[KV Connector] Opt DecodeBenchConnector into SupportsHMA

Previously the HMA (hybrid KV cache manager) layer refused to activate
when DecodeBenchConnector was in use, because the connector did not
advertise SupportsHMA. That forced decode-only benchmark recipes to
pass --disable-hybrid-kv-cache-manager, which collapsed hybrid-model
KV cache groups (SWA / MLA compress=4 / MLA compress=128 / sparse
indexer) into a single uniform page size via unify_kv_cache_spec_
page_size, throwing away the compression savings and capping concurrent
capacity on hybrid models (e.g. DeepSeek-V4 saw ~43 concurrent
8k/1k requests instead of the model's true ceiling).

This connector is a dummy fill that owns no external per-block state,
so the HMA path has nothing extra to do. Implementation is minimal:

- Inherit from SupportsHMA.
- Implement request_finished_all_groups: delegates to the same
  scheduler.request_finished() cleanup as the single-group variant,
  ignoring block_ids (no per-block state to release).

With this change, recipes can drop --disable-hybrid-kv-cache-manager
and let HMA size each KV cache group correctly.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Signed-off-by: Zijing Liu <liuzijing2014@gmail.com>

vllm/distributed/kv_transfer/kv_connector/v1/decode_bench_connector.py

@@ -40,7 +40,10 @@ from vllm.distributed.kv_transfer.kv_connector.v1 import (
     KVConnectorBase_V1,
     KVConnectorRole,
 )
-from vllm.distributed.kv_transfer.kv_connector.v1.base import KVConnectorMetadata
+from vllm.distributed.kv_transfer.kv_connector.v1.base import (
+    KVConnectorMetadata,
+    SupportsHMA,
+)
 from vllm.logger import init_logger
 from vllm.utils.math_utils import cdiv
 from vllm.v1.attention.backend import AttentionMetadata
@@ -71,7 +74,7 @@ class DecodeBenchConnectorMetadata(KVConnectorMetadata):
     reqs_to_fill: dict[str, tuple[tuple[list[int], ...], int]]
 
 
-class DecodeBenchConnector(KVConnectorBase_V1):
+class DecodeBenchConnector(KVConnectorBase_V1, SupportsHMA):
     """
     A KV Connector for decode instance performance testing.
 
@@ -164,6 +167,17 @@ class DecodeBenchConnector(KVConnectorBase_V1):
         self.connector_scheduler.request_finished(request)
         return False, None
 
+    def request_finished_all_groups(
+        self,
+        request: "Request",
+        block_ids: tuple[list[int], ...],
+    ) -> tuple[bool, dict[str, Any] | None]:
+        # HMA-enabled path: same cleanup as the single-group variant since
+        # this connector owns no external state per block.
+        assert self.connector_scheduler is not None
+        self.connector_scheduler.request_finished(request)
+        return False, None
+
 
 class DecodeBenchConnectorScheduler:
     """Scheduler-side implementation for DecodeBenchConnector."""