Neural codecs with residual vector quantization (RVQ) yield compact discrete representations that preserve high-fidelity speech. While flow matching-based TTS reports strong quality, directly applying discrete flow matching (DFM) to RVQ tokens under a single-code assumption underutilizes the codec hierarchy. We propose Hierarchical Discrete Flow Matching (H-DFM), a multi-head extension of DFM that aligns training with the RVQ hierarchy via a stochastic coarse/fine curriculum: coarse mode updates with fine masked, and fine mode updates conditioned on ground-truth coarse. Also, we use a coarse-biased inference schedule that allocates most of the refinement budget to coarse codebooks, with a small portion allocated to fine codebooks. Experimental results demonstrate superior improvements compared to the FM baselines, indicating that H-DFM integrates seamlessly with flow-matching-based TTS pipelines.