[Bugfix][KV Offload] Keep Mamba block span unscaled under DCP (#49964)

Signed-off-by: Jonguk Cheong <jdal3031@snu.ac.kr>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Jonguk Cheong
2026-07-28 17:41:55 +03:00
committed by GitHub
co-authored by OpenAI Codex
parent b6cbba8bc8
commit 1e81853afc
2 changed files with 66 additions and 2 deletions
@@ -8,10 +8,15 @@ from unittest.mock import MagicMock, patch
import pytest
import torch
from tests.v1.kv_connector.unit.offloading_connector.utils import MockOffloadingSpec
from vllm.config import KVTransferConfig, ParallelConfig, VllmConfig
from vllm.distributed.kv_transfer.kv_connector.v1.offloading.config import (
build_offloading_config,
)
from vllm.distributed.kv_transfer.kv_connector.v1.offloading.scheduler import (
SchedulerOffloadConfig,
is_store_reachable_swa_chunk,
)
from vllm.platforms import current_platform
from vllm.v1.kv_cache_interface import (
FullAttentionSpec,
@@ -179,6 +184,25 @@ def _make_hybrid_kv_cache_config() -> KVCacheConfig:
)
def _make_mamba_hybrid_kv_cache_config() -> KVCacheConfig:
return KVCacheConfig(
num_blocks=4,
kv_cache_tensors=[],
kv_cache_groups=[
KVCacheGroupSpec(["full_layer"], _full_attention_spec()),
KVCacheGroupSpec(
["mamba_layer"],
MambaSpec(
block_size=16,
shapes=((1, 1),),
dtypes=(torch.float32,),
mamba_cache_mode="align",
),
),
],
)
def _parallelism_agnostic(kv_cache_groups: list[KVCacheGroupSpec]) -> bool:
config = _make_vllm_config()
kv_cache_config = KVCacheConfig(
@@ -267,6 +291,38 @@ def test_prefill_context_parallelism_does_not_scale_group_blocks():
assert offloading_config.cache.blocks_per_chunk == 4
def test_dcp_scales_attention_but_not_mamba_group_blocks():
config = _make_vllm_config(tensor_parallel_size=2, decode_context_parallel_size=2)
config.speculative_config = None
offloading_config = build_offloading_config(
config, _make_mamba_hybrid_kv_cache_config()
)
assert tuple(group.tokens_per_block for group in offloading_config.groups) == (
32,
16,
)
scheduler_config = SchedulerOffloadConfig.from_spec(
MockOffloadingSpec(offloading_config),
config,
_make_mamba_hybrid_kv_cache_config(),
)
mamba_group = scheduler_config.kv_group_configs[1]
assert mamba_group.alignment_chunk_count == 2
assert [
chunk_idx
for chunk_idx in range(4)
if is_store_reachable_swa_chunk(
chunk_idx,
4,
mamba_group.alignment_chunk_count,
mamba_group.sliding_window_size_in_chunks,
mamba_group.is_eagle_group,
)
] == [1, 3]
def test_preserves_data_parallel_index():
config = _make_vllm_config()
config.parallel_config.data_parallel_index = 2
@@ -5,7 +5,11 @@
from typing import TYPE_CHECKING
from vllm.v1.core.kv_cache_utils import resolve_kv_cache_block_sizes
from vllm.v1.kv_cache_interface import FullAttentionSpec, MLAAttentionSpec
from vllm.v1.kv_cache_interface import (
AttentionSpec,
FullAttentionSpec,
MLAAttentionSpec,
)
from vllm.v1.kv_offload.config import (
OffloadingCacheConfig,
OffloadingConfig,
@@ -40,7 +44,11 @@ def build_offloading_config(
OffloadingGroupConfig(
tokens_per_block=(
group.kv_cache_spec.block_size
* parallel_config.decode_context_parallel_size
* (
parallel_config.decode_context_parallel_size
if isinstance(group.kv_cache_spec, AttentionSpec)
else 1
)
),
layer_names=tuple(group.layer_names),
)