# --- core stack (versions from the Bernini repo's own requirements.txt) ---
transformers==4.57.3
diffusers==0.35.2
accelerate
safetensors
torchvision
einops
numpy
Pillow
tqdm
ftfy
scipy
sentencepiece
packaging
psutil
hf_transfer

# --- video / image I/O ---
decord
imageio
imageio-ffmpeg

# --- veomni deps that its inference-side modules actually touch ---
# (the veomni package itself is vendored in ./veomni, Apache-2.0, v0.1.11,
#  because pip-installing it drags in datasets<=2.21.0 / torchdata / wandb)

# --- FlashAttention 2 ---
# bernini/models/modeling_qwen2_5_vl.py raises at import time unless flash_attn
# is importable, and the MLLM's vision tower asks for flash_attention_2.
# sm_120 (Blackwell) prebuilt wheel, cp312 / torch 2.11:
https://huggingface.co/datasets/multimodalart/zerogpu-blackwell-wheels/resolve/main/wheels/pt211-cu130-cp312/flash_attn-2.8.3-cp312-cp312-linux_x86_64.whl