snac-sweep

Benchmark sweep for SNAC's ONNX decoder inference: varies sessions (parallel sessions), chunk size (sequence length without context), and batch size to find the largest throughput (sessions × batch) under real-time factor (RTF) < 1.

Install

cargo install snac-sweep

Usage

snac-sweep --context-size <N> --max-chunk-size <M> --model <onnx-file>

• context-size (default 3): number of context windows added to sequence length.
• max-chunk-size (default 3): max chunk size to sweep (seq_len from 1..=M).
• model (default decoder_model_fp16.onnx): model file under https://huggingface.co/onnx-community/snac_24khz-ONNX/tree/main/onnx.

RTF calculation is based on 24kHz sample rate (each chunk is 2048 samples).

Example

snac-sweep --context-size 3 --max-chunk-size 5 --model decoder_model_fp16.onnx