Method decode

src/models/pocket_tts/mimi_decoder.cpp:1278–1985

    int stride,
    StreamingConvTranspose1dState & state) {
    auto output = runtime.run(input);
    const int64_t raw_frames = frames == 0 ? 0 : (frames - 1) * stride + kernel_size;
    if (state.partial_frames > 0) {
        for (int64_t channel = 0; channel < channels; ++channel) {
            for (int64_t t = 0; t < state.partial_frames; ++t) {
                output[static_cast<size_t>(channel * raw_frames + t)] +=
                    state.partial[static_cast<size_t>(channel * state.partial_frames + t)];
            }
        }
        state.partial = tail_bct(output, channels, raw_frames, state.partial_frames);
        output = head_bct(output, channels, raw_frames, raw_frames - state.partial_frames);
    }
    return output;
}

}  // namespace

struct MimiDecoder::RuntimeCache {
    const models::pocket_tts::PocketTTSAssets * manifest = nullptr;
    ggml_backend_t backend = nullptr;
    int threads = 0;
    size_t conv_graph_context_bytes = 0;
    size_t transformer_graph_context_bytes = 0;
    size_t tail_graph_context_bytes = 0;

    std::unique_ptr<DepthwiseConvTranspose1dRuntime> encoder_rate_upsample_runtime;
    int64_t encoder_rate_upsample_steps = -1;
    std::unique_ptr<Conv1dRuntime> quantizer_runtime;
    int64_t quantizer_steps = -1;
    std::unique_ptr<MimiTransformerRuntime> transformer_runtime;
    int64_t transformer_frames = -1;
    std::unique_ptr<Conv1dRuntime> input_projection_runtime;
    int64_t input_projection_frames = -1;
    std::unique_ptr<ConvTranspose1dRuntime> stage0_upsample_runtime;
    int64_t stage0_upsample_frames = -1;
    std::unique_ptr<Conv1dRuntime> stage0_conv1_runtime;
    std::unique_ptr<Conv1dRuntime> stage0_conv2_runtime;
    std::unique_ptr<ConvTranspose1dRuntime> stage1_upsample_runtime;
    int64_t stage1_upsample_frames = -1;
    std::unique_ptr<Conv1dRuntime> stage1_conv1_runtime;
    std::unique_ptr<Conv1dRuntime> stage1_conv2_runtime;
    std::unique_ptr<ConvTranspose1dRuntime> stage2_upsample_runtime;
    int64_t stage2_upsample_frames = -1;
    std::unique_ptr<Conv1dRuntime> stage2_conv1_runtime;
    std::unique_ptr<Conv1dRuntime> stage2_conv2_runtime;
    std::unique_ptr<Conv1dRuntime> output_projection_runtime;
    int64_t output_projection_frames = -1;
    std::array<int64_t, 3> resblock_conv1_frames = {-1, -1, -1};
    std::array<int64_t, 3> resblock_conv2_frames = {-1, -1, -1};

    std::unique_ptr<Conv1dRuntime> full_quantizer_runtime;
    int64_t full_quantizer_steps = -1;
    std::unique_ptr<DepthwiseConvTranspose1dRuntime> full_encoder_rate_upsample_runtime;
    int64_t full_encoder_steps = -1;
    std::unique_ptr<MimiTransformerRuntime> full_transformer_runtime;
    int64_t full_transformer_frames = -1;
    std::unique_ptr<MimiFullDecoderRuntime> full_decoder_runtime;
    int64_t full_decoder_steps = -1;