[Bugfix][Core] Keep every multimodal feature in the partial-block KV event (#58288)

Signed-off-by: haosenwang1018 <haosenwang1018@users.noreply.github.com>
Co-authored-by: haosenwang1018 <haosenwang1018@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Sense_wang
2026-09-24 23:25:01 +00:00
committed by GitHub
co-authored by haosenwang1018 Claude Opus 5
parent bd4cb3fe4e
commit f6aa2919bc
2 changed files with 78 additions and 1 deletions
@@ -7,6 +7,11 @@ import pytest
import vllm.v1.core.kv_cache_utils as kv_cache_utils import vllm.v1.core.kv_cache_utils as kv_cache_utils
from vllm.distributed.kv_events import BlockRemoved, BlockStored from vllm.distributed.kv_events import BlockRemoved, BlockStored
from vllm.multimodal.inputs import (
MultiModalFeatureSpec,
MultiModalKwargsItem,
PlaceholderRange,
)
from vllm.sampling_params import SamplingParams from vllm.sampling_params import SamplingParams
from vllm.utils.hashing import sha256 from vllm.utils.hashing import sha256
from vllm.v1.core.block_pool import BlockPool from vllm.v1.core.block_pool import BlockPool
@@ -34,12 +39,28 @@ def make_request(
hash_block_size: int, hash_block_size: int,
hash_fn: Callable, hash_fn: Callable,
session_id: str | None = None, session_id: str | None = None,
mm_positions: list[PlaceholderRange] | None = None,
mm_hashes: list[str] | None = None,
) -> Request: ) -> Request:
mm_features = []
if mm_positions is not None:
for j, position in enumerate(mm_positions):
identifier = mm_hashes[j] if mm_hashes else f"hash_{j}"
mm_features.append(
MultiModalFeatureSpec(
data=MultiModalKwargsItem.dummy(),
mm_position=position,
identifier=identifier,
modality="image",
)
)
sampling_params = SamplingParams(max_tokens=17) sampling_params = SamplingParams(max_tokens=17)
sampling_params.update_from_generation_config({}, eos_token_id=100) sampling_params.update_from_generation_config({}, eos_token_id=100)
return Request( return Request(
request_id=request_id, request_id=request_id,
prompt_token_ids=prompt_token_ids, prompt_token_ids=prompt_token_ids,
mm_features=mm_features if mm_features else None,
sampling_params=sampling_params, sampling_params=sampling_params,
pooling_params=None, pooling_params=None,
block_hasher=get_request_block_hasher(hash_block_size, hash_fn), block_hasher=get_request_block_hasher(hash_block_size, hash_fn),
@@ -173,6 +194,54 @@ def test_cache_partial_block_kv_cache_events():
assert removed_event.group_idx == kv_cache_group_id assert removed_event.group_idx == kv_cache_group_id
def test_cache_partial_block_event_keeps_every_mm_feature():
"""The event window can reach into more than one multimodal feature.
``cache_partial_block`` builds the ``BlockStored`` extra keys over
``[(num_hash_blocks - 1) * hash_block_size, num_tokens)``. Starting that
scan at the last feature drops any earlier one that also reaches into the
window, so two requests differing only in that feature would be published
under the same key.
"""
hash_block_size = 4
block_size = 8
num_tokens = 12
# The event window is tokens [8, 12): "A" ends inside it and "B" sits
# entirely within it.
mm_positions = [
PlaceholderRange(offset=6, length=3),
PlaceholderRange(offset=10, length=2),
]
pool = BlockPool(
num_gpu_blocks=2,
enable_caching=True,
hash_block_size=hash_block_size,
enable_kv_cache_events=True,
)
req = make_request(
"req_partial_mm",
prompt_token_ids=list(range(num_tokens)),
hash_block_size=hash_block_size,
hash_fn=sha256,
mm_positions=mm_positions,
mm_hashes=["A", "B"],
)
assert pool.cache_partial_block(
request=req,
block=pool.get_new_blocks(1)[0],
num_tokens=num_tokens,
kv_cache_group_id=0,
block_size=block_size,
)
events = pool.take_events()
assert len(events) == 1
assert isinstance(events[0], BlockStored)
assert events[0].extra_keys == [(("A", -2), ("B", 2))]
def test_partial_block_replacement_emits_remove_then_store_events(): def test_partial_block_replacement_emits_remove_then_store_events():
hash_block_size = 2 hash_block_size = 2
block_size = 6 block_size = 6
+9 -1
View File
@@ -11,6 +11,7 @@ from vllm.distributed.kv_events import (
KVCacheEvent, KVCacheEvent,
) )
from vllm.logger import init_logger from vllm.logger import init_logger
from vllm.multimodal.utils import get_mm_features_in_window
from vllm.v1.core.kv_cache_metrics import KVCacheMetricsCollector from vllm.v1.core.kv_cache_metrics import KVCacheMetricsCollector
from vllm.v1.core.kv_cache_utils import ( from vllm.v1.core.kv_cache_utils import (
BlockHash, BlockHash,
@@ -532,7 +533,14 @@ class BlockPool:
else None else None
) )
block_end = num_tokens block_end = num_tokens
curr_mm_idx = -1 if block_start > 0 else 0 curr_mm_idx = 0
mm_features = request.mm_features
if block_start > 0 and mm_features:
last_mm_pos = mm_features[-1].mm_position
if last_mm_pos.offset + last_mm_pos.length > block_start:
curr_mm_idx, _ = get_mm_features_in_window(
mm_features, block_start, block_end
)
extra_keys, _ = generate_block_hash_extra_keys( extra_keys, _ = generate_block_hash_extra_keys(
request, block_start, block_end, curr_mm_idx request, block_start, block_end, curr_mm_idx
) )