diff --git a/tests/v1/worker/test_dspark_utils.py b/tests/v1/worker/test_dspark_utils.py new file mode 100644 index 0000000000..94655f192c --- /dev/null +++ b/tests/v1/worker/test_dspark_utils.py @@ -0,0 +1,51 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +from dataclasses import dataclass, field + +from vllm.v1.worker.gpu.spec_decode.dspark.utils import _get_dspark_parallel_config + + +@dataclass +class _FakeEPLBConfig: + num_redundant_experts: int = 0 + + +@dataclass +class _FakeParallelConfig: + pipeline_parallel_size: int = 2 + tensor_parallel_size: int = 8 + enable_eplb: bool = True + eplb_config: _FakeEPLBConfig = field( + default_factory=lambda: _FakeEPLBConfig(num_redundant_experts=32) + ) + enable_elastic_ep: bool = True + + def __post_init__(self) -> None: + if not self.enable_eplb and self.eplb_config.num_redundant_experts: + raise ValueError("redundant experts require EPLB") + if self.enable_elastic_ep and not self.enable_eplb: + raise ValueError("elastic EP requires EPLB") + + +def test_dspark_parallel_config_disables_eplb_atomically(): + target_config = _FakeParallelConfig() + + draft_config = _get_dspark_parallel_config( + target_config, + tensor_parallel_size=4, + ) + + assert target_config.pipeline_parallel_size == 2 + assert target_config.tensor_parallel_size == 8 + assert target_config.enable_eplb + assert target_config.eplb_config.num_redundant_experts == 32 + assert target_config.enable_elastic_ep + + assert draft_config is not target_config + assert draft_config.pipeline_parallel_size == 1 + assert draft_config.tensor_parallel_size == 4 + assert not draft_config.enable_eplb + assert draft_config.eplb_config.num_redundant_experts == 0 + assert not draft_config.enable_elastic_ep + assert draft_config.eplb_config is not target_config.eplb_config diff --git a/vllm/v1/worker/gpu/spec_decode/dspark/utils.py b/vllm/v1/worker/gpu/spec_decode/dspark/utils.py index db1e17a9f2..bf1a5761b2 100644 --- a/vllm/v1/worker/gpu/spec_decode/dspark/utils.py +++ b/vllm/v1/worker/gpu/spec_decode/dspark/utils.py @@ -3,7 +3,7 @@ import torch.nn as nn -from vllm.config import ModelConfig, VllmConfig, replace +from vllm.config import ModelConfig, ParallelConfig, VllmConfig, replace from vllm.logger import init_logger from vllm.v1.attention.backends.registry import AttentionBackendEnum from vllm.v1.worker.gpu.spec_decode.utils import get_pp_safe_draft_load_config @@ -31,6 +31,29 @@ def _resolve_dspark_attention_backend( return None +def _get_dspark_parallel_config( + parallel_config: ParallelConfig, + tensor_parallel_size: int, +) -> ParallelConfig: + if parallel_config.enable_eplb: + logger.warning_once( + "EPLB is disabled for the DSpark draft model. EPLB remains enabled " + "for the target model." + ) + + return replace( + parallel_config, + pipeline_parallel_size=1, + tensor_parallel_size=tensor_parallel_size, + enable_eplb=False, + eplb_config=replace( + parallel_config.eplb_config, + num_redundant_experts=0, + ), + enable_elastic_ep=False, + ) + + def load_dspark_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Module: speculative_config = vllm_config.speculative_config assert speculative_config is not None @@ -54,12 +77,9 @@ def load_dspark_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Mo draft_vllm_config = replace( vllm_config, - parallel_config=replace( + parallel_config=_get_dspark_parallel_config( vllm_config.parallel_config, - pipeline_parallel_size=1, - tensor_parallel_size=( - speculative_config.draft_parallel_config.tensor_parallel_size - ), + speculative_config.draft_parallel_config.tensor_parallel_size, ), attention_config=replace( vllm_config.attention_config,