Function build_decoder_resblock

src/models/pocket_tts/mimi_decoder.cpp:1028–1067

        const int64_t step_elems = config_.num_heads * head_dim_;
        carry_key_sources_.assign(static_cast<size_t>(cache_steps_ + 1), {});
        carry_value_sources_.assign(static_cast<size_t>(cache_steps_ + 1), {});
        carry_key_destinations_.assign(static_cast<size_t>(cache_steps_ + 1), {});
        carry_value_destinations_.assign(static_cast<size_t>(cache_steps_ + 1), {});
        append_key_sources_.assign(static_cast<size_t>(cache_steps_ + 1), {});
        append_value_sources_.assign(static_cast<size_t>(cache_steps_ + 1), {});
        append_key_destinations_.assign(static_cast<size_t>(cache_steps_ + 1), {});
        append_value_destinations_.assign(static_cast<size_t>(cache_steps_ + 1), {});

        for (int64_t next_keep_steps = 1; next_keep_steps <= cache_steps_; ++next_keep_steps) {
            const int64_t append_steps = std::min<int64_t>(frames_, next_keep_steps);
            const int64_t carry_steps = next_keep_steps - append_steps;
            auto & carry_key_src = carry_key_sources_[static_cast<size_t>(next_keep_steps)];
            auto & carry_value_src = carry_value_sources_[static_cast<size_t>(next_keep_steps)];
            auto & carry_key_dst = carry_key_destinations_[static_cast<size_t>(next_keep_steps)];
            auto & carry_value_dst = carry_value_destinations_[static_cast<size_t>(next_keep_steps)];
            auto & append_key_src = append_key_sources_[static_cast<size_t>(next_keep_steps)];
            auto & append_value_src = append_value_sources_[static_cast<size_t>(next_keep_steps)];
            auto & append_key_dst = append_key_destinations_[static_cast<size_t>(next_keep_steps)];
            auto & append_value_dst = append_value_destinations_[static_cast<size_t>(next_keep_steps)];
            carry_key_src.reserve(static_cast<size_t>(config_.transformer_layers));
            carry_value_src.reserve(static_cast<size_t>(config_.transformer_layers));
            carry_key_dst.reserve(static_cast<size_t>(config_.transformer_layers));
            carry_value_dst.reserve(static_cast<size_t>(config_.transformer_layers));
            append_key_src.reserve(static_cast<size_t>(config_.transformer_layers));
            append_value_src.reserve(static_cast<size_t>(config_.transformer_layers));
            append_key_dst.reserve(static_cast<size_t>(config_.transformer_layers));
            append_value_dst.reserve(static_cast<size_t>(config_.transformer_layers));

            const size_t append_elems = static_cast<size_t>(append_steps * step_elems);
            const size_t append_src_offset = static_cast<size_t>((frames_ - append_steps) * step_elems) * sizeof(float);
            const size_t append_dst_offset = static_cast<size_t>((cache_steps_ - append_steps) * step_elems) * sizeof(float);
            const size_t carry_elems = static_cast<size_t>(carry_steps * step_elems);
            const size_t carry_src_offset = static_cast<size_t>((cache_steps_ - carry_steps) * step_elems) * sizeof(float);
            const size_t carry_dst_offset = static_cast<size_t>((cache_steps_ - next_keep_steps) * step_elems) * sizeof(float);

            for (size_t layer = 0; layer < prefix_keys_.size(); ++layer) {
                if (carry_steps > 0) {
                    carry_key_src.push_back(ggml_view_1d(ggml_ctx_, prefix_keys_[layer].tensor, carry_elems, carry_src_offset));
                    carry_value_src.push_back(ggml_view_1d(ggml_ctx_, prefix_values_[layer].tensor, carry_elems, carry_src_offset));
                    carry_key_dst.push_back(ggml_view_1d(ggml_ctx_, work_prefix_keys_[layer].tensor, carry_elems, carry_dst_offset));
                    carry_value_dst.push_back(ggml_view_1d(ggml_ctx_, work_prefix_values_[layer].tensor, carry_elems, carry_dst_offset));
                } else {
                    carry_key_src.push_back(nullptr);