[Bugfix][KV Offload] Track cache recency once per request (#51787)

Signed-off-by: mindungil <alswnsrlf12@naver.com>
Co-authored-by: Or Ozeri <oro@il.ibm.com>
This commit is contained in:
MINJUN GIL
2026-09-16 10:47:26 +03:00
committed by GitHub
co-authored by Or Ozeri
parent ab35354c21
commit f12fe10c5c
10 changed files with 916 additions and 169 deletions
@@ -258,6 +258,9 @@ def test_partial_tail_store_uses_attention_and_recurrent_cow_sources():
jobs = scheduler._build_partial_tail_store_jobs(output)
assert len(jobs) == 1
offered_keys = scheduler.manager.prepare_store.call_args.args[0]
req_context = scheduler._req_status["req"].req_context
assert all(req_context.get_offload_key_position(key) == 28 for key in offered_keys)
[job_id] = jobs
src_spec = jobs[job_id].src_spec
assert isinstance(src_spec, GPULoadStoreSpec)
@@ -380,6 +383,13 @@ def test_normal_store_excludes_align_mode_mamba_sources():
req_status.group_states[0].block_ids[:] = [11]
req_status.group_states[1].block_ids[:] = [99]
req_status.update_offload_keys()
for group_config, group_state in zip(
scheduler.config.kv_group_configs, req_status.group_states
):
assert (
req_status.req_context.get_offload_key_position(group_state.offload_keys[0])
== group_config.tokens_per_chunk
)
scheduler.manager.prepare_store.side_effect = lambda keys, req_context: (
generate_store_output(keys)
)
@@ -921,7 +931,6 @@ def test_offloading_connector(request_runner, async_scheduling: bool):
runner.run(decoded_tokens=[0] * (tokens_per_chunk + 1))
# 1 more block (+ token for kicking off offloading)
# now check touch was called with all 6 blocks
runner.manager.prepare_store.side_effect = lambda keys, req_context: (
generate_store_output(keys)
)
@@ -929,9 +938,6 @@ def test_offloading_connector(request_runner, async_scheduling: bool):
decoded_tokens=[0] * (tokens_per_chunk + 1),
expected_stored=(15, 16, 17),
)
runner.manager.touch.assert_called()
block_hashes1 = list(runner.manager.touch.call_args.args[0])
assert len(block_hashes1) == 6
# terminate request
runner.run(decoded_tokens=[EOS_TOKEN_ID])
@@ -939,13 +945,7 @@ def test_offloading_connector(request_runner, async_scheduling: bool):
# create a new request differing only on the last token
runner.new_request(token_ids=[0] * (tokens_per_chunk * 6 - 1) + [1])
runner.run(decoded_tokens=[0])
runner.manager.touch.assert_called()
block_hashes2 = list(runner.manager.touch.call_args.args[0])
assert len(block_hashes2) == 6
# verify hashes are the same, except for the last block
assert block_hashes1[:5] == block_hashes2[:5]
assert block_hashes1[5] != block_hashes2[5]
runner.manager.touch.assert_not_called()
# terminate request
runner.run(
@@ -1353,14 +1353,7 @@ def test_two_groups_full_and_sliding_window(request_runner, async_scheduling: bo
generate_store_output(keys)
)
runner.run(decoded_tokens=[0])
# _touch called from get_num_new_matched_tokens (2 groups) and
# _get_reqs_to_store (2 groups) → 4 touch calls total.
touch_calls = runner.manager.touch.call_args_list
assert len(touch_calls) == 4
assert len(touch_calls[0].args[0]) == 3
assert len(touch_calls[1].args[0]) == 3
assert len(touch_calls[2].args[0]) == 3
assert len(touch_calls[3].args[0]) == 3
runner.manager.touch.assert_not_called()
# store 3 more block
runner.manager.prepare_store.side_effect = lambda keys, req_context: (
@@ -1371,9 +1364,7 @@ def test_two_groups_full_and_sliding_window(request_runner, async_scheduling: bo
expected_stored=(0, 1, 2, 3, 4, 5),
)
# touch called from _get_reqs_to_store * 3 blocks, once for each group
touch_calls = runner.manager.touch.call_args_list
assert len(touch_calls) == 6
runner.manager.touch.assert_not_called()
# EOS lands in the last slot of the 7th block (offset 6). No forward pass
# writes that slot, so the finishing step declines the block.
@@ -1392,13 +1383,7 @@ def test_two_groups_full_and_sliding_window(request_runner, async_scheduling: bo
expected_loaded=((0, 0), (0, 1), (0, 2), (1, 1), (1, 2)),
)
# 2 touch calls from get_num_new_matched_tokens (2 groups)
touch_calls = runner.manager.touch.call_args_list
assert len(touch_calls) == 2
# full attention group touched all 3 blocks
assert len(touch_calls[0].args[0]) == 3
# sliding window group touched just the last 2 blocks
assert len(touch_calls[1].args[0]) == 2
runner.manager.touch.assert_not_called()
# 3 blocks are hit on GPU [0, 1, 2]
# 1 block loaded [3,]
@@ -1464,15 +1449,7 @@ def test_two_groups_different_block_sizes(request_runner, async_scheduling: bool
generate_store_output(keys)
)
runner.run(decoded_tokens=[0])
# _touch called from get_num_new_matched_tokens (2 groups) and
# _get_reqs_to_store (2 groups) → 4 touch calls total.
# Group 0 has 2 offload keys, group 1 has 1.
touch_calls = runner.manager.touch.call_args_list
assert len(touch_calls) == 4
assert len(touch_calls[0].args[0]) == 2
assert len(touch_calls[1].args[0]) == 1
assert len(touch_calls[2].args[0]) == 2
assert len(touch_calls[3].args[0]) == 1
runner.manager.touch.assert_not_called()
# Get to 31 tokens
# No further blocks offloaded
@@ -1482,11 +1459,7 @@ def test_two_groups_different_block_sizes(request_runner, async_scheduling: bool
# Group 0 blocks: [0, 1], ending_token_offset = 24
# Group 1 blocks: [0, 1], ending_token_offset = 32
runner.run(decoded_tokens=[0])
# _get_reqs_to_store touch: only group 1 has a new block to store
touch_calls = runner.manager.touch.call_args_list
assert len(touch_calls) == 2
assert len(touch_calls[0].args[0]) == 2
assert len(touch_calls[1].args[0]) == 2
runner.manager.touch.assert_not_called()
# Get to 35 tokens
# No further blocks offloaded
@@ -1496,11 +1469,7 @@ def test_two_groups_different_block_sizes(request_runner, async_scheduling: bool
# Group 0 blocks: [0, 1, 2], ending_token_offset = 36
# Group 1 blocks: [0, 1], ending_token_offset = 32
runner.run(decoded_tokens=[0])
# _get_reqs_to_store touch: only group 0 has a new block to store
touch_calls = runner.manager.touch.call_args_list
assert len(touch_calls) == 2
assert len(touch_calls[0].args[0]) == 3
assert len(touch_calls[1].args[0]) == 2
runner.manager.touch.assert_not_called()
# Get to 47 tokens
# No further blocks offloaded
@@ -1510,11 +1479,7 @@ def test_two_groups_different_block_sizes(request_runner, async_scheduling: bool
# Group 0 blocks: [0, 1, 2, 3], ending_token_offset = 4
# Group 1 blocks: [0, 1, 2], ending_token_offset = 48
runner.run(decoded_tokens=[0])
# _get_reqs_to_store touch: both groups have a new block, each with 1 key
touch_calls = runner.manager.touch.call_args_list
assert len(touch_calls) == 2
assert len(touch_calls[0].args[0]) == 4
assert len(touch_calls[1].args[0]) == 3
runner.manager.touch.assert_not_called()
runner.run(decoded_tokens=[0], expected_stored=((0, 3), (1, 2)))
+428 -8
View File
@@ -23,6 +23,7 @@ from vllm.v1.kv_offload.cpu.common import (
)
from vllm.v1.kv_offload.cpu.manager import CPUOffloadingManager
from vllm.v1.kv_offload.cpu.policies.arc import ARCCachePolicy
from vllm.v1.kv_offload.cpu.policies.lru import LRUCachePolicy
def make_req_context(
@@ -430,31 +431,34 @@ def test_cpu_manager():
assert cpu_manager.lookup(to_key(0), _EMPTY_REQ_CTX) is LookupResult.MISS
# prepare load [2, 3]
prepare_load_output = cpu_manager.prepare_load(to_keys([2, 3]), _EMPTY_REQ_CTX)
load_ctx = make_req_context("load-2-3")
prepare_load_output = cpu_manager.prepare_load(to_keys([2, 3]), load_ctx)
verify_load_output(prepare_load_output, [1, 2])
# prepare store with no space ([2, 3] is being loaded)
assert cpu_manager.prepare_store(to_keys([6, 7, 8]), _EMPTY_REQ_CTX) is None
# complete load [2, 3]. Load changes the eviction list, making 2, 3 recent.
cpu_manager.complete_load(to_keys([2, 3]), _EMPTY_REQ_CTX)
cpu_manager.complete_load(to_keys([2, 3]), load_ctx)
cpu_manager.on_request_finished(load_ctx)
# prepare store [6, 7, 8] -> evicts [4, 5, 2] (oldest)
# prepare store [6, 7, 8] -> evicts [4, 5, 3] (oldest). Within
# the accessed prefix [2, 3], the tail is less valuable than the head.
prepare_store_output = cpu_manager.prepare_store(to_keys([6, 7, 8]), _EMPTY_REQ_CTX)
verify_store_output(
prepare_store_output,
ExpectedPrepareStoreOutput(
keys_to_store=[6, 7, 8],
store_chunk_ids=[1, 0, 3],
evicted_keys=[4, 5, 2],
store_chunk_ids=[2, 0, 3],
evicted_keys=[4, 5, 3],
),
)
# complete store [6, 7, 8]
cpu_manager.complete_store(to_keys([6, 7, 8]), _EMPTY_REQ_CTX)
# touch [3, 6, 7] (move to end of LRU order)
cpu_manager.touch(to_keys([3, 6, 7]), _EMPTY_REQ_CTX)
# touch [2, 6, 7] (move to end of LRU order)
cpu_manager.touch(to_keys([2, 6, 7]), _EMPTY_REQ_CTX)
# prepare store [7, 9] -> evicts [8] (oldest following previous touch)
prepare_store_output = cpu_manager.prepare_store(to_keys([9]), _EMPTY_REQ_CTX)
@@ -477,7 +481,7 @@ def test_cpu_manager():
verify_events(
cpu_manager.take_events(),
expected_stores=({3, 4, 5}, {6, 7, 8}),
expected_evictions=({4, 5, 2}, {8}),
expected_evictions=({4, 5, 3}, {8}),
)
@@ -518,6 +522,65 @@ def test_prepare_load_preserves_key_order():
manager.complete_load([key_a, key_b, key_c], _EMPTY_REQ_CTX) # order irrelevant
def test_lru_batch_eviction_failure_is_atomic():
manager = make_cpu_manager(num_chunks=4, cache_policy="lru")
policy = manager._policy
assert isinstance(policy, LRUCachePolicy)
keys = to_keys([1, 2, 3, 4])
assert manager.prepare_store(keys, _EMPTY_REQ_CTX) is not None
manager.complete_store(keys, _EMPTY_REQ_CTX)
protected = set(keys[1:])
assert policy.evict(2, protected) is None
assert all(manager.lookup(key, _EMPTY_REQ_CTX) is LookupResult.HIT for key in keys)
evicted = policy.evict(1, protected)
assert evicted is not None
assert [key for key, _ in evicted] == [keys[0]]
def test_lru_repeated_pin_cycles_compact_lazy_heap_entries():
manager = make_cpu_manager(num_chunks=1, cache_policy="lru")
policy = manager._policy
assert isinstance(policy, LRUCachePolicy)
key = to_key(1)
assert manager.prepare_store([key], _EMPTY_REQ_CTX) is not None
manager.complete_store([key], _EMPTY_REQ_CTX)
for _ in range(128):
manager.prepare_load([key], _EMPTY_REQ_CTX)
manager.complete_load([key], _EMPTY_REQ_CTX)
assert len(policy._heap) < 64
evicted = policy.evict(1, set())
assert evicted is not None
assert [evicted_key for evicted_key, _ in evicted] == [key]
def test_reset_discards_stale_request_access_classification():
manager = make_cpu_manager(num_chunks=1, cache_policy="arc")
policy = manager._policy
assert isinstance(policy, ARCCachePolicy)
key = to_key(1)
resumed_ctx = make_req_context("resumed")
assert manager.prepare_store([key], resumed_ctx) is not None
manager.complete_store([key], resumed_ctx)
manager.reset_cache()
replacement_ctx = make_req_context("replacement")
assert manager.prepare_store([key], replacement_ctx) is not None
manager.complete_store([key], replacement_ctx)
manager.on_request_finished(replacement_ctx)
assert key in policy.t1
manager.prepare_load([key], resumed_ctx)
manager.complete_load([key], resumed_ctx)
manager.on_request_finished(resumed_ctx)
assert key in policy.t2
class TestARCPolicy:
"""Unit tests for CPUOffloadingManager with ARC eviction policy."""
@@ -1042,6 +1105,48 @@ def test_filter_reused_manager_oversized_offer_makes_progress():
assert stored_keys == set(keys)
def test_store_threshold_does_not_hide_ready_reuse():
manager = make_cpu_manager(
num_chunks=2,
cache_policy="lru",
store_threshold=2,
max_tracker_size=2,
)
head, tail, noise_1, noise_2, replacement = to_keys(list(range(5)))
seed_ctx = make_req_context("seed")
output = manager.prepare_store([head, tail], seed_ctx)
assert output is not None
assert not output.keys_to_store
output = manager.prepare_store([head, tail], seed_ctx)
assert output is not None
assert output.keys_to_store == [head, tail]
manager.complete_store([head, tail], seed_ctx)
manager.on_request_finished(seed_ctx)
# Age both resident keys out of the admission tracker without storing the
# noise keys. Reusing the tail recreates its counter at one, below the
# threshold, but must still refresh its cache recency.
output = manager.prepare_store([noise_1, noise_2], make_req_context("noise"))
assert output is not None
assert not output.keys_to_store
reuse_ctx = make_req_context("reuse")
skipped_before_reuse = manager.stores_skipped_in_current_batch
output = manager.prepare_store([tail], reuse_ctx)
assert output is not None
assert not output.keys_to_store
assert manager.stores_skipped_in_current_batch == skipped_before_reuse
manager.on_request_finished(reuse_ctx)
replacement_ctx = make_req_context("replacement")
output = manager.prepare_store([replacement], replacement_ctx)
assert output is not None
assert not output.keys_to_store
output = manager.prepare_store([replacement], replacement_ctx)
assert output is not None
assert output.evicted_keys == [head]
def test_evictable_cache_chunk_count():
"""
Verifies _num_evictable_cache_chunks is maintained correctly through the
@@ -1156,3 +1261,318 @@ def test_touch_forwards_req_context_to_policy(monkeypatch):
assert len(received) == 1
assert received[0][0] == keys
assert received[0][1] is ctx
@pytest.mark.parametrize("finish_before_completion", [False, True])
def test_request_finish_orders_lru_prefix_independent_of_store_completion(
finish_before_completion: bool,
):
"""A transfer's completion order must not make the prefix head LRU."""
manager = make_cpu_manager(num_chunks=4, cache_policy="lru")
prefix = to_keys([1, 2, 3, 4])
ctx = make_req_context("prefix")
output = manager.prepare_store(prefix, ctx)
assert output is not None
if finish_before_completion:
manager.on_request_finished(ctx)
manager.complete_store(set(prefix), ctx)
if not finish_before_completion:
manager.on_request_finished(ctx)
output = manager.prepare_store(to_keys([5]), make_req_context("evict"))
assert output is not None
assert output.evicted_keys == to_keys([4])
@pytest.mark.parametrize("finish_before_completion", [False, True])
def test_request_finish_orders_lru_prefix_independent_of_load_completion(
finish_before_completion: bool,
):
manager = make_cpu_manager(num_chunks=4, cache_policy="lru")
prefix = to_keys([1, 2, 3, 4])
seed_ctx = make_req_context("seed")
output = manager.prepare_store(prefix, seed_ctx)
assert output is not None
manager.complete_store(prefix, seed_ctx)
reuse_ctx = make_req_context("reuse")
manager.prepare_load(prefix, reuse_ctx)
if finish_before_completion:
manager.on_request_finished(reuse_ctx)
manager.complete_load(set(prefix), reuse_ctx)
if not finish_before_completion:
manager.on_request_finished(reuse_ctx)
output = manager.prepare_store(to_keys([5]), make_req_context("evict"))
assert output is not None
assert output.evicted_keys == to_keys([4])
def test_late_old_completion_does_not_override_newer_request_recency():
manager = make_cpu_manager(num_chunks=4, cache_policy="lru")
old_keys = to_keys([1, 2])
new_keys = to_keys([3, 4])
old_ctx = make_req_context("old")
assert manager.prepare_store(old_keys, old_ctx) is not None
manager.on_request_finished(old_ctx)
new_ctx = make_req_context("new")
assert manager.prepare_store(new_keys, new_ctx) is not None
manager.complete_store(new_keys, new_ctx)
manager.on_request_finished(new_ctx)
# Physical completion is late, but the logical access is still older.
manager.complete_store(old_keys, old_ctx)
output = manager.prepare_store(to_keys([5]), make_req_context("evict"))
assert output is not None
assert output.evicted_keys == [old_keys[-1]]
def test_arc_counts_reuse_once_and_keeps_insertions_in_t1():
manager = make_cpu_manager(num_chunks=4, cache_policy="arc")
policy = manager._policy
assert isinstance(policy, ARCCachePolicy)
prefix = to_keys([1, 2, 3, 4])
seed_ctx = make_req_context("seed")
assert manager.prepare_store(prefix, seed_ctx) is not None
manager.complete_store(prefix, seed_ctx)
# Repeated store offers from the same request must not become ARC hits.
assert manager.prepare_store(prefix, seed_ctx) is not None
# Nor should an internal read used to cascade the new chunks to a tier.
manager.prepare_load(prefix, seed_ctx)
manager.complete_load(prefix, seed_ctx)
manager.on_request_finished(seed_ctx)
assert list(policy.t1) == list(reversed(prefix))
assert not policy.t2
reuse_ctx = make_req_context("reuse")
manager.prepare_load(prefix, reuse_ctx)
manager.complete_load(prefix, reuse_ctx)
# A duplicate internal pin, as used by tiering cascades, is deduplicated.
manager.prepare_load(prefix, reuse_ctx)
manager.complete_load(prefix, reuse_ctx)
manager.on_request_finished(reuse_ctx)
assert not policy.t1
assert list(policy.t2) == list(reversed(prefix))
def test_arc_reuse_wins_if_same_request_later_reinserts_key():
manager = make_cpu_manager(num_chunks=1, cache_policy="arc")
policy = manager._policy
assert isinstance(policy, ARCCachePolicy)
key_1, key_2 = to_keys([1, 2])
seed_ctx = make_req_context("seed")
assert manager.prepare_store([key_1], seed_ctx) is not None
manager.complete_store([key_1], seed_ctx)
manager.on_request_finished(seed_ctx)
reuse_ctx = make_req_context("reuse-and-reinsert")
manager.prepare_load([key_1], reuse_ctx)
manager.complete_load([key_1], reuse_ctx)
other_ctx = make_req_context("evict-reused-key")
assert manager.prepare_store([key_2], other_ctx) is not None
manager.complete_store([key_2], other_ctx)
manager.on_request_finished(other_ctx)
assert manager.prepare_store([key_1], reuse_ctx) is not None
manager.complete_store([key_1], reuse_ctx)
manager.on_request_finished(reuse_ctx)
assert key_1 not in policy.t1
assert list(policy.t2) == [key_1]
def test_request_finish_forwards_grouped_deduplicated_accesses(monkeypatch):
manager = make_cpu_manager(num_chunks=5)
existing = make_offload_key(b"existing", 0)
group_0_new = make_offload_key(b"group-0-new", 0)
group_1_head = make_offload_key(b"group-1-head", 1)
group_1_tail = make_offload_key(b"group-1-tail", 1)
seed_ctx = make_req_context("seed")
assert manager.prepare_store([existing], seed_ctx) is not None
manager.complete_store([existing], seed_ctx)
ctx = make_req_context("grouped")
# Group 1 is observed first, but policy finalization follows group index.
offered = [group_1_head, existing, group_0_new]
assert manager.prepare_store(offered, ctx) is not None
# Later offers extend a group's prefix and may repeat earlier keys.
assert (
manager.prepare_store([group_0_new, group_1_head, group_1_tail], ctx)
is not None
)
received = []
def on_request_finished(key_groups, insertion_only_keys, reused_keys, req_context):
received.append((key_groups, insertion_only_keys, reused_keys, req_context))
monkeypatch.setattr(manager._policy, "on_request_finished", on_request_finished)
manager.on_request_finished(ctx)
manager.on_request_finished(ctx)
assert received == [
(
((existing, group_0_new), (group_1_head, group_1_tail)),
{group_0_new, group_1_head, group_1_tail},
{existing},
ctx,
)
]
def test_failed_store_records_access_once_and_uses_store_miss_hook(monkeypatch):
manager = make_cpu_manager(num_chunks=1)
resident, candidate = to_keys([1, 2])
seed_ctx = make_req_context("seed")
assert manager.prepare_store([resident], seed_ctx) is not None
manager.complete_store([resident], seed_ctx)
pin_ctx = make_req_context("pin")
manager.prepare_load([resident], pin_ctx)
record_calls = 0
original_record = manager._record_request_cache_access
def record_access(*args, **kwargs):
nonlocal record_calls
record_calls += 1
return original_record(*args, **kwargs)
store_miss_calls = []
def on_store_miss(keys, req_context):
store_miss_calls.append((list(keys), req_context))
finalized = []
def on_request_finished(key_groups, insertion_only_keys, reused_keys, req_context):
finalized.append((key_groups, insertion_only_keys, reused_keys, req_context))
monkeypatch.setattr(manager, "_record_request_cache_access", record_access)
monkeypatch.setattr(manager._policy, "on_store_miss", on_store_miss)
monkeypatch.setattr(manager._policy, "on_request_finished", on_request_finished)
ctx = make_req_context("failed-store")
assert manager.prepare_store([candidate, resident], ctx) is None
assert record_calls == 1
assert store_miss_calls == [([candidate], ctx)]
manager.on_request_finished(ctx)
assert finalized == [
(((candidate, resident),), set(), {resident}, ctx),
]
manager.complete_load([resident], pin_ctx)
def test_arc_ghost_hit_adapts_once_per_request_before_insertion():
manager = make_cpu_manager(num_chunks=2, cache_policy="arc")
policy = manager._policy
assert isinstance(policy, ARCCachePolicy)
keys = to_keys([1, 2, 3])
seed_ctx = make_req_context("seed")
assert manager.prepare_store(keys[:2], seed_ctx) is not None
manager.complete_store(keys[:2], seed_ctx)
manager.on_request_finished(seed_ctx)
evict_ctx = make_req_context("create-ghost")
output = manager.prepare_store([keys[2]], evict_ctx)
assert output is not None
[ghost_key] = output.evicted_keys
manager.complete_store([keys[2]], evict_ctx)
manager.on_request_finished(evict_ctx)
assert ghost_key in policy.b1
resident_keys = [key for key in keys if key != ghost_key]
pin_ctx = make_req_context("pin-residents")
manager.prepare_load(resident_keys, pin_ctx)
ghost_ctx = make_req_context("ghost-hit")
target_before = policy.target_t1_size
assert manager.prepare_store([ghost_key], ghost_ctx) is None
target_after_first_offer = policy.target_t1_size
assert target_after_first_offer > target_before
assert manager.prepare_store([ghost_key], ghost_ctx) is None
assert policy.target_t1_size == target_after_first_offer
manager.complete_load(resident_keys, pin_ctx)
output = manager.prepare_store([ghost_key], ghost_ctx)
assert output is not None
assert policy.target_t1_size == target_after_first_offer
manager.complete_store([ghost_key], ghost_ctx)
manager.on_request_finished(ghost_ctx)
assert ghost_key in policy.t1
assert ghost_key not in policy.t2
def test_arc_pending_chunk_is_not_a_frequency_hit():
manager = make_cpu_manager(num_chunks=2, cache_policy="arc")
policy = manager._policy
assert isinstance(policy, ARCCachePolicy)
key = to_key(1)
writer_ctx = make_req_context("writer")
assert manager.prepare_store([key], writer_ctx) is not None
observer_ctx = make_req_context("pending-observer")
output = manager.prepare_store([key], observer_ctx)
assert output is not None and not output.keys_to_store
manager.on_request_finished(observer_ctx)
chunk = policy.get(key)
assert chunk is not None and not chunk.is_ready
assert key in policy.t1
assert key not in policy.t2
def test_request_key_positions_override_store_observation_order():
manager = make_cpu_manager(num_chunks=2, cache_policy="lru")
head, tail = to_keys([1, 2])
ctx = make_req_context("out-of-order-store")
ctx.set_offload_key_position(head, 16)
ctx.set_offload_key_position(tail, 32)
# A backfill or partial-tail path may offer a later key first.
assert manager.prepare_store([tail], ctx) is not None
assert manager.prepare_store([head], ctx) is not None
manager.on_request_finished(ctx)
manager.complete_store({head, tail}, ctx)
output = manager.prepare_store([to_key(3)], make_req_context("evict-tail"))
assert output is not None
assert output.evicted_keys == [tail]
def test_request_key_positions_order_tails_across_kv_groups():
manager = make_cpu_manager(num_chunks=4, cache_policy="lru")
group_0_head = make_offload_key(b"group-0-head", 0)
group_0_tail = make_offload_key(b"group-0-tail", 0)
group_1_head = make_offload_key(b"group-1-head", 1)
group_1_tail = make_offload_key(b"group-1-tail", 1)
keys = [group_0_head, group_0_tail, group_1_head, group_1_tail]
ctx = make_req_context("hybrid-groups")
for key in (group_0_head, group_1_head):
ctx.set_offload_key_position(key, 16)
for key in (group_0_tail, group_1_tail):
ctx.set_offload_key_position(key, 32)
assert manager.prepare_store(keys, ctx) is not None
manager.complete_store(keys, ctx)
manager.on_request_finished(ctx)
output = manager.prepare_store(to_keys([10, 11]), make_req_context("evict"))
assert output is not None
assert set(output.evicted_keys) == {group_0_tail, group_1_tail}
@@ -40,6 +40,7 @@ from vllm.v1.kv_offload.base import (
TierMatcher,
make_offload_key,
)
from vllm.v1.kv_offload.cpu.policies.lru import LRUCachePolicy
from vllm.v1.kv_offload.tiering.base import (
JobResult,
SecondaryTierManager,
@@ -753,14 +754,64 @@ class TestTieringOffloadingManager:
# Touch chunks
self.manager.touch(chunks, _CTX)
# Verify touch was called on primary tier (check LRU order)
primary_keys = list(self.primary_tier._policy.evictable_chunks.keys())
assert primary_keys[-3:] == list(reversed(chunks))
# Verify touch was called on the primary tier (head is most recent).
policy = self.primary_tier._policy
assert isinstance(policy, LRUCachePolicy)
ranks = policy._ranks
assert ranks[chunks[2]] < ranks[chunks[1]] < ranks[chunks[0]]
# Verify touch was propagated to all secondary tiers
self.secondary_tier1.touch.assert_called_once_with(chunks, _CTX)
self.secondary_tier2.touch.assert_called_once_with(chunks, _CTX)
def test_request_order_survives_late_cascade_completion(self, manager_setup):
"""Cascade completion must not replace request order with I/O order."""
chunks = to_keys(range(5))
self._start_request()
output = self.manager.prepare_store(chunks, _CTX)
assert output is not None
self.manager.complete_store(chunks, _CTX, success=True)
# Finalize while both secondary-tier cascades still pin primary chunks.
self.manager.on_request_finished(_CTX)
self._simulate_on_schedule_end()
self._simulate_on_schedule_end()
evict_ctx = ReqContext(req_id="evict")
self._start_request(evict_ctx)
output = self.manager.prepare_store(to_keys([5]), evict_ctx)
assert output is not None
assert output.evicted_keys == [chunks[-1]]
def test_late_old_store_submission_does_not_override_newer_request_order(
self, manager_setup
):
"""Cascade submission after finish is not a new primary access."""
old_blocks = to_keys(range(2))
old_ctx = ReqContext(req_id="old")
self._start_request(old_ctx)
assert self.manager.prepare_store(old_blocks, old_ctx) is not None
self.manager.on_request_finished(old_ctx)
new_blocks = to_keys(range(2, 5))
new_ctx = ReqContext(req_id="new")
self._start_request(new_ctx)
assert self.manager.prepare_store(new_blocks, new_ctx) is not None
self.manager.complete_store(new_blocks, new_ctx, success=True)
self.manager.on_request_finished(new_ctx)
# The old GPU->primary write lands after the newer request finished.
# Its cascade pins must not turn it into the more recent request.
self.manager.complete_store(old_blocks, old_ctx, success=True)
self._simulate_on_schedule_end()
self._simulate_on_schedule_end()
evict_ctx = ReqContext(req_id="evict")
self._start_request(evict_ctx)
output = self.manager.prepare_store(to_keys([5]), evict_ctx)
assert output is not None
assert output.evicted_keys == [old_blocks[-1]]
def test_failed_store_no_cascade(self, manager_setup):
"""Test that failed GPU→primary store doesn't cascade."""
chunks = to_keys(range(3))
@@ -873,10 +924,10 @@ class TestTieringOffloadingManager:
job_metadata = self.secondary_tier1.submit_store.call_args.args[0]
assert job_metadata.req_context is ctx
def test_on_request_finished_delays_secondary_until_store_submitted(
def test_on_request_finished_delays_only_secondary_until_store_submitted(
self, manager_setup
):
"""Manager hook is eager; secondary hooks wait for cascade submission."""
"""Primary order commits immediately; secondary cleanup waits."""
chunks = to_keys(range(2))
ctx = ReqContext(req_id="req_delayed_secondary")
calls: list[tuple[str, str]] = []