[Core][KV events] Report prefix-cache-reused blocks in full report mode (#45261)
Signed-off-by: Lei Gong <gonglei25@huawei.com> Co-authored-by: Lei Gong <gonglei25@huawei.com> Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
co-authored by
Lei Gong
Claude
parent
95ed0feaa5
commit
e5588e49bc
@@ -12,7 +12,12 @@ import torch
|
||||
|
||||
import vllm.v1.core.kv_cache_manager as kv_cache_manager
|
||||
import vllm.v1.core.kv_cache_utils as kv_cache_utils
|
||||
from vllm.distributed.kv_events import AllBlocksCleared, BlockRemoved, BlockStored
|
||||
from vllm.distributed.kv_events import (
|
||||
MEDIUM_GPU,
|
||||
AllBlocksCleared,
|
||||
BlockRemoved,
|
||||
BlockStored,
|
||||
)
|
||||
from vllm.lora.request import LoRARequest
|
||||
from vllm.multimodal.inputs import (
|
||||
MultiModalFeatureSpec,
|
||||
@@ -2454,6 +2459,118 @@ def test_block_removed_event_group_idx(group_id: int):
|
||||
assert event.group_idx == group_id
|
||||
|
||||
|
||||
def test_emit_cached_block_events():
|
||||
"""emit_cached_block_events emits one BlockStored for already-cached
|
||||
(reused) prefix blocks, carrying the correct group_idx /
|
||||
parent_block_hash / token_ids, and without mutating block state."""
|
||||
block_size = 4
|
||||
num_cached_blocks = 3
|
||||
kv_cache_group_id = 1
|
||||
num_tokens = block_size * 4 # 4 full blocks; reuse the first 3
|
||||
|
||||
pool = BlockPool(
|
||||
num_gpu_blocks=8,
|
||||
enable_caching=True,
|
||||
hash_block_size=block_size,
|
||||
enable_kv_cache_events=True,
|
||||
)
|
||||
|
||||
req = make_request(
|
||||
"req_emit_cached",
|
||||
prompt_token_ids=list(range(num_tokens)),
|
||||
block_size=block_size,
|
||||
hash_fn=sha256,
|
||||
)
|
||||
assert len(req.block_hashes) >= num_cached_blocks
|
||||
|
||||
# Snapshot block state to prove emit_cached_block_events does not mutate it.
|
||||
free_before = pool.get_num_free_blocks()
|
||||
assert len(pool.cached_block_hash_to_block) == 0
|
||||
|
||||
pool.emit_cached_block_events(
|
||||
request=req,
|
||||
num_cached_blocks=num_cached_blocks,
|
||||
block_size=block_size,
|
||||
kv_cache_group_id=kv_cache_group_id,
|
||||
)
|
||||
|
||||
# No block-state mutation: nothing allocated, nothing inserted into the
|
||||
# prefix-cache map.
|
||||
assert pool.get_num_free_blocks() == free_before
|
||||
assert len(pool.cached_block_hash_to_block) == 0
|
||||
|
||||
events = pool.take_events()
|
||||
assert len(events) == 1
|
||||
event = events[0]
|
||||
assert isinstance(event, BlockStored)
|
||||
|
||||
expected_hashes = [
|
||||
kv_cache_utils.maybe_convert_block_hash(req.block_hashes[i])
|
||||
for i in range(num_cached_blocks)
|
||||
]
|
||||
assert event.block_hashes == expected_hashes
|
||||
# Reused blocks start from block 0, so there is no parent block hash.
|
||||
assert event.parent_block_hash is None
|
||||
assert event.token_ids == list(req.all_token_ids[: num_cached_blocks * block_size])
|
||||
assert event.group_idx == kv_cache_group_id
|
||||
assert event.block_size == block_size
|
||||
assert event.medium == MEDIUM_GPU
|
||||
assert event.lora_id is None
|
||||
assert event.lora_name is None
|
||||
|
||||
|
||||
def test_emit_cached_block_events_disabled():
|
||||
"""No events are emitted when enable_kv_cache_events is False."""
|
||||
block_size = 4
|
||||
pool = BlockPool(
|
||||
num_gpu_blocks=8,
|
||||
enable_caching=True,
|
||||
hash_block_size=block_size,
|
||||
enable_kv_cache_events=False,
|
||||
)
|
||||
req = make_request(
|
||||
"req_emit_disabled",
|
||||
prompt_token_ids=list(range(block_size * 4)),
|
||||
block_size=block_size,
|
||||
hash_fn=sha256,
|
||||
)
|
||||
|
||||
pool.emit_cached_block_events(
|
||||
request=req,
|
||||
num_cached_blocks=3,
|
||||
block_size=block_size,
|
||||
kv_cache_group_id=0,
|
||||
)
|
||||
|
||||
assert pool.take_events() == []
|
||||
|
||||
|
||||
def test_emit_cached_block_events_zero_cached():
|
||||
"""No events are emitted when num_cached_blocks == 0."""
|
||||
block_size = 4
|
||||
pool = BlockPool(
|
||||
num_gpu_blocks=8,
|
||||
enable_caching=True,
|
||||
hash_block_size=block_size,
|
||||
enable_kv_cache_events=True,
|
||||
)
|
||||
req = make_request(
|
||||
"req_emit_zero",
|
||||
prompt_token_ids=list(range(block_size * 4)),
|
||||
block_size=block_size,
|
||||
hash_fn=sha256,
|
||||
)
|
||||
|
||||
pool.emit_cached_block_events(
|
||||
request=req,
|
||||
num_cached_blocks=0,
|
||||
block_size=block_size,
|
||||
kv_cache_group_id=0,
|
||||
)
|
||||
|
||||
assert pool.take_events() == []
|
||||
|
||||
|
||||
def test_eagle_enabled_removes_last_block():
|
||||
"""Verify Eagle does NOT remove blocks when request
|
||||
length is divisible by block size."""
|
||||
|
||||
Reference in New Issue
Block a user