[Feat] NIXL: Support per-GPU UCX device selection (#43482)

This commit is contained in:
Jzz1943
2026-10-21 17:54:11 +08:00
committed by GitHub
parent 9c10e0dd61
commit 2e7c9d851c
4 changed files with 50 additions and 7 deletions
@@ -80,7 +80,7 @@ class Disagg(msgspec.Struct):
] = 8998
disaggregation_ib_device: A[
Optional[str],
'The InfiniBand devices for disaggregation transfer. Supports a single device (e.g., --disaggregation-ib-device mlx5_0), a shared comma-separated list (e.g., --disaggregation-ib-device mlx5_0,mlx5_1), a per-GPU JSON mapping (e.g., --disaggregation-ib-device \'{"0": "mlx5_0,mlx5_1", "1": "mlx5_2"}\'), or a path to a JSON file containing that mapping. Default is None, which triggers automatic device detection when mooncake backend is enabled.',
'The InfiniBand devices for disaggregation transfer. Supports a single device (e.g., --disaggregation-ib-device mlx5_0), a shared comma-separated list (e.g., --disaggregation-ib-device mlx5_0,mlx5_1), a per-GPU JSON mapping (e.g., --disaggregation-ib-device \'{"0": "mlx5_0,mlx5_1", "1": "mlx5_2"}\'), or a path to a JSON file containing that mapping. Supported by Mooncake and NIXL UCX. Mapping keys are process-visible GPU IDs (after CUDA_VISIBLE_DEVICES); use HCA names without port suffixes for NIXL UCX. Default is None, which preserves backend device selection.',
] = None
disaggregation_decode_enable_radix_cache: A[
bool,
@@ -179,6 +179,24 @@ NIXL_ERR_SETTLE_TIMEOUT_S = 5.0
NIXL_ERR_SETTLE_POLL_S = 0.001
def _resolve_ucx_device_list(ib_device: str, gpu_id: int) -> Optional[str]:
from sglang.srt.distributed.device_communicators.mooncake_transfer_engine import (
get_ib_devices_for_gpu,
)
devices = get_ib_devices_for_gpu(ib_device, gpu_id)
if devices is None:
return None
names = [name.strip() for name in devices.split(",")]
if any(not name or ":" in name for name in names):
raise ValueError(
"NIXL UCX --disaggregation-ib-device requires HCA names without "
"port suffixes (e.g. mlx5_0); NIXL appends port 1."
)
# NIXL's UCX device_list parser splits on the literal delimiter ", ".
return ", ".join(names)
def _normalize_kv_mem_kinds(kinds: Optional[List[str]], expected_len: int) -> List[str]:
if kinds is None:
return ["VRAM"] * expected_len
@@ -583,6 +601,31 @@ class NixlKVManager(StagingManagerMixin, CommonKVManager):
"SGLANG_DISAGGREGATION_NIXL_BACKEND_PARAMS must be a JSON object "
"with string keys and string values"
)
if backend == "UCX" and self.kv_args.ib_device:
device_list = _resolve_ucx_device_list(
self.kv_args.ib_device, self.kv_args.gpu_id
)
if device_list is not None:
existing = backend_params.get("device_list")
if (
existing is not None
and ", ".join(name.strip() for name in existing.split(","))
!= device_list
):
raise ValueError(
"Conflicting NIXL UCX device_list in "
"SGLANG_DISAGGREGATION_NIXL_BACKEND_PARAMS and "
"--disaggregation-ib-device. Configure only one "
"device selection or make them agree."
)
backend_params["device_list"] = device_list
logger.info(
"NIXL UCX device binding: gpu_id=%s, engine_rank=%s, "
"device_list=%s",
self.kv_args.gpu_id,
self.kv_args.engine_rank,
device_list,
)
# self.transfer_worker and self._start_bootstrap_thread runs concurrently
# so we cannot use sync_mode=None which is thread-unsafe.
agent_config = nixl_agent_config(
@@ -29,9 +29,8 @@ logger = logging.getLogger(__name__)
def configure_nixl_pd_backend(test_cls):
test_cls.transfer_backend = ["--disaggregation-transfer-backend", "nixl"]
# NIXL backend/network selection is driven by NIXL environment variables
# such as SGLANG_DISAGGREGATION_NIXL_BACKEND and backend params, not by the
# Mooncake-specific --disaggregation-ib-device argument.
# Keep these tests on the configured NIXL/UCX device selection rather than
# inheriting the Mooncake fixture's per-side RDMA device lists.
test_cls.rdma_devices = []
@@ -119,7 +118,7 @@ class PDDisaggregationServerBase(CustomTestCase):
@classmethod
def rdma_devices_for(cls, gpu_indices) -> list:
"""`--disaggregation-ib-device` args for a server pinned to these GPUs."""
if not is_in_ci():
if not is_in_ci() or not cls.rdma_devices:
return cls.rdma_devices
return ["--disaggregation-ib-device", get_rdma_devices_args(gpu_indices)]
@@ -532,8 +531,8 @@ def _maybe_set_ucx_net_devices(ib_devices) -> bool:
# createBackend and tcp is the only transport left that can serve it, so
# keep loopback here.
net_devices = ",".join([*(f"{d}:1" for d in devices), "lo"])
# NIXL ignores --disaggregation-ib-device; without this UCX opens every RDMA
# device on the host, and that full-device init can stall inside the driver.
# Limit default UCX device discovery; opening every RDMA device on the host
# can stall inside the driver.
os.environ["UCX_NET_DEVICES"] = net_devices
logger.warning("Set UCX_NET_DEVICES=%s for NIXL/UCX", net_devices)
return True