[Bugfix][Offloader] Preserve prefetch static-buffer slot ownership (#54975)

Signed-off-by: Big2Wheel <1404207590@qq.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
yulun
2026-09-07 10:58:44 +08:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent de69e821b7
commit 1f778486fc
2 changed files with 57 additions and 1 deletions
@@ -0,0 +1,37 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from vllm.model_executor.offloader.prefetch import _get_next_prefetch_index
pytestmark = pytest.mark.skip_global_cleanup
@pytest.mark.parametrize(
("module_count", "prefetch_step", "expected_targets"),
[
(3, 1, [1, 2, 0]),
(3, 2, [2, 1, 0]),
(4, 2, [2, 3, 0, 1]),
(5, 2, [2, 3, 4, 1, 0]),
(3, 3, [0, 1, 2]),
(3, 4, [1, 2, 0]),
],
)
def test_next_prefetch_index_preserves_slot_ownership(
module_count: int,
prefetch_step: int,
expected_targets: list[int],
) -> None:
targets = [
_get_next_prefetch_index(index, prefetch_step, module_count)
for index in range(module_count)
]
assert targets == expected_targets
if prefetch_step < module_count:
assert all(
target % prefetch_step == index % prefetch_step
for index, target in enumerate(targets)
)
+20 -1
View File
@@ -124,6 +124,21 @@ class StaticBufferPool:
return self._buffers[key][slot_idx % self.slot_capacity]
def _get_next_prefetch_index(
index: int,
prefetch_step: int,
module_count: int,
) -> int:
"""Return a refill target that preserves static-buffer slot ownership."""
next_index = (index + prefetch_step) % module_count
if (
prefetch_step < module_count
and next_index % prefetch_step != index % prefetch_step
):
next_index = index % prefetch_step
return next_index
class PrefetchOffloader(BaseOffloader):
"""Prefetching-based offloader with group-based layer selection.
@@ -231,7 +246,11 @@ class PrefetchOffloader(BaseOffloader):
# Start prefetch for next layer (circular)
# mutates_args on output_tensor creates ordering dependency
next_index = (index + self.prefetch_step) % len(self.module_offloaders)
next_index = _get_next_prefetch_index(
index,
self.prefetch_step,
len(self.module_offloaders),
)
# Handle tuple output (e.g., (hidden_states, residual))
if isinstance(output, tuple):
torch.ops.vllm.start_prefetch(output[0], next_index)