init src 0.9.2

vllm/distributed/kv_transfer/kv_connector/base.py

@@ -0,0 +1,128 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+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, 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 +1,133 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import importlib
+from typing import TYPE_CHECKING, Callable
+
+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=}")
+
+        kv_transfer_config = config.kv_transfer_config
+        connector_name = kv_transfer_config.kv_connector
+        if connector_name in cls._registry:
+            connector_cls = cls._registry[connector_name]()
+        else:
+            connector_module_path = kv_transfer_config.kv_connector_module_path
+            if connector_module_path is None:
+                raise ValueError(
+                    f"Unsupported connector type: {connector_name}")
+            connector_module = importlib.import_module(connector_module_path)
+            connector_cls = getattr(connector_module, connector_name)
+        assert issubclass(connector_cls, KVConnectorBase_V1)
+        logger.info("Creating v1 connector with name: %s and engine_id: %s",
+                    connector_name, kv_transfer_config.engine_id)
+        # 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(
+    "P2pNcclConnector",
+    "vllm.distributed.kv_transfer.kv_connector.v1.p2p.p2p_nccl_connector",
+    "P2pNcclConnector")
+
+KVConnectorFactory.register_connector(
+    "LMCacheConnectorV1",
+    "vllm.distributed.kv_transfer.kv_connector.v1.lmcache_connector",
+    "LMCacheConnectorV1")
+
+KVConnectorFactory.register_connector(
+    "NixlConnector",
+    "vllm.distributed.kv_transfer.kv_connector.v1.nixl_connector",
+    "NixlConnector")
+
+KVConnectorFactory.register_connector(
+    "MultiConnector",
+    "vllm.distributed.kv_transfer.kv_connector.v1.multi_connector",
+    "MultiConnector")

vllm/distributed/kv_transfer/kv_connector/lmcache_connector.py

@@ -0,0 +1,99 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+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, 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 +1,203 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+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, 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.kv_transfer_config = config.kv_transfer_config
+        self.kv_helper = kv_helper(config)
+        self.local_tp_rank = local_rank
+
+        # Init kv_store
+        if self.kv_transfer_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.kv_transfer_config)
+                self.kv_store = MooncakeStore(config)
+        else:
+            logger.error("Can not find %s",
+                         self.kv_transfer_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)

vllm/distributed/kv_transfer/kv_connector/simple_connector.py

@@ -0,0 +1,329 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Simple KV Cache Connector for Distributed Machine Learning Inference
+
+The SimpleConnector transfers KV caches between prefill vLLM worker (KV cache
+producer) and decode vLLM worker (KV cache consumer) using PyNcclPipe or
+MooncakePipe.
+
+But the logic can be extended to support other pipe and lookup buffer.
+"""
+from typing import TYPE_CHECKING, Optional, 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.distributed.kv_transfer.kv_lookup_buffer.simple_buffer import (
+    SimpleBuffer)
+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 SimpleConnector(KVConnectorBase):
+
+    def __init__(
+        self,
+        rank: int,
+        local_rank: int,
+        config: VllmConfig,
+    ):
+
+        self.config = config.kv_transfer_config
+        self.kv_helper = kv_helper(config)
+
+        if self.config.kv_connector == "PyNcclConnector":
+            from vllm.distributed.kv_transfer.kv_pipe.pynccl_pipe import (
+                PyNcclPipe)
+            logger.info(
+                "Initializing PyNcclConfig under kv_transfer_config %s",
+                self.config)
+        elif self.config.kv_connector == "MooncakeConnector":
+            # Check if MOONCAKE_CONFIG_PATH is set
+            import os
+            use_mooncake_distributed_pipe = os.getenv(
+                'MOONCAKE_CONFIG_PATH') is not None
+
+            if not use_mooncake_distributed_pipe:
+                raise ValueError(
+                    "To use MooncakeConnector, you need to pass the ENV: "
+                    "'MOONCAKE_CONFIG_PATH=/path/to/mooncake_config.json'.")
+            else:
+                from vllm.distributed.kv_transfer.kv_pipe.mooncake_pipe import (  # noqa: E501
+                    MooncakePipe)
+                logger.info(
+                    "Initializing MooncakeConfig under kv_transfer_config %s",
+                    self.config)
+
+        self.lookup_buffer_size = self.config.kv_buffer_size
+
+        self.producer_buffer: Optional[SimpleBuffer] = None
+        self.consumer_buffer: Optional[SimpleBuffer] = None
+
+        self.producer_data_pipe: Union[PyNcclPipe, MooncakePipe]
+        self.consumer_data_pipe: Union[PyNcclPipe, MooncakePipe]
+        self.producer_signal_pipe: Union[PyNcclPipe, MooncakePipe]
+        self.consumer_signal_pipe: Union[PyNcclPipe, MooncakePipe]
+
+        # 2 pipes for every rank in the world
+        port_offset_base = 2 * rank
+
+        # In disaggregated prefill, the prefill vLLM only uses send pipe
+        # and the decode vLLM only uses recv pipe
+        if self.config.is_kv_producer:
+
+            if self.config.kv_connector == "PyNcclConnector":
+                self.producer_data_pipe = PyNcclPipe(
+                    local_rank=local_rank,
+                    config=self.config,
+                    port_offset=port_offset_base,
+                )
+                self.producer_signal_pipe = PyNcclPipe(
+                    local_rank=local_rank,
+                    config=self.config,
+                    port_offset=port_offset_base + 1,
+                    device="cpu",
+                )
+            elif self.config.kv_connector == "MooncakeConnector":
+                self.producer_data_pipe = MooncakePipe(
+                    local_rank=local_rank,
+                    config=self.config,
+                )
+                # We only need to initialize MooncakePipe once
+                self.producer_signal_pipe = self.producer_data_pipe
+
+            self.producer_buffer = SimpleBuffer(self.producer_signal_pipe,
+                                                self.producer_data_pipe,
+                                                self.config.kv_buffer_size)
+
+        else:
+
+            # the current vLLM instance is KV consumer, so it needs to connect
+            # its recv pipe to the send pipe of KV producer
+            if self.config.kv_connector == "PyNcclConnector":
+                self.consumer_data_pipe = PyNcclPipe(
+                    local_rank=local_rank,
+                    config=self.config,
+                    port_offset=port_offset_base,
+                )
+                self.consumer_signal_pipe = PyNcclPipe(
+                    local_rank=local_rank,
+                    config=self.config,
+                    port_offset=port_offset_base + 1,
+                    device="cpu",
+                )
+            elif self.config.kv_connector == "MooncakeConnector":
+                self.consumer_data_pipe = MooncakePipe(
+                    local_rank=local_rank,
+                    config=self.config,
+                )
+                self.consumer_signal_pipe = self.consumer_data_pipe
+
+            self.consumer_buffer = SimpleBuffer(
+                self.consumer_signal_pipe,
+                self.consumer_data_pipe,
+                self.config.kv_buffer_size,
+            )
+
+    def select(self, input_tokens: Optional[torch.Tensor],
+               roi: Optional[torch.Tensor]) -> list[Optional[torch.Tensor]]:
+
+        assert self.consumer_buffer is not None, "Please initialize the "\
+            "consumer buffer before calling select."
+        return self.consumer_buffer.drop_select(input_tokens, roi)
+
+    def insert(self, input_tokens: torch.Tensor, roi: torch.Tensor,
+               key: torch.Tensor, value: torch.Tensor,
+               hidden: torch.Tensor) -> None:
+
+        assert self.producer_buffer is not None, "Please initialize the "\
+            "producer buffer before calling insert."
+
+        self.producer_buffer.insert(input_tokens, roi, key, value, hidden)
+
+    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()
+        num_prefill_tokens = model_input.attn_metadata.num_prefill_tokens
+        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)
+
+        # query_lens contains new KV caches that are added to vLLM.
+        # so we will send them to decode instance
+        # FIXME(Kuntai): This assume that all requests are prefill.
+        for idx, slen in enumerate(seq_lens):
+            start_pos = sum(seq_lens[:idx])
+            end_pos = start_pos + slen
+
+            if start_pos >= num_prefill_tokens:
+                # 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 have some decode requests while using "
+                               "SimpleConnector. Their KVCache won't be sent.")
+                break
+
+            current_tokens = input_tokens_tensor[start_pos:end_pos]
+
+            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)
+
+            self.insert(current_tokens,
+                        torch.ones_like(current_tokens,
+                                        dtype=bool), keys, values,
+                        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"]:
+
+        # When bypass_model_exec is set to False, it means that at least for one
+        # request its corresponding KV cache or hidden state is missing.
+        # In this case we need to do prefilling to recompute missing KV cache
+        # and hidden states.
+        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 = []
+
+        input_tokens_list = []
+        num_computed_tokens_list = []
+        start_pos_list = []
+
+        # enumerate different requests
+        # FIXME(Kuntai): This impl assumes that all requests are prefill.
+        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]
+            num_tokens = slen
+
+            # collecting data for rebuilding the input
+            input_tokens_list.append(current_tokens)
+            start_pos_list.append(start_pos)
+
+            ret = self.select(current_tokens,
+                              torch.ones_like(current_tokens, dtype=bool))
+            if ret[0] is None:
+                # didn't find any match.
+                bypass_model_exec = False
+                num_computed_tokens_list.append(0)
+                continue
+
+            roi: torch.Tensor = ret[1]
+            keys: torch.Tensor = ret[2]
+            values: torch.Tensor = ret[3]
+            hidden: torch.Tensor = ret[4]
+
+            num_computed_tokens = roi.shape[0]
+            num_computed_tokens_list.append(num_computed_tokens)
+
+            # check if both KV cache and the hidden states are received
+            # If not, need to redo the forwarding to compute missing states
+            if not all([(num_computed_tokens == num_tokens), hidden is not None
+                        ]):
+                bypass_model_exec = False
+
+            # update the end position based on how many tokens are cached.
+            end_pos = start_pos + num_computed_tokens
+
+            # put received KV caches into paged memory
+            for cur_layer in range(start_layer, end_layer):
+
+                layer_id = cur_layer - start_layer
+                kv_cache = kv_caches[layer_id]
+                layer = model_executable.model.layers[cur_layer]
+
+                # get remote kvcache
+                remote_k, remote_v = keys[layer_id], values[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:
+            # Some of the KV cache is not retrieved
+            # Here we will fall back to normal model forwarding
+            # But optionally you can adjust model_input so that you only do
+            # prefilling on those tokens that are missing KV caches.
+            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
+
+    def close(self):
+        self.producer_data_pipe.close()
+        self.consumer_data_pipe.close()
+        if self.config.kv_connector == "PyNcclConnector":
+            self.producer_signal_pipe.close()
+            self.consumer_signal_pipe.close()
+        elif self.config.kv_connector == "MooncakeConnector":
+            # MooncakePipe reuses data_pipe for signal_pipe, so we only have to
+            # close the data_pipe.
+            pass

vllm/distributed/kv_transfer/kv_connector/utils.py

@@ -0,0 +1,109 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+KV cache helper for store.
+"""
+import torch
+
+import vllm.envs as envs
+from vllm import _custom_ops as ops
+from vllm.config import VllmConfig, get_current_vllm_config
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class model_aware_kv_ops_helper:
+
+    def __init__(self, config: VllmConfig):
+        self.is_deepseek_mla = config.model_config.is_deepseek_mla
+        self.use_mla_opt = not envs.VLLM_MLA_DISABLE
+        self.tp_size = config.parallel_config.tensor_parallel_size
+
+    def get_model_args(self, model_executable: torch.nn.Module):
+
+        model_config = model_executable.model.config
+        self.model_executable = model_executable
+        num_heads = int(model_config.num_key_value_heads / self.tp_size)
+        hidden_size = model_config.hidden_size
+        num_attention_heads = model_config.num_attention_heads
+
+        # Deepseek's MLA (Multi-head Latent Attention) uses two different
+        # kv_cache shapes based on whether VLLM_MLA_DISABLE is set to 0.
+        # When VLLM_MLA_DISABLE=0 (default), forward absorb is applied,
+        # resulting in a kv_cache shape of [num_blks, blk_size, 1,
+        # kv_lora_rank + qk_rope_head_dim].
+        # When VLLM_MLA_DISABLE=1, standard FA is used instead, leading
+        # to a kv_cache shape of [2, num_blks, blk_size,
+        # num_key_value_heads / tp, qk_nope_head_dim + qk_rope_head_dim].
+        # For more details, see vllm/attention/backends/mla/common.py.
+        if self.is_deepseek_mla and self.use_mla_opt:
+            head_size = model_config.kv_lora_rank + \
+                model_config.qk_rope_head_dim
+            num_heads = 1
+        elif self.is_deepseek_mla and not self.use_mla_opt:
+            head_size = model_config.qk_nope_head_dim + \
+                model_config.qk_rope_head_dim
+        else:
+            head_size = getattr(model_config, "head_dim", None)
+            if head_size is None:
+                head_size = int(hidden_size // num_attention_heads)
+
+        return num_heads, head_size
+
+    def get_kv_from_cache(self, kv_cache, num_heads, head_size):
+        if self.is_deepseek_mla and self.use_mla_opt:
+            key_cache = kv_cache.reshape(-1, num_heads, head_size)
+            value_cache = kv_cache.reshape(-1, num_heads, head_size)
+        else:
+            key_cache = kv_cache[0].reshape(-1, num_heads, head_size)
+            value_cache = kv_cache[1].reshape(-1, num_heads, head_size)
+        return key_cache, value_cache
+
+    def put_kv_to_cache(self, model_executable: torch.nn.Module, keys, values,
+                        layer, kv_cache, slot_mapping, start_pos, end_pos):
+
+        model_config = model_executable.model.config
+
+        if self.is_deepseek_mla and self.use_mla_opt:
+            layer.self_attn.attn = layer.self_attn.mla_attn
+            k_c_normed_k_pe = keys.squeeze(1)
+            k_c_normed = k_c_normed_k_pe[:, :model_config.kv_lora_rank]
+            k_pe = k_c_normed_k_pe[:, model_config.kv_lora_rank:]
+            ops.concat_and_cache_mla(
+                k_c_normed.to(kv_cache.device),
+                k_pe.to(kv_cache.device),
+                kv_cache,
+                slot_mapping[start_pos:end_pos],
+                layer.self_attn.attn.kv_cache_dtype,
+                layer.self_attn.attn._k_scale,
+            )
+        else:
+            key_cache, value_cache = kv_cache[0], kv_cache[1]
+            ops.reshape_and_cache_flash(
+                keys.to(key_cache.device),
+                values.to(value_cache.device),
+                key_cache,
+                value_cache,
+                slot_mapping[start_pos:end_pos],
+                layer.self_attn.attn.kv_cache_dtype,
+                layer.self_attn.attn._k_scale,
+                layer.self_attn.attn._v_scale,
+            )
+
+
+def get_kv_connector_cache_layout():
+    # NOTE (NickLucche) When running disaggregated PD with NIXL, HND layout is
+    # used for faster transfer.
+    vllm_config = get_current_vllm_config()
+    kv_config = vllm_config.kv_transfer_config
+    if kv_config is not None and vllm_config.model_config is None:
+        logger.warning_once("Unable to detect current VLLM config. " \
+        "Defaulting to NHD kv cache layout.")
+    elif kv_config is not None:
+        use_mla = vllm_config.model_config.use_mla
+        if not use_mla and kv_config.kv_connector == "NixlConnector":
+            logger.info_once("NixlConnector detected. Setting KV cache " \
+            "layout to HND for better xfer performance.")
+            return "HND"
+    return "NHD"

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

@@ -0,0 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from vllm.distributed.kv_transfer.kv_connector.v1.base import (
+    KVConnectorBase_V1, KVConnectorRole)
+
+__all__ = ["KVConnectorRole", "KVConnectorBase_V1"]

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

@@ -0,0 +1,283 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+KVConnectorBase_V1 Class for Distributed KV Cache & Hidden State
+communication in vLLM v1
+
+The class provides the following primitives:
+    Scheduler-side: runs in the scheduler, binds metadata, which
+    is used by the worker-side to load/save KV cache.
+        get_num_new_matched_tokens() - get number of new tokens 
+            that exist in the remote KV cache. Might be called multiple
+            times for a given request and should be side-effect free.
+        update_state_after_alloc() - update KVConnector state after
+            temporary buffer alloc by the CacheManager.
+        request_finished() - called when a request is finished, with
+            the computed kv cache blocks for the request.
+            Returns whether KV cache should be freed now or will be
+            freed asynchronously and optionally returns KV transfer
+            params.
+
+    Worker-side: runs in each worker, loads/saves KV cache to/from
+    the Connector based on the metadata.
+        start_load_kv() - starts loading all KVs (maybe async)
+        wait_for_layer_load() - blocks until layer i load is done
+
+        save_kv_layer() - starts saving KV for layer i (maybe async)
+        wait_for_save() - blocks until all saves are done
+
+        get_finished() - called with ids of finished requests, returns
+            ids of requests that have completed async sending/recving.
+"""
+
+import enum
+from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING, Any, Optional
+
+import torch
+
+from vllm.logger import init_logger
+from vllm.v1.core.sched.output import SchedulerOutput
+
+if TYPE_CHECKING:
+    from vllm.attention.backends.abstract import AttentionMetadata
+    from vllm.config import VllmConfig
+    from vllm.forward_context import ForwardContext
+    from vllm.v1.core.kv_cache_manager import KVCacheBlocks
+    from vllm.v1.request import Request
+
+logger = init_logger(__name__)
+
+
+class KVConnectorRole(enum.Enum):
+    # Connector running in the scheduler process
+    SCHEDULER = 0
+
+    # Connector running in the worker process
+    WORKER = 1
+
+
+class KVConnectorMetadata:
+    """
+    Abstract Metadata used to communicate between the
+    Scheduler KVConnector and Worker KVConnector.
+    """
+    pass
+
+
+class KVConnectorBase_V1(ABC):
+
+    def __init__(self, vllm_config: "VllmConfig", role: KVConnectorRole):
+        logger.warning(
+            "Initializing KVConnectorBase_V1. This API is experimental and "
+            "subject to change in the future as we iterate the design.")
+        self._connector_metadata = KVConnectorMetadata()
+        self._vllm_config = vllm_config
+        self._role = role
+
+    @property
+    def role(self) -> KVConnectorRole:
+        return self._role
+
+    # ==============================
+    # Worker-side methods
+    # ==============================
+
+    def bind_connector_metadata(
+            self, connector_metadata: KVConnectorMetadata) -> None:
+        """Set the connector metadata from the scheduler.
+
+        This function should be called by the model runner every time 
+        before the model execution. The metadata will be used for runtime
+        KV cache loading and saving.
+
+        Args:
+            connector_metadata (dict): the connector metadata.
+        """
+        self._connector_metadata = connector_metadata
+
+    def clear_connector_metadata(self) -> None:
+        """Clear the connector metadata.
+
+        This function should be called by the model runner every time 
+        after the model execution.
+        """
+        self._connector_metadata = KVConnectorMetadata()
+
+    def _get_connector_metadata(self) -> KVConnectorMetadata:
+        """Get the connector metadata.
+
+        This function should only be called inside the connector.
+
+        Returns:
+            ConnectorMetadata: the connector metadata.
+        """
+        return self._connector_metadata
+
+    def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
+        """
+        Initialize with the KV caches. Useful for pre-registering the
+        KV Caches in the KVConnector (e.g. for NIXL).
+
+        Args: kv_caches:
+            dictionary of layer names, kv cache
+        """
+        return
+
+    @abstractmethod
+    def start_load_kv(self, forward_context: "ForwardContext",
+                      **kwargs) -> None:
+        """
+        Start loading the KV cache from the connector to vLLM's paged
+        KV buffer. This is called from the forward context before the
+        forward pass to enable async loading during model execution.
+
+        Args:
+            forward_context (ForwardContext): the forward context.
+            **kwargs: additional arguments for the load operation
+
+        Note:
+            The number of elements in kv_caches and layer_names should be 
+            the same.
+            
+        """
+        pass
+
+    @abstractmethod
+    def wait_for_layer_load(self, layer_name: str) -> None:
+        """
+        Block until the KV for a specific layer is loaded into vLLM's
+        paged buffer. This is called from within attention layer to ensure
+        async copying from start_load_kv is complete.
+        
+        This interface will be useful for layer-by-layer pipelining.
+
+        Args:
+            layer_name: the name of that layer
+        """
+        pass
+
+    @abstractmethod
+    def save_kv_layer(self, layer_name: str, kv_layer: torch.Tensor,
+                      attn_metadata: "AttentionMetadata", **kwargs) -> None:
+        """
+        Start saving a layer of KV cache from vLLM's paged buffer 
+        to the connector. This is called from within attention layer to
+        enable async copying during execution.
+
+        Args:
+            layer_name (str): the name of the layer.
+            kv_layer (torch.Tensor): the paged KV buffer of the current 
+                layer in vLLM.
+            attn_metadata (AttentionMetadata): the attention metadata.
+            **kwargs: additional arguments for the save operation.
+        """
+        pass
+
+    @abstractmethod
+    def wait_for_save(self):
+        """
+        Block until all the save operations is done. This is called
+        as the forward context exits to ensure that the async saving
+        from save_kv_layer is complete before finishing the forward.
+
+        This prevents overwrites of paged KV buffer before saving done.
+        """
+        pass
+
+    def get_finished(
+        self, finished_req_ids: set[str]
+    ) -> tuple[Optional[set[str]], Optional[set[str]]]:
+        """
+        Notifies worker-side connector ids of requests that have
+        finished generating tokens.
+
+        Returns:
+            ids of requests that have finished asynchronous transfer
+            (requests that previously returned True from request_finished()),
+            tuple of (sending/saving ids, recving/loading ids).
+            The finished saves/sends req ids must belong to a set provided in a
+            call to this method (this call or a prior one).
+        """
+        return None, None
+
+    # ==============================
+    # Scheduler-side methods
+    # ==============================
+
+    @abstractmethod
+    def get_num_new_matched_tokens(
+        self,
+        request: "Request",
+        num_computed_tokens: int,
+    ) -> tuple[int, bool]:
+        """
+        Get number of new tokens that can be loaded from the
+        external KV cache beyond the num_computed_tokens.
+        
+        Args:
+            request (Request): the request object.
+            num_computed_tokens (int): the number of locally
+                computed tokens for this request
+
+        Returns:
+            A tuple with the following elements:
+                - The number of tokens that can be loaded from the 
+                  external KV cache beyond what is already computed.
+                - `True` if external KV cache tokens will be loaded
+                  asynchronously (between scheduler steps). Must be
+                  'False' if the first element is 0.
+        """
+        pass
+
+    @abstractmethod
+    def update_state_after_alloc(self, request: "Request",
+                                 blocks: "KVCacheBlocks",
+                                 num_external_tokens: int):
+        """
+        Update KVConnector state after block allocation.
+
+        If get_num_new_matched_tokens previously returned True for a
+        request, this function may be called twice for that same request -
+        first when blocks are allocated for the connector tokens to be
+        asynchronously loaded into, and second when any additional blocks
+        are allocated, after the load/transfer is complete.
+
+        Args:
+            request (Request): the request object.
+            blocks (KVCacheBlocks): the blocks allocated for the request.
+            num_external_tokens (int): the number of tokens that will be
+                loaded from the external KV cache.
+        """
+        pass
+
+    @abstractmethod
+    def build_connector_meta(
+            self, scheduler_output: SchedulerOutput) -> KVConnectorMetadata:
+        """
+        Build the connector metadata for this step.
+
+        This function should NOT modify fields in the scheduler_output.
+        Also, calling this function will reset the state of the connector.
+
+        Args:
+            scheduler_output (SchedulerOutput): the scheduler output object.
+        """
+        pass
+
+    def request_finished(
+        self,
+        request: "Request",
+        block_ids: list[int],
+    ) -> tuple[bool, Optional[dict[str, Any]]]:
+        """
+        Called when a request has finished, before its blocks are freed.
+
+        Returns:
+            True if the request is being saved/sent asynchronously and blocks
+            should not be freed until the request_id is returned from
+            get_finished().
+            Optional KVTransferParams to be included in the request outputs
+            returned by the engine.
+        """
+        return False, None

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

@@ -0,0 +1,167 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import TYPE_CHECKING, Any, Optional
+
+import torch
+from lmcache.integration.vllm.vllm_v1_adapter import LMCacheConnectorV1Impl
+
+from vllm.config import VllmConfig
+from vllm.distributed.kv_transfer.kv_connector.v1.base import (
+    KVConnectorBase_V1, KVConnectorMetadata, KVConnectorRole)
+from vllm.logger import init_logger
+from vllm.v1.core.sched.output import SchedulerOutput
+
+if TYPE_CHECKING:
+    from vllm.attention.backends.abstract import AttentionMetadata
+    from vllm.forward_context import ForwardContext
+    from vllm.v1.core.kv_cache_manager import KVCacheBlocks
+    from vllm.v1.request import Request
+
+logger = init_logger(__name__)
+
+
+class LMCacheConnectorV1(KVConnectorBase_V1):
+
+    def __init__(self, vllm_config: "VllmConfig", role: KVConnectorRole):
+        super().__init__(vllm_config=vllm_config, role=role)
+        self._lmcache_engine = LMCacheConnectorV1Impl(vllm_config, role, self)
+
+    # ==============================
+    # Worker-side methods
+    # ==============================
+    def start_load_kv(self, forward_context: "ForwardContext",
+                      **kwargs) -> None:
+        """
+        Start loading the KV cache from the connector to vLLM's paged
+        KV buffer. This is called from the forward context before the
+        forward pass to enable async loading during model execution.
+
+        Args:
+            forward_context (ForwardContext): the forward context.
+            **kwargs: additional arguments for the load operation
+
+        Note:
+            The number of elements in kv_caches and layer_names should be 
+            the same.
+            
+        """
+        self._lmcache_engine.start_load_kv(forward_context, **kwargs)
+
+    def wait_for_layer_load(self, layer_name: str) -> None:
+        """
+        Block until the KV for a specific layer is loaded into vLLM's
+        paged buffer. This is called from within attention layer to ensure
+        async copying from start_load_kv is complete.
+        
+        This interface will be useful for layer-by-layer pipelining.
+
+        Args:
+            layer_name: the name of that layer
+        """
+        self._lmcache_engine.wait_for_layer_load(layer_name)
+
+    def save_kv_layer(self, layer_name: str, kv_layer: torch.Tensor,
+                      attn_metadata: "AttentionMetadata", **kwargs) -> None:
+        """
+        Start saving the a layer of KV cache from vLLM's paged buffer 
+        to the connector. This is called from within attention layer to
+        enable async copying during execution.
+
+        Args:
+            layer_name (str): the name of the layer.
+            kv_layer (torch.Tensor): the paged KV buffer of the current 
+                layer in vLLM.
+            attn_metadata (AttentionMetadata): the attention metadata.
+            **kwargs: additional arguments for the save operation.
+        """
+        self._lmcache_engine.save_kv_layer(layer_name, kv_layer, attn_metadata,
+                                           **kwargs)
+
+    def wait_for_save(self):
+        """
+        Block until all the save operations is done. This is called
+        as the forward context exits to ensure that the async saving
+        from save_kv_layer is complete before finishing the forward.
+
+        This prevents overwrites of paged KV buffer before saving done.
+        """
+        self._lmcache_engine.wait_for_save()
+
+    def get_finished(
+        self, finished_req_ids: set[str]
+    ) -> tuple[Optional[set[str]], Optional[set[str]]]:
+        """
+        Notifies worker-side connector ids of requests that have
+        finished generating tokens.
+
+        Returns:
+            ids of requests that have finished asynchronous transfer
+            (requests that previously returned True from request_finished()),
+            tuple of (sending/saving ids, recving/loading ids).
+            The finished saves/sends req ids must belong to a set provided in a
+            call to this method (this call or a prior one).
+        """
+        return self._lmcache_engine.get_finished(finished_req_ids)
+
+    # ==============================
+    # Scheduler-side methods
+    # ==============================
+    def get_num_new_matched_tokens(
+        self,
+        request: "Request",
+        num_computed_tokens: int,
+    ) -> tuple[int, bool]:
+        """
+        Get number of new tokens that can be loaded from the
+        external KV cache beyond the num_computed_tokens.
+        
+        Args:
+            request (Request): the request object.
+            num_computed_tokens (int): the number of locally
+                computed tokens for this request
+
+        Returns:
+            the number of tokens that can be loaded from the 
+            external KV cache beyond what is already computed.
+        """
+        return self._lmcache_engine.get_num_new_matched_tokens(
+            request, num_computed_tokens), False
+
+    def update_state_after_alloc(self, request: "Request",
+                                 blocks: "KVCacheBlocks",
+                                 num_external_tokens: int):
+        """
+        Update KVConnector state after block allocation.
+        """
+        self._lmcache_engine.update_state_after_alloc(request,
+                                                      num_external_tokens)
+
+    def build_connector_meta(
+            self, scheduler_output: SchedulerOutput) -> KVConnectorMetadata:
+        """
+        Build the connector metadata for this step.
+
+        This function should NOT modify fields in the scheduler_output.
+        Also, calling this function will reset the state of the connector.
+
+        Args:
+            scheduler_output (SchedulerOutput): the scheduler output object.
+        """
+        return self._lmcache_engine.build_connector_meta(scheduler_output)
+
+    def request_finished(
+        self,
+        request: "Request",
+        block_ids: list[int],
+    ) -> tuple[bool, Optional[dict[str, Any]]]:
+        """
+        Called when a request has finished, before its blocks are freed.
+
+        Returns:
+            True if the request is being saved/sent asynchronously and blocks
+            should not be freed until the request_id is returned from
+            get_finished().
+            Optional KVTransferParams to be included in the request outputs
+            returned by the engine.
+        """
+        return self._lmcache_engine.request_finished(request, block_ids)

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

@@ -0,0 +1,201 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import copy
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Optional
+
+import torch
+
+from vllm.config import KVTransferConfig, VllmConfig
+from vllm.distributed.kv_transfer.kv_connector.factory import (
+    KVConnectorFactory)
+from vllm.distributed.kv_transfer.kv_connector.v1.base import (
+    KVConnectorBase_V1, KVConnectorMetadata, KVConnectorRole)
+from vllm.logger import init_logger
+from vllm.v1.core.kv_cache_manager import KVCacheBlocks
+from vllm.v1.core.sched.output import SchedulerOutput
+
+if TYPE_CHECKING:
+    from vllm.attention.backends.abstract import AttentionMetadata
+    from vllm.forward_context import ForwardContext
+    from vllm.v1.request import Request
+
+logger = init_logger(__name__)
+
+
+@dataclass
+class MultiKVConnectorMetadata(KVConnectorMetadata):
+    metadata: tuple[KVConnectorMetadata, ...]
+    extra_async_saves: Optional[dict[str, int]] = None
+
+
+class MultiConnector(KVConnectorBase_V1):
+    """
+    A wrapper for using multiple KVConnectors at the same time.
+
+    The current logic is:
+    - Load KV from the first connector that advertises available tokens from
+      get_num_new_matched_tokens(), based on the order in the config.
+    - Save to all connectors.
+    """
+
+    def __init__(self, vllm_config: "VllmConfig", role: KVConnectorRole):
+        super().__init__(vllm_config=vllm_config, role=role)
+        self._connectors: list[KVConnectorBase_V1] = []
+        ktcs = vllm_config.kv_transfer_config.kv_connector_extra_config.get(
+            "connectors")
+        assert ktcs is not None
+        for ktc in ktcs:
+            temp_config = copy.copy(vllm_config)
+            temp_config.kv_transfer_config = KVTransferConfig(**ktc)
+            self._connectors.append(
+                KVConnectorFactory.create_connector_v1(temp_config, role))
+
+        # A mapping from request id to the index of the connector chosen to
+        # load the request from (if any).
+        self._requests_to_connector: dict[str, int] = {}
+
+        # Keeps track of *additional* remaining async saves (beyond 1) to be
+        # finished per request. Not needed for async loads since we only allow
+        # a single connector to load.
+        # Propagated from scheduler to worker side via the connector metadata.
+        self._extra_async_saves: dict[str, int] = {}
+
+    def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
+        for c in self._connectors:
+            c.register_kv_caches(kv_caches)
+
+    # We must override the base class method here because we need to bind
+    # the metadata to each connector in the order of the connectors in the
+    # MultiKVConnectorMetadata.
+    def bind_connector_metadata(
+            self, connector_metadata: KVConnectorMetadata) -> None:
+        assert isinstance(connector_metadata, MultiKVConnectorMetadata)
+        if connector_metadata.extra_async_saves:
+            self._extra_async_saves.update(
+                connector_metadata.extra_async_saves)
+        for c, cm in zip(self._connectors, connector_metadata.metadata):
+            c.bind_connector_metadata(cm)
+
+    def clear_connector_metadata(self) -> None:
+        for c in self._connectors:
+            c.clear_connector_metadata()
+
+    # ==============================
+    # Worker-side methods
+    # ==============================
+    def start_load_kv(self, forward_context: "ForwardContext",
+                      **kwargs) -> None:
+        for c in self._connectors:
+            c.start_load_kv(forward_context, **kwargs)
+
+    def wait_for_layer_load(self, layer_name: str) -> None:
+        for c in self._connectors:
+            c.wait_for_layer_load(layer_name)
+
+    def save_kv_layer(self, layer_name: str, kv_layer: torch.Tensor,
+                      attn_metadata: "AttentionMetadata", **kwargs) -> None:
+        for c in self._connectors:
+            c.save_kv_layer(layer_name, kv_layer, attn_metadata, **kwargs)
+
+    def wait_for_save(self):
+        for c in self._connectors:
+            c.wait_for_save()
+
+    def get_finished(
+        self, finished_req_ids: set[str]
+    ) -> tuple[Optional[set[str]], Optional[set[str]]]:
+        finished_sending: set[str] = set()
+        finished_recving: set[str] = set()
+        for c in self._connectors:
+            sending, recving = c.get_finished(finished_req_ids)
+            if not recving and not sending:
+                continue
+            # Aggregate finished recving request ids.
+            finished_recving.update(recving or ())
+            # Aggregate finished sending request ids - only include
+            # once we've drained the "extra" count (for cases where
+            # more than one connector is async-saving the same request).
+            for req_id in sending or ():
+                extra_pending = self._extra_async_saves.get(req_id)
+                if extra_pending is None:
+                    finished_sending.add(req_id)
+                    continue
+                assert extra_pending > 0
+                if extra_pending == 1:
+                    del self._extra_async_saves[req_id]
+                else:
+                    self._extra_async_saves[req_id] = extra_pending - 1
+
+        return finished_sending or None, finished_recving or None
+
+    # ==============================
+    # Scheduler-side methods
+    # ==============================
+    def get_num_new_matched_tokens(
+        self,
+        request: "Request",
+        num_computed_tokens: int,
+    ) -> tuple[int, bool]:
+        to_return = (0, False)
+        for i, c in enumerate(self._connectors):
+            toks, load_async = c.get_num_new_matched_tokens(
+                request, num_computed_tokens)
+            # The first connector that has new matched tokens will be assigned
+            # to this request.
+            if to_return[0] == 0 and toks > 0:
+                self._requests_to_connector[request.request_id] = i
+                to_return = (toks, load_async)
+        return to_return
+
+    def update_state_after_alloc(self, request: "Request",
+                                 blocks: "KVCacheBlocks",
+                                 num_external_tokens: int):
+        chosen_connector = self._requests_to_connector.get(
+            request.request_id, -1)
+        empty_blocks = blocks.new_empty()
+        for i, c in enumerate(self._connectors):
+            if i == chosen_connector:
+                # Forward call to the chosen connector (if any).
+                c.update_state_after_alloc(request, blocks,
+                                           num_external_tokens)
+            else:
+                # Call with empty blocks for other connectors.
+                c.update_state_after_alloc(request, empty_blocks, 0)
+
+    def build_connector_meta(
+            self,
+            scheduler_output: SchedulerOutput) -> MultiKVConnectorMetadata:
+        metadata = MultiKVConnectorMetadata(metadata=tuple(
+            c.build_connector_meta(scheduler_output)
+            for c in self._connectors))
+        if self._extra_async_saves:
+            metadata.extra_async_saves = self._extra_async_saves
+            self._extra_async_saves = {}
+        return metadata
+
+    def request_finished(
+        self,
+        request: "Request",
+        blocks: list[int],
+    ) -> tuple[bool, Optional[dict[str, Any]]]:
+        async_saves = 0
+        kv_txfer_params = None
+        for c in self._connectors:
+            async_save, txfer_params = c.request_finished(request, blocks)
+            if async_save:
+                async_saves += 1
+            if txfer_params is not None:
+                if kv_txfer_params is not None:
+                    #TODO we can probably change this to merge the dicts here,
+                    # checking for key clashes.
+                    raise RuntimeError(
+                        "Only one connector can produce KV transfer params")
+                kv_txfer_params = txfer_params
+        if async_saves > 1:
+            self._extra_async_saves[request.request_id] = async_saves - 1
+
+        # Clean up other state for this request.
+        self._requests_to_connector.pop(request.request_id, None)
+
+        return async_saves > 0, kv_txfer_params

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py

@@ -0,0 +1,505 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Optional
+
+import regex as re
+import torch
+
+from vllm.config import VllmConfig
+from vllm.distributed.kv_transfer.kv_connector.v1.base import (
+    KVConnectorBase_V1, KVConnectorMetadata, KVConnectorRole)
+from vllm.distributed.kv_transfer.kv_connector.v1.p2p.p2p_nccl_engine import (
+    P2pNcclEngine)
+from vllm.distributed.parallel_state import get_world_group
+from vllm.forward_context import get_forward_context
+from vllm.logger import init_logger
+from vllm.v1.attention.backends.mla.common import MLACommonMetadata
+from vllm.v1.core.sched.output import SchedulerOutput
+
+if TYPE_CHECKING:
+    from vllm.attention.backends.abstract import AttentionMetadata
+    from vllm.forward_context import ForwardContext
+    from vllm.v1.core.kv_cache_manager import KVCacheBlocks
+    from vllm.v1.request import Request
+
+logger = init_logger(__name__)
+
+
+@dataclass
+class ReqMeta:
+    # Request Id
+    request_id: str
+    # Request tokens
+    token_ids: torch.Tensor
+    # Slot mappings, should have the same length as token_ids
+    slot_mapping: torch.Tensor
+
+    @staticmethod
+    def make_meta(request_id: str, token_ids: list[int], block_ids: list[int],
+                  block_size: int) -> "ReqMeta":
+        valid_num_tokens = len(token_ids)
+        token_ids_tensor = torch.tensor(token_ids)
+        block_ids_tensor = torch.tensor(block_ids)
+        num_blocks = block_ids_tensor.shape[0]
+        block_offsets = torch.arange(0, block_size)
+        slot_mapping = block_offsets.reshape((1, block_size)) + \
+                block_ids_tensor.reshape((num_blocks, 1)) * block_size
+        slot_mapping = slot_mapping.flatten()[:valid_num_tokens]
+
+        return ReqMeta(
+            request_id=request_id,
+            token_ids=token_ids_tensor,
+            slot_mapping=slot_mapping,
+        )
+
+
+@dataclass
+class P2pNcclConnectorMetadata(KVConnectorMetadata):
+    requests: list[ReqMeta]
+
+    def __init__(self):
+        self.requests = []
+
+    def add_request(
+        self,
+        request_id: str,
+        token_ids: list[int],
+        block_ids: list[int],
+        block_size: int,
+    ) -> None:
+        self.requests.append(
+            ReqMeta.make_meta(request_id, token_ids, block_ids, block_size))
+
+
+class P2pNcclConnector(KVConnectorBase_V1):
+
+    def __init__(self, vllm_config: "VllmConfig", role: KVConnectorRole):
+        super().__init__(vllm_config=vllm_config, role=role)
+        self._block_size = vllm_config.cache_config.block_size
+        self._requests_need_load: dict[str, Any] = {}
+        self.config = vllm_config.kv_transfer_config
+        self.is_producer = self.config.is_kv_producer
+        self.chunked_prefill: dict[str, Any] = {}
+
+        self._rank = get_world_group().rank \
+            if role == KVConnectorRole.WORKER else 0
+        self._local_rank = get_world_group().local_rank \
+            if role == KVConnectorRole.WORKER else 0
+
+        self.p2p_nccl_engine = P2pNcclEngine(
+            local_rank=self._local_rank,
+            config=self.config,
+            hostname="",
+            port_offset=self._rank,
+        ) if role == KVConnectorRole.WORKER else None
+
+    # ==============================
+    # Worker-side methods
+    # ==============================
+
+    def start_load_kv(self, forward_context: "ForwardContext",
+                      **kwargs) -> None:
+        """Start loading the KV cache from the connector buffer to vLLM's
+        paged KV buffer.
+
+        Args:
+            forward_context (ForwardContext): the forward context.
+            **kwargs: additional arguments for the load operation
+
+        Note:
+            The number of elements in kv_caches and layer_names should be
+            the same.
+        """
+
+        # Only consumer/decode loads KV Cache
+        if self.is_producer:
+            return
+
+        assert self.p2p_nccl_engine is not None
+
+        attn_metadata = forward_context.attn_metadata
+        if attn_metadata is None:
+            return
+
+        def inject_kv_into_layer(
+            dst_kv_cache_layer: torch.Tensor,
+            src_kv_cache: torch.Tensor,
+            slot_mapping: torch.Tensor,
+            request_id: str,
+        ) -> None:
+            """Inject the KV cache into the layer.
+
+            Args:
+                dst_kv_cache_layer (torch.Tensor): the destination KV cache
+                    layer. In shape [2, num_pages, page_size, xxx] if not
+                    using MLA, [num_pages, page_size, xxx] otherwise.
+                src_kv_cache (torch.Tensor): the source KV cache. In shape
+                    [2, num_tokens, xxx] if not using MLA, [num_tokens, xxx]
+                    otherwise.
+                slot_mapping (torch.Tensor): the slot mapping. In shape
+                    [num_tokens].
+                request_id (str): request id for log
+            """
+            dst_kv_cache_layer_shape = dst_kv_cache_layer.shape
+            if isinstance(attn_metadata, MLACommonMetadata) or all(isinstance(value, MLACommonMetadata) for value in attn_metadata.values()):
+                num_pages = dst_kv_cache_layer_shape[0]
+                page_size = dst_kv_cache_layer_shape[1]
+                dst_kv_cache_layer = dst_kv_cache_layer.reshape(
+                    num_pages * page_size, -1)
+                self.check_tensors_except_dim(dst_kv_cache_layer, src_kv_cache,
+                                              0)
+                num_token = src_kv_cache.shape[0]
+                if len(slot_mapping) == num_token:
+                    dst_kv_cache_layer[slot_mapping, ...] = src_kv_cache
+                else:
+                    dst_kv_cache_layer[slot_mapping[:num_token],
+                                       ...] = src_kv_cache
+                    logger.warning(
+                        "🚧src_kv_cache does not match, num_slot:%d, "
+                        "num_token:%d, request_id:%s", len(slot_mapping),
+                        num_token, request_id)
+
+                dst_kv_cache_layer.reshape(dst_kv_cache_layer_shape)
+            else:
+                num_pages = dst_kv_cache_layer_shape[1]
+                page_size = dst_kv_cache_layer_shape[2]
+                dst_kv_cache_layer = dst_kv_cache_layer.reshape(
+                    2, num_pages * page_size, -1)
+                self.check_tensors_except_dim(dst_kv_cache_layer, src_kv_cache,
+                                              1)
+                num_token = src_kv_cache.shape[1]
+                if len(slot_mapping) == num_token:
+                    dst_kv_cache_layer[:, slot_mapping, ...] = src_kv_cache
+                else:
+                    dst_kv_cache_layer[:, slot_mapping[:num_token],
+                                       ...] = src_kv_cache
+                    logger.warning(
+                        "🚧src_kv_cache does not match, num_slot:%d, "
+                        "num_token:%d, request_id:%s", len(slot_mapping),
+                        num_token, request_id)
+
+                dst_kv_cache_layer.reshape(dst_kv_cache_layer_shape)
+
+        # Get the metadata
+        metadata: KVConnectorMetadata = \
+            self._get_connector_metadata()
+        assert isinstance(metadata, P2pNcclConnectorMetadata)
+
+        if metadata is None:
+            return
+
+        # Load the KV for each request each layer
+        for request in metadata.requests:
+            for layer_name in forward_context.no_compile_layers:
+                layer = forward_context.no_compile_layers[layer_name]
+
+                # Only process layers that have kv_cache
+                # attribute (attention layers) Skip non-attention
+                # layers like FusedMoE
+                kv_cache = getattr(layer, 'kv_cache', None)
+                if kv_cache is None:
+                    continue
+
+                kv_cache_layer = kv_cache[ \
+                    forward_context.virtual_engine]
+
+                kv_cache = self.p2p_nccl_engine.recv_tensor(
+                    request.request_id + "#" + layer_name)
+
+                if kv_cache is None:
+                    logger.warning("🚧src_kv_cache is None, %s",
+                                   request.request_id)
+                    continue
+
+                inject_kv_into_layer(kv_cache_layer, kv_cache,
+                                     request.slot_mapping, request.request_id)
+                tensor_id = request.request_id + "#" + layer_name
+                if tensor_id in self.p2p_nccl_engine.recv_store:
+                    tensor = self.p2p_nccl_engine.recv_store.pop(tensor_id, None)
+                    self.p2p_nccl_engine.send_request_id_to_tensor_ids.pop(
+                                request.request_id, None)
+                    self.p2p_nccl_engine.recv_request_id_to_tensor_ids.pop(
+                                request.request_id, None)
+                    addr = 0
+                    if isinstance(tensor, tuple):
+                        addr, _, _ = tensor
+                        self.p2p_nccl_engine.pool.free(addr)
+
+
+    def wait_for_layer_load(self, layer_name: str) -> None:
+        """Blocking until the KV for a specific layer is loaded into vLLM's
+        paged buffer.
+
+        This interface will be useful for layer-by-layer pipelining.
+
+        Args:
+            layer_name: the name of that layer
+        """
+        return
+
+    def save_kv_layer(self, layer_name: str, kv_layer: torch.Tensor,
+                      attn_metadata: "AttentionMetadata", **kwargs) -> None:
+        """Start saving the KV cache of the layer from vLLM's paged buffer
+        to the connector.
+
+        Args:
+            layer_name (str): the name of the layer.
+            kv_layer (torch.Tensor): the paged KV buffer of the current
+                layer in vLLM.
+            attn_metadata (AttentionMetadata): the attention metadata.
+            **kwargs: additional arguments for the save operation.
+        """
+
+        # Only producer/prefill saves KV Cache
+        if not self.is_producer:
+            return
+
+        assert self.p2p_nccl_engine is not None
+
+        def extract_kv_from_layer(
+            layer: torch.Tensor,
+            slot_mapping: torch.Tensor,
+        ) -> torch.Tensor:
+            """Extract the KV cache from the layer.
+
+            Assume the shape of the layer is (2, num_pages, page_size, xxx)
+            if MLA is not used, and (num_pages, page_size, xxx) otherwise.
+            """
+            if isinstance(attn_metadata, MLACommonMetadata):
+                num_pages, page_size = layer.shape[0], layer.shape[1]
+                return layer.reshape(num_pages * page_size, -1)[slot_mapping,
+                                                                ...]
+            num_pages, page_size = layer.shape[1], layer.shape[2]
+            return layer.reshape(2, num_pages * page_size, -1)[:, slot_mapping,
+                                                               ...]
+
+        connector_metadata = self._get_connector_metadata()
+        assert isinstance(connector_metadata, P2pNcclConnectorMetadata)
+        for request in connector_metadata.requests:
+            request_id = request.request_id
+            ip, port = self.parse_request_id(request_id, True)
+            remote_address = ip + ":" + str(port + self._rank)
+            kv_cache = extract_kv_from_layer(kv_layer, request.slot_mapping)
+            self.p2p_nccl_engine.send_tensor(request_id + "#" + layer_name,
+                                             kv_cache, remote_address)
+
+    def wait_for_save(self):
+        if self.is_producer:
+            assert self.p2p_nccl_engine is not None
+            self.p2p_nccl_engine.wait_for_sent()
+
+    def get_finished(
+            self, finished_req_ids: set[str],
+            **kwargs) -> tuple[Optional[set[str]], Optional[set[str]]]:
+        """
+        Notifies worker-side connector ids of requests that have
+        finished generating tokens.
+
+        Returns:
+            ids of requests that have finished asynchronous transfer,
+            tuple of (sending/saving ids, recving/loading ids).
+            The finished saves/sends req ids must belong to a set provided in a
+            call to this method (this call or a prior one).
+        """
+
+        assert self.p2p_nccl_engine is not None
+
+        forward_context: ForwardContext = get_forward_context()
+        return self.p2p_nccl_engine.get_finished(finished_req_ids,
+                                                 forward_context)
+
+    # ==============================
+    # Scheduler-side methods
+    # ==============================
+
+    def get_num_new_matched_tokens(
+        self,
+        request: "Request",
+        num_computed_tokens: int,
+    ) -> tuple[int, bool]:
+        """
+        Get number of new tokens that can be loaded from the
+        external KV cache beyond the num_computed_tokens.
+
+        Args:
+            request (Request): the request object.
+            num_computed_tokens (int): the number of locally
+                computed tokens for this request
+
+        Returns:
+            the number of tokens that can be loaded from the
+            external KV cache beyond what is already computed.
+        """
+        if self.is_producer:
+            return 0, False
+
+        num_external_tokens = (len(request.prompt_token_ids) - 1 -
+                               num_computed_tokens)
+
+        if num_external_tokens < 0:
+            num_external_tokens = 0
+
+        return num_external_tokens, False
+
+    def update_state_after_alloc(self, request: "Request",
+                                 blocks: "KVCacheBlocks",
+                                 num_external_tokens: int):
+        """
+        Update KVConnector state after block allocation.
+        """
+        if not self.is_producer and num_external_tokens > 0:
+            self._requests_need_load[request.request_id] = (
+                request, blocks.get_block_ids()[0])
+
+    def build_connector_meta(
+        self,
+        scheduler_output: SchedulerOutput,
+    ) -> KVConnectorMetadata:
+        """Build the connector metadata for this step.
+
+        This function should NOT modify any fields in the scheduler_output.
+        Also, calling this function will reset the state of the connector.
+
+        Args:
+            scheduler_output (SchedulerOutput): the scheduler output object.
+        """
+
+        meta = P2pNcclConnectorMetadata()
+
+        for new_req in scheduler_output.scheduled_new_reqs:
+            if self.is_producer:
+                num_scheduled_tokens = (
+                    scheduler_output.num_scheduled_tokens)[new_req.req_id]
+                num_tokens = num_scheduled_tokens + new_req.num_computed_tokens
+                # the request's prompt is chunked prefill
+                if num_tokens < len(new_req.prompt_token_ids):
+                    # 'CachedRequestData' has no attribute 'prompt_token_ids'
+                    self.chunked_prefill[new_req.req_id] = (
+                        new_req.block_ids[0], new_req.prompt_token_ids)
+                    continue
+                # the request's prompt is not chunked prefill
+                meta.add_request(request_id=new_req.req_id,
+                                 token_ids=new_req.prompt_token_ids,
+                                 block_ids=new_req.block_ids[0],
+                                 block_size=self._block_size)
+                continue
+            if new_req.req_id in self._requests_need_load:
+                meta.add_request(request_id=new_req.req_id,
+                                 token_ids=new_req.prompt_token_ids,
+                                 block_ids=new_req.block_ids[0],
+                                 block_size=self._block_size)
+                self._requests_need_load.pop(new_req.req_id)
+
+        cached_reqs = scheduler_output.scheduled_cached_reqs
+        for i, req_id in enumerate(cached_reqs.req_ids):
+            num_computed_tokens = cached_reqs.num_computed_tokens[i]
+            new_block_ids = cached_reqs.new_block_ids[i]
+            resumed_from_preemption = cached_reqs.resumed_from_preemption[i]
+
+            if self.is_producer:
+                num_scheduled_tokens = (
+                    scheduler_output.num_scheduled_tokens)[req_id]
+                num_tokens = (num_scheduled_tokens + num_computed_tokens)
+                assert req_id in self.chunked_prefill
+                block_ids = new_block_ids[0]
+                if not resumed_from_preemption:
+                    block_ids = (self.chunked_prefill[req_id][0] + block_ids)
+                prompt_token_ids = self.chunked_prefill[req_id][1]
+                # the request's prompt is chunked prefill again
+                if num_tokens < len(prompt_token_ids):
+                    self.chunked_prefill[req_id] = (block_ids,
+                                                    prompt_token_ids)
+                    continue
+                # the request's prompt is all prefilled finally
+                meta.add_request(request_id=req_id,
+                                 token_ids=prompt_token_ids,
+                                 block_ids=block_ids,
+                                 block_size=self._block_size)
+                self.chunked_prefill.pop(req_id, None)
+                continue
+
+            # NOTE(rob): here we rely on the resumed requests being
+            # the first N requests in the list scheduled_cache_reqs.
+            if not resumed_from_preemption:
+                break
+            if req_id in self._requests_need_load:
+                request, _ = self._requests_need_load.pop(req_id)
+                total_tokens = num_computed_tokens + 1
+                token_ids = request.all_token_ids[:total_tokens]
+
+                # NOTE(rob): For resumed req, new_block_ids is all
+                # of the block_ids for the request.
+                block_ids = new_block_ids[0]
+
+                meta.add_request(request_id=req_id,
+                                 token_ids=token_ids,
+                                 block_ids=block_ids,
+                                 block_size=self._block_size)
+
+        # Requests loaded asynchronously are not in the scheduler_output.
+        # for request_id in self._requests_need_load:
+        #     request, block_ids = self._requests_need_load[request_id]
+        #     meta.add_request(request_id=request.request_id,
+        #                      token_ids=request.prompt_token_ids,
+        #                      block_ids=block_ids,
+        #                      block_size=self._block_size)
+
+        self._requests_need_load.clear()
+        return meta
+
+    def request_finished(
+        self,
+        request: "Request",
+        block_ids: list[int],
+    ) -> tuple[bool, Optional[dict[str, Any]]]:
+        """
+        Called when a request has finished, before its blocks are freed.
+
+        Returns:
+            True if the request is being saved/sent asynchronously and blocks
+            should not be freed until the request_id is returned from
+            get_finished().
+            Optional KVTransferParams to be included in the request outputs
+            returned by the engine.
+        """
+
+        self.chunked_prefill.pop(request.request_id, None)
+
+        return False, None
+
+    # ==============================
+    # Static methods
+    # ==============================
+
+    @staticmethod
+    def parse_request_id(request_id: str, is_prefill=True) -> tuple[str, int]:
+        # Regular expression to match the string hostname and integer port
+        if is_prefill:
+            pattern = r"___decode_addr_(.*):(\d+)"
+        else:
+            pattern = r"___prefill_addr_(.*):(\d+)___"
+
+        # Use re.search to find the pattern in the request_id
+        match = re.search(pattern, request_id)
+        if match:
+            # Extract the ranks
+            ip = match.group(1)
+            port = int(match.group(2))
+
+            return ip, port
+        raise ValueError(
+            f"Request id {request_id} does not contain hostname and port")
+
+    @staticmethod
+    def check_tensors_except_dim(tensor1, tensor2, dim):
+        shape1 = tensor1.size()
+        shape2 = tensor2.size()
+
+        if len(shape1) != len(shape2) or not all(
+                s1 == s2
+                for i, (s1, s2) in enumerate(zip(shape1, shape2)) if i != dim):
+            raise NotImplementedError(
+                "Currently, only symmetric TP is supported. Asymmetric TP, PP,"
+                "and others will be supported in future PRs.")

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_engine.py

@@ -0,0 +1,529 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import logging
+import os
+import threading
+import time
+import typing
+from collections import deque
+from contextlib import contextmanager
+from typing import TYPE_CHECKING, Any, Optional
+
+import msgpack
+import torch
+import zmq
+
+from vllm.config import KVTransferConfig
+from vllm.distributed.device_communicators.pynccl_wrapper import (
+    NCCLLibrary, buffer_type, cudaStream_t, ncclComm_t, ncclDataTypeEnum)
+from vllm.distributed.kv_transfer.kv_connector.v1.p2p.tensor_memory_pool import (  # noqa: E501
+    TensorMemoryPool)
+from vllm.utils import current_stream, get_ip
+
+if TYPE_CHECKING:
+    from vllm.forward_context import ForwardContext
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_MEM_POOL_SIZE_GB = 32
+
+
+@contextmanager
+def set_p2p_nccl_context(num_channels: str):
+    original_values: dict[str, Any] = {}
+    env_vars = [
+        'NCCL_MAX_NCHANNELS',
+        'NCCL_MIN_NCHANNELS',
+        'NCCL_CUMEM_ENABLE',
+        'NCCL_BUFFSIZE',
+        'NCCL_PROTO',  # LL,LL128,SIMPLE
+        'NCCL_ALGO',  # RING,TREE
+    ]
+
+    for var in env_vars:
+        original_values[var] = os.environ.get(var)
+
+    logger.info("set_p2p_nccl_context, original_values: %s", original_values)
+
+    try:
+        os.environ['NCCL_MAX_NCHANNELS'] = num_channels
+        os.environ['NCCL_MIN_NCHANNELS'] = num_channels
+        os.environ['NCCL_CUMEM_ENABLE'] = '1'
+        yield
+    finally:
+        for var in env_vars:
+            if original_values[var] is not None:
+                os.environ[var] = original_values[var]
+            else:
+                os.environ.pop(var, None)
+
+
+class P2pNcclEngine:
+
+    def __init__(self,
+                 local_rank: int,
+                 config: KVTransferConfig,
+                 hostname: str = "",
+                 port_offset: int = 0,
+                 library_path: Optional[str] = None) -> None:
+        self.config = config
+        self.rank = port_offset
+        self.local_rank = local_rank
+        self.device = torch.device(f"cuda:{self.local_rank}")
+        self.nccl = NCCLLibrary(library_path)
+
+        if not hostname:
+            hostname = get_ip()
+        port = int(self.config.kv_port) + port_offset
+        if port == 0:
+            raise ValueError("Port cannot be 0")
+        self._hostname = hostname
+        self._port = port
+
+        # Each card corresponds to a ZMQ address.
+        self.zmq_address = f"{self._hostname}:{self._port}"
+
+        # The `http_port` must be consistent with the port of OpenAI.
+        self.http_address = (
+            f"{self._hostname}:"
+            f"{self.config.kv_connector_extra_config['http_port']}")
+
+        # If `proxy_ip` or `proxy_port` is `""`,
+        # then the ping thread will not be enabled.
+        proxy_ip = self.config.get_from_extra_config("proxy_ip", "")
+        proxy_port = self.config.get_from_extra_config("proxy_port", "")
+        if proxy_ip == "" or proxy_port == "":
+            self.proxy_address = ""
+        else:
+            self.proxy_address = proxy_ip + ":" + proxy_port
+
+        self.context = zmq.Context()
+        self.router_socket = self.context.socket(zmq.ROUTER)
+        self.router_socket.bind(f"tcp://{self.zmq_address}")
+
+        self.poller = zmq.Poller()
+        self.poller.register(self.router_socket, zmq.POLLIN)
+
+        self.send_store_cv = threading.Condition()
+        self.send_queue_cv = threading.Condition()
+        self.recv_store_cv = threading.Condition()
+
+        self.send_stream = torch.cuda.Stream()
+        self.recv_stream = torch.cuda.Stream()
+
+        mem_pool_size_gb = self.config.get_from_extra_config(
+            "mem_pool_size_gb", DEFAULT_MEM_POOL_SIZE_GB)
+        self.pool = TensorMemoryPool(max_block_size=int(mem_pool_size_gb) *
+                                     1024**3)  # GB
+
+        # The sending type includes tree mutually exclusive options:
+        # PUT, GET, PUT_ASYNC.
+        self.send_type = self.config.get_from_extra_config("send_type", "PUT")
+        if self.send_type == "GET":
+            # tensor_id: torch.Tensor
+            self.send_store: dict[str, torch.Tensor] = {}
+        else:
+            # PUT or PUT_ASYNC
+            # tensor_id: torch.Tensor
+            self.send_queue: deque[list[Any]] = deque()
+            self.send_request_id_to_tensor_ids: dict[str, set[str]] = {}
+            if self.send_type == "PUT_ASYNC":
+                self._send_thread = threading.Thread(target=self._send_async,
+                                                     daemon=True)
+                self._send_thread.start()
+
+        # tensor_id: torch.Tensor/(addr, dtype, shape)
+        self.recv_store: dict[str, Any] = {}
+        self.recv_request_id_to_tensor_ids: dict[str, set[str]] = {}
+        self.socks: dict[str, Any] = {}  # remote_address: client socket
+        self.comms: dict[str, Any] = {}  # remote_address: (ncclComm_t, rank)
+
+        self.buffer_size = 0
+        self.buffer_size_threshold = float(self.config.kv_buffer_size)
+
+        self.nccl_num_channels = self.config.get_from_extra_config(
+            "nccl_num_channels", "8")
+
+        self._listener_thread = threading.Thread(
+            target=self._listen_for_requests, daemon=True)
+        self._listener_thread.start()
+
+        self._ping_thread = None
+        if port_offset == 0 and self.proxy_address != "":
+            self._ping_thread = threading.Thread(target=self._ping,
+                                                 daemon=True)
+            self._ping_thread.start()
+
+        logger.info(
+            "💯P2pNcclEngine init, rank:%d, local_rank:%d, http_address:%s, "
+            "zmq_address:%s, proxy_address:%s, send_type:%s, buffer_size_"
+            "threshold:%.2f, nccl_num_channels:%s", self.rank, self.local_rank,
+            self.http_address, self.zmq_address, self.proxy_address,
+            self.send_type, self.buffer_size_threshold, self.nccl_num_channels)
+
+    def _create_connect(self, remote_address: typing.Optional[str] = None):
+        assert remote_address is not None
+        if remote_address not in self.socks:
+            sock = self.context.socket(zmq.DEALER)
+            sock.setsockopt_string(zmq.IDENTITY, self.zmq_address)
+            sock.connect(f"tcp://{remote_address}")
+            self.socks[remote_address] = sock
+            if remote_address in self.comms:
+                logger.info("👋comm exists, remote_address:%s, comms:%s",
+                            remote_address, self.comms)
+                return sock, self.comms[remote_address]
+
+            unique_id = self.nccl.ncclGetUniqueId()
+            data = {"cmd": "NEW", "unique_id": bytes(unique_id.internal)}
+            sock.send(msgpack.dumps(data))
+
+            with torch.cuda.device(self.device):
+                rank = 0
+                with set_p2p_nccl_context(self.nccl_num_channels):
+                    comm: ncclComm_t = self.nccl.ncclCommInitRank(
+                        2, unique_id, rank)
+                self.comms[remote_address] = (comm, rank)
+                logger.info("🤝ncclCommInitRank Success, %s👉%s, MyRank: %s",
+                            self.zmq_address, remote_address, rank)
+
+        return self.socks[remote_address], self.comms[remote_address]
+
+    def send_tensor(
+        self,
+        tensor_id: str,
+        tensor: torch.Tensor,
+        remote_address: typing.Optional[str] = None,
+    ) -> bool:
+        if remote_address is None:
+            with self.recv_store_cv:
+                self.recv_store[tensor_id] = tensor
+                self.recv_store_cv.notify()
+            return True
+        else:
+            if self.send_type == "PUT":
+                return self._send_sync(tensor_id, tensor, remote_address)
+            elif self.send_type == "PUT_ASYNC":
+                with self.send_queue_cv:
+                    self.send_queue.append([tensor_id, remote_address, tensor])
+                    self.send_queue_cv.notify()
+            else:  # GET
+                with self.send_store_cv:
+                    tensor_size = tensor.element_size() * tensor.numel()
+                    while (self.buffer_size + tensor_size
+                           > self.buffer_size_threshold):
+                        oldest_tenser_id = next(iter(self.send_store))
+                        oldest_tenser = self.send_store.pop(oldest_tenser_id)
+                        oldest_tenser_size = oldest_tenser.element_size(
+                        ) * oldest_tenser.numel()
+                        self.buffer_size -= oldest_tenser_size
+                        logger.info(
+                            "⛔[GET]Send to %s, tensor_id:%s, tensor_size:%d,"
+                            " buffer_size:%d, oldest_tenser_size:%d, rank:%d",
+                            remote_address, tensor_id, tensor_size,
+                            self.buffer_size, oldest_tenser_size, self.rank)
+
+                    self.send_store[tensor_id] = tensor
+                    self.buffer_size += tensor_size
+                    logger.debug(
+                        "🔵[GET]Send to %s, tensor_id:%s, tensor_size:%d, "
+                        "shape:%s, rank:%d, buffer_size:%d(%.2f%%)",
+                        remote_address, tensor_id, tensor_size, tensor.shape,
+                        self.rank, self.buffer_size,
+                        self.buffer_size / self.buffer_size_threshold * 100)
+
+        return True
+
+    def recv_tensor(
+        self,
+        tensor_id: str,
+        remote_address: typing.Optional[str] = None,
+    ) -> torch.Tensor:
+        if self.send_type == "PUT" or self.send_type == "PUT_ASYNC":
+            start_time = time.time()
+            with self.recv_store_cv:
+                while tensor_id not in self.recv_store:
+                    self.recv_store_cv.wait()
+                tensor = self.recv_store[tensor_id]
+
+            if tensor is not None:
+                if isinstance(tensor, tuple):
+                    addr, dtype, shape = tensor
+                    tensor = self.pool.load_tensor(addr, dtype, shape,
+                                                   self.device)
+                else:
+                    self.buffer_size -= (tensor.element_size() *
+                                         tensor.numel())
+            else:
+                duration = time.time() - start_time
+                logger.warning(
+                    "🔴[PUT]Recv From %s, tensor_id:%s, duration:%.3fms, "
+                    "rank:%d", remote_address, tensor_id, duration * 1000,
+                    self.rank)
+            return tensor
+
+        # GET
+        if remote_address is None:
+            return None
+
+        if remote_address not in self.socks:
+            self._create_connect(remote_address)
+
+        sock = self.socks[remote_address]
+        comm, rank = self.comms[remote_address]
+
+        data = {"cmd": "GET", "tensor_id": tensor_id}
+        sock.send(msgpack.dumps(data))
+
+        message = sock.recv()
+        data = msgpack.loads(message)
+        if data["ret"] != 0:
+            logger.warning("🔴[GET]Recv From %s, tensor_id: %s, ret: %d",
+                           remote_address, tensor_id, data["ret"])
+            return None
+
+        tensor = torch.empty(data["shape"],
+                             dtype=getattr(torch, data["dtype"]),
+                             device=self.device)
+
+        self._recv(comm, tensor, rank ^ 1, self.recv_stream)
+
+        return tensor
+
+    def _listen_for_requests(self):
+        while True:
+            socks = dict(self.poller.poll())
+            if self.router_socket in socks:
+                remote_address, message = self.router_socket.recv_multipart()
+                data = msgpack.loads(message)
+                if data["cmd"] == "NEW":
+                    unique_id = self.nccl.unique_id_from_bytes(
+                        bytes(data["unique_id"]))
+                    with torch.cuda.device(self.device):
+                        rank = 1
+                        with set_p2p_nccl_context(self.nccl_num_channels):
+                            comm: ncclComm_t = self.nccl.ncclCommInitRank(
+                                2, unique_id, rank)
+                        self.comms[remote_address.decode()] = (comm, rank)
+                        logger.info(
+                            "🤝ncclCommInitRank Success, %s👈%s, MyRank:%s",
+                            self.zmq_address, remote_address.decode(), rank)
+                elif data["cmd"] == "PUT":
+                    tensor_id = data["tensor_id"]
+                    try:
+                        with torch.cuda.stream(self.recv_stream):
+                            tensor = torch.empty(data["shape"],
+                                                 dtype=getattr(
+                                                     torch, data["dtype"]),
+                                                 device=self.device)
+                        self.router_socket.send_multipart(
+                            [remote_address, b"0"])
+                        comm, rank = self.comms[remote_address.decode()]
+                        self._recv(comm, tensor, rank ^ 1, self.recv_stream)
+                        tensor_size = tensor.element_size() * tensor.numel()
+                        if (self.buffer_size + tensor_size
+                                > self.buffer_size_threshold):
+                            # Store Tensor in memory pool
+                            addr = self.pool.store_tensor(tensor)
+                            tensor = (addr, tensor.dtype, tensor.shape)
+                        else:
+                            self.buffer_size += tensor_size
+
+                    except torch.cuda.OutOfMemoryError:
+                        self.router_socket.send_multipart(
+                            [remote_address, b"1"])
+                        tensor = None
+                        logger.warning(
+                            "🔴[PUT]Recv Tensor, Out Of Memory, %s👈%s, "
+                            "data:%s", self.zmq_address,
+                            remote_address.decode(), data)
+
+                    with self.recv_store_cv:
+                        self.recv_store[tensor_id] = tensor
+                        self._have_received_tensor_id(tensor_id)
+                        self.recv_store_cv.notify()
+
+                elif data["cmd"] == "GET":
+                    tensor_id = data["tensor_id"]
+                    with self.send_store_cv:
+                        tensor = self.send_store.pop(tensor_id, None)
+                        if tensor is not None:
+                            data = {
+                                "ret": 0,
+                                "shape": tensor.shape,
+                                "dtype":
+                                str(tensor.dtype).replace("torch.", "")
+                            }
+                            # LRU
+                            self.send_store[tensor_id] = tensor
+                            self._have_sent_tensor_id(tensor_id)
+                        else:
+                            data = {"ret": 1}
+
+                    self.router_socket.send_multipart(
+                        [remote_address, msgpack.dumps(data)])
+
+                    if data["ret"] == 0:
+                        comm, rank = self.comms[remote_address.decode()]
+                        self._send(comm, tensor.to(self.device), rank ^ 1,
+                                   self.send_stream)
+                else:
+                    logger.warning(
+                        "🚧Unexpected, Received message from %s, data:%s",
+                        remote_address, data)
+
+    def _have_sent_tensor_id(self, tensor_id: str):
+        request_id = tensor_id.split('#')[0]
+        if request_id not in self.send_request_id_to_tensor_ids:
+            self.send_request_id_to_tensor_ids[request_id] = set()
+        self.send_request_id_to_tensor_ids[request_id].add(tensor_id)
+
+    def _have_received_tensor_id(self, tensor_id: str):
+        request_id = tensor_id.split('#')[0]
+        if request_id not in self.recv_request_id_to_tensor_ids:
+            self.recv_request_id_to_tensor_ids[request_id] = set()
+        self.recv_request_id_to_tensor_ids[request_id].add(tensor_id)
+
+    def _send_async(self):
+        while True:
+            with self.send_queue_cv:
+                while not self.send_queue:
+                    self.send_queue_cv.wait()
+                tensor_id, remote_address, tensor = self.send_queue.popleft()
+                if not self.send_queue:
+                    self.send_queue_cv.notify()
+            self._send_sync(tensor_id, tensor, remote_address)
+
+    def wait_for_sent(self):
+        if self.send_type == "PUT_ASYNC":
+            start_time = time.time()
+            with self.send_queue_cv:
+                while self.send_queue:
+                    self.send_queue_cv.wait()
+            duration = time.time() - start_time
+            logger.debug(
+                "🚧[PUT_ASYNC]It took %.3fms to wait for the send_queue"
+                " to be empty, rank:%d", duration * 1000, self.rank)
+
+    def _send_sync(
+        self,
+        tensor_id: str,
+        tensor: torch.Tensor,
+        remote_address: typing.Optional[str] = None,
+    ) -> bool:
+        if remote_address is None:
+            return False
+        if remote_address not in self.socks:
+            self._create_connect(remote_address)
+
+        sock = self.socks[remote_address]
+        comm, rank = self.comms[remote_address]
+        data = {
+            "cmd": "PUT",
+            "tensor_id": tensor_id,
+            "shape": tensor.shape,
+            "dtype": str(tensor.dtype).replace("torch.", "")
+        }
+        sock.send(msgpack.dumps(data))
+
+        response = sock.recv()
+        if response != b"0":
+            logger.error(
+                "🔴Send Tensor, Peer Out Of Memory/Threshold, %s 👉 %s, "
+                "MyRank:%s, data:%s, tensor:%s, size:%fGB, response:%s",
+                self.zmq_address, remote_address, rank, data, tensor.shape,
+                tensor.element_size() * tensor.numel() / 1024**3,
+                response.decode())
+            return False
+
+        self._send(comm, tensor.to(self.device), rank ^ 1, self.send_stream)
+
+        if self.send_type == "PUT_ASYNC":
+            self._have_sent_tensor_id(tensor_id)
+
+        return True
+
+    def get_finished(
+        self, finished_req_ids: set[str], forward_context: "ForwardContext"
+    ) -> tuple[Optional[set[str]], Optional[set[str]]]:
+        """
+        Notifies worker-side connector ids of requests that have
+        finished generating tokens.
+
+        Returns:
+            ids of requests that have finished asynchronous transfer,
+            tuple of (sending/saving ids, recving/loading ids).
+            The finished saves/sends req ids must belong to a set provided in a
+            call to this method (this call or a prior one).
+        """
+
+        # Clear the buffer upon request completion.
+        for request_id in finished_req_ids:
+            for layer_name in forward_context.no_compile_layers:
+                tensor_id = request_id + "#" + layer_name
+                if tensor_id in self.recv_store:
+                    with self.recv_store_cv:
+                        tensor = self.recv_store.pop(tensor_id, None)
+                        self.send_request_id_to_tensor_ids.pop(
+                            request_id, None)
+                        self.recv_request_id_to_tensor_ids.pop(
+                            request_id, None)
+                    addr = 0
+                    if isinstance(tensor, tuple):
+                        addr, _, _ = tensor
+                        self.pool.free(addr)
+
+        # TODO:Retrieve requests that have already sent the KV cache.
+        finished_sending: set[str] = set()
+
+        # TODO:Retrieve requests that have already received the KV cache.
+        finished_recving: set[str] = set()
+
+        return finished_sending or None, finished_recving or None
+
+    def _ping(self):
+        sock = self.context.socket(zmq.DEALER)
+        sock.setsockopt_string(zmq.IDENTITY, self.zmq_address)
+        logger.debug("ping start, zmq_address:%s", self.zmq_address)
+        sock.connect(f"tcp://{self.proxy_address}")
+        data = {
+            "type": "P" if self.config.is_kv_producer else "D",
+            "http_address": self.http_address,
+            "zmq_address": self.zmq_address
+        }
+        while True:
+            sock.send(msgpack.dumps(data))
+            time.sleep(3)
+
+    def _send(self, comm, tensor: torch.Tensor, dst: int, stream=None):
+        assert tensor.device == self.device, (
+            f"this nccl communicator is created to work on {self.device}, "
+            f"but the input tensor is on {tensor.device}")
+        if stream is None:
+            stream = current_stream()
+
+        with torch.cuda.stream(stream):
+            self.nccl.ncclSend(buffer_type(tensor.data_ptr()), tensor.numel(),
+                               ncclDataTypeEnum.from_torch(tensor.dtype), dst,
+                               comm, cudaStream_t(stream.cuda_stream))
+        stream.synchronize()
+
+    def _recv(self, comm, tensor: torch.Tensor, src: int, stream=None):
+        assert tensor.device == self.device, (
+            f"this nccl communicator is created to work on {self.device}, "
+            f"but the input tensor is on {tensor.device}")
+        if stream is None:
+            stream = current_stream()
+
+        with torch.cuda.stream(stream):
+            self.nccl.ncclRecv(buffer_type(tensor.data_ptr()), tensor.numel(),
+                               ncclDataTypeEnum.from_torch(tensor.dtype), src,
+                               comm, cudaStream_t(stream.cuda_stream))
+        stream.synchronize()
+
+    def close(self) -> None:
+        self._listener_thread.join()
+        if self.send_type == "PUT_ASYNC":
+            self._send_thread.join()
+        if self._ping_thread is not None:
+            self._ping_thread.join()

vllm/distributed/kv_transfer/kv_connector/v1/p2p/tensor_memory_pool.py

@@ -0,0 +1,265 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import atexit
+import ctypes
+import math
+from dataclasses import dataclass
+
+import torch
+
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+@dataclass
+class MemoryBlock:
+    size: int
+    addr: int
+
+
+"""A memory pool for managing pinned host memory allocations for tensors.
+
+This class implements a buddy allocation system to efficiently manage pinned
+host memory for tensor storage. It supports allocation, deallocation, and
+tensor storage/retrieval operations.
+
+Key Features:
+- Uses power-of-two block sizes for efficient buddy allocation
+- Supports splitting and merging of memory blocks
+- Provides methods to store CUDA tensors in pinned host memory
+- Allows loading tensors from pinned memory back to device
+- Automatically cleans up memory on destruction
+
+Attributes:
+    max_block_size (int): Maximum block size (rounded to nearest power of two)
+    min_block_size (int): Minimum block size (rounded to nearest power of two)
+    free_lists (dict): Dictionary of free memory blocks by size
+    allocated_blocks (dict): Dictionary of currently allocated blocks
+    base_tensor (torch.Tensor): Base pinned memory tensor
+    base_address (int): Base memory address of the pinned memory region
+
+Example:
+    >>> pool = TensorMemoryPool(max_block_size=1024*1024)
+    >>> tensor = torch.randn(100, device='cuda')
+    >>> addr = pool.store_tensor(tensor)
+    >>> loaded_tensor = pool.load_tensor(addr, tensor.dtype,
+    ...                                  tensor.shape, 'cuda')
+    >>> pool.free(addr)
+"""
+
+
+class TensorMemoryPool:
+    """Initializes the memory pool with given size constraints.
+
+    Args:
+        max_block_size (int): Maximum size of memory blocks to manage
+        min_block_size (int, optional): Minimum size of memory blocks
+            to manage. Defaults to 512.
+
+    Raises:
+        ValueError: If block sizes are invalid or max_block_size is less
+            than min_block_size
+    """
+
+    def __init__(self, max_block_size: int, min_block_size: int = 512):
+        if max_block_size <= 0 or min_block_size <= 0:
+            raise ValueError("Block sizes must be positive")
+        if max_block_size < min_block_size:
+            raise ValueError(
+                "Max block size must be greater than min block size")
+
+        self.max_block_size = self._round_to_power_of_two(max_block_size)
+        self.min_block_size = self._round_to_power_of_two(min_block_size)
+
+        self.free_lists: dict[int, dict[int, MemoryBlock]] = {}
+        self.allocated_blocks: dict[int, MemoryBlock] = {}
+
+        self._initialize_free_lists()
+        self._allocate_pinned_memory()
+
+        atexit.register(self.cleanup)
+
+    def _round_to_power_of_two(self, size: int) -> int:
+        return 1 << (size - 1).bit_length()
+
+    def _initialize_free_lists(self):
+        size = self.max_block_size
+        while size >= self.min_block_size:
+            self.free_lists[size] = {}
+            size //= 2
+
+    def _allocate_pinned_memory(self):
+        self.base_tensor = torch.empty(self.max_block_size // 4,
+                                       dtype=torch.float32,
+                                       pin_memory=True)
+        self.base_address = self.base_tensor.data_ptr()
+        initial_block = MemoryBlock(size=self.max_block_size,
+                                    addr=self.base_address)
+        self.free_lists[self.max_block_size][
+            initial_block.addr] = initial_block
+        logger.debug("TensorMemoryPool, base_address:", self.base_address,
+                     self.base_address % self.max_block_size)
+
+    def allocate(self, size: int) -> int:
+        """Allocates a memory block of at least the requested size.
+
+        Args:
+            size (int): Minimum size of memory to allocate
+
+        Returns:
+            int: Address of the allocated memory block
+
+        Raises:
+            ValueError: If size is invalid or insufficient memory is available
+        """
+        if size <= 0:
+            raise ValueError("Allocation size must be positive")
+
+        required_size = self._round_to_power_of_two(
+            max(size, self.min_block_size))
+        if required_size > self.max_block_size:
+            raise ValueError("Requested size exceeds maximum block size")
+
+        current_size = required_size
+        while current_size <= self.max_block_size:
+            if self.free_lists[current_size]:
+                _, block = self.free_lists[current_size].popitem()
+                self._split_block(block, required_size)
+                self.allocated_blocks[block.addr] = block
+                return block.addr
+            current_size *= 2
+
+        raise ValueError("Insufficient memory")
+
+    def _split_block(self, block: MemoryBlock, required_size: int):
+        while (block.size > required_size
+               and block.size // 2 >= self.min_block_size):
+            buddy_size = block.size // 2
+            buddy_addr = block.addr + buddy_size
+
+            buddy = MemoryBlock(size=buddy_size, addr=buddy_addr)
+            block.size = buddy_size
+
+            self.free_lists[buddy_size][buddy.addr] = buddy
+
+    def free(self, addr: int):
+        """Frees an allocated memory block.
+
+        Args:
+            addr (int): Address of the block to free
+
+        Raises:
+            ValueError: If address is invalid or not allocated
+        """
+        if addr not in self.allocated_blocks:
+            raise ValueError("Invalid address to free")
+
+        block = self.allocated_blocks.pop(addr)
+        self._merge_buddies(block)
+
+    def _merge_buddies(self, block: MemoryBlock):
+        MAX_MERGE_DEPTH = 30
+        depth = 0
+
+        while depth < MAX_MERGE_DEPTH:
+            buddy_offset = block.size if (block.addr - self.base_address) % (
+                2 * block.size) == 0 else -block.size
+            buddy_addr = block.addr + buddy_offset
+            buddy = self.free_lists[block.size].get(buddy_addr)
+            if buddy:
+                del self.free_lists[buddy.size][buddy.addr]
+                merged_addr = min(block.addr, buddy.addr)
+                merged_size = block.size * 2
+                block = MemoryBlock(size=merged_size, addr=merged_addr)
+                depth += 1
+            else:
+                break
+        self.free_lists[block.size][block.addr] = block
+
+    def store_tensor(self, tensor: torch.Tensor) -> int:
+        """Stores a CUDA tensor in pinned host memory.
+
+        Args:
+            tensor (torch.Tensor): CUDA tensor to store
+
+        Returns:
+            int: Address where the tensor is stored
+
+        Raises:
+            ValueError: If tensor is not on CUDA or allocation fails
+        """
+        if not tensor.is_cuda:
+            raise ValueError("Only CUDA tensors can be stored")
+
+        size = tensor.element_size() * tensor.numel()
+        addr = self.allocate(size)
+        block = self.allocated_blocks[addr]
+
+        if block.size < size:
+            self.free(addr)
+            raise ValueError(
+                f"Allocated block size {block.size} is smaller than "
+                f"required size {size}")
+
+        try:
+            buffer = (ctypes.c_byte * block.size).from_address(block.addr)
+            cpu_tensor = torch.frombuffer(buffer,
+                                          dtype=tensor.dtype,
+                                          count=tensor.numel()).reshape(
+                                              tensor.shape)
+        except ValueError as err:
+            self.free(addr)
+            raise ValueError(f"Failed to create tensor view: {err}") from err
+
+        cpu_tensor.copy_(tensor)
+
+        return addr
+
+    def load_tensor(self, addr: int, dtype: torch.dtype,
+                    shape: tuple[int, ...], device) -> torch.Tensor:
+        """Loads a tensor from pinned host memory to the specified device.
+
+        Args:
+            addr (int): Address where tensor is stored
+            dtype (torch.dtype): Data type of the tensor
+            shape (tuple[int, ...]): Shape of the tensor
+            device: Target device for the loaded tensor
+
+        Returns:
+            torch.Tensor: The loaded tensor on the specified device
+
+        Raises:
+            ValueError: If address is invalid or sizes don't match
+        """
+        if addr not in self.allocated_blocks:
+            raise ValueError("Invalid address to load")
+
+        block = self.allocated_blocks[addr]
+        num_elements = math.prod(shape)
+        dtype_size = torch.tensor([], dtype=dtype).element_size()
+        required_size = num_elements * dtype_size
+
+        if required_size > block.size:
+            raise ValueError("Requested tensor size exceeds block size")
+
+        buffer = (ctypes.c_byte * block.size).from_address(block.addr)
+        cpu_tensor = torch.frombuffer(buffer, dtype=dtype,
+                                      count=num_elements).reshape(shape)
+
+        cuda_tensor = torch.empty(shape, dtype=dtype, device=device)
+
+        cuda_tensor.copy_(cpu_tensor)
+
+        return cuda_tensor
+
+    def cleanup(self):
+        """Cleans up all memory resources and resets the pool state."""
+        self.free_lists.clear()
+        self.allocated_blocks.clear()
+        if hasattr(self, 'base_tensor'):
+            del self.base_tensor
+
+    def __del__(self):
+        self.cleanup()

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

@@ -0,0 +1,389 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import hashlib
+import os
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+import safetensors
+import torch
+
+from vllm.config import VllmConfig
+from vllm.distributed.kv_transfer.kv_connector.v1.base import (
+    KVConnectorBase_V1, KVConnectorMetadata, KVConnectorRole)
+from vllm.logger import init_logger
+from vllm.v1.attention.backends.mla.common import MLACommonMetadata
+from vllm.v1.core.sched.output import SchedulerOutput
+
+if TYPE_CHECKING:
+    from vllm.attention.backends.abstract import AttentionMetadata
+    from vllm.forward_context import ForwardContext
+    from vllm.v1.core.kv_cache_manager import KVCacheBlocks
+    from vllm.v1.request import Request
+
+logger = init_logger(__name__)
+
+
+@dataclass
+class ReqMeta:
+    # Request tokens
+    token_ids: torch.Tensor
+    # Slot mappings, should have the same length as token_ids
+    slot_mapping: torch.Tensor
+    # Is store or load
+    is_store: bool
+
+    @staticmethod
+    def make_meta(token_ids: list[int], block_ids: list[int], block_size: int,
+                  is_store: bool) -> "ReqMeta":
+        valid_num_tokens = align_to_block_size(len(token_ids), block_size)
+        token_ids_tensor = torch.tensor(token_ids)[:valid_num_tokens]
+        block_ids_tensor = torch.tensor(block_ids)
+        num_blocks = block_ids_tensor.shape[0]
+        block_offsets = torch.arange(0, block_size)
+        slot_mapping = block_offsets.reshape((1, block_size)) + \
+                block_ids_tensor.reshape((num_blocks, 1)) * block_size
+        slot_mapping = slot_mapping.flatten()[:valid_num_tokens]
+        return ReqMeta(
+            token_ids=token_ids_tensor,
+            slot_mapping=slot_mapping,
+            is_store=is_store,
+        )
+
+
+@dataclass
+class SharedStorageConnectorMetadata(KVConnectorMetadata):
+    requests: list[ReqMeta]
+
+    def __init__(self):
+        self.requests = []
+
+    def add_request(
+        self,
+        token_ids: list[int],
+        block_ids: list[int],
+        block_size: int,
+        is_store: bool,
+    ) -> None:
+        self.requests.append(
+            ReqMeta.make_meta(token_ids, block_ids, block_size, is_store))
+
+
+class SharedStorageConnector(KVConnectorBase_V1):
+    # NOTE: This is Simple debug implementation of the KV connector.
+    # It save / load the KV cache to / from the disk.
+    # It does extra work which will overwrite the existing prefix-cache in GPU
+    # - to remove the overhead, need to add some "mask" in the ReqMeta class
+
+    def __init__(self, vllm_config: "VllmConfig", role: KVConnectorRole):
+        super().__init__(vllm_config=vllm_config, role=role)
+        self._block_size = vllm_config.cache_config.block_size
+        self._requests_need_load: dict[str, Request] = {}
+        transfer_config = vllm_config.kv_transfer_config
+        self._storage_path = transfer_config.get_from_extra_config(
+            "shared_storage_path", "/tmp")
+        logger.info(vllm_config.kv_transfer_config)
+        logger.info("Shared storage path is %s", self._storage_path)
+
+    def start_load_kv(self, forward_context: "ForwardContext",
+                      **kwargs) -> None:
+        """Start loading the KV cache from the connector buffer to vLLM's 
+        paged KV buffer.
+
+        Args:
+            forward_context (ForwardContext): the forward context.
+            **kwargs: additional arguments for the load operation
+
+        Note:
+            The number of elements in kv_caches and layer_names should be 
+            the same.
+        """
+        attn_metadata = forward_context.attn_metadata
+
+        def inject_kv_into_layer(
+            dst_kv_cache_layer: torch.Tensor,
+            src_kv_cache: torch.Tensor,
+            slot_mapping: torch.Tensor,
+        ) -> None:
+            """Inject the KV cache into the layer.
+
+            Args:
+                dst_kv_cache_layer (torch.Tensor): the destination KV cache 
+                    layer. In shape [2, num_pages, page_size, xxx] if not 
+                    using MLA, [num_pages, page_size, xxx] otherwise.
+                src_kv_cache (torch.Tensor): the source KV cache. In shape
+                    [2, num_tokens, xxx] if not using MLA, [num_tokens, xxx] 
+                    otherwise.
+                slot_mapping (torch.Tensor): the slot mapping. In shape 
+                    [num_tokens].
+            """
+            dst_kv_cache_layer_shape = dst_kv_cache_layer.shape
+            if isinstance(attn_metadata, MLACommonMetadata):
+                num_pages = dst_kv_cache_layer_shape[0]
+                page_size = dst_kv_cache_layer_shape[1]
+                dst_kv_cache_layer = dst_kv_cache_layer.reshape(
+                    num_pages * page_size, -1)
+                dst_kv_cache_layer[slot_mapping, ...] = src_kv_cache
+                dst_kv_cache_layer.reshape(dst_kv_cache_layer_shape)
+            else:
+                num_pages = dst_kv_cache_layer_shape[1]
+                page_size = dst_kv_cache_layer_shape[2]
+                dst_kv_cache_layer = dst_kv_cache_layer.reshape(
+                    2, num_pages * page_size, -1)
+                dst_kv_cache_layer[:, slot_mapping, ...] = src_kv_cache
+                dst_kv_cache_layer.reshape(dst_kv_cache_layer_shape)
+
+        # Get the metadata
+        metadata: KVConnectorMetadata = self._get_connector_metadata()
+        assert isinstance(metadata, SharedStorageConnectorMetadata)
+
+        if metadata is None:
+            logger.warning(
+                "In connector.start_load_kv, but the connector metadata is None"
+            )
+            return
+
+        attn_metadata = forward_context.attn_metadata
+        if attn_metadata is None:
+            logger.warning(
+                "In connector.start_load_kv, but the attn_metadata is None")
+            return
+
+        # Load the KV for each request each layer
+        for request in metadata.requests:
+            if request.is_store:
+                continue
+            logger.info("Inject KV cache of %d tokens to the paged memory",
+                        len(request.slot_mapping))
+            for layer_name in forward_context.no_compile_layers:
+                attn_layer = forward_context.no_compile_layers[layer_name]
+                kv_cache_layer = attn_layer.kv_cache[\
+                        forward_context.virtual_engine]
+
+                filename = self._generate_filename_debug(
+                    layer_name, request.token_ids)
+                kv_cache = safetensors.torch.load_file(
+                    filename)["kv_cache"].cuda()
+                inject_kv_into_layer(kv_cache_layer, kv_cache,
+                                     request.slot_mapping)
+
+    def wait_for_layer_load(self, layer_name: str) -> None:
+        """Blocking until the KV for a specific layer is loaded into vLLM's
+        paged buffer. 
+        
+        This interface will be useful for layer-by-layer pipelining.
+
+        Args:
+            layer_name: the name of that layer
+        """
+        return
+
+    def save_kv_layer(self, layer_name: str, kv_layer: torch.Tensor,
+                      attn_metadata: "AttentionMetadata", **kwargs) -> None:
+        """Start saving the KV cache of the layer from vLLM's paged buffer 
+        to the connector.
+
+        Args:
+            layer_name (str): the name of the layer.
+            kv_layer (torch.Tensor): the paged KV buffer of the current 
+                layer in vLLM.
+            attn_metadata (AttentionMetadata): the attention metadata.
+            **kwargs: additional arguments for the save operation.
+        """
+
+        def extract_kv_from_layer(
+            layer: torch.Tensor,
+            slot_mapping: torch.Tensor,
+        ) -> torch.Tensor:
+            """Extract the KV cache from the layer.
+
+            Assume the shape of the layer is (2, num_pages, page_size, xxx)
+            if MLA is not used, and (num_pages, page_size, xxx) otherwise.
+            """
+            if isinstance(attn_metadata, MLACommonMetadata):
+                num_pages, page_size = layer.shape[0], layer.shape[1]
+                return layer.reshape(num_pages * page_size, -1)[slot_mapping,
+                                                                ...]
+            num_pages, page_size = layer.shape[1], layer.shape[2]
+            return layer.reshape(2, num_pages * page_size, -1)[:, slot_mapping,
+                                                               ...]
+
+        connector_metadata = self._get_connector_metadata()
+        assert isinstance(connector_metadata, SharedStorageConnectorMetadata)
+        for request in connector_metadata.requests:
+            if request.is_store:
+                filename = self._generate_filename_debug(
+                    layer_name, request.token_ids)
+                kv_cache = extract_kv_from_layer(kv_layer,
+                                                 request.slot_mapping)
+                tensors = {"kv_cache": kv_cache.detach().cpu()}
+                safetensors.torch.save_file(tensors, filename)
+
+    def wait_for_save(self):
+        return
+
+    def get_num_new_matched_tokens(
+        self,
+        request: "Request",
+        num_computed_tokens: int,
+    ) -> tuple[int, bool]:
+        """
+        Get number of new tokens that can be loaded from the
+        external KV cache beyond the num_computed_tokens.
+        
+        Args:
+            request (Request): the request object.
+            num_computed_tokens (int): the number of locally
+                computed tokens for this request
+
+        Returns:
+            the number of tokens that can be loaded from the 
+            external KV cache beyond what is already computed.
+        """
+        # NOTE: in this debug implementation, we assume that the prompt is
+        # cached_prompt + newly_generated_single_token
+        # Therefore, we use prompt_token_ids[:-1] to determine the folder name
+
+        # NOTE: in current v1 scheduler, the num_computed_tokens is aligned
+        # with the block granularity. And it expects the returned blocks and
+        # num_computed_tokens to also be aligned with the block granularity.
+        if not self._found_match_for_request(request):
+            return 0, False
+
+        logger.info("External Cache Hit!")
+
+        # Now, first num_tokens_to_check tokens are hit, we need to prepare
+        # the metadata for the worker connector to correctly load the KV
+        num_tokens_to_check = align_to_block_size(
+            len(request.prompt_token_ids) - 1, self._block_size)
+
+        return num_tokens_to_check - num_computed_tokens, False
+
+    def update_state_after_alloc(self, request: "Request",
+                                 blocks: "KVCacheBlocks",
+                                 num_external_tokens: int):
+        """
+        Update KVConnector state after block allocation.
+
+        If blocks were allocated, add to _requests_need_load,
+        such that we load the KVs in the next forward pass.
+        """
+        if num_external_tokens > 0:
+            self._requests_need_load[request.request_id] = request
+
+    def build_connector_meta(
+        self,
+        scheduler_output: SchedulerOutput,
+    ) -> KVConnectorMetadata:
+        """Build the connector metadata for this step.
+
+        This function should NOT modify any fields in the scheduler_output.
+        Also, calling this function will reset the state of the connector.
+
+        Args:
+            scheduler_output (SchedulerOutput): the scheduler output object.
+        """
+        meta = SharedStorageConnectorMetadata()
+
+        total_need_load = 0
+        for new_req in scheduler_output.scheduled_new_reqs:
+            if new_req.req_id in self._requests_need_load:
+                meta.add_request(token_ids=new_req.prompt_token_ids,
+                                 block_ids=new_req.block_ids[0],
+                                 block_size=self._block_size,
+                                 is_store=False)
+                total_need_load += 1
+            else:
+                # NOTE: here, we set the store and load being exclusive,
+                # but a single request can have both store and load.
+                # NOTE(rob): for this debug implementation, we only cache
+                # the original prompt tokens.
+                if not self._found_match_for_request(new_req):
+                    meta.add_request(token_ids=new_req.prompt_token_ids,
+                                     block_ids=new_req.block_ids[0],
+                                     block_size=self._block_size,
+                                     is_store=True)
+
+        cached_reqs = scheduler_output.scheduled_cached_reqs
+        for i, req_id in enumerate(cached_reqs.req_ids):
+            num_computed_tokens = cached_reqs.num_computed_tokens[i]
+            num_new_tokens = scheduler_output.num_scheduled_tokens[req_id]
+            new_block_ids = cached_reqs.new_block_ids[i]
+            resumed_from_preemption = cached_reqs.resumed_from_preemption[i]
+
+            # NOTE(rob): here we rely on the resumed requests being
+            # the first N requests in the list scheduled_cache_reqs.
+            if not resumed_from_preemption:
+                break
+            if req_id in self._requests_need_load:
+                # NOTE(rob): cached_req_data does not have the full
+                # list of token ids (only new tokens). So we look it
+                # up in the actual request object.
+                request = self._requests_need_load[req_id]
+                total_tokens = num_computed_tokens + num_new_tokens
+                token_ids = request.all_token_ids[:total_tokens]
+
+                # NOTE(rob): For resumed req, new_block_ids is all
+                # of the block_ids for the request.
+                block_ids = new_block_ids[0]
+
+                meta.add_request(token_ids=token_ids,
+                                 block_ids=block_ids,
+                                 block_size=self._block_size,
+                                 is_store=False)
+                total_need_load += 1
+
+        assert total_need_load == len(self._requests_need_load)
+        self._requests_need_load.clear()
+        return meta
+
+    # ==============================
+    # Helper functions
+    # ==============================
+
+    def _found_match_for_request(
+        self,
+        request: "Request",
+    ) -> bool:
+        """Check if the cache is hit for the request.
+        """
+        num_tokens_to_check = align_to_block_size(
+            len(request.prompt_token_ids) - 1, self._block_size)
+        foldername = self._generate_foldername_debug(torch.tensor(
+            request.prompt_token_ids)[:num_tokens_to_check],
+                                                     create_folder=False)
+        return os.path.exists(foldername)
+
+    def _generate_foldername_debug(
+        self,
+        input_ids: torch.Tensor,
+        create_folder=False,
+    ) -> str:
+        """Generate a folder name based on the hash of the bytes of the input 
+        ids.
+        """
+        input_ids_bytes = input_ids.numpy().tobytes()
+        input_ids_hash = hashlib.md5(input_ids_bytes,
+                                     usedforsecurity=False).hexdigest()
+        foldername = os.path.join(self._storage_path, input_ids_hash)
+        if create_folder:
+            os.makedirs(foldername, exist_ok=True)
+        return foldername
+
+    def _generate_filename_debug(
+        self,
+        layer_name: str,
+        input_ids: torch.Tensor,
+    ) -> str:
+        """Generate a file name based on the layer name and the hash 
+        of the bytes of the input ids.
+        """
+        foldername = self._generate_foldername_debug(input_ids,
+                                                     create_folder=True)
+        return os.path.join(foldername, f"{layer_name}.safetensors")
+
+
+def align_to_block_size(num_tokens: int, block_size) -> int:
+    """Align the number of tokens to the block size.
+    """
+    return (num_tokens - 1) // block_size * block_size