| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970 |
- syntax = "proto3";
- package node_service;
- service NodeService {
- rpc SendPrompt (PromptRequest) returns (Tensor) {}
- rpc SendTensor (TensorRequest) returns (Tensor) {}
- rpc ResetShard (ResetShardRequest) returns (Empty) {}
- rpc GetInferenceResult (GetInferenceResultRequest) returns (InferenceResult) {}
- rpc CollectTopology (CollectTopologyRequest) returns (Topology) {}
- }
- message Shard {
- string model_id = 1;
- int32 start_layer = 2;
- int32 end_layer = 3;
- int32 n_layers = 4;
- }
- message PromptRequest {
- Shard shard = 1;
- string prompt = 2;
- optional string request_id = 3;
- }
- message TensorRequest {
- Shard shard = 1;
- Tensor tensor = 2;
- optional string request_id = 3;
- }
- message GetInferenceResultRequest {
- string request_id = 1;
- }
- message InferenceResult {
- optional Tensor tensor = 1;
- bool is_finished = 2;
- }
- message Tensor {
- bytes tensor_data = 1;
- repeated int32 shape = 2;
- string dtype = 3;
- }
- message ResetShardRequest {
- Shard shard = 1;
- }
- message CollectTopologyRequest {
- int32 max_depth = 1;
- }
- message Topology {
- map<string, DeviceCapabilities> nodes = 1;
- map<string, Peers> peer_graph = 2;
- }
- message Peers {
- repeated string peer_ids = 1;
- }
- message DeviceCapabilities {
- string model = 1;
- string chip = 2;
- int32 memory = 3;
- }
- message Empty {}
|