emb1ter/ACE-Step-v1.5-5Hz-LM-4B-ONNX-WebGPU

ACE-Step 1.5 5 Hz LM 4B — FP16 ONNX WebGPU

Experimental browser conversion of the official ACE-Step/acestep-5Hz-lm-4B music planner for Transformers.js and ONNX Runtime Web.

This is the ACE-Step planner, not the audio generator. It produces metadata and 5 Hz semantic audio-code IDs which must be decoded by the ACE-Step semantic detokenizer and supplied to a compatible ACE-Step DiT.

Conversion

• Exact 36-layer ACE-Step 4B fine-tuned weights, converted from BF16 to FP16
• FP16 activations and KV cache
• Five external-data chunks, each below 2 GB
• ONNX graph plus weight data: 8,389,939,364 bytes
• Browser filename: onnx/model_fp16.onnx

A direct FP16 conversion overflows in the residual stream on the tested runtime. This graph applies a residual scale of 1/256 to the embedding and attention/MLP output projections, and scales residual RMSNorm epsilon by the square of that factor. See conversion-metadata.json.

Validation status

• A native ONNX Runtime cached-generation smoke test completed 150 semantic steps with finite logits.
• That smoke run produced 142 unique codes and 149 transitions, with no identical adjacent codes.
• Against the official MLX checkpoint on one 30-second vocal prompt, first-step top-20 overlap was 60% conditional, 85% unconditional, and 50% after CFG 2.0. Centered logit correlations were 0.993942, 0.990590, and 0.990873.

This is an experimental numerical workaround, not an exact conversion. End-to-end browser audio quality must not be inferred from successful model loading or the finite-logit smoke test.

License and attribution

The source ACE-Step checkpoint is MIT licensed. The repository LICENSE applies to this derivative conversion and ACE-Step attribution is preserved. ONNX Runtime and Transformers.js retain their respective licenses.