Min Dalle Versions Save

• Fixed a criticial CUDA runtime error that occurred when generating tokens larger than the VQGAN's vocabulary
• Added generate_images_stream and generate_images to generate individual images. Is in active use in discord bot.
• Faster inference, can generate a 9x9 grid in 38 seconds on an A100
• Added temperature, top_k, and supercondition_factor parameters
• Added a simple TKinter UI (thanks to @20kdc)
• Added an option to tiles images in token space instead of pixel space. This creates a seamless effect where the borders between images are blended.

• added is_reusable parameter. Turning it off saves memory (e.g. for command line script) and keeping it on makes multiple calls to generate_image faster
• added log2_k parameter to control top-k image token sampling
• added log2_supercondition_factor parameter to control the super conditioning amount
• added log2_mid_count and generate_image_stream to stream intermediate outputs. Incomplete tokens are detokenized to an image multiple times during the decoding process. This adds very little time to the overall run time
• added dtype parameter to autocast operations to float32, float16, or bfloat16
• a grid size of 8x8 now generates in 35 seconds on an A100

• Added to PyPI so now the entire setup process is pip install min-dalle
• Pre-converted PyTorch weights are downloaded when needed from a Hugging Face hub, no more converting from flax

Breaking Changes

• MinDalleTorch is now MinDalle
• MinDalleFlax and flax-to-torch conversion code have been moved to a different repository

Important Bug Fixes

• Image tokens were mistakenly being computed twice in command line script when using torch
• Tokenizer was not working correctly on some machines previously (e.g. windows). Files are now read with ut8-encoding.

New Features

• is_expendable argument reduces memory usage for command line script by loading then unloading encoder/decoder/detokenizer when needed
• simpler 4D attention_state replacing 5D keys_values_state and faster inference time

MinDalleTorch and MinDalleFlax classes to initialize model once and run multiple times