# `diffusers` is installed from **main**, pinned to a commit rather than tracking the branch: a floating branch means # a future rebuild can break this Space for reasons unrelated to anything here, and a recorded commit still is # "latest main" as of this deploy. `h3_a2v_blocks.py` subclasses these block classes, so re-check it against the # block names — `MiniMaxH3FL2VACoreDenoiseStep`, `MiniMaxH3PrepareConditionLatentsStep`, # `MiniMaxH3FL2VAPrepareLatentsStep` — whenever the pin moves. # # Main rather than the MiniMax-H3 follow-up PR (#14371, which this Space used to pin at `665f578`): that PR branched # before the LoRA work landed and carries no `MiniMaxH3LoraLoaderMixin` at all, so the turbo LoRA cannot be read # against it. # # d5baa4fb548294f47dbca49890abd4b291204c60 = huggingface/diffusers main at the time of this deploy --extra-index-url https://download.pytorch.org/whl/cu130 diffusers @ git+https://github.com/huggingface/diffusers.git@d5baa4fb548294f47dbca49890abd4b291204c60 torch==2.11.0 torchvision==0.26.0 # An uploaded soundtrack that is not already at the audio VAE's 32 kHz is resampled with `torchaudio` — which is # almost every upload, so this is not optional here. torchaudio==2.11.0 # The Qwen3-VL processor decides the vision patch count, so a different minor changes the conditioning. transformers==5.8.0 accelerate==1.14.0 # The turbo LoRA is injected with peft, the version the MiniMax-H3 LoRA roundup and the a2v sweep both ran on. peft==0.18.1 # diffusers pins <2. huggingface-hub==1.24.0 gradio==6.20.0 spaces==0.51.1 # No `kernels` pin on purpose: the Hub attention backends want `kernels>=0.12.3`, and that version breaks # transformers 5.8.0 at import. # PyAV decodes the uploaded soundtrack and muxes the result (`MiniMaxH3AudioReference.from_file`, `encode_video`). av pillow numpy requests safetensors>=0.8.0