vllm.config.kv_transfer ¶
Classes:
-
KVTransferConfig–Configuration for distributed KV cache transfer.
KVTransferConfig ¶
Configuration for distributed KV cache transfer.
Methods:
-
compute_hash–WARNING: Whenever a new field is added to this config,
-
has_connector–Whether
connector_nameis configured, directly or in MultiConnector.
Attributes:
-
enable_permute_local_kv(bool) –Experiment feature flag to enable HND to NHD KV Transfer
-
engine_id(str | None) –The engine id for KV transfers.
-
kv_buffer_device(str) –The device used by kv connector to buffer the KV cache. Choices are
-
kv_buffer_size(float) –The buffer size for TorchDistributedConnector. Measured in number of
-
kv_connector(str | None) –The KV connector for vLLM to transmit KV caches between vLLM instances.
-
kv_connector_extra_config(dict[str, Any]) –any extra config that the connector may need.
-
kv_connector_module_path(str | None) –The Python module path to dynamically load the KV connector from.
-
kv_ip(str) –The KV connector ip, used to build distributed connection.
-
kv_load_failure_policy(Literal['recompute', 'fail']) –Policy for handling KV cache load failures.
-
kv_parallel_size(int) –The number of parallel instances for KV cache transfer.
-
kv_port(int) –The KV connector port, used to build distributed connection.
-
kv_rank(int | None) –The rank of this vLLM instance in the KV cache transfer. Typical value:
-
kv_role(KVRole | None) –Whether this vLLM instance produces, consumes KV cache, or both. Choices
Source code in vllm/config/kv_transfer.py
56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 | |
enable_permute_local_kv = False class-attribute instance-attribute ¶
Experiment feature flag to enable HND to NHD KV Transfer
engine_id = None class-attribute instance-attribute ¶
The engine id for KV transfers.
kv_buffer_device = field(default_factory=kv_buffer_device_default_factory) class-attribute instance-attribute ¶
The device used by kv connector to buffer the KV cache. Choices are 'cuda', 'cpu' and 'xpu'.
kv_buffer_size = 1000000000.0 class-attribute instance-attribute ¶
The buffer size for TorchDistributedConnector. Measured in number of bytes. Recommended value: 1e9 (about 1GB).
kv_connector = None class-attribute instance-attribute ¶
The KV connector for vLLM to transmit KV caches between vLLM instances.
kv_connector_extra_config = field(default_factory=dict) class-attribute instance-attribute ¶
any extra config that the connector may need.
kv_connector_module_path = None class-attribute instance-attribute ¶
The Python module path to dynamically load the KV connector from. Only supported in V1.
kv_ip = '127.0.0.1' class-attribute instance-attribute ¶
The KV connector ip, used to build distributed connection.
kv_load_failure_policy = 'fail' class-attribute instance-attribute ¶
Policy for handling KV cache load failures. 'recompute': reschedule the request to recompute failed blocks 'fail': immediately fail the request with an error finish reason (default)
kv_parallel_size = 1 class-attribute instance-attribute ¶
The number of parallel instances for KV cache transfer.
kv_port = 14579 class-attribute instance-attribute ¶
The KV connector port, used to build distributed connection.
kv_rank = None class-attribute instance-attribute ¶
The rank of this vLLM instance in the KV cache transfer. Typical value: 0 for prefill instance, 1 for decode instance. Currently only 1P1D is supported.
kv_role = None class-attribute instance-attribute ¶
Whether this vLLM instance produces, consumes KV cache, or both. Choices are 'kv_producer', 'kv_consumer', and 'kv_both'.
compute_hash() ¶
WARNING: Whenever a new field is added to this config, ensure that it is included in the factors list if it affects the computation graph.
Provide a hash that uniquely identifies all the configs that affect the structure of the computation graph from input ids/embeddings to the final hidden states, excluding anything before input ids/embeddings and after the final hidden states.
Source code in vllm/config/kv_transfer.py
has_connector(connector_name) ¶
Whether connector_name is configured, directly or in MultiConnector.