13 +219,10 @@ def test_hash_request_tokens(hash_fn):
request = make_request(
request_id=0,
prompt_token_ids=[_ for _ in range(6)],
mm_positions=[{
"offset": 0,
"length": 3
}, {
"offset": 3,
"length": 3
}],
mm_positions=[
PlaceholderRange(offset=0, length=3),
PlaceholderRange(offset=3, length=3),
],
mm_hashes=["hash1", "hash2"],
)
@@ -253,25 +247,19 @@ def test_hash_tokens_different_mm_input(hash_fn):
request1 = make_request(
request_id=0,
prompt_token_ids=[_ for _ in range(6)],
mm_positions=[{
"offset": 0,
"length": 3
}, {
"offset": 3,
"length": 3
}],
mm_positions=[
PlaceholderRange(offset=0, length=3),
PlaceholderRange(offset=3, length=3),
],
mm_hashes=["hash1", "hash2"],
)
request2 = make_request(
request_id=1,
prompt_token_ids=[_ for _ in range(6)],
mm_positions=[{
"offset": 0,
"length": 3
}, {
"offset": 3,
"length": 3
}],
mm_positions=[
PlaceholderRange(offset=0, length=3),
PlaceholderRange(offset=3, length=3),
],
mm_hashes=["hash3", "hash2"],
)
block_size = 3