moshiko-full-duplex-onnx-catalogue

Moshiko full-duplex ONNX package

Real-weight four-component Kyutai Moshiko speech-to-speech package: Mimi encoder, 7B temporal transformer, 8-step depformer, and Mimi decoder. The LM MatMul and constant Gather weights use ONNX Runtime symmetric INT4 RTN (block size 128, Gather axis -1); Mimi remains in its exported precision. This substitutes for inaccessible gated PersonaPlex while preserving the canonical delayed multi-stream/recurrent full-duplex workflow.

Pinned source

• kyutai/moshiko-pytorch-bf16@2bfc9ae6e89079a5cc7ed2a68436010d91a3d289
• CC BY 4.0, public and ungated.
• Original creator: Kyutai. This derivative converts the checkpoint to ONNX and applies symmetric INT4 RTN quantization to MatMul and constant Gather weights.
• Mobius exporter: onnxruntime/mobius@80e11c45cb47b876bc2b4c6b183ebf38427c121b.
• PersonaPlex substitution and quantization rationale: provenance.json.
• Hashless canonical workflow metadata: inference_metadata.yaml.
• Component hashes and real graph ports: graph_report.json.

Download

hf download justinchuby/moshiko-full-duplex-onnx-catalogue \
  --repo-type model --local-dir moshiko-full-duplex
cd moshiko-full-duplex

Run deterministic CUDA duplex generation

pip install 'onnxruntime-gpu==1.28.0' 'numpy==2.3.3' 'soundfile==0.13.1' \
  'nvidia-cudnn-cu12==9.10.2.21' 'nvidia-cublas-cu12==12.6.4.1' \
  'nvidia-cuda-runtime-cu12==12.6.77'
bash run_cuda.sh --seed 0 --dep-q 8 --frames 5 --save-to reproduced

The checked evidence uses five 80 ms silence frames encoded by the real Mimi encoder. output.npz contains generated text tokens, 8-stream audio codes, and decoded waveform; assistant.wav is the decoded sample. On H200, compact INT4 steady-state LM frame time averaged 19.97 ms versus the 80 ms budget. Exact per-frame timings are in runtime_output.json.

Graph/state contract

The workflow consumes a 24 kHz audio chunk and session ID. Temporal KV caches and the 17-channel delay ring are session scoped; the 8-step depformer cache resets each frame. The output is simultaneously generated text tokens and 24 kHz assistant audio.

<!-- inference-metadata-annotation:start -->

Annotated inference metadata

Review inference_metadata.annotated.yaml for inline explanations of this package's workflow, tensor/state/cache contracts, and fail-closed omissions. inference_metadata.yaml remains the canonical machine-authored contract; automated validation confirms both files parse to the same metadata object.

<!-- inference-metadata-annotation:end -->