Method run

src/models/vibevoice/decoder.cpp:421–440

    }

    VibeVoiceTokenEmbeddings run(const std::vector<int32_t> & input_ids) {
        const auto & config = runtime_->assets().config.decoder;
        if (static_cast<int64_t>(input_ids.size()) != steps_) {
            throw std::runtime_error("VibeVoice decoder embedding input size mismatch");
        }
        ggml_backend_tensor_set(input_ids_, input_ids.data(), 0, input_ids.size() * sizeof(int32_t));
        core::set_backend_threads(runtime_->backend(), runtime_->threads());
        const ggml_status status = engine::core::compute_backend_graph(runtime_->backend(), graph_);
        ggml_backend_synchronize(runtime_->backend());
        if (status != GGML_STATUS_SUCCESS) {
            throw std::runtime_error("VibeVoice decoder embedding graph compute failed");
        }

        VibeVoiceTokenEmbeddings out;
        out.steps = steps_;
        out.hidden_size = config.hidden_size;
        out.values.resize(static_cast<size_t>(steps_ * config.hidden_size));
        ggml_backend_tensor_get(output_, out.values.data(), 0, out.values.size() * sizeof(float));
        return out;
    }

private:
    const VibeVoiceDecoderWeightsRuntime * runtime_ = nullptr;