Problems with new gen rtx

[MerzouuuGH]

I tried to install it on my rtx 5070, but the pytorch version was wrong, but after the change other dependiences fell apart.
It's good to note that some pytorch versions can't be installed on newer python.

Btw, I have a question, does it support TV color range? Because the original subtitle remover has some problem with it and the colors aren't matching.

[SysAdminDoc]

Thanks for the detailed report — you're right on both counts, and the deps issue is fixed in v3.16.0 (d96e72d).

RTX 50-series (Blackwell): the installer was pinning the CUDA 11.8 (cu118) PyTorch wheels for every NVIDIA card. Your 5070 is compute capability sm_120, which has no kernels in cu118/cu121 builds — hence the failure and the dependency cascade when you bumped torch by hand. setup.py now auto-detects 50-series cards and installs PyTorch from the CUDA 12.8 (cu128) index (torch 2.7+, which supports Python 3.9–3.13), with PaddlePaddle from cu126. 20/30/40-series cards keep the existing path, so no regression there. I also fixed a stale torch==2.7.0 line in the README that conflicted with the >=2.10.0 requirement — that mismatch was likely part of what broke for you.

To repair your current venv without reinstalling:

.\venv\Scripts\activate
pip uninstall -y torch torchvision
pip install "torch>=2.10.0" "torchvision>=0.25.0" --index-url https://download.pytorch.org/whl/cu128

TV color range: good news — this fork already handles it. There's a preserve_color_metadata pass (on by default) that probes the source's color_primaries / color_transfer / color_space / color_range via ffprobe and re-applies those tags on the final encode, so limited/TV-range (BT.601/709) clips don't get the wash-out the original project produces by dropping those tags. One caveat: the internal pipeline is 8-bit BGR, so genuine 10-bit HDR sources get tone-mapped to SDR (tagged correctly, just not 10-bit). If you still see a color mismatch on an SDR clip after updating, paste the ffprobe color fields of the source and I'll dig in.

Could you confirm whether the cu128 path installs cleanly on your 5070? I don't have Blackwell hardware to verify end-to-end, so a thumbs-up (or the error) would help.

[SysAdminDoc]

There's now a tagged v3.16.0 release with the Blackwell install path. Two notes for your 5070:

• Run from source at this tag for GPU — the attached ...-windows.zip is a CPU-only PyInstaller build and won't use your card. The cu128 auto-detect lives in setup.py, or use the manual pip install ... --index-url .../whl/cu128 from my earlier comment.
• If you grabbed the ZIP and your AV flagged it: that's the unsigned-PyInstaller-bundle false positive (same pattern as IT"S AN INFO STEALER #2), not malware — weights are SHA-256 verified.

Let me know if the cu128 path installs cleanly on the 5070 and I'll close this out.

[SysAdminDoc]

Closing — the Blackwell cu128 install path and the README/color-range fixes shipped in v3.16.0. If the cu128 path doesn't install cleanly on your 5070, please reopen with the error and I'll dig in.