[EPLB] Reject NCCL-based EPLB communicators with async EPLB (#44978)

Signed-off-by: Markov Ilya <markovilya197@gmail.com>
This commit is contained in:
Ilya Markov
2026-06-10 19:51:27 +00:00
committed by GitHub
parent 3d300aecb1
commit 6471ec75bd
8 changed files with 37 additions and 32 deletions
+4
View File
@@ -78,6 +78,8 @@ def test_elastic_ep_scaling():
"--enable-eplb",
"--eplb-config.num_redundant_experts",
"0",
"--eplb-config.use_async",
"false",
"--data-parallel-backend",
"ray",
"--data-parallel-size",
@@ -151,6 +153,8 @@ def test_elastic_ep_scaling_uneven():
"--enable-eplb",
"--eplb-config.num_redundant_experts",
"0",
"--eplb-config.use_async",
"false",
"--data-parallel-backend",
"ray",
"--data-parallel-size",
+1 -3
View File
@@ -644,9 +644,7 @@ def _test_rearrange_expert_weights_no_change(env, world_size) -> None:
(2, 2, 2, 3),
],
)
@pytest.mark.parametrize(
"eplb_communicator", ["torch_nccl", "torch_gloo", "pynccl", "nixl"]
)
@pytest.mark.parametrize("eplb_communicator", ["torch_gloo", "nixl"])
def test_async_transfer_layer_without_mtp(
world_size: int,
num_layers: int,
+3
View File
@@ -1285,6 +1285,9 @@ def _test_body_eplb(
expert_weights = [list(eplb_moe_layer.get_expert_weights())]
expert_buffer = [torch.empty_like(w) for w in expert_weights[0]]
assert vllm_config.parallel_config.eplb_config.communicator is not None, (
"EPLB communicator backend must be set by ParallelConfig"
)
communicator = create_eplb_communicator(
group_coordinator=get_eplb_group(),
backend=vllm_config.parallel_config.eplb_config.communicator,