[Sleep] Retain draft-private weights with existing parameter selector (#61046)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
aoshen02
2026-10-11 23:30:56 +08:00
committed by GitHub
co-authored by Codex
parent 3e8e1364ff
commit 239a3fa3a8
4 changed files with 97 additions and 6 deletions
+42
View File
@@ -19,6 +19,48 @@ from vllm.v1.worker.startup_plan import (
)
def test_sleep_preservation_restores_private_draft_not_shared_target():
target = torch.nn.Linear(2, 2, bias=False)
draft = torch.nn.Module()
draft.private = torch.nn.Linear(2, 2, bias=False)
draft.shared = target.weight
expected = draft.private.weight.detach().clone()
pointer = draft.private.weight.data_ptr()
worker = object.__new__(gpu_worker.Worker)
worker.get_draft_model = lambda: draft
assert set(worker._get_sleep_parameters(target)) == {
"weight",
"draft.private.weight",
}
worker._sleep_saved_parameters = {"draft.private.weight": expected}
with torch.no_grad():
target.weight.fill_(3.0)
draft.private.weight.zero_()
worker._restore_sleep_parameters(target)
torch.testing.assert_close(draft.private.weight, expected)
assert draft.private.weight.data_ptr() == pointer
assert torch.all(target.weight == 3.0)
assert worker._sleep_saved_parameters == {}
@pytest.mark.parametrize("has_shared_draft", [False, True])
def test_sleep_preservation_allows_empty_draft_but_rejects_target_typo(
has_shared_draft,
):
target = torch.nn.Linear(2, 2, bias=False)
draft = torch.nn.Module() if has_shared_draft else None
if draft is not None:
draft.shared = target.weight
worker = object.__new__(gpu_worker.Worker)
worker.get_draft_model = lambda: draft
worker.model_config = SimpleNamespace(sleep_preserve_parameter_names=["draft.*"])
worker._save_sleep_parameters(target)
assert worker._sleep_saved_parameters == {}
worker.model_config.sleep_preserve_parameter_names = ["missing"]
with pytest.raises(ValueError, match="No parameter matches sleep retention"):
worker._save_sleep_parameters(target)
def test_load_model_preserves_compiled_graphs_at_runtime(monkeypatch):
"""Profiling must use serving's thread count to keep Dynamo guards valid."""
from torch._dynamo.testing import CompileCounter