mouii/Jasna_engines

Jasna TensorRT Engines - RTX 5080 (Blackwell SM 12.0)

Pre-compiled TensorRT engine files for Jasna video restoration tool, built on NVIDIA RTX 5080.

Compatibility

These engines are ONLY compatible with GPUs sharing the same architecture:

• GPU Architecture: Blackwell (SM 12.0)
• Compatible GPUs: RTX 5070 Ti, RTX 5080, RTX 5090
• TensorRT: 10.14.1.48
• CUDA: 13.1
• PyTorch: 2.10.0a0+nv26.01
• Driver: 590.48.01+
• Jasna: 0.4.1+

Will NOT work on Ada Lovelace (RTX 4090), Ampere (RTX 3090), or older GPUs.

Files

• lada_mosaic_restoration_model_generic_v1.2_clip10.trt_fp16.linux.engine (250 MB) - BasicVSR++ restoration model (clip_size=10, FP16)
• lada_mosaic_restoration_model_generic_v1.2_clip40.trt_fp16.linux.engine (967 MB) - BasicVSR++ restoration model (clip_size=40, FP16)
• rfdetr-v3.bs4.fp16.linux.engine (77 MB) - RF-DETR detection model (batch_size=4, FP16)

Build Environment

• GPU: NVIDIA GeForce RTX 5080 (16 GB VRAM)
• OS: Ubuntu 22.04 (Docker container)
• Platform: Vast.ai
• NVIDIA Driver: 590.48.01
• CUDA: 13.1 (V13.1.115)
• TensorRT: 10.14.1.48
• PyTorch: 2.10.0a0+a36e1d39eb.nv26.01

Usage

Download the engine files and place them in jasna/model_weights/, then run:

./jasna --input "input.mp4" --output "output.mp4" --fp16 --max-clip-size 40 --log-level info

Notes

• clip10 engine uses less VRAM (~4 GB), suitable for GPUs with limited memory
• clip40 engine uses more VRAM (~8 GB) but processes faster
• clip90 compilation failed with OOM on 16 GB VRAM; may work with more swap space
• If engines don't work on your setup, delete them and let Jasna recompile automatically

Known Issues

On Vast.ai, multi-GPU hosts with single GPU allocation may cause NVDEC/NVENC failures due to nvidia-container-toolkit#1249 (https://github.com/NVIDIA/nvidia-container-toolkit/issues/1249). Use single-GPU hosts as a workaround.