@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 09a7cc5ba9
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 4, 03:45 UTC

[KV Connector] Opt DecodeBenchConnector into SupportsHMA (#41770)

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

Zijing Liu committed May 7, 2026 at 16:10 UTC 09a7cc5ba94c5c1b1c908a29f28ed41c793f1e9f
1 file changed +16 -2
vllm/distributed/kv_transfer/kv_connector/v1/decode_bench_connector.py
+16 -2
@@ -40,7 +40,10 @@ from vllm.distributed.kv_transfer.kv_connector.v1 import (
40 KVConnectorBase_V1,
41 KVConnectorRole,
42 )
43 -from vllm.distributed.kv_transfer.kv_connector.v1.base import KVConnectorMetadata
43 +from vllm.distributed.kv_transfer.kv_connector.v1.base import (
44 + KVConnectorMetadata,
45 + SupportsHMA,
46 +)
47 from vllm.logger import init_logger
48 from vllm.utils.math_utils import cdiv
49 from vllm.v1.attention.backend import AttentionMetadata
@@ -71,7 +74,7 @@ class DecodeBenchConnectorMetadata(KVConnectorMetadata):
74 reqs_to_fill: dict[str, tuple[tuple[list[int], ...], int]]
75
76
74 -class DecodeBenchConnector(KVConnectorBase_V1):
77 +class DecodeBenchConnector(KVConnectorBase_V1, SupportsHMA):
78 """
79 A KV Connector for decode instance performance testing.
80
@@ -164,6 +167,17 @@ class DecodeBenchConnector(KVConnectorBase_V1):
167 self.connector_scheduler.request_finished(request)
168 return False, None
169
170 + def request_finished_all_groups(
171 + self,
172 + request: "Request",
173 + block_ids: tuple[list[int], ...],
174 + ) -> tuple[bool, dict[str, Any] | None]:
175 + # HMA-enabled path: same cleanup as the single-group variant since
176 + # this connector owns no external state per block.
177 + assert self.connector_scheduler is not None
178 + self.connector_scheduler.request_finished(request)
179 + return False, None
180 +
181
182 class DecodeBenchConnectorScheduler:
183 """Scheduler-side implementation for DecodeBenchConnector."""