preemo_text-generation-inference/proto/generate.proto

syntax = "proto3";

package generate.v1;

service TextGenerationService {
    /// Service discovery
    rpc ServiceDiscovery (ServiceDiscoveryRequest) returns (ServiceDiscoveryResponse) {}
    /// Empties batch cache
    rpc ClearCache (ClearCacheRequest) returns (ClearCacheResponse);
    /// Generate tokens for a batch
    rpc Generate (GenerateRequest) returns (GenerateResponse);
    /// Generate tokens for a list of cached batches
    rpc GenerateWithCache (GenerateWithCacheRequest) returns (GenerateWithCacheResponse);
    /// Generate tokens until the text of at least one request of the batch is generated
    rpc GenerateUntilFinished (GenerateUntilFinishedRequest) returns (GenerateUntilFinishedResponse);
    /// Generate tokens until the text of at least one request of the cached batches i finished
    rpc GenerateUntilFinishedWithCache (GenerateUntilFinishedWithCacheRequest) returns (GenerateUntilFinishedWithCacheResponse);
}

/// Empty request
message ServiceDiscoveryRequest {}

message ServiceDiscoveryResponse {
    /// Other shards urls
    repeated string urls = 1;
}

/// Empty request
message ClearCacheRequest {}

/// Empty response
message ClearCacheResponse {}

message LogitsWarperParameters {
    float temperature = 1;
    uint32 top_k = 2;
    float top_p = 3;
    bool do_sample = 4;
}

message Request {
    /// Request ID
    uint64 id = 1;
    /// The generation context
    string inputs = 2;
    /// The number of tokens inside inputs
    uint32 input_length = 3;
    /// Logits Warper Parameters
    LogitsWarperParameters parameters = 4;
    /// Stopping criteria
    uint32 max_new_tokens = 5;
}

message Batch {
    /// Batch ID
    uint64 id = 1;
    /// Individual requests
    repeated Request requests = 2;
    /// Batch size (==len(requests))
    uint32 size = 3;
    /// Length of the longest sequence within the batch (used for padding)
    uint32 max_sequence_length = 4;
}

message GeneratedText {
    /// Request
    Request request = 1;
    /// Output
    string output = 2;
}

message GenerateRequest {
    /// Batch
    Batch batch = 1;
}

message GenerateResponse {
    /// Finished requests
    repeated GeneratedText generated_texts = 1;
    /// Next batch (cached)
    optional Batch batch = 2;
}

message GenerateWithCacheRequest {
    /// Cached batches
    repeated Batch batches = 1;
}

message GenerateWithCacheResponse {
    /// Finished requests
    repeated GeneratedText generated_texts = 1;
    /// Next batch (cached)
    optional Batch batch = 2;
}

message GenerateUntilFinishedRequest {
    /// Batch
    Batch batch = 1;
}

message GenerateUntilFinishedResponse {
    /// Finished requests
    repeated GeneratedText generated_texts = 1;
    /// Next batch (cached)
    optional Batch batch = 2;
}

message GenerateUntilFinishedWithCacheRequest {
    /// Cached batches
    repeated Batch batches = 1;
}

message GenerateUntilFinishedWithCacheResponse {
    /// Finished requests
    repeated GeneratedText generated_texts = 1;
    /// Next batch (cached)
    optional Batch batch = 2;
}