Skip to content

stages

LTX-2 family pipeline stages.

Classes

fastvideo.pipelines.basic.ltx2.stages.LTX2AudioDecodingStage 

LTX2AudioDecodingStage(audio_decoder, vocoder)

Bases: PipelineStage

Decode LTX-2 audio latents into a waveform.

Source code in fastvideo/pipelines/basic/ltx2/stages/ltx2_audio_decoding.py 
28
29
30
31
def __init__(self, audio_decoder, vocoder) -> None:
    super().__init__()
    self.audio_decoder = audio_decoder
    self.vocoder = vocoder

fastvideo.pipelines.basic.ltx2.stages.LTX2DenoisingStage 

LTX2DenoisingStage(transformer)

Bases: PipelineStage

Run the LTX-2 denoising loop over the sigma schedule.

Source code in fastvideo/pipelines/basic/ltx2/stages/ltx2_denoising.py 
82
83
84
def __init__(self, transformer) -> None:
    super().__init__()
    self.transformer = transformer

fastvideo.pipelines.basic.ltx2.stages.LTX2LatentPreparationStage 

LTX2LatentPreparationStage(transformer)

Bases: PipelineStage

Prepare initial LTX-2 latents without relying on a diffusers scheduler.

Source code in fastvideo/pipelines/basic/ltx2/stages/ltx2_latent_preparation.py 
25
26
27
def __init__(self, transformer) -> None:
    super().__init__()
    self.transformer = transformer

fastvideo.pipelines.basic.ltx2.stages.LTX2TextEncodingStage 

LTX2TextEncodingStage(text_encoders, tokenizers)

Bases: TextEncodingStage

LTX2 text encoding stage with sequence parallelism support.

When SP is enabled (sp_world_size > 1), only rank 0 runs the text encoder and broadcasts embeddings to other ranks. This avoids I/O contention from all ranks loading the Gemma model simultaneously, which can cause text encoding to take 100+ seconds instead of ~5 seconds.

Source code in fastvideo/pipelines/stages/text_encoding.py 
28
29
30
31
32
33
34
35
36
37
38
39
def __init__(self, text_encoders, tokenizers) -> None:
    """
    Initialize the prompt encoding stage.

    Args:
        enable_logging: Whether to enable logging for this stage.
        is_secondary: Whether this is a secondary text encoder.
    """
    super().__init__()
    self.tokenizers = tokenizers
    self.text_encoders = text_encoders
    self._last_audio_embeds: list[torch.Tensor] | None = None

Modules

fastvideo.pipelines.basic.ltx2.stages.ltx2_audio_decoding

Audio decoding stage for LTX-2 pipelines.

Classes

fastvideo.pipelines.basic.ltx2.stages.ltx2_audio_decoding.LTX2AudioDecodingStage 
LTX2AudioDecodingStage(audio_decoder, vocoder)

Bases: PipelineStage

Decode LTX-2 audio latents into a waveform.

Source code in fastvideo/pipelines/basic/ltx2/stages/ltx2_audio_decoding.py 
28
29
30
31
def __init__(self, audio_decoder, vocoder) -> None:
    super().__init__()
    self.audio_decoder = audio_decoder
    self.vocoder = vocoder

Functions

fastvideo.pipelines.basic.ltx2.stages.ltx2_denoising

LTX-2 denoising stage using the native sigma schedule.

Classes

fastvideo.pipelines.basic.ltx2.stages.ltx2_denoising.LTX2DenoisingStage 
LTX2DenoisingStage(transformer)

Bases: PipelineStage

Run the LTX-2 denoising loop over the sigma schedule.

Source code in fastvideo/pipelines/basic/ltx2/stages/ltx2_denoising.py 
82
83
84
def __init__(self, transformer) -> None:
    super().__init__()
    self.transformer = transformer

Functions

fastvideo.pipelines.basic.ltx2.stages.ltx2_latent_preparation

Latent preparation stage for LTX-2 pipelines.

Classes

fastvideo.pipelines.basic.ltx2.stages.ltx2_latent_preparation.LTX2LatentPreparationStage 
LTX2LatentPreparationStage(transformer)

Bases: PipelineStage

Prepare initial LTX-2 latents without relying on a diffusers scheduler.

Source code in fastvideo/pipelines/basic/ltx2/stages/ltx2_latent_preparation.py 
25
26
27
def __init__(self, transformer) -> None:
    super().__init__()
    self.transformer = transformer

Functions

fastvideo.pipelines.basic.ltx2.stages.ltx2_text_encoding

LTX2-specific text encoding stage with sequence parallelism broadcast support.

When running with sequence parallelism (SP), the Gemma text encoder is only executed on rank 0, and the embeddings are broadcast to all other ranks. This avoids I/O contention from all ranks loading the Gemma model simultaneously.

Classes

fastvideo.pipelines.basic.ltx2.stages.ltx2_text_encoding.LTX2TextEncodingStage 
LTX2TextEncodingStage(text_encoders, tokenizers)

Bases: TextEncodingStage

LTX2 text encoding stage with sequence parallelism support.

When SP is enabled (sp_world_size > 1), only rank 0 runs the text encoder and broadcasts embeddings to other ranks. This avoids I/O contention from all ranks loading the Gemma model simultaneously, which can cause text encoding to take 100+ seconds instead of ~5 seconds.

Source code in fastvideo/pipelines/stages/text_encoding.py 
28
29
30
31
32
33
34
35
36
37
38
39
def __init__(self, text_encoders, tokenizers) -> None:
    """
    Initialize the prompt encoding stage.

    Args:
        enable_logging: Whether to enable logging for this stage.
        is_secondary: Whether this is a secondary text encoder.
    """
    super().__init__()
    self.tokenizers = tokenizers
    self.text_encoders = text_encoders
    self._last_audio_embeds: list[torch.Tensor] | None = None

Functions