Pre-converted, vanilla-PyTorch weights for running Khala — a high-fidelity, unified acoustic-token song generator — on Apple Silicon (MPS) or CPU, with no NVIDIA / Megatron / TransformerEngine / FlashAttention stack.

These are format conversions of the original weights, not a retrain. They are produced by gathering the upstream Megatron torch_dist checkpoint and the DAC-RVQ decoder, then renaming tensors to the de-Megatron KhalaModel layout. The numerics match the original (backbone greedy decode is bit-identical to the CUDA reference, 64/64 tokens).

Files

File	What it is
`khala_backbone.safetensors`	Backbone GPT (q0/q1 coarse acoustic tokens), `KhalaModel` naming
`khala_superres.safetensors`	Super-resolution GPT (expands to q0…q63), non-causal
`decoder_weights.pt`	DAC-RVQ decoder (generator-only), stereo, 64 quantizers
`backbone_megatron_args.json`	Backbone config (consumed by `KhalaConfig.from_megatron_args`)
`superres_megatron_args.json`	Super-res config
`decoder_config.yaml`	Decoder config

Usage

Download into the directory the vanilla runtime reads (_cuda_artifacts/, or set KHALA_VANILLA_WEIGHTS):

hf download Vinpolar/Khala-MusicGeneration-v1.0-MPS --local-dir _cuda_artifacts

Then run the CLI generator or the Mac web stack from the Khala repo (Apple Silicon section of the README):

bash

# one-off track from the command line
KHALA_BACKEND=vanilla .venv-mac/bin/python -u tools/generate_vanilla.py \
    --duration 3 --tags "upbeat, pop, piano"

# or the full web UI (worker + API)
bash backend/run_backend_mac.sh        # --device cpu also supported

Attribution & license

Original model, weights, and research: liujiafeng / the Khala team — liujiafeng/Khala-MusicGeneration-v1.0, paper.
Released under CC BY-NC 4.0, the same license as the original weights. This repository only re-packages those weights for Apple-Silicon inference.
⚠️ The upstream team noted (2026-05-07) a possible inference-quality precision issue under investigation; treat generation quality as not yet final.