dia tested by Argusic Runner

What the test found

Package nari-tts installs and all modules import, config loads from HF, CLI help parses, and audio utility functions work correctly, but the 1.6B-parameter model cannot load on this container's 755 MB of RAM (needs at least 3.2 GB).

Installed nari-tts 0.1.0 in .venv with CPU-only PyTorch 2.6.0. All Python modules import successfully (dia.model, dia.layers, dia.config, dia.audio, dia.state; plus dac, torch, torchaudio, soundfile, safetensors, gradio, huggingface_hub, librosa, numpy, pydantic). CLI --help and app.py --help parse correctly. HF config downloads and deserializes correctly (model_type=dia, 1024-dim encoder, 2048-dim decoder, 12 encoder layers, 18 decoder layers). Audio delay/build/revert utility functions verified via round-trip test. The 1.6B-parameter model requires ~3.2 GB RAM minimum and this container has only 755 MB total. Running example/simple-cpu.py was OOM-killed after downloading the HF config. Attempting app.py was OOM-killed at model load. No third-party keys or network services are required; the limiting factor is RAM. A potential broadcasting bug in SelfAttention/RotaryEmbedding (line 134 of dia/layers.py, dimension mismatch between input tokens and full position range) exists but only manifests in manual sub-component tests, not in the production code path where Dia._pad_text_input pads all inputs to max_position_embeddings.

Argusic installed dia in 3.2 minutes; verification reached Nothing verified. 2 of 2 recorded errors were worked through (see the timeline below).

Did not runall runs of dia

This summary is drawn from the agent's recorded report for this run. Every figure traces to the log and recording above; nothing here is authored.