Method _sample_in_tool

src/common/AutoModel/modeling_gpt_oss.cpp:329–367

}

int GPT_OSS::_sample_in_tool(buffer<bf16>& logits) {
    // to=functions.get_current_weather <|constrain|>json<|message|>{"location":"San Francisco"}<|call|>
    std::vector<int> allowed_tokens;
    switch (current_state) {
        case STATE_EXPECT_TO_FUNCTIONS: {
            allowed_tokens = get_to_functions_tokens();
            break;
        }
        case STATE_EXPECT_FUNCTION_NAME: {
            allowed_tokens = get_function_name_tokens();
            break;
        }
        case STATE_EXPECT_CONSTRAIN: {
            auto constrain_tokens = tokenizer->encode(" <|constrain|>");
            auto message_tokens = tokenizer->encode("json");
            allowed_tokens.insert(allowed_tokens.end(),
                constrain_tokens.begin(), constrain_tokens.end());
            allowed_tokens.insert(allowed_tokens.end(),
                message_tokens.begin(), message_tokens.end());
            break;
        }
        case STATE_EXPECT_MESSAGE: {
            auto tokens = tokenizer->encode("<|message|>");
            allowed_tokens = { tokens[0] };
            break;
        }
        //case STATE_IN_JSON: {
        //    header_print("JSONSSS", "SSSS");
        //    return this->sampler->sample(logits);
        //    //allowed_tokens = constrain_by_json_schema(logits);
        //    //break;
        //}
    }
    mask_logits(logits, allowed_tokens);
    int token_id = this->sampler->sample(logits);
    update_state(token_id);
    return token_id;
    //return this->sampler->sample(logits);
}

std::vector<int> GPT_OSS::get_to_functions_tokens() {
    const std::string target = " to=functions.";