Skip to content

lingbot_video_pipeline

Stage-composed LingBot-Video Dense and MoE/refiner T2V pipeline.

Classes

fastvideo.pipelines.basic.lingbot_video.lingbot_video_pipeline.LingBotVideoPipeline 

LingBotVideoPipeline(*args, **kwargs)

Bases: LoRAPipeline, ComposedPipelineBase

T2V pipeline with optional released MoE pixel-space refinement.

Source code in fastvideo/pipelines/lora_pipeline.py 
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
def __init__(self, *args, **kwargs) -> None:
    super().__init__(*args, **kwargs)
    self.device = get_local_torch_device()
    self.lora_adapter_paths = {}
    # build list of trainable transformers
    for transformer_name in self.trainable_transformer_names:
        if (transformer_name in self.modules and self.modules[transformer_name] is not None):
            self.trainable_transformer_modules[transformer_name] = (self.modules[transformer_name])
        # check for transformer_2 in case of Wan2.2 MoE or fake_score_transformer_2
        if transformer_name.endswith("_2"):
            raise ValueError(
                f"trainable_transformer_name override in pipelines should not include _2 suffix: {transformer_name}"
            )

        secondary_transformer_name = transformer_name + "_2"
        if (secondary_transformer_name in self.modules and self.modules[secondary_transformer_name] is not None):
            self.trainable_transformer_modules[secondary_transformer_name] = self.modules[
                secondary_transformer_name]

    logger.info(
        "trainable_transformer_modules: %s",
        self.trainable_transformer_modules.keys(),
    )

    for (
            transformer_name,
            transformer_module,
    ) in self.trainable_transformer_modules.items():
        self.exclude_lora_layers[transformer_name] = (transformer_module.config.arch_config.exclude_lora_layers)
    self.lora_target_modules = self.fastvideo_args.lora_target_modules
    self.lora_path = self.fastvideo_args.lora_path
    self.lora_nickname = self.fastvideo_args.lora_nickname
    self.training_mode = self.fastvideo_args.training_mode
    if self.training_mode and getattr(self.fastvideo_args, "lora_training", False):
        assert isinstance(self.fastvideo_args, TrainingArgs)
        if self.fastvideo_args.lora_alpha is None:
            self.fastvideo_args.lora_alpha = self.fastvideo_args.lora_rank
        self.lora_rank = self.fastvideo_args.lora_rank  # type: ignore
        self.lora_alpha = self.fastvideo_args.lora_alpha  # type: ignore
        logger.info(
            "Using LoRA training with rank %d and alpha %d",
            self.lora_rank,
            self.lora_alpha,
        )
        if self.lora_target_modules is None:
            self.lora_target_modules = [
                "q_proj",
                "k_proj",
                "v_proj",
                "o_proj",
                "to_q",
                "to_k",
                "to_v",
                "to_out",
                "to_qkv",
                "to_gate_compress",
            ]
            logger.info(
                "Using default lora_target_modules for all transformers: %s",
                self.lora_target_modules,
            )
        else:
            logger.warning(
                "Using custom lora_target_modules for all transformers, which may not be intended: %s",
                self.lora_target_modules,
            )

        self.convert_to_lora_layers()
    # Inference
    elif not self.training_mode and self.lora_path is not None:
        self.convert_to_lora_layers()
        self.set_lora_adapter(
            self.lora_nickname,  # type: ignore
            self.lora_path,
        )  # type: ignore

Methods:

fastvideo.pipelines.basic.lingbot_video.lingbot_video_pipeline.LingBotVideoPipeline.create_pipeline_stages 
create_pipeline_stages(fastvideo_args: FastVideoArgs) -> None

Create base generation and the optional decoded-video refiner stages.

Source code in fastvideo/pipelines/basic/lingbot_video/lingbot_video_pipeline.py 
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
def create_pipeline_stages(self, fastvideo_args: FastVideoArgs) -> None:
    """Create base generation and the optional decoded-video refiner stages."""
    refiner = self.get_module("transformer_2")
    self.add_stage(
        "input_validation_stage",
        LingBotVideoInputValidationStage(refiner_enabled=refiner is not None),
    )
    self.add_stage(
        "prompt_encoding_stage",
        TextEncodingStage(
            text_encoders=[self.get_module("text_encoder")],
            tokenizers=[self.get_module("tokenizer")],
        ),
    )
    self.add_stage("conditioning_stage", ConditioningStage())
    self.add_stage(
        "timestep_preparation_stage",
        TimestepPreparationStage(scheduler=self.get_module("scheduler")),
    )
    self.add_stage(
        "latent_preparation_stage",
        LingBotVideoLatentPreparationStage(transformer=self.get_module("transformer")),
    )
    self.add_stage(
        "denoising_stage",
        LingBotVideoDenoisingStage(
            transformer=self.get_module("transformer"),
            scheduler=self.get_module("scheduler"),
        ),
    )
    self.add_stage(
        "decoding_stage",
        DecodingStage(vae=self.get_module("vae"), pipeline=self),
    )
    if refiner is not None:
        self.add_stage(
            "refiner_preparation_stage",
            LingBotVideoRefinerPreparationStage(
                vae=self.get_module("vae"),
                scheduler=self.get_module("scheduler"),
            ),
        )
        self.add_stage(
            "refiner_denoising_stage",
            LingBotVideoDenoisingStage(
                transformer=refiner,
                scheduler=self.get_module("scheduler"),
                refiner=True,
            ),
        )
        self.add_stage(
            "refiner_decoding_stage",
            DecodingStage(vae=self.get_module("vae"), pipeline=self),
        )
fastvideo.pipelines.basic.lingbot_video.lingbot_video_pipeline.LingBotVideoPipeline.initialize_pipeline 
initialize_pipeline(fastvideo_args: FastVideoArgs) -> None

Apply the released runtime flow shift to the loaded scheduler.

Source code in fastvideo/pipelines/basic/lingbot_video/lingbot_video_pipeline.py 
43
44
45
46
47
48
def initialize_pipeline(self, fastvideo_args: FastVideoArgs) -> None:
    """Apply the released runtime flow shift to the loaded scheduler."""
    shift = fastvideo_args.pipeline_config.flow_shift
    if shift is None:
        raise ValueError("LingBot-Video requires a flow shift")
    self.get_module("scheduler").set_shift(float(shift))
fastvideo.pipelines.basic.lingbot_video.lingbot_video_pipeline.LingBotVideoPipeline.load_modules 
load_modules(fastvideo_args: FastVideoArgs, loaded_modules: dict[str, Any] | None = None) -> dict[str, Any]

Load the optional refiner DiT and the VAE encoder only when declared.

Source code in fastvideo/pipelines/basic/lingbot_video/lingbot_video_pipeline.py 
28
29
30
31
32
33
34
35
36
37
38
39
40
41
def load_modules(
    self,
    fastvideo_args: FastVideoArgs,
    loaded_modules: dict[str, Any] | None = None,
) -> dict[str, Any]:
    """Load the optional refiner DiT and the VAE encoder only when declared."""
    model_index = self._load_config(self.model_path)
    required = list(type(self)._required_config_modules)
    load_refiner = "transformer_2" in model_index and getattr(fastvideo_args, "refine_enabled", None) is not False
    if load_refiner:
        required.append("transformer_2")
        fastvideo_args.pipeline_config.vae_config.load_encoder = True
    self._required_config_modules = required
    return super().load_modules(fastvideo_args, loaded_modules)