0.2.7

Fixes

• Fixed broken prompt conditioning in both in-process and subprocess modes.

  • The HuggingFace/RoBERTa cache patch was pointing at the wrong cache level.
  • In bad cache states, RoBERTa could load an invalid tokenizer, causing different prompts to tokenize almost identically.
  • This could make Flow/DFlow ignore prompts, produce robotic/static output, or produce near-silent subprocess output.

• Fixed DFlow/DVFlow checkpoint loading regression.

  • Legacy VideoKontext key remapping is now scoped only to VFlow/VideoKontext checkpoints.
  • Distilled FlowMap models are no longer remapped through the VFlow compatibility path.

• Aligned Flow/VFlow ODE solver dtype handling with SonyResearch/Woosh test runners.

  • CUDA/CPU use the upstream float64 solver option.
  • MPS stays float32 to avoid Apple Silicon float64 errors.

• Improved model memory cleanup.

  • Changing the selected generative model now fully releases the previous model reference.
  • Changing Woosh TextConditioning mode now unloads the previous text conditioner.
  • force_offload also releases the cached Synchformer feature model.
  • CUDA/MPS/Comfy cache cleanup is handled more safely.

• Added model/checkpoint type validation.

  • The loader now errors if the selected model_type does not match the checkpoint config.