From eb0fdeb1e83443e1b901cd97dd38cff803488bba Mon Sep 17 00:00:00 2001 From: zhanqiuhu <49648934+ZhanqiuHu@users.noreply.github.com> Date: Wed, 17 Jun 2026 11:17:14 -0400 Subject: [PATCH] [Bugfix][PD] Fix DSV4 disaggregated serving (#45831) Signed-off-by: ZhanqiuHu --- .../kv_transfer/kv_connector/v1/nixl/base_worker.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py index e587b0cd1fa..74980dabf71 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py @@ -71,6 +71,7 @@ from vllm.v1.kv_cache_interface import ( FullAttentionSpec, MambaSpec, MLAAttentionSpec, + SlidingWindowMLASpec, UniformTypeKVCacheSpecs, ) from vllm.v1.worker.block_table import BlockTable @@ -962,7 +963,9 @@ class NixlBaseConnectorWorker: ) else: self.block_len_per_layer.append(physical_page_size) - is_mla_region = isinstance(layer_spec, MLAAttentionSpec) + is_mla_region = isinstance( + layer_spec, (MLAAttentionSpec, SlidingWindowMLASpec) + ) self._region_is_mla.append(is_mla_region) if not is_mla_region: