[KV Offload] Expose data-parallel topology to offloading backends (#51879)

Signed-off-by: Ziqi Fan <ziqif@nvidia.com>
This commit is contained in:
Ziqi Fan
2026-08-12 20:27:26 -07:00
committed by GitHub
parent 6e502b62cb
commit f9a0f629bd
8 changed files with 23 additions and 1 deletions
@@ -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():
@@ -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,
),
)
+4
View File
@@ -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,
+2
View File
@@ -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),
@@ -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,
@@ -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,
@@ -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,
+4
View File
@@ -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