mirror of
https://github.com/huggingface/diffusers.git
synced 2026-10-24 20:59:27 +03:00
284419b352
* Cosmos3 ModularPipeline initial commit * Fix from_pretrained for modular pipleine without modular_model_index.json * Isolate text prep inside Cosmos3TextEncoderStep; add Cosmos3ActionTextStep * Cosmos3 modular encode/decode flow * Drop device/dtype required InputParams; use components._execution_device and model dtype * Add lazy pipeline properties and requires_safety_checker-based safety gating * Drop callback_on_step_end from Cosmos3 denoise loop * Move container/auto blocks into modular_blocks * Move action_gen check, remove unused params * Declare video_processor as ComponentSpec; drop unused pipeline state * Split VAE encode block by modality (image and video) * Remove Cosmos3AssembleResultStep; add Cosmos3ExportStep * Remove unused action_mode * Refactor Cosmos3 modular denoising by modality * Document Cosmos3 modular output selection * Address Cosmos3 modular review feedback * Remaining local-variable device cleanup * Address Cosmos3 modular review feedback * Address final Cosmos3 modular feedback --------- Co-authored-by: Atharva Joshi <atjoshi@nvidia.com>