Non-distilled models (xl-base / xl-sft) produce garbled audio on Apple Silicon — MLX and PyTorch-MPS; distilled turbo models fine #1259

Setup: Apple Silicon (M5 Max, 128 GB, macOS), ACE-Step 1.5 via uv sync (torch-MPS + mlx), generation through cli.py with the full LM pipeline.

Symptom: acestep-v15-xl-base and acestep-v15-xl-sft reproducibly generate garbled/mushy audio (spectral flatness ~0.25 vs. 0.07–0.10 for the distilled models), regardless of steps (40–120) and guidance_scale (1.0–10.0). All distilled turbo models (2B and XL) sound correct with identical prompts and pipeline.

Ruled out:

• Corrupted download: all 4 xl-base safetensors shards are sha256-identical to the HuggingFace source (HfApi().get_paths_info() vs. local shasum -a 256).
• Wrong weights: base ≠ turbo verified via byte compare of the shards (only the index JSON is identical, as expected for the shared architecture).
• MLX DiT backend: same garbled result with use_mlx_dit=false → PyTorch MPS (xl-sft, 50 steps, guidance_scale=7, thinking=false, exit 0, no errors).

Observation: xl-base and xl-sft share modeling_acestep_v15_xl_base.py and fail identically, while all distilled models work — so the problem appears to live in the shared non-distilled CFG/sampler path on Apple Silicon, not in the weights or a single backend.

Possibly related: #1220 (closed as stale, unanswered).

Happy to provide audio samples, configs, or to test patches.