For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python class

AudioGenerationPipeline

AudioGenerationPipeline

class max.pipelines.audio.AudioGenerationPipeline(pipeline_config, pipeline_model)

source

Bases: Pipeline[AudioGenerationInputs[AudioGenerationContextType], GenerationOutput], Generic[AudioGenerationContextType]

Audio generation pipeline, driven by an AudioExecutor.

Unlike text generation, one request is one call: the executor runs every stage the model has – autoregressive, denoising, vocoding – and returns a finished waveform. The pipeline’s own job is only to hand the batch to the executor and turn its samples into a response.

Parameters:

• pipeline_config (PipelineConfig) – Configuration for the pipeline and runtime behavior.
• pipeline_model (type[AudioExecutor[Any, Any, Any]]) – The audio executor class to instantiate.

execute()

execute(inputs)

source

Generates one waveform per request in the batch.

Parameters:

inputs (AudioGenerationInputs[AudioGenerationContextType])

Return type:

dict[RequestID, GenerationOutput]

max_batch_size

property max_batch_size: int

source

audio generation runs one request at a time.

Type:

Returns 1

pipeline_config

property pipeline_config: PipelineConfig

source

Returns the pipeline configuration.

prepare_batch()

prepare_batch(batch)

source

Delegates input preparation to the executor.

Parameters:

batch (dict[RequestID, AudioGenerationContextType]) – Maps each request ID to its AudioContext.

Returns:

The executor’s inputs, and the batch flattened to (request_id, context) pairs in the order the executor sees them, which is the order its waveforms come back in.

Raises:

ValueError – If the batch holds more than one request.

Return type:

tuple[Any, list[tuple[RequestID, AudioGenerationContextType]]]

release()

release(request_id)

source

Releases resources held for a request.

Nothing to do: an audio request holds no state between calls, since each one is generated in a single execute().

Parameters:

request_id (RequestID)

Return type:

None