requirements.txt

# Pinned to the versions the adapter was trained and evaluated with
# (see saved_metrics/train_config/ in the project repo).
transformers==4.57.3
peft==0.18.1
accelerate

# Qwen3-VL video sampling / pixel budgeting (pulls in `av` for decoding).
qwen-vl-utils
torchvision

# Training-time preprocessing:
#   mediapipe -> pose-guided signer crop
#   rtmlib    -> RTMPose Wholebody landmarks (pulls in onnxruntime + opencv)
mediapipe
rtmlib

# rtmlib brings the CPU build of onnxruntime, which is what this app wants.
# It runs the `balanced` wholebody model: measured on a local RTX 3060 that is
# 47 ms/frame on CPU, *faster* than the larger `performance` model on the GPU
# (105 ms/frame), because it is the same backbone at 192x256 instead of 288x384.
#
# onnxruntime-gpu is deliberately NOT used. Sharing the GPU between ONNX Runtime
# and PyTorch caused two failures locally: the ORT CUDA arena competed with the
# model for memory, and releasing the session mid-request left PyTorch unable to
# find cuDNN kernels ("GET was unable to find an engine to execute this
# computation"). Keeping ONNX Runtime on the CPU leaves the GPU to the model.

# Browser-playable H.264 preview of the processed clip.
imageio
imageio-ffmpeg

numpy