Initial DiT and VAE changes for LTX inference #783
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Collaborator
|
Could you run pre-commit on this PR? |
| ) | ||
|
|
||
| generator = VideoGenerator.from_pretrained( | ||
| model_path="data/Lightricks/LTX-Video", |
Collaborator
There was a problem hiding this comment.
VideoGenerator.from_pretrained() should download the model for you, meaning directly passing Lightricks/LTX-Video as the model_path should work. Could you test and simply this example accordingly? thanks
Comment on lines
+30
to
+44
| # TODO: fix all of the configs so it's exact match | ||
|
|
||
| # # Text encoder configuration | ||
| # text_encoder_configs: tuple[EncoderConfig, ...] = field( | ||
| # # todo: set max length later | ||
| # #def ltx_t5_config(): | ||
| # # config = T5Config() | ||
| # # config.tokenizer_kwargs["max_length"] = 128 | ||
| # # return config | ||
|
|
||
| # # @dataclass | ||
| # # class LTXConfig(PipelineConfig): | ||
| # # text_encoder_configs: tuple[EncoderConfig, ...] = field( | ||
| # # default_factory=lambda: (ltx_t5_config(), )) | ||
| # default_factory=lambda: (T5Config(), )) |
Collaborator
There was a problem hiding this comment.
Are these not needed? if so please remove
Comment on lines
+104
to
+120
| # TODO: load differently for each config | ||
| # Text-to-Video: Only needs the decoder (to decode latents to video) | ||
| # Image-to-Video: Needs both encoder (to encode input image) and decoder | ||
| # @dataclass | ||
| # class LTXT2VConfig(LTXConfig): | ||
| # def __post_init__(self): | ||
| # super().__post_init__() | ||
| # self.vae_config.load_encoder = False | ||
| # self.vae_config.load_decoder = True | ||
|
|
||
| # @dataclass | ||
| # class LTXI2VConfig(LTXConfig): | ||
| # def __post_init__(self): | ||
| # super().__post_init__() | ||
| # self.vae_config.load_encoder = True | ||
| # self.vae_config.load_decoder = True | ||
|
|
Collaborator
There was a problem hiding this comment.
Also here, please clean up comments
Comment on lines
+8
to
+14
| class LTXSamplingParam(SamplingParam): | ||
| # Video parameters | ||
| height: int = 512 | ||
| width: int = 704 | ||
|
|
||
| # Most defaults set in pipeline config | ||
| num_inference_steps: int = 50 |
Collaborator
There was a problem hiding this comment.
What's the default number of frames that the official repo generates? Could you add it here as well?
| from diffusers.utils.torch_utils import maybe_allow_in_graph | ||
| # from ..attention import FeedForward | ||
| from fastvideo.attention import DistributedAttention, LocalAttention | ||
| #from diffusers.attention_processor import Attention |
Collaborator
There was a problem hiding this comment.
clean up comments please
Comment on lines
+46
to
+51
| # Add ImageVAEEncodingStage for I2V (conditional based on input) | ||
| # Before LatentPreparation for I2V | ||
| # if fastvideo_args.pipeline_config.ltx_i2v_mode: | ||
| # self.add_stage( | ||
| # stage_name="image_vae_encoding_stage", | ||
| # stage=LTXImageVAEEncodingStage(vae=self.get_module("vae"))) |
Collaborator
There was a problem hiding this comment.
remove if not needed
17 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
DiT and VAE changes for LTX inference pipeline, will add more optimizations in upcoming PRs including full pipeline