emb1ter/ACE-Step-v1.5-5Hz-LM-4B-FP32-ONNX-WebGPU

ACE-Step 1.5 5 Hz LM 4B · ONNX WebGPU bodies

Experimental browser exports of the ACE-Step 1.5 5 Hz 4B planning language model for ai-music-js.

The repository keeps two transformer-body variants:

• int8-fp32/onnx/model.onnx: browser-qualified weight-only INT8 body with FP32 activations, residuals, normalization, and KV cache;
• onnx/model.onnx: original unquantized FP32 reference body.

Both variants evaluate token embeddings and the 64,000-row audio-code output head outside ONNX Runtime. This avoids placing the complete 217,204 × 2,560 tied FP32 matrix inside one browser ORT session.

Repository layout:

• int8-fp32/onnx/model.onnx plus five .data.* files: qualified hybrid transformer body;
• onnx/model.onnx plus eight model.onnx_data* files: unquantized FP32 reference body;
• fp32-head/audio_head_*.bin: row-aligned FP32 audio-code-head shards;
• fp32-embedding/model.onnx_data: tied embedding source read sparsely with HTTP Range requests;
• tokenizer files at the repository root.

Hybrid precision

• 252 asymmetric com.microsoft:MatMulNBits INT8 projection nodes
• block size 64, accuracy level 1
• FP32 activations, residual stream, normalization, and KV cache
• FP32 sparse token embeddings and sharded audio-code output head
• FP32 CFG and sampling

This is weight-only INT8 with FP32 compute, not an INT8-compute model.

Browser qualification

Desktop Chromium on Apple Metal 3:

Body	Session	150 semantic codes	Unique	Conditional top-20	Unconditional top-20	CFG top-20
INT8 weights / FP32 compute	2.5 s	9.0 s	148	100%	80%	60%
Unquantized FP32 reference	12.9 s	16.0 s	144	100%	80%	55%

The ranking columns compare the first browser step with native unquantized FP32 reference logits. The qualification gate requires at least 50% top-20 overlap for the conditional, unconditional, and CFG branches, finite logits at every step, and completion without semantic-code collapse.

The original FP32 plan also completed downstream ACE-Step XL Turbo DiT/VAE generation and produced intelligible audio. The packaged hybrid planner-to-audio path should still be listening-tested on the target machine; successful model loading alone is not an audio-quality claim.

License

The original ACE-Step project and these conversions are distributed under the MIT license. This repository contains unofficial experimental conversions and is not endorsed by the ACE-Step authors.