| # d3 runtime. Validated with transformers 5.17.0 on ROCm (torch 2.13). | |
| torch>=2.8 | |
| transformers==5.17.0 | |
| accelerate>=1.10 | |
| safetensors>=0.6 | |
| huggingface_hub>=1.0 | |
| # GPU kernels for the Gated DeltaNet layers; without them transformers runs its slower PyTorch reference. | |
| flash-linear-attention==0.5.2 | |
| einops | |
| # Image inputs: PIL decodes the images; the processor's torchvision backend resizes them as in evaluation | |
| # (install the torchvision build that matches torch; without it transformers falls back to a PIL resize). | |
| pillow>=10 | |
| torchvision | |
| # Video files: OpenCV decodes them with FFmpeg (frame arrays need no decoder). | |
| opencv-python-headless | |
| # d3_server.py needs nothing beyond the above (Python standard library HTTP server). | |