ACE-Step-v1.5-XL-Turbo-ONNX-WebGPU

ACE-Step 1.5 XL Turbo ONNX WebGPU

Experimental browser-oriented INT4 and INT8 ONNX exports of the official ACE-Step/acestep-v15-xl-turbo 4.169B diffusion transformer, its checkpoint-specific condition encoder, and the official Oobleck VAE decoder upcast directly from BF16 to FP32.

The q4 files are the standard profile and the q8 files are the high-precision profile of the ai-music-js browser library. Large external data is split for streaming into browser Origin Private File System storage and for ONNX Runtime Web session creation. The JavaScript package, Worker, ONNX Runtime Web WASM runtime, tokenizer, text encoder, lyric embedding, semantic-code detokenizer, and silence conditioning asset are not duplicated in this repository.

This is an unofficial, experimental conversion. It is not affiliated with or endorsed by the ACE-Step authors.

Files

XL-specific download:

• INT4 standard: 3,401,248,219 bytes (3.401 GB / 3.168 GiB)
• INT8 high precision: 5,778,846,562 bytes (5.779 GB / 5.382 GiB)

The DiT data files are initializer-aligned, sub-1 GB chunks. Larger external data files exceeded Chromium's single-buffer path during ONNX Runtime Web session creation. Repacking only changed external-data locations and offsets; it did not change tensor values.

Browser usage

npm install ai-music-js

import { AceStepWebGpu } from "ai-music-js";

const ace = new AceStepWebGpu();

const result = await ace.generate({
  prompt:
    "Warm analog synthwave instrumental, steady electronic drums, pulsing bass, cinematic pads, memorable lead melody, polished studio mix",
  audioQuality: "high",
  seed: 42,
  durationSeconds: 10,
});

const audio = new Audio(URL.createObjectURL(result.wav));
await audio.play();

The URL above is pinned to the verified model-file commit. Use main only when you intentionally want future artifact revisions.

The standard complete browser model set is approximately 5.626 GB. Shared model and support assets are resolved by the npm library from pinned upstream Hugging Face revisions. The intended runtime is current desktop Chrome or Edge with WebGPU. The FP32 VAE runs on strict WebGPU first and restarts all chunks on WASM only when compatibility fallback is enabled and WebGPU fails.

Export configuration

• Official ACE-Step runtime commit: 6d53286b9183b720b2b86f338a06b2a614836871
• Starting ONNX conversion: shreyask/ACE-Step-v1.5-ONNX@bdabfb5684fd70fcc76f98cbb51bb9ebc47ee342
• XL Turbo DiT: 32 layers, hidden size 2560, 4,168,897,088 parameters
• Standard quantization: asymmetric INT4 MatMulNBits, block size 64, accuracy_level=1
• High-precision quantization: asymmetric INT8 MatMulNBits, block size 64, accuracy_level=1
• VAE source: official AutoencoderOobleck BF16 checkpoint weights upcast directly to FP32; dynamic latent length, opset 18
• Batch size: 1
• Inference: instrumental, no planning LM, no reference audio, no CFG
• Sampler: eight shift-3 Euler flow-matching evaluations
• Output: 48 kHz stereo
• Validated duration shapes: 10–120 seconds
• ONNX Runtime native: 1.23.2
• ONNX Runtime Web: 1.27.0

The condition and DiT graphs are fresh checkpoint-specific exports. The unofficial shreyask/ACE-Step-v1.5-ONNX repository supplied the experimental starting point and shared browser assets, but its published Turbo DiT uses the smaller 24-layer, hidden-size-2048 architecture rather than XL.

Numerical validation

The complete condition → eight-step DiT → VAE chains were compared with the official Python implementation using the same prompt, seed 42, learned silence latent, and deterministic browser-compatible noise generator.

The INT8 first DiT velocity cosine was 0.999859 and its mean absolute error was 0.014079. The resulting ten-second waveform passed the numerical parity threshold. The 20-second INT4 run remained finite and non-collapsed, but its divergence compounds with duration and exceeds the chosen parity gate. Successful model loading alone is not evidence of successful music generation. A browser WebGPU listening run of the new INT8 profile remains an acceptance gate.

The corresponding FP32 ONNX exports matched the Python implementation closely:

• condition encoder maximum absolute error: 0.00000182
• condition encoder cosine: effectively 1.0
• first DiT velocity maximum absolute error: 0.00031088
• first DiT velocity cosine: 0.9999999997
• VAE native-ORT/PyTorch cosine: 1.0
• VAE native-ORT/PyTorch maximum absolute difference: 0.00001886

The FP32 VAE was separately qualified in desktop Chromium against the same FP32 graph on pthread/WASM. Strict WebGPU was 8.12 times faster, with waveform cosine 1.0000000000, maximum absolute difference 0.00002072, and mean absolute difference 0.00000049.

WebGPU notes

Both profiles use com.microsoft:MatMulNBits. Static inspection and native ONNX Runtime validation found no known strict-WebGPU blocker in either fresh XL graph. Each INT8 external-data shard is smaller than 1 GB.

The shared upstream Qwen3 graph contains ai.onnx:IsNaN, so the browser pipeline requests WebGPU with WASM compatibility fallback. The former FP16 VAE WebGPU path is rejected because it produced waveform cosine 0.1748364475 against WASM. The FP32 VAE replaces it and is browser-qualified on WebGPU.

Limitations

Unsupported: vocals, planning LM, reference audio, source audio, cover/repaint/lego/extract, audio-code hints, CFG, base/SFT checkpoints, DCW, Heun/SDE samplers, durations beyond 120 seconds, tiled long-form generation, batching, LoRA, mobile, Safari, and Firefox.

Peak browser memory and browser wall time vary substantially by browser, adapter, duration, and cache state and are not represented by the native validation timings.

Attribution

• Original model and implementation: ACE-Step/acestep-v15-xl-turbo
• Experimental ONNX starting point and shared assets: shreyask/ACE-Step-v1.5-ONNX

Review the original repositories and their licenses before redistribution or commercial use.