vllm-cpu 0.8.5.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/distributed/device_communicators/xpu_communicator.py

@@ -0,0 +1,54 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional
+
+import torch
+import torch.distributed as dist
+from torch.distributed import ProcessGroup
+
+from .base_device_communicator import DeviceCommunicatorBase
+
+
+class XpuCommunicator(DeviceCommunicatorBase):
+
+    def __init__(self,
+                 cpu_group: ProcessGroup,
+                 device: Optional[torch.device] = None,
+                 device_group: Optional[ProcessGroup] = None,
+                 unique_name: str = ""):
+        super().__init__(cpu_group, device, device_group, unique_name)
+
+    def all_reduce(self, input_) -> torch.Tensor:
+        dist.all_reduce(input_, group=self.device_group)
+        return input_
+
+    def gather(self,
+               input_: torch.Tensor,
+               dst: int = 0,
+               dim: int = -1) -> Optional[torch.Tensor]:
+        assert -input_.dim() <= dim < input_.dim(), (
+            f"Invalid dim ({dim}) for input tensor with shape {input_.size()}")
+        if dim < 0:
+            # Convert negative dim to positive.
+            dim += input_.dim()
+        # For xpu path, gather doesn't work properly together with ray
+        # cluster so we use all_gather instead for now.
+        input_size = input_.size()
+        # Allocate output tensor.
+        output_tensor = torch.empty((self.world_size, ) + input_size,
+                                    dtype=input_.dtype,
+                                    device=input_.device)
+        # All-gather.
+        dist.all_gather_into_tensor(output_tensor,
+                                    input_,
+                                    group=self.device_group)
+        if self.rank_in_group == dst:
+            # Reshape
+            output_tensor = output_tensor.movedim(0, dim)
+            output_tensor = output_tensor.reshape(input_size[:dim] +
+                                                  (self.world_size *
+                                                   input_size[dim], ) +
+                                                  input_size[dim + 1:])
+        else:
+            output_tensor = None
+        return output_tensor

vllm/distributed/kv_transfer/README.md

@@ -0,0 +1,29 @@
+
+# Distributed KV cache transfer
+
+This folder implements distributed KV cache transfer across vLLM instances.
+Currently the main usecase is for disaggregated prefilling.
+
+## Abstractions
+
+The KV cache transfer contains three layer of abstractions:
+
+- KV pipe: a FIFO pipe for torch.tensor transmission. Key APIs: `send_tensor` and `recv_tensor`.
+- KV lookup buffer: a lookup buffer for KV caches. Key: the tokens, value: the KV caches (and/or hidden states). Key APIs: `insert` and `drop_select` (similar to SQL semantics).
+- KV connector: a connector that connects the KV pipe and KV lookup buffer to vLLM. Key APIs: `send_kv_caches_and_hidden_states` and `recv_kv_caches_and_hidden_states`.
+
+Why we need KV lookup buffer: FIFO pipe itself is not enough as prefill vLLM worker may process requests in a different order compared to decode vLLM worker. Say the QPS is really high, prefill worker may handle requests in order A -> B -> C, but the decode worker may process request C first. This is not the case that can be naturally handled by FIFO pipe, so we provide KV lookup buffer to help translate a FIFO pipe to a lookup buffer.
+
+NOTE: KV pipe layer is bypassible: you can skip this layer if your distributed
+communication service already supports key-value-based lookup (like redis or
+RDMA database).
+
+NOTE: If you want to not only transfer KV caches, but adjust the model execution flow of vLLM as well (for example, allow vLLM to receive KV caches on some tokens and do prefill on the remaining tokens), you can bypass both KV pipe layer and KV lookup buffer layer, and directly implement on KV connector layer. Bear in mind that as vLLM's model input is constantly changing, this implementation will likely be broken when vLLM has new updates.
+
+## Disaggregated prefilling
+
+The example usage is in [this file](../../../examples/online_serving/disaggregated_prefill.sh).
+
+Here is the diagram of how we run disaggregated prefilling.
+
+![Disaggregated prefill workflow](./disagg_prefill_workflow.jpg)

vllm/distributed/kv_transfer/__init__.py

@@ -0,0 +1,11 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from vllm.distributed.kv_transfer.kv_transfer_state import (
+    ensure_kv_transfer_initialized, get_kv_transfer_group,
+    has_kv_transfer_group, is_v1_kv_transfer_group)
+
+__all__ = [
+    "get_kv_transfer_group", "has_kv_transfer_group",
+    "is_v1_kv_transfer_group", "ensure_kv_transfer_initialized",
+    "KVConnectorBaseType"
+]

vllm/distributed/kv_transfer/kv_connector/base.py

@@ -0,0 +1,127 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+KVConnectorBase Class for Distributed KV Cache & Hidden State communication
+
+The class provides two primary abstract methods:
+1. send_kv_caches_and_hidden_states(): Send KV caches and hidden states
+2. recv_kv_caches_and_hidden_states(): Recv KV caches and hidden states
+"""
+
+from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING, List, Tuple, Union
+
+import torch
+
+from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorBase_V1
+from vllm.sequence import IntermediateTensors
+
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
+    from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
+
+
+class KVConnectorBase(ABC):
+    """
+    Abstract base class for a KV connector.
+
+    The class provides two primary abstract methods:
+    1. send_kv_caches_and_hidden_states(): Send KV caches and hidden states
+    2. recv_kv_caches_and_hidden_states(): Recv KV caches and hidden states
+    """
+
+    @abstractmethod
+    def __init__(
+        self,
+        rank: int,
+        local_rank: int,
+        config: "VllmConfig",
+    ):
+        raise NotImplementedError
+
+    @abstractmethod
+    def close(self) -> None:
+        """Close the buffer and release resources.
+
+        This method is responsible for cleaning up resources related to the 
+        connector when it is no longer needed.
+
+        Raises:
+            NotImplementedError: This method must be implemented in subclasses.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def send_kv_caches_and_hidden_states(
+        self,
+        model_executable: torch.nn.Module,
+        model_input: "ModelInputForGPUWithSamplingMetadata",
+        kv_caches: List[torch.Tensor],
+        hidden_or_intermediate_states: Union[torch.Tensor,
+                                             IntermediateTensors],
+    ) -> None:
+        """
+        Send KV caches and hidden states to the connector.
+
+        This method processes the input tokens, KV caches, and 
+        hidden/intermediate states for a given model and sends the data to the 
+        decode instance.
+
+        Args:
+            model_executable (torch.nn.Module): The model executable containing 
+                start and end layer information.
+            model_input (ModelInputForGPUWithSamplingMetadata): The input
+                metadata from vLLM.
+            kv_caches (List[torch.Tensor]): List of KV caches (keys and values) 
+                for each layer.
+            hidden_or_intermediate_states (Union[torch.Tensor, 
+            IntermediateTensors]): 
+                The hidden or intermediate states associated with the tokens.
+
+        Returns:
+            None
+
+        """
+
+        raise NotImplementedError
+
+    @abstractmethod
+    def recv_kv_caches_and_hidden_states(
+        self, model_executable: torch.nn.Module,
+        model_input: "ModelInputForGPUWithSamplingMetadata",
+        kv_caches: List[torch.Tensor]
+    ) -> Tuple[Union[torch.Tensor, IntermediateTensors], bool,
+               "ModelInputForGPUWithSamplingMetadata"]:
+        """
+        Receive KV caches and hidden states from the connector.
+
+        This method attempts to retrieve KV caches and hidden states for input
+        tokens. If all required KV caches and hidden states are received, it
+        will bypass model input, else it will fall back to normal vLLM model 
+        forwarding.
+
+        Args:
+            model_executable (torch.nn.Module): 
+                The model executable from vLLM modelrunner.
+            model_input (ModelInputForGPUWithSamplingMetadata): 
+                The model input from vLLM modelrunner.
+            kv_caches (List[torch.Tensor]): 
+                List of KV caches for each layer.
+
+        Returns:
+            - hidden_or_intermediate_states (torch.Tensor or
+            IntermediateTensors): 
+                Concatenated hidden states if all required data is retrieved, 
+                otherwise `None`.
+            - bypass_model_exec (bool): 
+                Indicates whether the model execution can be skipped (True) or 
+                needs to be redone (False).
+            - model_input (ModelInputForGPUWithSamplingMetadata): 
+                Optionally adjusted input metadata for re-execution when 
+                `bypass_model_exec=False`.
+
+        """
+
+        raise NotImplementedError
+
+
+KVConnectorBaseType = Union[KVConnectorBase, KVConnectorBase_V1]

vllm/distributed/kv_transfer/kv_connector/factory.py

@@ -0,0 +1,107 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import importlib
+from typing import TYPE_CHECKING, Callable, Dict, Type
+
+import vllm.envs as envs
+from vllm.distributed.kv_transfer.kv_connector.base import KVConnectorBaseType
+from vllm.distributed.kv_transfer.kv_connector.v1 import (KVConnectorBase_V1,
+                                                          KVConnectorRole)
+from vllm.logger import init_logger
+
+from .base import KVConnectorBase
+
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
+
+logger = init_logger(__name__)
+
+
+class KVConnectorFactory:
+    _registry: Dict[str, Callable[[], Type[KVConnectorBaseType]]] = {}
+
+    @classmethod
+    def register_connector(cls, name: str, module_path: str,
+                           class_name: str) -> None:
+        """Register a connector with a lazy-loading module and class name."""
+        if name in cls._registry:
+            raise ValueError(f"Connector '{name}' is already registered.")
+
+        def loader() -> Type[KVConnectorBaseType]:
+            module = importlib.import_module(module_path)
+            return getattr(module, class_name)
+
+        cls._registry[name] = loader
+
+    @classmethod
+    def create_connector_v0(cls, rank: int, local_rank: int,
+                            config: "VllmConfig") -> KVConnectorBase:
+        if envs.VLLM_USE_V1:
+            raise ValueError("Attempting to initialize a V0 Connector, "
+                             f"but found {envs.VLLM_USE_V1=}")
+
+        connector_name = config.kv_transfer_config.kv_connector
+        if connector_name not in cls._registry:
+            raise ValueError(f"Unsupported connector type: {connector_name}")
+
+        connector_cls = cls._registry[connector_name]()
+        assert issubclass(connector_cls, KVConnectorBase)
+        return connector_cls(rank, local_rank, config)
+
+    @classmethod
+    def create_connector_v1(
+        cls,
+        config: "VllmConfig",
+        role: KVConnectorRole,
+    ) -> KVConnectorBase_V1:
+        if not envs.VLLM_USE_V1:
+            raise ValueError("Attempting to initialize a V1 Connector, "
+                             f"but found {envs.VLLM_USE_V1=}")
+
+        connector_name = config.kv_transfer_config.kv_connector
+        connector_cls = cls._registry[connector_name]()
+        assert issubclass(connector_cls, KVConnectorBase_V1)
+        logger.info("Creating v1 connector with name: %s", connector_name)
+        # NOTE(Kuntai): v1 connector is explicitly separated into two roles.
+        # Scheduler connector:
+        # - Co-locate with scheduler process
+        # - Should only be used inside the Scheduler class
+        # Worker connector:
+        # - Co-locate with worker process
+        # - Should only be used inside the forward context & attention layer
+        # We build separately to enforce strict separation
+        return connector_cls(config, role)
+
+
+# Register various connectors here.
+# The registration should not be done in each individual file, as we want to
+# only load the files corresponding to the current connector.
+KVConnectorFactory.register_connector(
+    "PyNcclConnector",
+    "vllm.distributed.kv_transfer.kv_connector.simple_connector",
+    "SimpleConnector")
+
+KVConnectorFactory.register_connector(
+    "MooncakeConnector",
+    "vllm.distributed.kv_transfer.kv_connector.simple_connector",
+    "SimpleConnector")
+
+KVConnectorFactory.register_connector(
+    "LMCacheConnector",
+    "vllm.distributed.kv_transfer.kv_connector.lmcache_connector",
+    "LMCacheConnector")
+
+KVConnectorFactory.register_connector(
+    "MooncakeStoreConnector",
+    "vllm.distributed.kv_transfer.kv_connector.mooncake_store_connector",
+    "MooncakeStoreConnector")
+
+KVConnectorFactory.register_connector(
+    "SharedStorageConnector",
+    "vllm.distributed.kv_transfer.kv_connector.v1.shared_storage_connector",
+    "SharedStorageConnector")
+
+KVConnectorFactory.register_connector(
+    "LMCacheConnectorV1",
+    "vllm.distributed.kv_transfer.kv_connector.v1.lmcache_connector",
+    "LMCacheConnectorV1")

vllm/distributed/kv_transfer/kv_connector/lmcache_connector.py

@@ -0,0 +1,98 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+LMCache KV Cache Connector for Distributed Machine Learning Inference
+
+The LMCacheConnector can (1) transfer KV caches between prefill vLLM worker
+(KV cache producer) and decode vLLM worker (KV cache consumer) using LMCache;
+(2) offload and share KV caches.
+"""
+
+from typing import TYPE_CHECKING, List, Tuple, Union
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.distributed.kv_transfer.kv_connector.base import KVConnectorBase
+from vllm.logger import init_logger
+from vllm.sequence import IntermediateTensors
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
+
+logger = init_logger(__name__)
+
+
+class LMCacheConnector(KVConnectorBase):
+
+    def __init__(
+        self,
+        rank: int,
+        local_rank: int,
+        config: VllmConfig,
+    ):
+
+        self.transfer_config = config.kv_transfer_config
+        self.vllm_config = config
+
+        from lmcache.experimental.cache_engine import LMCacheEngineBuilder
+        from lmcache.integration.vllm.utils import ENGINE_NAME
+        from lmcache.integration.vllm.vllm_adapter import (
+            RetrieveStatus, StoreStatus, init_lmcache_engine,
+            lmcache_retrieve_kv, lmcache_should_retrieve, lmcache_should_store,
+            lmcache_store_kv)
+        logger.info("Initializing LMCacheConfig under kv_transfer_config %s",
+                    self.transfer_config)
+
+        # TODO (Jiayi): Find model_config, parallel_config, and cache_config
+        self.engine = init_lmcache_engine(config.model_config,
+                                          config.parallel_config,
+                                          config.cache_config)
+        self.lmcache_engine_name = ENGINE_NAME
+        self.lmcache_engine_builder = LMCacheEngineBuilder
+
+        self.model_config = config.model_config
+        self.parallel_config = config.parallel_config
+        self.cache_config = config.cache_config
+        self.lmcache_retrieve_kv = lmcache_retrieve_kv
+        self.lmcache_store_kv = lmcache_store_kv
+        self.lmcache_should_retrieve = lmcache_should_retrieve
+        self.lmcache_should_store = lmcache_should_store
+        self.store_status = StoreStatus
+        self.retrieve_status = RetrieveStatus
+
+    def recv_kv_caches_and_hidden_states(
+        self, model_executable: torch.nn.Module,
+        model_input: "ModelInputForGPUWithSamplingMetadata",
+        kv_caches: List[torch.Tensor]
+    ) -> Tuple[Union[torch.Tensor, IntermediateTensors], bool,
+               "ModelInputForGPUWithSamplingMetadata"]:
+
+        retrieve_status = self.lmcache_should_retrieve(model_input)
+        model_input, bypass_model_exec, hidden_or_intermediate_states =\
+            self.lmcache_retrieve_kv(
+                model_executable, model_input, self.cache_config, kv_caches,
+                retrieve_status)
+        return hidden_or_intermediate_states, bypass_model_exec, model_input
+
+    def send_kv_caches_and_hidden_states(
+        self,
+        model_executable: torch.nn.Module,
+        model_input: "ModelInputForGPUWithSamplingMetadata",
+        kv_caches: List[torch.Tensor],
+        hidden_or_intermediate_states: Union[torch.Tensor,
+                                             IntermediateTensors],
+    ) -> None:
+
+        store_status = self.lmcache_should_store(model_input)
+        self.lmcache_store_kv(
+            self.model_config,
+            self.parallel_config,
+            self.cache_config,
+            model_executable,
+            model_input,
+            kv_caches,
+            store_status,
+        )
+
+    def close(self):
+        self.lmcache_engine_builder.destroy(self.lmcache_engine_name)

vllm/distributed/kv_transfer/kv_connector/mooncake_store_connector.py

@@ -0,0 +1,201 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+MooncakeStore Connector for Distributed Machine Learning Inference
+The MooncakeStoreConnector transfers KV caches between prefill vLLM workers
+(KV cache producer) and decode vLLM workers (KV cache consumer) using a
+database-style KVStore.
+"""
+import hashlib
+from typing import TYPE_CHECKING, List, Tuple, Union
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.distributed.kv_transfer.kv_connector.base import KVConnectorBase
+from vllm.distributed.kv_transfer.kv_connector.utils import (
+    model_aware_kv_ops_helper as kv_helper)
+from vllm.logger import init_logger
+from vllm.sequence import IntermediateTensors
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
+
+logger = init_logger(__name__)
+
+
+class MooncakeStoreConnector(KVConnectorBase):
+
+    def __init__(
+        self,
+        rank: int,
+        local_rank: int,
+        config: VllmConfig,
+    ):
+        self.config = config.kv_transfer_config
+        self.kv_helper = kv_helper(config)
+        self.local_tp_rank = local_rank
+
+        # Init kv_store
+        if self.config.kv_connector == "MooncakeStoreConnector":
+            # Check if MOONCAKE_CONFIG_PATH is set
+            import os
+            use_mooncake_store = os.getenv('MOONCAKE_CONFIG_PATH') is not None
+
+            if not use_mooncake_store:
+                raise ValueError(
+                    "To use MooncakeStoreConnector, you need to pass the ENV: "
+                    "'MOONCAKE_CONFIG_PATH=/path/to/mooncake_config.json'.")
+            else:
+                from vllm.distributed.kv_transfer.kv_lookup_buffer.mooncake_store import (  # noqa: E501
+                    MooncakeStore)
+                logger.info(
+                    "Initializing KVStoreConnector under kv_transfer_config %s",
+                    self.config)
+                self.kv_store = MooncakeStore(config)
+        else:
+            logger.error("Can not find %s", self.config.kv_connector)
+
+        assert self.kv_store is not None
+
+    def close(self) -> None:
+        """Close the buffer and release resources.
+        This method is responsible for cleaning up resources related to the 
+        connector when it is no longer needed.
+        Raises:
+            NotImplementedError: This method must be implemented in subclasses.
+        """
+        self.kv_store.close()
+
+    def send_kv_caches_and_hidden_states(
+        self,
+        model_executable: torch.nn.Module,
+        model_input: "ModelInputForGPUWithSamplingMetadata",
+        kv_caches: List[torch.Tensor],
+        hidden_or_intermediate_states: Union[torch.Tensor,
+                                             IntermediateTensors],
+    ) -> None:
+        input_tokens_tensor = model_input.input_tokens
+        seq_lens = model_input.attn_metadata.seq_lens
+        slot_mapping_flat = model_input.attn_metadata.slot_mapping.flatten()
+        start_layer = model_executable.model.start_layer
+        end_layer = model_executable.model.end_layer
+        num_heads, head_size = self.kv_helper.get_model_args(model_executable)
+
+        for idx, slen in enumerate(seq_lens):
+            start_pos = sum(seq_lens[:idx])
+            end_pos = start_pos + slen
+
+            current_tokens = input_tokens_tensor[start_pos:end_pos]
+            store_key_prefix = self.tensor_hash(current_tokens)
+            keys, values = [], []
+
+            for layer_id in range(start_layer, end_layer):
+                kv_cache = kv_caches[layer_id - start_layer]
+                key_cache, value_cache = self.kv_helper.get_kv_from_cache(
+                    kv_cache, num_heads, head_size)
+                current_slot_mapping = slot_mapping_flat[start_pos:end_pos]
+
+                keys.append(key_cache[current_slot_mapping].unsqueeze(0))
+                values.append(value_cache[current_slot_mapping].unsqueeze(0))
+
+            keys = torch.cat(keys, dim=0)
+            values = torch.cat(values, dim=0)
+            kvcache_to_sent = torch.stack((keys, values), dim=0)
+            store_kvcache_key = f"{store_key_prefix}_{self.local_tp_rank}"
+            self.kv_store.put(store_kvcache_key, kvcache_to_sent)
+
+            hidden_key = f"{store_key_prefix}_hidden_{self.local_tp_rank}"
+            self.kv_store.put(hidden_key,
+                              hidden_or_intermediate_states[start_pos:end_pos])
+
+        logger.debug("[rank%d]: KV send DONE.", torch.distributed.get_rank())
+
+    def recv_kv_caches_and_hidden_states(
+        self, model_executable: torch.nn.Module,
+        model_input: "ModelInputForGPUWithSamplingMetadata",
+        kv_caches: List[torch.Tensor]
+    ) -> Tuple[Union[torch.Tensor, IntermediateTensors], bool,
+               "ModelInputForGPUWithSamplingMetadata"]:
+        bypass_model_exec = True
+        input_tokens_tensor = model_input.input_tokens
+        seq_lens = model_input.attn_metadata.seq_lens
+        num_prefill_tokens = model_input.attn_metadata.num_prefill_tokens
+        slot_mapping = model_input.attn_metadata.slot_mapping.flatten()
+        start_layer = model_executable.model.start_layer
+        end_layer = model_executable.model.end_layer
+        hidden_or_intermediate_states_for_one_req = []
+
+        for idx, slen in enumerate(seq_lens):
+            start_pos = sum(seq_lens[:idx])
+            end_pos = start_pos + slen
+
+            if start_pos >= num_prefill_tokens:
+                # This can happen during inflight batching. See:
+                # vllm/worker/model_runner.py::_prepare_model_input_tensors:
+                # - input_tokens[:num_prefill_tokens] contains prefill tokens.
+                # - input_tokens[num_prefill_tokens:] contains decode tokens.
+                logger.warning("You should set --enable_chunked_prefill=False "
+                               "and --max_num_batched_tokens "
+                               "should be equal to max_seq_len_to_capture")
+                bypass_model_exec = False
+                assert start_pos == num_prefill_tokens
+                break
+
+            current_tokens = input_tokens_tensor[start_pos:end_pos]
+
+            # get roi for current seq
+            load_key_prefix = self.tensor_hash(current_tokens)
+            load_kvcache_key = f"{load_key_prefix}_{self.local_tp_rank}"
+            remote_kv = self.kv_store.get(load_kvcache_key)
+            hidden_key = f"{load_key_prefix}_hidden_{self.local_tp_rank}"
+            hidden = self.kv_store.get(hidden_key)
+
+            if remote_kv is None or hidden is None:
+                # didn't find any match.
+                bypass_model_exec = False
+                continue
+
+            num_computed_tokens = current_tokens.shape[0]
+
+            # update the end position based on how many tokens are cached.
+            end_pos = start_pos + num_computed_tokens
+
+            # call self.kv_store to get kv layer by layer
+            for layer_id in range(start_layer, end_layer):
+                layer = model_executable.model.layers[layer_id]
+                # get kvcache object
+                kv_cache = kv_caches[layer_id - start_layer]
+
+                # get remote kvcache
+                remote_k, remote_v = remote_kv[0][layer_id], remote_kv[1][
+                    layer_id]
+
+                self.kv_helper.put_kv_to_cache(model_executable, remote_k,
+                                               remote_v, layer, kv_cache,
+                                               slot_mapping, start_pos,
+                                               end_pos)
+
+            hidden_or_intermediate_states_for_one_req.append(hidden)
+
+        if not bypass_model_exec:
+            logger.warning(
+                "[rank%d]: Failed to receive all KVs and hidden "
+                "states, redo model forwarding.", torch.distributed.get_rank())
+            hidden_or_intermediate_states = None
+
+        else:
+            logger.debug(
+                "[rank%d]: Successfully received all KVs and hidden "
+                "states, skip model forwarding.", torch.distributed.get_rank())
+            hidden_or_intermediate_states = torch.cat(
+                hidden_or_intermediate_states_for_one_req, dim=0)
+
+        return hidden_or_intermediate_states, bypass_model_exec, model_input
+
+    @staticmethod
+    def tensor_hash(tensor: torch.Tensor) -> int:
+        """Calculate the hash value of the tensor."""
+        tensor_bytes = tensor.clone().detach().cpu().numpy().tobytes()
+        hash_object = hashlib.blake2b(tensor_bytes)
+        hash_hex = hash_object.hexdigest()
+        return int(hash_hex[:16], 16)