[Core][Kernel] Share persistent workspaces for Marlin and Humming (#57421)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: ErinYin <liang.yin@daocloud.io>
This commit is contained in:
Misha Goin
2026-09-20 17:28:33 +08:00
committed by GitHub
co-authored by ErinYin
parent 0e110f696d
commit e5fce7b56b
21 changed files with 598 additions and 211 deletions
+17 -3
View File
@@ -1384,24 +1384,33 @@ def test_humming_global_valid_shape_m(
assert result == expected
def test_humming_permute_scratch_is_keyed_by_runtime_topk(
def test_humming_permute_scratch_is_shared_by_config(
monkeypatch: pytest.MonkeyPatch,
):
from types import SimpleNamespace
from unittest.mock import Mock
import vllm.model_executor.layers.fused_moe.experts.fused_humming_moe as humming
import vllm.model_executor.layers.fused_moe.moe_permute_unpermute as permute
import vllm.v1.worker.workspace as workspace
manager = workspace.WorkspaceManager(torch.device("cpu"))
monkeypatch.setattr(workspace, "_manager", manager)
scratch_topk6 = Mock()
scratch_topk1 = Mock()
scratch_type = Mock(side_effect=[scratch_topk6, scratch_topk1])
monkeypatch.setattr(humming, "moe_permute_unpermute_supported", lambda: True)
monkeypatch.setattr(humming, "MoEPermuteScratch", scratch_type)
monkeypatch.setattr(permute, "MoEPermuteScratch", scratch_type)
moe_config = make_dummy_moe_config(max_num_tokens=512, experts_per_token=6)
moe_config.moe_parallel_config.dp_size = 2
experts = SimpleNamespace(_permute_scratch={}, moe_config=moe_config)
experts = SimpleNamespace(moe_config=moe_config)
other_layer = SimpleNamespace(moe_config=moe_config)
assert humming.HummingExpertsBase._get_permute_scratch(experts, 6) is scratch_topk6
assert (
humming.HummingExpertsBase._get_permute_scratch(other_layer, 6) is scratch_topk6
)
assert humming.HummingExpertsBase._get_permute_scratch(experts, 1) is scratch_topk1
assert humming.HummingExpertsBase._get_permute_scratch(experts, 6) is scratch_topk6
@@ -1412,6 +1421,11 @@ def test_humming_permute_scratch_is_keyed_by_runtime_topk(
assert second_call.kwargs["max_num_tokens"] == 6144
assert second_call.kwargs["topk"] == 1
manager.lock()
assert humming.HummingExpertsBase._get_permute_scratch(experts, 6) is scratch_topk6
with pytest.raises(AssertionError, match="was not allocated during warmup"):
humming.HummingExpertsBase._get_permute_scratch(experts, 2)
def test_humming_delegates_to_instance_activation():
from types import SimpleNamespace