Code2 Wav Runner#

class Code2WavRunner#

Runner for Qwen3-Omni Code2Wav vocoder.

This class handles Code2Wav vocoder inference, converting RVQ codec codes to audio waveform. It follows the same pattern as AudioRunner and QwenViTRunner for consistency.

Public Functions

Code2WavRunner(
std::string const &engineDir,
cudaStream_t stream,
std::string const &checkpointDir = ""
)#

Constructor for Code2WavRunner.

Parameters:

  • engineDir – [in] Directory containing the Code2Wav engine

  • stream – [in] CUDA stream for execution

  • checkpointDir – [in] Original checkpoint for runtime weight inputs

~Code2WavRunner() noexcept = default#
bool generateWaveform(
std::vector<std::vector<int32_t>> const &codes,
rt::audioUtils::AudioData &outputAudio,
cudaStream_t stream
)#

Generate audio waveform from RVQ codes (single sample)

Parameters:

  • codes – [in] RVQ codec codes [numQuantizers][seqLen]

  • outputAudio – [out] Output audio waveform data

  • stream – [in] CUDA stream for execution

Returns:

True if generation succeeded, false otherwise

inline Code2WavConfig const &getConfig() const#

Get Code2Wav configuration.

Returns:

Reference to Code2Wav config

inline int64_t getExpectedWaveformLength(int64_t codeLen) const#

Get expected waveform length for given code length.

Parameters:

codeLen – [in] Number of codec frames

Returns:

Expected waveform length in samples

struct Code2WavConfig#

Configuration for Qwen3-Omni Code2Wav vocoder All values initialized to 0 and must be read from config.json. Do NOT use hardcoded defaults that may not match the actual model.

Public Members

int32_t numQuantizers = {0}#

Number of RVQ quantizer layers.

int32_t codebookSize = {0}#

Codebook vocabulary size per layer.

int32_t hiddenSize = {0}#

Code2Wav hidden dimension.

int32_t decoderDim = {0}#

Decoder base dimension.

int32_t upsampleRate = {0}#

Total upsampling rate (codes → samples)

int32_t sampleRate = {24000}#

Output audio sample rate (24kHz for Qwen3-Omni)

int32_t chunkSize = {300}#

Chunk size for long sequences (in codec frames)

int32_t leftContextSize = {25}#

Overlap size to avoid boundary artifacts.