// Copyright 2023 The ODML Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
syntax = "proto3";
package odml.infra.proto;
import "mediapipe/tasks/cc/genai/inference/proto/prompt_template.proto";
import "mediapipe/tasks/cc/genai/inference/proto/transformer_params.proto";
option java_package = "com.google.odml.infra.proto";
option java_outer_classname = "LLMParametersProto";
// Naming convention is as following: LLM_MODEL_TYPE_<model_name>_<model_size>.
enum LlmModelType {
// Unknown LLM model type
LLM_MODEL_TYPE_UNKNOWN = 0;
reserved 1, 2, 3, 4, 7, 9, 10, 13, 14, 15, 16;
// FALCON RefinedWeb with 1B parameters.
// https://huggingface.co/tiiuae/falcon-rw-1b
LLM_MODEL_TYPE_FALCON_RW_1B = 5;
// GEMMA with 2B parameters
LLM_MODEL_TYPE_GEMMA_2B = 6;
// GEMMA with 7B parameters
LLM_MODEL_TYPE_GEMMA_7B = 12;
// StableLM 4E1T with 3B parameters
LLM_MODEL_TYPE_STABLELM_4E1T_3B = 8;
// Phi-2
// https://huggingface.co/microsoft/phi-2
LLM_MODEL_TYPE_PHI_2 = 11;
// Arbitrary TfLite model via ODML stack.
LLM_MODEL_TYPE_TF_LITE = 100;
}
// Parameters for Large Language Models (LLM).
message LlmParameters {
TransformerParameters transformer_parameters = 1;
// Size of vocabulary.
int32 vocab_size = 2;
// Was used for disable_kv_cache.
reserved 3;
// Start token prepended to the beginning of input sequence.
oneof start_token_union {
int32 start_token_id = 4;
string start_token = 6;
}
// Stop tokens to determine the end of output stream.
repeated string stop_tokens = 5;
enum InputOutputNormalization {
INPUT_OUTPUT_NORMALIZATION_UNKNOWN = 0;
// Enables GPT2 stype bytes to unicode mapping normalization. For more
// details, see:
// https://github.com/openai/gpt-2/blob/master/src/encoder.py#L9
INPUT_OUTPUT_NORMALIZATION_BYTES_TO_UNICODE = 1;
}
repeated InputOutputNormalization input_output_normalizations = 7;
PromptTemplate prompt_template = 8;
// Number of tokens to draft in a single step when using speculative decoding.
optional int32 num_draft_tokens = 9;
}