Method require_f32

src/framework/assets/tensor_source.cpp:371–378

    const int64_t count = static_cast<int64_t>(values.size());
    const int64_t chunks = (count + kF32ConvertChunkElements - 1) / kF32ConvertChunkElements;
    #pragma omp parallel for schedule(static)
    for (int64_t chunk = 0; chunk < chunks; ++chunk) {
        const int64_t offset = chunk * kF32ConvertChunkElements;
        const int64_t length = std::min(kF32ConvertChunkElements, count - offset);
        ggml_fp32_to_bf16_row(values.data() + offset, converted.data() + offset, length);
    }
    set_tensor_bytes(tensor, converted.data(), converted.size() * sizeof(ggml_bf16_t), name);
    if (retained_bytes) {
        const auto * begin = reinterpret_cast<const std::byte *>(converted.data());
        retained_bytes->assign(begin, begin + converted.size() * sizeof(ggml_bf16_t));
    }