lmcache-cli 0.5.7rc4.dev0__py3-none-any.whl → 0.5.8.dev0__py3-none-any.whl

lmcache/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.5.7rc4.dev0'
-__version_tuple__ = version_tuple = (0, 5, 7, 'rc4', 'dev0')
+__version__ = version = '0.5.8.dev0'
+__version_tuple__ = version_tuple = (0, 5, 8, 'dev0')
 
-__commit_id__ = commit_id = 'g45a0bb420'
+__commit_id__ = commit_id = 'g236f3eec3'

lmcache/cli/commands/coordinator.py

@@ -13,6 +13,7 @@ import json
 # First Party
 from lmcache.cli.commands.base import BaseCommand
 from lmcache.logging import init_logger
+from lmcache.security.api_key import add_api_key_args
 
 logger = init_logger(__name__)
 
@@ -244,6 +245,10 @@ class CoordinatorCommand(BaseCommand):
                 "--event-transport kafka."
             ),
         )
+        # Shared authentication option (--api-key-file / LMCACHE_API_KEY*). An
+        # unset flag leaves the coordinator open; supplying a key closes every
+        # route except /healthz.
+        add_api_key_args(parser)
 
     def execute(self, args: argparse.Namespace) -> None:
         """Build the coordinator config and serve the app with uvicorn.
@@ -272,6 +277,9 @@ class CoordinatorCommand(BaseCommand):
                 KafkaCacheEventSourceConfig,
                 MPCoordinatorConfig,
             )
+            from lmcache.v1.mp_coordinator.http_auth import (
+                CoordinatorAuthenticationConfig,
+            )
             from lmcache.v1.mp_coordinator.observability import (
                 init_coordinator_metrics,
             )
@@ -324,7 +332,11 @@ class CoordinatorCommand(BaseCommand):
         config = MPCoordinatorConfig(**fields)
 
         init_coordinator_metrics(config)
-        app = create_app(config)
+        # ``api_key_file`` is registered by add_api_key_args; getattr keeps
+        # hand-built Namespaces (e.g. unit tests) working when they omit it.
+        key_files = getattr(args, "api_key_file", None) or []
+        http_authentication = CoordinatorAuthenticationConfig.from_sources(key_files)
+        app = create_app(config, http_authentication=http_authentication)
         uvicorn.run(
             app,
             host=config.host,

lmcache/inprocess/integration/base_service_factory.py

@@ -21,6 +21,7 @@ from lmcache.logging import init_logger
 
 if TYPE_CHECKING:
     # First Party
+    from lmcache.inprocess.observability import PrometheusLogger
     from lmcache.inprocess.v1.cache_engine import LMCacheEngine
     from lmcache.inprocess.v1.internal_api_server.api_server import InternalAPIServer
     from lmcache.inprocess.v1.lookup_client.abstract_client import LookupClientInterface
@@ -33,7 +34,6 @@ if TYPE_CHECKING:
     from lmcache.inprocess.v1.manager import LMCacheManager
     from lmcache.inprocess.v1.metadata import LMCacheMetadata
     from lmcache.inprocess.v1.offload_server.zmq_server import ZMQOffloadServer
-    from lmcache.observability import PrometheusLogger
     from lmcache.v1.plugin.runtime_plugin_launcher import RuntimePluginLauncher
 
 logger = init_logger(__name__)
@@ -107,7 +107,7 @@ class BaseServiceFactory(ABC):
         Shared implementation used by subclass maybe_create_health_monitor.
         """
         # First Party
-        from lmcache.observability import PrometheusLogger
+        from lmcache.inprocess.observability import PrometheusLogger
         from lmcache.v1.periodic_thread import (
             PeriodicThreadRegistry,
             ThreadLevel,

lmcache/inprocess/integration/vllm/vllm_service_factory.py

@@ -174,7 +174,7 @@ class VllmServiceFactory(BaseServiceFactory):
         ):
             # Create PrometheusLogger for scheduler without engine
             # First Party
-            from lmcache.observability import PrometheusLogger
+            from lmcache.inprocess.observability import PrometheusLogger
 
             PrometheusLogger.GetOrCreate(
                 self.metadata,

lmcache/inprocess/integration/vllm/vllm_v1_adapter.py

@@ -40,13 +40,13 @@ from lmcache.inprocess.integration.vllm.utils import (
 from lmcache.inprocess.integration.vllm.vllm_service_factory import (
     VllmServiceFactory,
 )
+from lmcache.inprocess.observability import LMCStatsMonitor, PrometheusLogger
 from lmcache.inprocess.v1.cache_engine import LMCacheEngine
 from lmcache.inprocess.v1.compute.blend import LMCBlenderBuilder
 from lmcache.inprocess.v1.config import LMCacheEngineConfig
 from lmcache.inprocess.v1.config_base import validate_and_set_config_value
 from lmcache.inprocess.v1.manager import LMCacheManager
 from lmcache.logging import init_logger
-from lmcache.observability import LMCStatsMonitor, PrometheusLogger
 from lmcache.utils import CacheStoreEvent, _lmcache_nvtx_annotate, cdiv
 
 if TYPE_CHECKING:

lmcache/{observability.py → inprocess/observability.py}

@@ -25,8 +25,8 @@ from prometheus_client import REGISTRY
 import prometheus_client
 
 # First Party
+from lmcache.inprocess.usage_telemetry.continuous import ContinuousUsageContext
 from lmcache.logging import init_logger
-from lmcache.usage_telemetry import ContinuousUsageContext
 from lmcache.utils import thread_safe
 
 if TYPE_CHECKING:

lmcache/inprocess/usage_telemetry/__init__.py

@@ -0,0 +1 @@
+# SPDX-License-Identifier: Apache-2.0

lmcache/{usage_telemetry → inprocess/usage_telemetry}/context.py

@@ -1,40 +1,29 @@
 # SPDX-License-Identifier: Apache-2.0
-"""Context reporting: snapshot messages describing a deployment.
+"""Context reporting for the single-process (InProcess) engine.
 
-Context reports are sent once each (at startup, or at KV-cache registration
-for the MP path in :mod:`.mp`), in contrast to the periodic interval
-counters in :mod:`.continuous`.
+Context reports are sent once at engine startup, in contrast to the
+periodic interval counters in :mod:`.continuous`. The MP cache server's
+one-shot reporter lives in :mod:`lmcache.usage_telemetry.mp` and shares
+:class:`lmcache.usage_telemetry.base.UsageContextBase` with this module.
 """
 
 # Future
 from __future__ import annotations
 
 # Standard
-from abc import ABC, abstractmethod
 from datetime import datetime
 from typing import TYPE_CHECKING
 import threading
 
 # First Party
+from lmcache.inprocess.usage_telemetry.messages import EngineMessage, MetadataMessage
 from lmcache.logging import init_logger
+from lmcache.usage_telemetry.base import UsageContextBase
 from lmcache.usage_telemetry.env_probe import collect_env_message
 from lmcache.usage_telemetry.guard import swallow_telemetry_errors
-from lmcache.usage_telemetry.identity import (
-    get_usage_identity,
-    is_usage_tracking_enabled,
-)
-from lmcache.usage_telemetry.messages import (
-    DeploymentMode,
-    EngineMessage,
-    MetadataMessage,
-    UsageMessage,
-)
-from lmcache.usage_telemetry.transport import (
-    DEFAULT_SENDER,
-    UsageMessageSender,
-    build_usage_payload,
-    usage_server_url,
-)
+from lmcache.usage_telemetry.identity import is_usage_tracking_enabled
+from lmcache.usage_telemetry.messages import DeploymentMode, UsageMessage
+from lmcache.usage_telemetry.transport import UsageMessageSender
 
 if TYPE_CHECKING:
     # First Party
@@ -44,65 +33,6 @@ if TYPE_CHECKING:
 logger = init_logger(__name__)
 
 
-class UsageContextBase(ABC):
-    """Shared plumbing for one-shot usage reporting.
-
-    Subclasses define which messages the report contains via
-    :meth:`_collect_messages`; this base owns the identity, transport, and
-    optional local logging. Each message is POSTed to the endpoint it
-    declares in the schema (:mod:`lmcache.usage_telemetry.messages`).
-    """
-
-    def __init__(
-        self,
-        mode: DeploymentMode,
-        local_log: str | None,
-        sender: UsageMessageSender | None,
-    ) -> None:
-        """Initialize shared reporting state.
-
-        Args:
-            mode: Deployment mode stamped on every payload this reporter
-                sends.
-            local_log: Path of a human-readable local log of every sent
-                payload; ``None`` disables local logging.
-            sender: Message transport; ``None`` selects the default HTTP
-                sender.
-        """
-        self._mode = mode
-        self._local_log = local_log
-        self._sender = sender if sender is not None else DEFAULT_SENDER
-        self._identity = get_usage_identity()
-        self._start_time = datetime.now()
-
-    @abstractmethod
-    def _collect_messages(self) -> list[UsageMessage]:
-        """Return the messages of the one-shot report."""
-        raise NotImplementedError
-
-    @swallow_telemetry_errors
-    def report_once(self) -> None:
-        """Collect and send every one-shot message on the calling thread.
-
-        Failures are swallowed; never raises.
-        """
-        for message in self._collect_messages():
-            payload = build_usage_payload(message, self._identity, self._mode)
-            self._sender.send(usage_server_url(message.ENDPOINT), payload)
-            self._write_local(payload)
-
-    def _write_local(self, payload: dict[str, object]) -> None:
-        """Append *payload* to the local log file, if one is configured."""
-        if self._local_log is None:
-            return
-        text = "".join(f"{key}: {value}\n" for key, value in payload.items()) + "\n"
-        try:
-            with open(self._local_log, "a") as f:
-                f.write(text)
-        except OSError:
-            logger.debug("Unable to write usage log to %s", self._local_log)
-
-
 class UsageContext(UsageContextBase):
     """One-shot usage reporter for the single-process LMCacheEngine path.
 

lmcache/{usage_telemetry → inprocess/usage_telemetry}/continuous.py

@@ -13,6 +13,7 @@ import time
 import numpy as np
 
 # First Party
+from lmcache.inprocess.usage_telemetry.messages import CacheLifespanMessage
 from lmcache.logging import init_logger
 from lmcache.usage_telemetry.guard import swallow_telemetry_errors
 from lmcache.usage_telemetry.identity import (
@@ -20,7 +21,6 @@ from lmcache.usage_telemetry.identity import (
     is_usage_tracking_enabled,
 )
 from lmcache.usage_telemetry.messages import (
-    CacheLifespanMessage,
     ContinuousContextMessage,
     DeploymentMode,
 )
@@ -33,8 +33,8 @@ from lmcache.usage_telemetry.transport import (
 
 if TYPE_CHECKING:
     # First Party
+    from lmcache.inprocess.observability import LMCacheStats
     from lmcache.inprocess.v1.metadata import LMCacheMetadata
-    from lmcache.observability import LMCacheStats
 
 logger = init_logger(__name__)
 

lmcache/inprocess/usage_telemetry/messages.py

@@ -0,0 +1,111 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Wire schema for the single-process (InProcess) usage-telemetry messages.
+
+The single source of truth for the whole wire schema lives in
+:mod:`lmcache.usage_telemetry.messages`; this module only holds the
+messages that the InProcess engine can send, so that MP-side code never
+pulls InProcess types through the shared schema module.
+"""
+
+# Future
+from __future__ import annotations
+
+# Standard
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, ClassVar
+
+# Third Party
+import torch
+
+# First Party
+from lmcache import torch_device_type
+from lmcache.usage_telemetry.messages import UsageMessage
+
+if TYPE_CHECKING:
+    # First Party
+    from lmcache.inprocess.v1.config import LMCacheEngineConfig
+    from lmcache.inprocess.v1.metadata import LMCacheMetadata
+
+
+@dataclass
+class EngineMessage(UsageMessage):
+    """Configuration snapshot of a single-process LMCacheEngine.
+
+    Sent once at engine startup (single-process integrations only).
+    """
+
+    chunksize: int
+    local_device: str
+    max_local_cache_size: int
+    remote_url: str | None
+    remote_serde: str | None
+    pipelined_backend: bool
+    save_decode_cache: bool
+    enable_blending: bool
+    blend_recompute_ratio: float
+    blend_min_tokens: int
+    model_name: str
+    world_size: int
+    worker_id: int
+    kv_dtype: torch.dtype
+    kv_shape: tuple[int, int, int, int, int]
+
+    @classmethod
+    def from_config(
+        cls, config: LMCacheEngineConfig, metadata: LMCacheMetadata
+    ) -> EngineMessage:
+        """Build the message from the engine configuration and metadata.
+
+        Args:
+            config: The engine configuration to snapshot.
+            metadata: The engine metadata (model, world size, kv layout).
+
+        Returns:
+            The populated message.
+        """
+        return cls(
+            chunksize=config.chunk_size,
+            local_device="cpu" if config.local_cpu else torch_device_type,
+            max_local_cache_size=int(config.max_local_cpu_size),
+            remote_url=config.remote_url,
+            remote_serde=config.remote_serde,
+            pipelined_backend=False,
+            save_decode_cache=config.save_decode_cache,
+            enable_blending=config.enable_blending,
+            blend_recompute_ratio=0.15,
+            blend_min_tokens=config.blend_min_tokens,
+            model_name=metadata.model_name,
+            world_size=metadata.world_size,
+            worker_id=metadata.worker_id,
+            kv_dtype=metadata.kv_dtype,
+            kv_shape=metadata.kv_shape,
+        )
+
+
+@dataclass
+class MetadataMessage(UsageMessage):
+    """Process start time and uptime at send time.
+
+    Sent once at engine startup (single-process integrations only).
+    """
+
+    start_time: str
+    duration: float
+
+
+@dataclass
+class CacheLifespanMessage(UsageMessage):
+    """Cache-entry lifespan histogram flushed with each continuous report.
+
+    Single-process mode only; measures the store-to-reuse gap per reused
+    chunk.
+    """
+
+    ENDPOINT: ClassVar[str] = "cache-lifespan"
+
+    cache_lifespan_histogram: dict[float, int]
+    """Bucket lower bound (minutes) to sample count. Numeric keys become
+    strings on the wire (JSON object keys)."""
+    sequence_number: int
+    uptime_seconds: float
+    """Seconds since the reporting process started."""

lmcache/inprocess/v1/cache_engine.py

@@ -29,6 +29,8 @@ import torch
 
 # First Party
 from lmcache import torch_dev, torch_device_type
+from lmcache.inprocess.observability import LMCacheStatsLogger, LMCStatsMonitor
+from lmcache.inprocess.usage_telemetry.context import InitializeUsageContext
 from lmcache.inprocess.v1.config import LMCacheEngineConfig
 from lmcache.inprocess.v1.event_manager import EventManager, EventStatus, EventType
 from lmcache.inprocess.v1.gpu_connector.gpu_connectors import GPUConnectorInterface
@@ -46,8 +48,6 @@ from lmcache.inprocess.v1.token_database import (
     TokenDatabase,
 )
 from lmcache.logging import init_logger
-from lmcache.observability import LMCacheStatsLogger, LMCStatsMonitor
-from lmcache.usage_telemetry import InitializeUsageContext
 from lmcache.utils import (
     CacheEngineKey,
     CacheStoreEvent,

lmcache/inprocess/v1/health_monitor/checks/remote_backend_check.py

@@ -13,6 +13,7 @@ import time
 import torch
 
 # First Party
+from lmcache.inprocess.observability import LMCStatsMonitor
 from lmcache.inprocess.v1.health_monitor.base import HealthCheck
 from lmcache.inprocess.v1.health_monitor.constants import (
     DEFAULT_FALLBACK_POLICY,
@@ -32,7 +33,6 @@ from lmcache.inprocess.v1.storage_backend.connector.audit_connector import (
     AuditConnector,
 )
 from lmcache.logging import init_logger
-from lmcache.observability import LMCStatsMonitor
 from lmcache.utils import CacheEngineKey
 
 if TYPE_CHECKING:

lmcache/inprocess/v1/lookup_client/chunk_statistics_lookup_client.py

@@ -5,13 +5,13 @@ import threading
 import time
 
 # First Party
+from lmcache.inprocess.observability import PrometheusLogger
 from lmcache.inprocess.v1.lookup_client.abstract_client import LookupClientInterface
 from lmcache.inprocess.v1.lookup_client.record_strategies import (
     AsyncRecorder,
     create_record_strategy,
 )
 from lmcache.logging import init_logger
-from lmcache.observability import PrometheusLogger
 
 if TYPE_CHECKING:
     # Third Party

lmcache/inprocess/v1/storage_backend/batched_message_sender.py

@@ -5,6 +5,7 @@ import queue
 import threading
 
 # First Party
+from lmcache.inprocess.observability import PrometheusLogger
 from lmcache.inprocess.v1.cache_controller.message import (
     BatchedKVOperationMsg,
     KVOpEvent,
@@ -13,7 +14,6 @@ from lmcache.inprocess.v1.cache_controller.message import (
 from lmcache.inprocess.v1.config import LMCacheEngineConfig
 from lmcache.inprocess.v1.metadata import LMCacheMetadata
 from lmcache.logging import init_logger
-from lmcache.observability import PrometheusLogger
 
 if TYPE_CHECKING:
     # First Party

lmcache/inprocess/v1/storage_backend/cache_policy/lru.py

@@ -5,12 +5,12 @@ from typing import Any, Dict
 import time
 
 # First Party
+from lmcache.inprocess.observability import LMCStatsMonitor
 from lmcache.inprocess.v1.storage_backend.cache_policy.base_policy import (
     BaseCachePolicy,
     KeyType,
 )
 from lmcache.logging import init_logger
-from lmcache.observability import LMCStatsMonitor
 from lmcache.utils import CacheEngineKey
 
 logger = init_logger(__name__)

lmcache/inprocess/v1/storage_backend/connector/instrumented_connector.py

@@ -4,11 +4,11 @@ from typing import List, Optional
 import time
 
 # First Party
+from lmcache.inprocess.observability import LMCStatsMonitor
 from lmcache.inprocess.v1.storage_backend.connector.base_connector import (
     RemoteConnector,
 )
 from lmcache.logging import init_logger
-from lmcache.observability import LMCStatsMonitor
 from lmcache.utils import CacheEngineKey
 from lmcache.v1.memory_management import MemoryObj
 

lmcache/inprocess/v1/storage_backend/connector/sagemaker_hyperpod_connector.py

@@ -14,6 +14,7 @@ import aiohttp
 import torch
 
 # First Party
+from lmcache.inprocess.observability import LMCStatsMonitor
 from lmcache.inprocess.v1.protocol import RemoteMetadata
 from lmcache.inprocess.v1.storage_backend.connector.base_connector import (
     RemoteConnector,
@@ -23,7 +24,6 @@ from lmcache.inprocess.v1.storage_backend.job_executor.pq_executor import (
 )
 from lmcache.inprocess.v1.storage_backend.local_cpu_backend import LocalCPUBackend
 from lmcache.logging import init_logger
-from lmcache.observability import LMCStatsMonitor
 from lmcache.utils import CacheEngineKey
 from lmcache.v1.memory_management import MemoryObj
 

lmcache/inprocess/v1/storage_backend/local_cpu_backend.py

@@ -11,6 +11,7 @@ import torch
 
 # First Party
 from lmcache import torch_dev, torch_device_type
+from lmcache.inprocess.observability import LMCStatsMonitor, PrometheusLogger
 from lmcache.inprocess.v1.cache_controller.message import OpType
 from lmcache.inprocess.v1.config import LMCacheEngineConfig
 from lmcache.inprocess.v1.memory_allocators.paged_cpu_gpu_memory_allocator import (
@@ -27,7 +28,6 @@ from lmcache.inprocess.v1.storage_backend.cache_policy import get_cache_policy
 from lmcache.inprocess.v1.system_detection import NUMADetector, SystemMemoryDetector
 from lmcache.integration.vllm.utils import get_size_bytes
 from lmcache.logging import init_logger
-from lmcache.observability import LMCStatsMonitor, PrometheusLogger
 from lmcache.utils import CacheEngineKey, _lmcache_nvtx_annotate
 from lmcache.v1.memory_allocators.mixed_memory_allocator import MixedMemoryAllocator
 from lmcache.v1.memory_management import (

lmcache/inprocess/v1/storage_backend/local_disk_backend.py

@@ -13,6 +13,7 @@ import torch
 
 # First Party
 from lmcache import torch_dev, torch_device_type
+from lmcache.inprocess.observability import LMCStatsMonitor
 from lmcache.inprocess.v1.cache_controller.message import OpType
 from lmcache.inprocess.v1.config import LMCacheEngineConfig
 from lmcache.inprocess.v1.metadata import LMCacheMetadata
@@ -29,7 +30,6 @@ from lmcache.inprocess.v1.storage_backend.job_executor.pq_executor import (
 from lmcache.inprocess.v1.storage_backend.local_cpu_backend import LocalCPUBackend
 from lmcache.inprocess.v1.storage_backend.path_sharder import PathSharder
 from lmcache.logging import init_logger
-from lmcache.observability import LMCStatsMonitor
 from lmcache.utils import CacheEngineKey, DiskCacheMetadata, _lmcache_nvtx_annotate
 from lmcache.v1.memory_management import MemoryFormat, MemoryObj
 

lmcache/inprocess/v1/storage_backend/p2p_backend.py

@@ -12,6 +12,7 @@ import zmq
 import zmq.asyncio
 
 # First Party
+from lmcache.inprocess.observability import LMCStatsMonitor
 from lmcache.inprocess.v1.cache_controller.message import (
     BatchedP2PLookupMsg,
     BatchedP2PLookupRetMsg,
@@ -38,7 +39,6 @@ from lmcache.inprocess.v1.transfer_channel.transfer_utils import (
     P2PInitSideRetMsg,
 )
 from lmcache.logging import init_logger
-from lmcache.observability import LMCStatsMonitor
 from lmcache.utils import CacheEngineKey
 from lmcache.v1.memory_management import (
     MemoryFormat,

lmcache/inprocess/v1/storage_backend/plugins/rust_raw_block_backend.py

@@ -25,7 +25,6 @@ from lmcache.v1.distributed.l2_adapters.raw_block import (
     encode_legacy_key,
     normalize_raw_block_io_engine,
     round_up,
-    validate_raw_block_io_options,
 )
 
 if TYPE_CHECKING:
@@ -146,7 +145,7 @@ class RustRawBlockBackend(StoragePluginInterface):
         )
         if self._core.io_engine == "io_uring":
             try:
-                self._core.register_fixed_buffers_from_allocator(
+                self._core.register_io_buffers_from_allocator(
                     self.local_cpu_backend.get_memory_allocator()
                 )
             except Exception as e:
@@ -207,13 +206,25 @@ class RustRawBlockBackend(StoragePluginInterface):
 
     @property
     def _raw(self) -> Any:
-        """Return the raw device handle for legacy test compatibility."""
-        return self._core.raw_device()
+        """Return the raw device handle for legacy test compatibility.
+
+        The device stays private on the engine; this exposes the lazily opened
+        handle so tests can inspect or replace it without reaching into the
+        engine's public I/O interface.
+        """
+        assert self._core._io_engine is not None
+        return self._core._io_engine._raw_device()  # type: ignore[attr-defined]
 
     @_raw.setter
     def _raw(self, raw_device: Any) -> None:
-        """Replace the raw device handle for legacy test compatibility."""
-        self._core.set_raw_device_for_testing(raw_device)
+        """Replace the raw device handle for legacy test compatibility.
+
+        The injected handle is installed directly on the engine's private
+        ``_raw`` attribute so recovery stays routed through the public I/O
+        interface rather than a leaked device accessor.
+        """
+        assert self._core._io_engine is not None
+        self._core._io_engine._raw = raw_device  # type: ignore[attr-defined]
 
     def lock_refcount(self, encoded_key: str) -> int:
         """Return the L2 lock refcount for a legacy encoded key."""
@@ -257,9 +268,6 @@ class RustRawBlockBackend(StoragePluginInterface):
         max_data_transfer_size = int(
             extra.get("rust_raw_block.max_data_transfer_size", 0)
         )
-        validate_raw_block_io_options(
-            iouring_queue_depth=iouring_queue_depth,
-        )
         meta_total_bytes = int(
             extra.get("rust_raw_block.meta_total_bytes", 128 * 1024 * 1024)
         )

lmcache/inprocess/v1/storage_backend/remote_backend.py

@@ -8,6 +8,7 @@ import time
 
 # First Party
 from lmcache import torch_device_type
+from lmcache.inprocess.observability import LMCStatsMonitor, PrometheusLogger
 from lmcache.inprocess.v1.config import LMCacheEngineConfig
 from lmcache.inprocess.v1.metadata import LMCacheMetadata
 from lmcache.inprocess.v1.storage_backend.abstract_backend import (
@@ -20,7 +21,6 @@ from lmcache.inprocess.v1.storage_backend.connector.base_connector import (
 from lmcache.inprocess.v1.storage_backend.local_cpu_backend import LocalCPUBackend
 from lmcache.inprocess.v1.storage_backend.naive_serde import CreateSerde
 from lmcache.logging import init_logger
-from lmcache.observability import LMCStatsMonitor, PrometheusLogger
 from lmcache.utils import CacheEngineKey, _lmcache_nvtx_annotate
 from lmcache.v1.exceptions import IrrecoverableException
 from lmcache.v1.memory_management import MemoryObj

lmcache/inprocess/v1/storage_backend/storage_manager.py

@@ -24,6 +24,7 @@ import torch
 
 # First Party
 from lmcache import torch_dev, torch_device_type
+from lmcache.inprocess.observability import PrometheusLogger
 from lmcache.inprocess.v1.config import LMCacheEngineConfig
 from lmcache.inprocess.v1.event_manager import EventManager, EventStatus, EventType
 from lmcache.inprocess.v1.metadata import LMCacheMetadata
@@ -34,7 +35,6 @@ from lmcache.inprocess.v1.storage_backend.abstract_backend import (
 )
 from lmcache.inprocess.v1.storage_backend.local_cpu_backend import LocalCPUBackend
 from lmcache.logging import init_logger
-from lmcache.observability import PrometheusLogger
 from lmcache.utils import (
     CacheEngineKey,
     _lmcache_nvtx_annotate,

lmcache/integration/sglang/unified_lmcache_mp_connector.py

@@ -551,7 +551,6 @@ class UnifiedLMCacheMPConnector:
                 self.kv_world_size,
                 self.blocks_in_chunk,
                 self._mq_timeout,
-                layout_hints={"kv_list_layout": "unified"},
                 engine_group_infos=engine_group_infos,
                 engine_type=EngineType.SGLANG,
                 lmcache_chunk_tokens=self.chunk_size,

lmcache/integration/vllm/kv_cache_groups.py

@@ -398,10 +398,11 @@ def create_engine_group_infos_from_vllm(
     # Within one vLLM engine group, layers can have different hidden dimensions
     # (e.g. a different head count), which require different GPU copy kernels.
     # ``group_layers_by_identity`` splits each engine group further by physical
-    # transfer identity (kv_size, num_heads, head_size, block_size, dtype), so
-    # every resulting LMCache group can be served by a single copy kernel. It is
-    # the shared, engine-neutral primitive the server reuses to reproduce the
-    # same grouping from the registered tensors.
+    # transfer identity (kv_size, num_heads, head_size, block_size,
+    # engine_group_idx, dtype, engine_kv_format, block_stride_elems), so every
+    # resulting LMCache group can be served by a single copy kernel. It is the
+    # shared, engine-neutral primitive the server reuses to reproduce the same
+    # grouping from the registered tensors.
     return [
         EngineGroupInfo(
             engine_group_id=identity.engine_group_idx,