[Refactor] Use dynamic MM cache in processor (#57674)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-09-19 15:38:00 +00:00
committed by GitHub
parent 41c4a3ed4e
commit 751f6807d9
23 changed files with 343 additions and 401 deletions
+34 -35
View File
@@ -1,8 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import multiprocessing as mp
from concurrent.futures import ThreadPoolExecutor
from functools import partial
from types import SimpleNamespace
from unittest.mock import Mock
@@ -37,10 +35,13 @@ from vllm.multimodal.inputs import (
PlaceholderRange,
)
from vllm.multimodal.processing import PromptInsertion
from vllm.renderers.base import BaseRenderer
from vllm.renderers import renderer_from_config
from vllm.utils.async_utils import make_async
from vllm.utils.mem_constants import GiB_bytes, MiB_bytes
from vllm.v1.engine.async_llm import AsyncLLM
from vllm.v1.engine.llm_engine import LLMEngine
from ..models.utils import build_model_context
from ..utils import create_new_process_for_each_test
pytestmark = pytest.mark.cpu_test
@@ -773,12 +774,14 @@ def test_processor_cache_shared_across_loras():
@pytest.mark.asyncio
async def test_release_kv_cache_resends_mm_payload(use_async, release_error):
"""Release must not leave a sender hit pointing at a cleared receiver."""
from vllm.v1.engine.async_llm import AsyncLLM
from vllm.v1.engine.llm_engine import LLMEngine
model_config = SimpleNamespace(
get_multimodal_config=lambda: MultiModalConfig(mm_processor_cache_gb=1)
ctx = build_model_context(
"llava-hf/llava-v1.6-mistral-7b-hf",
mm_processor_kwargs=None,
limit_mm_per_prompt={"image": 1},
mm_processor_cache_gb=1,
)
model_config = ctx.model_config
sender = MultiModalProcessorSenderCache(model_config)
receiver = MultiModalReceiverCache(model_config)
item = _dummy_item({"pixel_values": 16})
@@ -787,43 +790,39 @@ async def test_release_kv_cache_resends_mm_payload(use_async, release_error):
assert receiver.get_and_update_item(payload, mm_hash) is item
assert sender.get_and_update_item(None, mm_hash)[0] is None
renderer = SimpleNamespace(mm_processor_cache=sender, _mm_cache_stats=None)
renderer.clear_mm_cache = partial(BaseRenderer.clear_mm_cache, renderer)
renderer = renderer_from_config(VllmConfig(model_config=model_config))
renderer._mm_processor_cache = sender
def release():
if release_error:
raise RuntimeError(release_error)
receiver.clear_cache()
with ThreadPoolExecutor(max_workers=1) as executor:
renderer._clear_mm_cache_async = make_async(
renderer.clear_mm_cache, executor=executor
)
renderer.clear_mm_cache_async = partial(
BaseRenderer.clear_mm_cache_async, renderer
)
engine = SimpleNamespace(
renderer=renderer,
engine_core=SimpleNamespace(
release_kv_cache_memory=release,
release_kv_cache_memory_async=make_async(release, executor=executor),
engine = SimpleNamespace(
renderer=renderer,
engine_core=SimpleNamespace(
release_kv_cache_memory=release,
release_kv_cache_memory_async=make_async(
release,
executor=renderer._executor,
),
logger_manager=Mock(),
)
),
logger_manager=Mock(),
)
async def call_release():
if use_async:
await AsyncLLM.release_kv_cache_memory(engine)
else:
LLMEngine.release_kv_cache_memory(engine)
if release_error:
with pytest.raises(RuntimeError, match=release_error):
await call_release()
engine.logger_manager.record_sleep_state.assert_not_called()
async def call_release():
if use_async:
await AsyncLLM.release_kv_cache_memory(engine)
else:
LLMEngine.release_kv_cache_memory(engine)
if release_error:
with pytest.raises(RuntimeError, match=release_error):
await call_release()
engine.logger_manager.record_sleep_state.assert_called_once_with(1, 0)
engine.logger_manager.record_sleep_state.assert_not_called()
else:
await call_release()
engine.logger_manager.record_sleep_state.assert_called_once_with(1, 0)
payload, _ = sender.get_and_update_item((item, []), mm_hash)
assert payload is item
+36 -1
View File
@@ -7,7 +7,7 @@ from contextlib import nullcontext
import numpy as np
import pytest
from vllm.config import ModelConfig
from vllm.config import ModelConfig, SchedulerConfig
from vllm.exceptions import VLLMValidationError
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.hasher import MultiModalHasher
@@ -32,6 +32,7 @@ from vllm.multimodal.processing.processor import (
)
from vllm.utils.collection_utils import flatten_2d_lists
from ..models.utils import build_model_context
from .utils import random_image
pytestmark = pytest.mark.cpu_test
@@ -1361,3 +1362,37 @@ def test_processor_inputs_hashes_distinguish_kwargs_shapes(left, right):
).get_mm_hashes("test-model", "blake3")["image"][0]
assert hash_with(left) != hash_with(right)
@pytest.mark.parametrize(
("chunked_prefill", "max_model_len", "expected_seq_len"),
[(None, 491520, 491520), (True, 491520, 8192), (True, 128, 128), (False, 128, 128)],
)
def test_dummy_inputs_scheduler_budget(
chunked_prefill, max_model_len, expected_seq_len
):
ctx = build_model_context(
"llava-hf/llava-v1.6-mistral-7b-hf",
mm_processor_kwargs=None,
limit_mm_per_prompt={"image": 1},
)
ctx.model_config.max_model_len = max_model_len
processor = MULTIMODAL_REGISTRY.create_processor(
ctx.model_config,
tokenizer=ctx.tokenizer,
)
processor.apply = lambda *args, **kwargs: {"prompt_token_ids": [7]}
kwargs = {}
if chunked_prefill is not None:
kwargs["scheduler_config"] = SchedulerConfig(
max_model_len=max_model_len,
is_encoder_decoder=False,
max_num_batched_tokens=8192,
max_num_seqs=1,
enable_chunked_prefill=chunked_prefill,
)
result = processor.get_dummy_mm_inputs({"image": 1}, **kwargs)
assert len(result["prompt_token_ids"]) == expected_seq_len
-30
View File
@@ -5,11 +5,9 @@ Qwen2.5-VL visual component loading behavior.
"""
from types import SimpleNamespace
from unittest.mock import MagicMock
import pytest
from vllm.config import SchedulerConfig
from vllm.multimodal import MULTIMODAL_REGISTRY
from ..models.utils import build_model_context
@@ -17,34 +15,6 @@ from ..models.utils import build_model_context
pytestmark = pytest.mark.cpu_test
@pytest.mark.parametrize(
("chunked_prefill", "max_model_len", "expected_seq_len"),
[(None, 491520, 491520), (True, 491520, 8192), (True, 128, 128), (False, 128, 128)],
)
def test_dummy_inputs_scheduler_budget(
chunked_prefill, max_model_len, expected_seq_len
):
model_config = MagicMock()
model_config.max_model_len = max_model_len
processor = MagicMock()
processor.apply.return_value = {"prompt_token_ids": [7]}
kwargs = {}
if chunked_prefill is not None:
kwargs["scheduler_config"] = SchedulerConfig(
max_model_len=max_model_len,
is_encoder_decoder=False,
max_num_batched_tokens=8192,
max_num_seqs=1,
enable_chunked_prefill=chunked_prefill,
)
result = MULTIMODAL_REGISTRY.get_dummy_mm_inputs(
model_config, {"image": 1}, processor=processor, **kwargs
)
get_inputs = processor.dummy_inputs.get_dummy_processor_inputs
assert get_inputs.call_args.kwargs["seq_len"] == expected_seq_len
assert len(result["prompt_token_ids"]) == expected_seq_len
@pytest.mark.parametrize(
"model_id,limit_mm_per_prompt,expected",
[