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

AudioContext

AudioContext

class max.pipelines.context.AudioContext(*, tokens, request_id=<factory>, model_name='', negative_tokens=None, audio_duration, num_inference_steps, guidance_scale=None, seed=None, audio_format='wav', waveform=None, status=GenerationStatus.ACTIVE)

source

Bases: object

A model-ready context for audio generation requests.

Like PixelContext, this carries only what the model executes against: the caller’s text has already been through the tokenizer, and what is left is token ids plus the numbers that size the generation.

Parameters:

• tokens (TokenBuffer) – The conditional prompt’s token ids.
• request_id (RequestID) – A unique identifier for this generation request.
• model_name (str) – Name of the model being used.
• negative_tokens (TokenBuffer | None) – The unconditional prompt’s token ids, for models that generate with classifier-free guidance.
• audio_duration (float) – Upper bound on the generated audio, in seconds. A model may stop earlier, so the waveform’s own length is the authority on what was produced. Required, because what a request leaves unset is the checkpoint’s default rather than any value this framework could pick, and the tokenizer resolves it.
• num_inference_steps (int) – Denoising steps, for models whose audio comes from a diffusion or flow-matching stage. Required, for the same reason as audio_duration.
• guidance_scale (float | None) – Classifier-free guidance scale, or None to leave it to the model. Audio models often bake distinct scales into distinct stages, and a single request-level number cannot say which one it means.
• seed (int | None) – RNG seed for the sampling the model does.
• audio_format (str) – Container the waveform is encoded into for the response.
• waveform (ndarray[tuple[Any, ...], dtype[float32]] | None) – The generated samples, once there are any.
• status (GenerationStatus)

audio_duration

audio_duration: float

source

audio_format

audio_format: str = 'wav'

source

compute_num_available_steps()

compute_num_available_steps(max_seq_len)

source

Returns the denoising steps, for scheduler compatibility.

Parameters:

max_seq_len (int)

Return type:

int

guidance_scale

guidance_scale: float | None = None

source

is_done

property is_done: bool

source

Whether the request has completed generation.

model_name

model_name: str = ''

source

negative_tokens

negative_tokens: TokenBuffer | None = None

source

Unconditional prompt tokens. None for models that do not guide.

num_inference_steps

num_inference_steps: int

source

request_id

request_id: RequestID

source

reset()

reset()

source

Resets the context’s state.

Return type:

None

seed

seed: int | None = None

source

status

status: GenerationStatus = 'active'

source

tokens

tokens: TokenBuffer

source

Conditional prompt tokens.

update()

update(waveform)

source

Stores the generated samples on the context.

Parameters:

waveform (ndarray[tuple[Any, ...], dtype[float32]])

Return type:

None

waveform

waveform: ndarray[tuple[Any, ...], dtype[float32]] | None = None

source

Generated samples in (channels, samples) layout, in [-1, 1].