diff --git a/tests/v1/kv_connector/unit/offloading_connector/test_config.py b/tests/v1/kv_connector/unit/offloading_connector/test_config.py index 0b0b6b874f..091b50fa1a 100644 --- a/tests/v1/kv_connector/unit/offloading_connector/test_config.py +++ b/tests/v1/kv_connector/unit/offloading_connector/test_config.py @@ -326,13 +326,17 @@ def test_dcp_scales_attention_but_not_mamba_group_blocks(): ] == [1, 3] -def test_preserves_data_parallel_index(): +def test_preserves_data_parallel_config(): config = _make_vllm_config() config.parallel_config.data_parallel_index = 2 + config.parallel_config.data_parallel_size = 4 + config.parallel_config.data_parallel_rank_local = 1 offloading_config = build_offloading_config(config, _make_kv_cache_config()) assert offloading_config.parallel.data_parallel_index == 2 + assert offloading_config.parallel.data_parallel_size == 4 + assert offloading_config.parallel.data_parallel_rank_local == 1 def test_resolves_heterogeneous_hybrid_block_sizes(): diff --git a/tests/v1/kv_connector/unit/offloading_connector/test_worker.py b/tests/v1/kv_connector/unit/offloading_connector/test_worker.py index b65cfe8150..8a1f3723c4 100644 --- a/tests/v1/kv_connector/unit/offloading_connector/test_worker.py +++ b/tests/v1/kv_connector/unit/offloading_connector/test_worker.py @@ -191,6 +191,8 @@ def _offloading_config(rank: int = 0) -> OffloadingConfig: pcp_size=1, dcp_size=1, data_parallel_index=0, + data_parallel_size=1, + data_parallel_rank_local=None, is_parallelism_agnostic=False, ), ) diff --git a/tests/v1/kv_offload/test_factory.py b/tests/v1/kv_offload/test_factory.py index ecd706828c..6d1c103fdf 100644 --- a/tests/v1/kv_offload/test_factory.py +++ b/tests/v1/kv_offload/test_factory.py @@ -49,6 +49,8 @@ def _make_offloading_config( pcp_size: int = 1, dcp_size: int = 1, data_parallel_index: int = 0, + data_parallel_size: int = 1, + data_parallel_rank_local: int | None = None, is_parallelism_agnostic: bool = False, replicated_layout: bool = False, extra_config: dict[str, Any] | None = None, @@ -81,6 +83,8 @@ def _make_offloading_config( pcp_size=pcp_size, dcp_size=dcp_size, data_parallel_index=data_parallel_index, + data_parallel_size=data_parallel_size, + data_parallel_rank_local=data_parallel_rank_local, is_parallelism_agnostic=is_parallelism_agnostic, ), replicated_layout=replicated_layout, diff --git a/tests/v1/kv_offload/test_file_mapper.py b/tests/v1/kv_offload/test_file_mapper.py index 0c8416b0c0..9ba1e56487 100644 --- a/tests/v1/kv_offload/test_file_mapper.py +++ b/tests/v1/kv_offload/test_file_mapper.py @@ -49,6 +49,8 @@ def make_mapper_from_offloading_spec(**kwargs) -> FileMapper: pcp_size=kwargs.get("pcp_size", 1), dcp_size=kwargs.get("dcp_size", 1), data_parallel_index=0, + data_parallel_size=1, + data_parallel_rank_local=None, is_parallelism_agnostic=kwargs.get("is_parallelism_agnostic", False), ), replicated_layout=kwargs.get("replicated_layout", False), diff --git a/tests/v1/kv_offload/tiering/test_fs_tier.py b/tests/v1/kv_offload/tiering/test_fs_tier.py index f007eb49e1..efeb59274f 100644 --- a/tests/v1/kv_offload/tiering/test_fs_tier.py +++ b/tests/v1/kv_offload/tiering/test_fs_tier.py @@ -81,6 +81,8 @@ def _make_offloading_spec( pcp_size=1, dcp_size=1, data_parallel_index=0, + data_parallel_size=1, + data_parallel_rank_local=None, is_parallelism_agnostic=is_parallelism_agnostic, ), replicated_layout=replicated_layout, diff --git a/tests/v1/kv_offload/tiering/test_obj_tier.py b/tests/v1/kv_offload/tiering/test_obj_tier.py index 54ec4c1f9f..d0c4d4cb04 100644 --- a/tests/v1/kv_offload/tiering/test_obj_tier.py +++ b/tests/v1/kv_offload/tiering/test_obj_tier.py @@ -74,6 +74,8 @@ def _make_offloading_config( pcp_size=1, dcp_size=1, data_parallel_index=0, + data_parallel_size=1, + data_parallel_rank_local=None, is_parallelism_agnostic=is_parallelism_agnostic, ), replicated_layout=replicated_layout, diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/offloading/config.py b/vllm/distributed/kv_transfer/kv_connector/v1/offloading/config.py index 2bf140f1e3..a38b2ca7a4 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/offloading/config.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/offloading/config.py @@ -213,6 +213,8 @@ def build_offloading_config( pcp_size=parallel_config.prefill_context_parallel_size, dcp_size=parallel_config.decode_context_parallel_size, data_parallel_index=parallel_config.data_parallel_index, + data_parallel_size=parallel_config.data_parallel_size, + data_parallel_rank_local=parallel_config.data_parallel_rank_local, is_parallelism_agnostic=is_parallelism_agnostic, ), replicated_layout=replicated_layout, diff --git a/vllm/v1/kv_offload/config.py b/vllm/v1/kv_offload/config.py index fdbce491e0..26408d2992 100644 --- a/vllm/v1/kv_offload/config.py +++ b/vllm/v1/kv_offload/config.py @@ -48,6 +48,10 @@ class OffloadingParallelConfig: dcp_size: int # Data parallel replica index of this engine. data_parallel_index: int + # Number of data parallel replicas. + data_parallel_size: int + # Local rank of the data parallel group, set only in SPMD mode. + data_parallel_rank_local: int | None # True when the bytes that will be persisted for a block are portable # across parallelism configurations: for the direct layout, concatenating # the block's data across all workers in rank order yields the same bytes