Skip to main content

/

DocsCode

Python module

core

TTSContext​

class max.pipelines.core.TTSContext(*, max_length, tokens, request_id=<factory>, eos_token_ids=<factory>, eos_sequences=<factory>, log_probabilities=0, log_probabilities_echo=False, ignore_eos=False, json_schema=None, sampling_params=<factory>, model_name='', _matcher=None, status=GenerationStatus.ACTIVE, _log_probabilities_data=<factory>, _is_initial_prompt=True, _draft_offset=0, target_endpoint=None, audio_prompt_tokens=<factory>, buffer_speech_tokens=None, audio_buffer=None, prev_samples_beyond_offset=0, streaming=False, _speech_token_size=128, _speech_token_end_idx=0, _speech_tokens=<factory>, decoded_index=0, _block_counter=0, _arrival_time=<factory>, audio_generation_status=GenerationStatus.ACTIVE)

A context for Text-to-Speech (TTS) model inference.

This class extends TextContext to handle speech token generation and management. It maintains buffers for audio prompt tokens and generated speech tokens, along with tracking indices for decoding progress.

Parameters:

audio_buffer​

audio_buffer: ndarray[tuple[Any, ...], dtype[floating[Any]]] | None = None

audio_generation_status​

audio_generation_status: GenerationStatus = 'active'

audio_prompt_tokens​

audio_prompt_tokens: ndarray[tuple[Any, ...], dtype[integer[Any]]]

block_counter​

property block_counter: int

buffer_speech_tokens​

buffer_speech_tokens: ndarray[tuple[Any, ...], dtype[integer[Any]]] | None = None

decoded_index​

decoded_index: int = 0

is_done​

property is_done: bool

next_speech_tokens()​

next_speech_tokens(audio_chunk_size=None, buffer=None)

Returns a chunk of the next unseen speech tokens.

Calling this function will not update the index of the last seen token. This must be done by setting decoded_index after the chunk is processed.

Parameters:

  • audio_chunk_size (int | None) – The number of speech tokens to return.
  • buffer (int | None) – The number of previous speech tokens to pass to the audio decoder on each generation step.

Returns:

A tuple of (chunk of speech tokens, buffer).

Return type:

tuple[ndarray[tuple[Any, …], dtype[integer[Any]]], int]

prev_samples_beyond_offset​

prev_samples_beyond_offset: int = 0

speech_tokens​

property speech_tokens: ndarray[tuple[Any, ...], dtype[integer[Any]]]

streaming​

streaming: bool = False

update_speech_tokens()​

update_speech_tokens(new_tokens)

Updates the next_tokens

Parameters:

new_tokens (ndarray[tuple[Any, ...], dtype[integer[Any]]])

Return type:

None

TextAndVisionContext​

class max.pipelines.core.TextAndVisionContext(*, max_length, tokens, request_id=<factory>, eos_token_ids=<factory>, eos_sequences=<factory>, log_probabilities=0, log_probabilities_echo=False, ignore_eos=False, json_schema=None, sampling_params=<factory>, model_name='', _matcher=None, status=GenerationStatus.ACTIVE, _log_probabilities_data=<factory>, _is_initial_prompt=True, _draft_offset=0, target_endpoint=None, vision_token_ids, images=<factory>, extra_model_args=<factory>)

A base class for model context, specifically for Vision model variants.

For example:

- <vision_start_token_id> = 97
- <vision_token_id> = 98
- <vision_end_token_id> = 99

Token array:

-       idx: [  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 ]
- token_ids: [ 51 52 53 54 97 98 98 98 98 99 55 56 57 58 97 98 98 98 98 99 59 60 61 62 ]
                              ^-- img0 --^                  ^-- img1 --^
                                                 ^ start_idx=11 (image_idx=1)

Then we would have:

- ImageMetadata(start_idx=5, end_idx=9, ...)  # img0
- ImageMetadata(start_idx=15, end_idx=19, ...)  # img1

These image ranges should be non-overlapping.

The image_idx is determined based on the value of start_idx. It is the idx of the first image that is not yet encoded. For example in the above diagram when start_idx=11, this implies that image_idx=1.

Currently we restrict start_idx and current_position from being in the middle of an image! This is verified in _validate_state methods that are called before and after mutating methods like _bump_token_indices.

Parameters:

compute_image_aligned_idx()​

compute_image_aligned_idx(idx)

Possibly aligns a index value downward if it lies in the middle of an image.

Parameters:

idx (int)

Return type:

int

extra_model_args​

extra_model_args: dict[str, ndarray[tuple[Any, ...], dtype[Any]]]

Extra model arguments for the vision model. These are model specific arguments.

image_idx​

property image_idx: int

Index of the next unencoded image in the prompt.

images​

images: list[ImageMetadata]

Metadata about each image in the prompt.

needs_vision_encoding​

property needs_vision_encoding: bool

Returns whether vision encoding is needed for this context.

next_images​

property next_images: list[ImageMetadata]

Returns the images that are not yet encoded.

update()​

update(new_token, log_probabilities=None)

Updates the next_tokens and extends existing tokens to include all generated tokens.

Parameters:

Return type:

None

vision_token_ids​

vision_token_ids: list[int]

The value of the <vision_token_id> special token. The reason this is a list is primarily due to Pixtral which also has a image_break_token_id.

TextContext​

class max.pipelines.core.TextContext(*, max_length, tokens, request_id=<factory>, eos_token_ids=<factory>, eos_sequences=<factory>, log_probabilities=0, log_probabilities_echo=False, ignore_eos=False, json_schema=None, sampling_params=<factory>, model_name='', _matcher=None, status=GenerationStatus.ACTIVE, _log_probabilities_data=<factory>, _is_initial_prompt=True, _draft_offset=0, target_endpoint=None)

A base class for model context, specifically for Text model variants.

This class manages the state and processing of text generation, including token management, caching, and generation parameters.

Parameters:

  • max_length (int) – Maximum allowed length of the generated sequence
  • tokens (TokenBuffer) – NumPy array containing the token IDs
  • request_id (RequestID) – A unique identifier for this sequence.
  • eos_token_ids (set[int]) – Set of token IDs that indicate end of sequence
  • eos_sequences (list[list[int]])
  • log_probabilities (int) – Whether to return token log probabilities
  • log_probabilities_echo (bool) – Whether to return log probabilities for prompt tokens
  • ignore_eos (bool) – Whether to ignore end of sequence tokens and continue generating
  • json_schema (str | None) – Optional JSON schema for structured output
  • sampling_params (SamplingParams) – Parameters controlling the token sampling strategy
  • model_name (str)
  • _matcher (Any | None)
  • status (GenerationStatus)
  • _log_probabilities_data (dict[int, LogProbabilities]) – Token log probabilities data
  • _is_initial_prompt (bool) – Whether this is the initial prompt encoding
  • _draft_offset (int) – Offset for draft decoding
  • target_endpoint (str | None) – Optional target endpoint identifier for routing requests

apply_processing_offset()​

apply_processing_offset(offset)

Parameters:

offset (int)

Return type:

None

compute_num_available_steps()​

compute_num_available_steps(max_seq_len)

Compute the max number of steps we can execute for a given context without exceeding the max_seq_len.

Parameters:

max_seq_len (int)

Return type:

int

eos_sequences​

eos_sequences: list[list[int]]

eos_token_ids​

eos_token_ids: set[int]

get_min_token_logit_mask()​

get_min_token_logit_mask(num_steps)

Returns a set of indices for the tokens in the output that should be masked.

This is primarily used for the min_tokens setting, where we mask eos tokens in the logits to avoid generating them before we reach min_tokens.

Returns:

A set of indices for the tokens in the output that should be masked.

Parameters:

num_steps (int)

Return type:

list[ndarray[tuple[Any, …], dtype[int32]]]

ignore_eos​

ignore_eos: bool = False

is_done​

property is_done: bool

is_initial_prompt​

property is_initial_prompt: bool

Returns true if the context has not been updated with tokens.

json_schema​

json_schema: str | None = None

jump_ahead()​

jump_ahead(new_token)

Updates the token array, while ensuring the new token is returned to the user.

Parameters:

new_token (int)

Return type:

None

log_probabilities​

log_probabilities: int = 0

log_probabilities_echo​

log_probabilities_echo: bool = False

matcher​

property matcher: LLMatcher | None

max_length​

max_length: int

min_tokens​

property min_tokens: int

The minimum number of new tokens to generate.

model_name​

model_name: str = ''

request_id​

request_id: RequestID

reset()​

reset()

Resets the context’s state by combining all tokens into a new prompt.

Return type:

None

sampling_params​

sampling_params: SamplingParams

set_matcher()​

set_matcher(matcher)

Parameters:

matcher (LLMatcher)

Return type:

None

status​

status: GenerationStatus = 'active'

target_endpoint​

target_endpoint: str | None = None

to_generation_output()​

to_generation_output()

Get completion tokens that are ready to be returned to the user.

This method retrieves tokens that have been generated but not yet delivered to the user, along with their associated log probability data.

Returns:

The completion tokens and their associated log probabilities, if available.

Return type:

TextGenerationOutput

tokens​

tokens: TokenBuffer

update()​

update(new_token, log_probabilities=None)

Updates the next_tokens and extends existing tokens to include all generated tokens.

Parameters:

Return type:

None

reserve_token_space_for_batch()​

max.pipelines.core.reserve_token_space_for_batch(batch, num_tokens)

Temporarily reserves token space for each context in a batch by incrementing the _active_idx and _end_idx attributes by num_tokens for the duration of the context. These indices are restored to their original values upon exit. :param batch: List of TextContext objects to reserve space for. :param num_tokens: Number of tokens to reserve for each context.

Yields:

None

Parameters:

Return type:

Iterator[None]

validate_aspect_ratio_args()​

max.pipelines.core.validate_aspect_ratio_args(context)

Validates that required aspect ratio arguments are present for vision input.

Parameters:

context (TextContext | TextAndVisionContext) – The context to validate.

Raises:

InputError – If required aspect ratio arguments are missing.

Return type:

None

validate_image_grid_thw_args()​

max.pipelines.core.validate_image_grid_thw_args(context)

Validates that image_grid_thw is present when vision encoding is needed.

Parameters:

context (TextContext | TextAndVisionContext) – The context to validate.

Raises:

InputError – If image_grid_thw is missing from extra_model_args when vision encoding is needed.

Return type:

None

validate_image_shape_5d()​

max.pipelines.core.validate_image_shape_5d(context)

Validates that images have the expected 5-dimensional shape.

Parameters:

context (TextContext | TextAndVisionContext) – The context to validate.

Raises:

InputError – If the image shape is not 5-dimensional.

Return type:

None

validate_initial_prompt_has_image()​

max.pipelines.core.validate_initial_prompt_has_image(context)

Validates that initial prompts contain an image for vision models.

Parameters:

context (TextContext | TextAndVisionContext) – The context to validate.

Raises:

InputError – If the initial prompt doesn’t contain an image.

Return type:

None

validate_only_one_image()​

max.pipelines.core.validate_only_one_image(context)

Validates that at most one image is provided in the context.

Parameters:

context (TextContext | TextAndVisionContext) – The context to validate.

Raises:

InputError – If more than one image is provided.

Return type:

None

validate_requires_vision_context()​

max.pipelines.core.validate_requires_vision_context(context)

Validates that the context is a TextAndVisionContext.

Parameters:

context (TextContext | TextAndVisionContext) – The context to validate.

Raises:

InputError – If the context is not a TextAndVisionContext.

Return type:

None

validate_vision_position_ids()​

max.pipelines.core.validate_vision_position_ids(context)

Validates that vision_position_ids is present when vision encoding is needed.

Parameters:

context (TextContext | TextAndVisionContext) – The context to validate.

Raises:

InputError – If vision_position_ids is missing from extra_model_args when vision encoding is needed.

Return type:

None

View source

View source

Was this page helpful?

Models

Cookbook

Puzzles

Blog

Community

Contact