Commit History
-
Add MLX base-model DiT weights for LoRA training
Verified signaturedit_{sm-music,sm-sfx,medium}-base_f16.npz — the *-base (rectified_flow) DiT + baked conditioner in MLX layout, for LoRA training via optimized/mlx lora_train_mlx.py (registered in weights.TRAINING_BASE for lazy download).be6b035
-
Add MLX weights for SA3 inference
“Cortexelus”Eight FP16/FP32 npz files for pure-MLX Stable Audio 3 inference on Apple Silicon (~8.4 GB total). - dit_sm-music_f16.npz / dit_sm-sfx_f16.npz / dit_medium_f16.npz DiT diffusion transformers (FP16), each with its conditioner baked under a "cond." key prefix. - same_s_encoder_f32.npz / same_s_decoder_f32.npz (FP32) Shared codec for the two small DiTs. - same_l_encoder_f32.npz / same_l_decoder_f32.npz (FP32) Codec paired with the medium DiT. - t5gemma_f16.npz Text encoder (FP16) with SentencePiece tokenizer bytes embedded.2879a89 -
Create MLX/README.md
Verified signaturebcb907f