ales27pm/musicmaker-t5-base-coreml

T5-base masked FP32 Core ML encoder

This repository contains a Core ML conversion of the encoder from google-t5/t5-base at the exact source revision:

a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1

The conversion keeps the explicit T5 attention mask and zeros output positions whose mask value is zero. It uses a fixed 64-token interface for local iOS and macOS inference.

Artifact

StableAudioT5EncoderMaskedFP32.mlpackage.zip

• Inputs:
  • input_ids: Int32 [1, 64]
  • attention_mask: Int32 [1, 64], with 1 for valid tokens
• Output:
  • text_embeddings: Float32 [1, 64, 768]
• Storage precision: Float32
• Minimum deployment target: iOS 17 / macOS 14
• Size: 228,196,156 bytes
• SHA-256: 66f69dca44056cfa0b2779a70ffe406d0b31215ff7ebb0161a3b469051ac1975

The artifact filename reflects the downstream interface for which it was prepared. The repository contains only Google T5-base-derived weights. It does not contain Stable Audio, Stability AI, DiT, autoencoder, or other audio-model weights.

Conversion

convert_masked_t5.py downloads only the pinned public T5-base revision, exports the masked encoder with Core ML Tools, and compares Core ML output with the PyTorch reference on valid and padded token positions.

The published package records torch==2.2.2 and coremltools==9.0 in its Core ML metadata. Run conversion on macOS with Python 3.11 and compatible transformers, numpy, and sentencepiece releases.

License

This T5-only derivative is provided under the Apache License 2.0, matching the upstream Google T5-base model. Users remain responsible for reviewing the upstream model card and license.