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 6, 20:30 UTC
[Refactor] Remove dead code in KV connector (#36424)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Mar 11, 2026 at 15:40 UTC
35bdca5431e652b4c00267489a632c1bf5522103
3 files changed
+5
-15
vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+1
-7
@@ -50,7 +50,6 @@ from vllm.distributed.kv_transfer.kv_connector.v1.metrics import (
50
from vllm.distributed.parallel_state import (
51
get_tensor_model_parallel_rank,
52
get_tensor_model_parallel_world_size,
53
- get_tp_group,
53
)
54
from vllm.forward_context import ForwardContext
55
from vllm.logger import init_logger
@@ -564,7 +563,6 @@ class NixlConnectorScheduler:
563
564
# Background thread for handling new handshake requests.
565
self._nixl_handshake_listener_t: threading.Thread | None = None
567
- self._encoded_xfer_handshake_metadata: dict[int, Any] = {}
566
self._stop_event = threading.Event()
567
568
# Requests that need to start recv/send.
@@ -650,7 +648,6 @@ class NixlConnectorScheduler:
648
tp_rank,
649
str(len(encoded_data[tp_rank])),
650
)
653
- self._encoded_xfer_handshake_metadata = encoded_data
651
652
# Only start the listener when we have metadata to serve.
653
if self._nixl_handshake_listener_t is None:
@@ -995,7 +992,7 @@ class NixlConnectorWorker:
992
self.engine_id: EngineId = engine_id
993
self.tp_rank = get_tensor_model_parallel_rank()
994
self.world_size = get_tensor_model_parallel_world_size()
998
- self.tp_group = get_tp_group()
995
+
996
self.num_blocks = kv_cache_config.num_blocks
997
self.enable_permute_local_kv = False
998
@@ -1064,7 +1061,6 @@ class NixlConnectorWorker:
1061
# Number of NIXL regions. Currently one region per cache
1062
# (so 1 per layer for MLA, otherwise 2 per layer)
1063
self.num_regions = 0
1067
- self.num_layers = 0
1064
1065
# nixl_prepped_dlist_handle.
1066
self.src_xfer_handles_by_block_size: dict[int, int] = {}
@@ -1108,7 +1104,6 @@ class NixlConnectorWorker:
1104
1105
self.block_size = vllm_config.cache_config.block_size
1106
self.model_config = vllm_config.model_config
1111
- self.cache_config = vllm_config.cache_config
1107
1108
self.use_mla = self.model_config.use_mla
1109
@@ -1540,7 +1535,6 @@ class NixlConnectorWorker:
1535
1536
self.kv_caches_base_addr[self.engine_id][self.tp_rank] = seen_base_addresses
1537
self.num_regions = len(caches_data)
1543
- self.num_layers = len(xfer_buffers.keys())
1538
1539
descs = self.nixl_wrapper.get_reg_descs(caches_data, self.nixl_memory_type)
1540
logger.debug("Registering descs: %s", caches_data)
vllm/v1/core/sched/scheduler.py
+3
-5
@@ -184,13 +184,11 @@ class Scheduler(SchedulerInterface):
184
185
# Encoder-related.
186
# Calculate encoder cache size if applicable
187
- self.supports_mm_inputs = mm_registry.supports_multimodal_inputs(
187
+ supports_mm_inputs = mm_registry.supports_multimodal_inputs(
188
vllm_config.model_config
189
)
190
- self.mm_budget = mm_budget = (
191
- MultiModalBudget(vllm_config, mm_registry)
192
- if self.supports_mm_inputs
193
- else None
190
+ mm_budget = (
191
+ MultiModalBudget(vllm_config, mm_registry) if supports_mm_inputs else None
192
)
193
194
# NOTE: Text-only encoder-decoder models are implemented as
vllm/v1/engine/core.py
+1
-3
@@ -148,7 +148,7 @@ class EngineCore:
148
if self.scheduler.connector is not None: # type: ignore
149
self.model_executor.init_kv_output_aggregator(self.scheduler.connector) # type: ignore
150
151
- self.mm_registry = mm_registry = MULTIMODAL_REGISTRY
151
+ mm_registry = MULTIMODAL_REGISTRY
152
self.mm_receiver_cache = mm_registry.engine_receiver_cache_from_config(
153
vllm_config
154
)
@@ -800,8 +800,6 @@ class EngineCoreProc(EngineCore):
800
vllm_config,
801
client_handshake_address,
802
) as addresses:
803
- self.client_count = len(addresses.outputs)
804
-
803
# Set up data parallel environment.
804
self.has_coordinator = addresses.coordinator_output is not None
805
self.frontend_stats_publish_address = (