[EPLB] Add nixl-based eplb communicator (#36276)

Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Markov Ilya <markovilya19@gmail.com>
This commit is contained in:
Ilya Markov
2026-04-20 10:24:23 +00:00
committed by GitHub
parent f774ba028a
commit 50dd4cb427
10 changed files with 556 additions and 80 deletions
+14 -3
View File
@@ -9,7 +9,10 @@ import torch
import torch.distributed
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.distributed.eplb.eplb_communicator import create_eplb_communicator
from vllm.distributed.eplb.eplb_communicator import (
create_eplb_communicator,
has_nixl,
)
from vllm.distributed.eplb.rebalance_execute import (
move_from_buffer,
rearrange_expert_weights_inplace,
@@ -527,7 +530,9 @@ def _test_rearrange_expert_weights_with_redundancy(
(4, 8, 8, 16),
],
)
@pytest.mark.parametrize("eplb_communicator", ["torch_nccl", "torch_gloo", "pynccl"])
@pytest.mark.parametrize(
"eplb_communicator", ["torch_nccl", "torch_gloo", "pynccl", "nixl"]
)
def test_rearrange_expert_weights_with_redundancy(
world_size,
num_layers,
@@ -537,6 +542,8 @@ def test_rearrange_expert_weights_with_redundancy(
):
"""Test the functionality of rearranging expert weights with redundancy."""
if eplb_communicator == "nixl" and not has_nixl():
pytest.skip("NIXL is not available")
if torch.accelerator.device_count() < world_size:
pytest.skip(f"Need at least {world_size} GPUs to run the test")
distributed_run(
@@ -633,7 +640,9 @@ 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"])
@pytest.mark.parametrize(
"eplb_communicator", ["torch_nccl", "torch_gloo", "pynccl", "nixl"]
)
def test_async_transfer_layer_without_mtp(
world_size: int,
num_layers: int,
@@ -643,6 +652,8 @@ def test_async_transfer_layer_without_mtp(
):
"""Exercise async EPLB transfer path without MTP/spec decode."""
if eplb_communicator == "nixl" and not has_nixl():
pytest.skip("NIXL is not available")
if torch.accelerator.device_count() < world_size:
pytest.skip(f"Need at least {world_size} GPUs to run the test")