[Misc] Avoid importing nixl_ep on every vllm serve config (#50879)

Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
This commit is contained in:
Nicolò Lucchesi
2026-08-04 07:56:54 -07:00
committed by GitHub
parent 122b3d46b6
commit 7b50d2c0bc
@@ -30,32 +30,9 @@ from vllm.model_executor.layers.fused_moe.prepare_finalize.flashinfer_nvlink_two
FlashInferNVLinkTwoSidedPrepareAndFinalize,
)
from vllm.platforms import current_platform
from vllm.utils.import_utils import (
has_deep_ep,
has_deep_ep_v2,
has_mori,
has_nixl_ep,
)
logger = init_logger(__name__)
if current_platform.is_cuda_alike():
if has_deep_ep():
from .prepare_finalize.deepep_ht import DeepEPHTPrepareAndFinalize
from .prepare_finalize.deepep_ll import (
DEEPEP_QUANT_BLOCK_SHAPE,
DeepEPLLPrepareAndFinalize,
)
if has_deep_ep_v2():
from .prepare_finalize.deepep_v2 import DeepEPV2PrepareAndFinalize
if has_mori():
from .prepare_finalize.mori import MoriPrepareAndFinalize
if has_nixl_ep():
from .prepare_finalize.nixl_ep import (
NIXL_EP_QUANT_BLOCK_SHAPE,
NixlEPPrepareAndFinalize,
)
def get_ep_all2all_manager(eep_stage: bool = False) -> Any:
if eep_stage:
@@ -93,21 +70,29 @@ def maybe_roundup_layer_hidden_size(
Original hidden size otherwise.
"""
if moe_parallel_config.use_deepep_ht_kernels:
from .prepare_finalize.deepep_ht import DeepEPHTPrepareAndFinalize
hidden_size = DeepEPHTPrepareAndFinalize.maybe_roundup_layer_hidden_size(
hidden_size, act_dtype
)
if moe_parallel_config.use_deepep_ll_kernels:
from .prepare_finalize.deepep_ll import DeepEPLLPrepareAndFinalize
hidden_size = DeepEPLLPrepareAndFinalize.maybe_roundup_layer_hidden_size(
hidden_size
)
if moe_parallel_config.use_deepep_v2_kernels:
from .prepare_finalize.deepep_v2 import DeepEPV2PrepareAndFinalize
hidden_size = DeepEPV2PrepareAndFinalize.maybe_roundup_layer_hidden_size(
hidden_size, act_dtype
)
if moe_parallel_config.use_nixl_ep_kernels:
from .prepare_finalize.nixl_ep import NixlEPPrepareAndFinalize
hidden_size = NixlEPPrepareAndFinalize.maybe_roundup_layer_hidden_size(
hidden_size
)
@@ -171,6 +156,8 @@ def maybe_make_prepare_finalize(
prepare_finalize: FusedMoEPrepareAndFinalize | None = None
if moe.use_deepep_ht_kernels:
from .prepare_finalize.deepep_ht import DeepEPHTPrepareAndFinalize
assert moe.dp_size == all2all_manager.dp_world_size
all_to_all_args: dict[str, Any] = dict()
@@ -183,6 +170,11 @@ def maybe_make_prepare_finalize(
)
elif moe.use_deepep_ll_kernels:
from .prepare_finalize.deepep_ll import (
DEEPEP_QUANT_BLOCK_SHAPE,
DeepEPLLPrepareAndFinalize,
)
assert quant_config is not None
global_to_physical = physical_to_global = local_expert_global_ids = None
if routing_tables is not None:
@@ -217,6 +209,8 @@ def maybe_make_prepare_finalize(
local_expert_global_ids=local_expert_global_ids,
)
elif moe.use_deepep_v2_kernels:
from .prepare_finalize.deepep_v2 import DeepEPV2PrepareAndFinalize
assert moe.dp_size == all2all_manager.dp_world_size
use_fp8_dispatch = (
@@ -247,6 +241,8 @@ def maybe_make_prepare_finalize(
)
elif moe.use_mori_kernels:
from .prepare_finalize.mori import MoriPrepareAndFinalize
assert quant_config is not None
# Note: We may want to use FP8 dispatch just to reduce
@@ -334,6 +330,11 @@ def maybe_make_prepare_finalize(
)
elif moe.use_nixl_ep_kernels:
from .prepare_finalize.nixl_ep import (
NIXL_EP_QUANT_BLOCK_SHAPE,
NixlEPPrepareAndFinalize,
)
assert quant_config is not None
global_to_physical = physical_to_global = local_expert_global_ids = None
if routing_tables is not None: