Fix: Native CUDA 13 support and dynamic CUDA_HOME fetching for CuPy

[wzgrx]

Fix: Native CUDA 13 support and dynamic CUDA_HOME fetching for CuPy

Description
Motivation & Context
With the release of the new RTX 50-series GPUs and the adoption of CUDA 13.x, users are experiencing installation and runtime failures with CuPy. Previously, the installation script either attempted a risky fallback to CUDA 12.x or forced a source compilation that often failed. Additionally, CuPy's JIT compiler (RawModule) frequently failed to locate the necessary CUDA headers in CUDA 13 environments when CUDA_HOME was not explicitly set.

Since CuPy now officially provides cupy-cuda13x wheels, this PR updates the installation logic and JIT compiler paths to natively support CUDA 13.

Changes Included:

install.py:

Updated get_cuda_ver_from_dir to correctly detect CUDA 13 and return '13x', fetching the official cupy-cuda13x wheel instead of falling back to 12.x.

Added cupy-cuda13x to the pip uninstall cleanup list to prevent environment conflicts.

requirements-with-cupy.txt:

Updated the generic cupy-wheel (which acts as an empty shell) to cupy-cuda13x to streamline manual installations on modern setups.

vfi_models/ops/cupy_ops/utils.py:

Enhanced cuda_launch to dynamically fetch and inject CUDA_HOME and CUDA_PATH using get_cuda_home_path(). This ensures that CuPy's NVRTC compiler can successfully locate the PyTorch-bundled or system CUDA components, preventing RawModule runtime crashes on unconfigured environments.

How to Test

Run install.py on a machine with CUDA 13.x (e.g., equipped with an RTX 5090).

Verify that cupy-cuda13x is installed automatically without triggering a source build.

Run a frame interpolation node requiring CuPy (e.g., RIFE) and verify that the CUDA kernels compile and launch successfully without missing header errors.

[essence25]

And... it is still not fixed/updated into the main release. I posted this exact fix in the issues section also.

@wzgrx
Inside "vfi_utils.py" please also change this (the first 2 links are dead):

BASE_MODEL_DOWNLOAD_URLS = [
    "https://github.com/styler00dollar/VSGAN-tensorrt-docker/releases/download/models/",
    "https://github.com/Fannovel16/ComfyUI-Frame-Interpolation/releases/download/models/",
    "https://github.com/dajes/frame-interpolation-pytorch/releases/download/v1.0.0/"
]

To this:

BASE_MODEL_DOWNLOAD_URLS = [
    "https://github.com/dajes/frame-interpolation-pytorch/releases/download/v1.0.2/"
]

That will update film_net_fp32.pt to v1.0.2 and get rid of this warning that we always get with v1.0.0:

python_embeded\Lib\site-packages\torch\nn\modules\module.py:1790: UserWarning: Using padding='same' with even kernel lengths and odd dilation may require a zero-padded copy of the input be created (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\pytorch\aten\src\ATen\native\Convolution.cpp:1025.)

Here are the details for the updated model:

v1.0.2 Newer torch version, improved support for dt argument Latest
The model is re-exported with torch 2.1.1

Includes minor fixes:

Fix UserWarning: Using padding='same' with even kernel lengths and odd dilation may require a zero-padded copy of the input be created
Now it doesn't ignore dt argument thanks to @niqodea (however it seems that authors of the model recommend to stick with .5)
Supports batched inference

https://github.com/dajes/frame-interpolation-pytorch/releases