A high-throughput and memory-efficient inference and serving engine for LLMs
[Bugfix][PD] Fix DSV4 disaggregated serving (#45831)
Signed-off-by: ZhanqiuHu <zhu@redhat.com>
zhanqiuhu committed
Jun 17, 2026 at 11:17 UTC
eb0fdeb1e83443e1b901cd97dd38cff803488bba
1 file changed
+4
-1
vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py
+4
-1
@@ -71,6 +71,7 @@ from vllm.v1.kv_cache_interface import (
71
FullAttentionSpec,
72
MambaSpec,
73
MLAAttentionSpec,
74
+ SlidingWindowMLASpec,
75
UniformTypeKVCacheSpecs,
76
)
77
from vllm.v1.worker.block_table import BlockTable
@@ -962,7 +963,9 @@ class NixlBaseConnectorWorker:
963
)
964
else:
965
self.block_len_per_layer.append(physical_page_size)
965
- is_mla_region = isinstance(layer_spec, MLAAttentionSpec)
966
+ is_mla_region = isinstance(
967
+ layer_spec, (MLAAttentionSpec, SlidingWindowMLASpec)
968
+ )
969
self._region_is_mla.append(is_mla_region)
970
971
if not is_mla_region: