[Bugfix][HiSparse] Preserve host prefix publication after request completion (#59007)

Signed-off-by: Yueh-Ting Chen <yueh.ting.chen@gmail.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Yueh-Ting (eop) Chen
2026-09-30 16:14:51 +00:00
committed by GitHub
co-authored by Matthew Bonanni
parent 279012f94d
commit ff1b87cca2
3 changed files with 223 additions and 0 deletions
+75
View File
@@ -1178,6 +1178,81 @@ def test_hisparse_finished_request_leaves_free_copies():
assert not coordinator.spills_to_send
@pytest.mark.parametrize(
"status,expected_cached_pages",
[
(RequestStatus.FINISHED_STOPPED, 2),
(RequestStatus.FINISHED_ABORTED, 2),
(RequestStatus.RUNNING, 0),
],
ids=["stopped", "aborted", "running"],
)
def test_hisparse_cleanup_publishes_only_finalized_terminal_pages(
status, expected_cached_pages
):
"""Only terminal cleanup may publish completed, finalized pages."""
manager = make_hisparse_kv_cache_manager(32, 16, enable_caching=True)
coordinator = get_hisparse_coordinator(manager)
request = make_request("terminal", list(range(33)), HISPARSE_BLOCK_SIZE, sha256)
request.spec_token_ids = list(range(33, 49))
assert manager.allocate_slots(request, 49, num_lookahead_tokens=16) is not None
host_blocks = list(manager.get_blocks(request.request_id).blocks[0])
command = coordinator.build_offload_command()
counts = {transfer.transfer_id: 1 for transfer in command.page_transfers}
assert len(counts) == 2
request.status = status
manager.free(request)
coordinator.update_spills(counts, counts)
assert all(
block.block_hash is not None for block in host_blocks[:expected_cached_pages]
)
assert all(
block.block_hash is None for block in host_blocks[expected_cached_pages:]
)
assert all(block.ref_cnt == 0 for block in host_blocks)
assert not coordinator.has_pending_work()
def test_hisparse_terminal_prefix_waits_for_all_workers_and_preserves_identity():
"""Old completions retain their pages through pressure and request-ID reuse."""
manager = make_hisparse_kv_cache_manager(32, 8, enable_caching=True)
coordinator = get_hisparse_coordinator(manager)
host_pool = coordinator.get_host_block_pool()
tokens = list(range(2 * HISPARSE_BLOCK_SIZE + 1))
request = make_request("reused", tokens, HISPARSE_BLOCK_SIZE, sha256)
assert manager.allocate_slots(request, len(tokens)) is not None
old_blocks = list(manager.get_blocks(request.request_id).blocks[0])
command = coordinator.build_offload_command()
counts = {transfer.transfer_id: 2 for transfer in command.page_transfers}
partial = dict.fromkeys(counts, 1)
first_completions = partial.copy()
first_completions[next(iter(counts))] = 2
coordinator.update_spills(counts, first_completions)
request.status = RequestStatus.FINISHED_LENGTH_CAPPED
manager.free(request)
replacement = make_request(
"reused", list(range(100, 117)), HISPARSE_BLOCK_SIZE, sha256
)
assert manager.allocate_slots(replacement, 17) is not None
new_blocks = list(manager.get_blocks(replacement.request_id).blocks[0])
pressure = host_pool.get_new_blocks(host_pool.get_num_free_blocks())
assert not {block.block_id for block in old_blocks[:2]} & {
block.block_id for block in pressure + new_blocks
}
assert all(block.block_hash is None for block in old_blocks)
coordinator.update_spills({}, partial)
repeated = make_request("probe", tokens, HISPARSE_BLOCK_SIZE, sha256)
assert manager.get_computed_blocks(repeated)[1] == 2 * HISPARSE_BLOCK_SIZE
assert all(block.block_hash is None for block in new_blocks)
assert all(block.ref_cnt == 0 for block in old_blocks[:2])
coordinator.update_spills({}, partial)
assert all(block.ref_cnt == 0 for block in old_blocks[:2])
host_pool.free_blocks(pressure)
manager.free(replacement)
_publish_hisparse_pages(manager)
assert not coordinator.has_pending_work()
def test_hisparse_reset_prefix_cache_drops_copies():
"""Reset must forget GPU copies along with the host hashes they mirror."""
manager = make_hisparse_kv_cache_manager(32, 16, enable_caching=True)
+99
View File
@@ -1,6 +1,8 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import time
import prometheus_client
import pytest
import torch
@@ -18,6 +20,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.multi_connector import (
MultiConnector,
)
from vllm.platforms import current_platform
from vllm.sampling_params import SamplingParams
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
from vllm.v1.hisparse.coordinator import get_hisparse_coordinator
@@ -301,3 +304,99 @@ def test_hisparse_host_exhaustion_defers_requests(
assert checked_steps > 0
assert refusals > 0
assert actual == expected
@pytest.mark.skipif(
not current_platform.is_cuda(), reason="HiSparse requires NVIDIA CUDA"
)
@fork_new_process_for_each_test
def test_hisparse_terminal_prefix_reuse(
monkeypatch: pytest.MonkeyPatch,
vllm_runner: type[VllmRunner],
):
"""Finished requests publish mirrored host KV after a late acknowledgement."""
capability = current_platform.get_device_capability()
if capability is None or capability.major < 9:
pytest.skip("Sparse MLA requires Hopper or newer")
monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
monkeypatch.setenv("VLLM_DEEP_GEMM_WARMUP", "skip")
target = [1000 + i % 64 for i in range(257)]
sampling = SamplingParams(temperature=0, max_tokens=1, ignore_eos=True)
with vllm_runner(
MODEL,
load_format="dummy",
hf_overrides=_shrink_config,
attention_config=AttentionConfig(
hisparse_config=HiSparseConfig(
device_buffer_size=512, eager_host_mirror=True
)
),
kv_transfer_config=KVTransferConfig(
kv_connector="HiSparseConnector",
kv_role="kv_both",
kv_connector_extra_config={"host_pool_gib": 0.01},
),
block_size=64,
max_model_len=320,
max_num_batched_tokens=512,
max_num_seqs=2,
num_gpu_blocks_override=128,
enable_prefix_caching=True,
enable_chunked_prefill=True,
enforce_eager=True,
) as runner:
engine = runner.llm.llm_engine.engine_core.engine_core
coordinator = get_hisparse_coordinator(engine.scheduler.kv_cache_manager)
worker = _get_hisparse_worker(runner)
original_updates = worker.take_transfer_updates
original_free = coordinator.free
held_completions: list[int] = []
update_calls = 0
terminal_pending_pages = 0
def take_updates():
nonlocal update_calls
with gpu_sync_allowed():
torch.accelerator.synchronize()
enqueued, completed = original_updates()
held_completions.extend(completed)
update_calls += 1
if update_calls <= 2:
return enqueued, []
completed = held_completions.copy()
held_completions.clear()
return enqueued, completed
def free(request_id):
nonlocal terminal_pending_pages
state = coordinator.request_states.get(request_id)
if state is not None:
terminal_pending_pages += len(state.pending_pages)
original_free(request_id)
def drain_pending_work():
deadline = time.monotonic() + 10
while coordinator.has_pending_work():
assert time.monotonic() < deadline, "HiSparse transfers did not drain"
runner.llm.llm_engine.step()
monkeypatch.setattr(worker, "take_transfer_updates", take_updates)
monkeypatch.setattr(coordinator, "free", free)
[first] = runner.llm.generate(
[{"prompt_token_ids": target}], sampling, use_tqdm=False
)
assert terminal_pending_pages == 4
assert first.num_cached_tokens == 0
assert coordinator.has_pending_work()
drain_pending_work()
assert not coordinator.has_pending_work()
assert not held_completions
assert update_calls > 2
[repeated] = runner.llm.generate(
[{"prompt_token_ids": target}], sampling, use_tqdm=False
)
assert repeated.num_cached_tokens == 256
assert repeated.outputs[0].token_ids == first.outputs[0].token_ids
drain_pending_work()
assert not coordinator.has_pending_work()
+49
View File
@@ -44,6 +44,8 @@ class _PendingPublication:
num_pages: int
retention_interval: int | None
replay_boundaries: Sequence[int]
detached_blocks: list[KVCacheBlock] | None = None
num_cached_blocks: int = 0
@dataclass
@@ -724,6 +726,10 @@ class HiSparseCoordinator:
if pending.restore:
self._record_copy(request_id, page_idx, pending.host_block)
completed_request_ids.add(request_id)
elif pending.request_state.publication is not None:
if not pending.restore:
pending.request_state.valid_pages.add(page_idx)
self._publish_detached_blocks(pending.request_state)
assert self.host_manager is not None
self.host_manager.block_pool.free_blocks([pending.host_block])
for request_id in completed_request_ids:
@@ -732,12 +738,55 @@ class HiSparseCoordinator:
state.ready_prefix_pages += 1
self._publish_host_blocks_if_ready(request_id)
def _publish_detached_blocks(self, state: _HiSparseRequestState) -> None:
publication = state.publication
if publication is None or state.pending_pages:
return
blocks = publication.detached_blocks
if blocks is None:
return
assert self.host_manager is not None
while state.ready_prefix_pages in state.valid_pages:
state.ready_prefix_pages += 1
# The host group uses full-attention retention. Only sealed pages with
# completed copies can outlive the request, including on cancellation.
self.host_manager.block_pool.cache_full_blocks(
request=publication.request,
blocks=blocks,
num_cached_blocks=publication.num_cached_blocks,
num_full_blocks=min(state.ready_prefix_pages, publication.num_pages),
block_size=self.host_manager.block_size,
kv_cache_group_id=self.host_manager.kv_cache_group_id,
)
self.host_manager.block_pool.free_blocks(reversed(blocks))
state.publication = None
def free(self, request_id: str) -> None:
"""Detach the request; its clean pages stay readable copies in the pool."""
self._pending_imports.pop(request_id, None)
state = self.request_states.pop(request_id, None)
if state is None:
return
publication = state.publication
if (
publication is not None
and publication.request.is_finished()
and state.pending_pages
):
assert self.host_manager is not None
publication.detached_blocks = self.host_manager.req_to_blocks[request_id][
: publication.num_pages
]
publication.num_cached_blocks = self.host_manager.num_cached_block.get(
request_id, 0
)
# Completed pages need leases too: another allocation may evict
# them before the remaining copies finish and the prefix publishes.
self.host_manager.block_pool.touch(publication.detached_blocks)
else:
# Preemption can cancel a scheduled forward after its spills were
# planned, so its copies do not prove that the KV was computed.
state.publication = None
for manager in self.resident_managers:
blocks = manager.req_to_blocks.get(request_id)
if not blocks: