Moonshine Streaming Model(s) Recipe

[nenad1002]

No description provided.

[Copilot]

Pull request overview

Adds a new CPU recipe under usefulSensors-moonshine-streaming/ to export the Useful Sensors Moonshine Streaming ASR model (tiny/small) into the five ONNX Runtime GenAI streaming components, with an end-to-end pipeline script and optional INT8 k-quant for encoder/decoder_kv.

Changes:

• Introduces an end-to-end exporter pipeline (cpu/optimize.py) that runs five Olive conversions, generates GenAI runtime configs, exports tokenizer files, and downloads Silero VAD.
• Adds model wrapper modules + dummy inputs for exporting the five-component streaming graph contract (cpu/moonshine_model_load.py), plus a standalone non-Olive exporter for debugging (cpu/export_moonshine_streaming.py).
• Adds Olive JSON configs for FP32 and optional k-quant flows, along with documentation and metadata for the new recipe.

Reviewed changes

Copilot reviewed 13 out of 14 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
usefulSensors-moonshine-streaming/cpu/requirements.txt	Declares Python dependencies needed to run the CPU Moonshine streaming export pipeline.
usefulSensors-moonshine-streaming/cpu/README.md	Documents the recipe purpose, setup, run commands, quantization options, and expected outputs.
usefulSensors-moonshine-streaming/cpu/optimize.py	Implements the full pipeline: Olive runs per component, config generation, tokenizer export, and VAD download.
usefulSensors-moonshine-streaming/cpu/moonshine_model_load.py	Provides model loaders and thin wrapper modules used by Olive/standalone export to match the GenAI streaming I/O contract.
usefulSensors-moonshine-streaming/cpu/moonshine_frontend_fp32_cpu.json	Olive config to export the frontend component to ONNX (FP32).
usefulSensors-moonshine-streaming/cpu/moonshine_encoder_fp32_cpu.json	Olive config to export the encoder component to ONNX (FP32).
usefulSensors-moonshine-streaming/cpu/moonshine_adapter_fp32_cpu.json	Olive config to export the adapter component to ONNX (FP32).
usefulSensors-moonshine-streaming/cpu/moonshine_cross_kv_fp32_cpu.json	Olive config to export the cross-KV component to ONNX (FP32).
usefulSensors-moonshine-streaming/cpu/moonshine_decoder_kv_fp32_cpu.json	Olive config to export the decoder-KV component to ONNX (FP32).
usefulSensors-moonshine-streaming/cpu/moonshine_encoder_kquant8_cpu.json	Olive config to export+quantize the encoder using INT8 k-quant.
usefulSensors-moonshine-streaming/cpu/moonshine_decoder_kv_kquant8_cpu.json	Olive config to export+quantize the decoder-KV using INT8 k-quant.
usefulSensors-moonshine-streaming/cpu/info.yaml	Registers recipe metadata and lists produced artifacts/scripts.
usefulSensors-moonshine-streaming/cpu/export_moonshine_streaming.py	Standalone TorchDynamo ONNX exporter for debugging outside Olive.
usefulSensors-moonshine-streaming/cpu/init.py	Marks cpu/ as a Python package for imports used by the recipe.

Comments suppressed due to low confidence (1)

usefulSensors-moonshine-streaming/cpu/moonshine_model_load.py:115

• sample_buffer and sample_len are accepted as inputs but never used. This means any buffered sub-frame audio passed in (e.g., a flush call after a non-multiple-of-frame_len chunk) would be silently dropped, contradicting the module’s stateful interface. If the runtime contract guarantees sample_len == 0, it’s safer to enforce that explicitly to avoid producing incorrect features silently.

        # In the genai streaming pipeline chunk_samples is a multiple of
        # frame_len, so the carried sample_buffer is always empty on input
        # (sample_len == 0).  We therefore frame the chunk directly.  The
        # correct leftover state is still emitted so a final (flush) chunk of
        # non-multiple length threads/records its remainder before reset.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 13 out of 14 changed files in this pull request and generated 2 comments.