5070Ti cuda error

[Lucky-xf]

The current workstation GPU is a 5070ti, and the currently installed CUDA version is 12.8
Now I want to install a torch package compatible with CUDA 12.8; the command I use to install is:

pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

After installation, I then built and installed RFdiffusion. After installation, when running computational tasks, I encountered an error with the torchdata package. Currently I have torchdata version 0.11.0 installed; in this version the datapipes module was removed or moved into the core library, but the DGL library I am using still imports from from torchdata.datapipes.iter, so I changed the torchdata version to 0.7.1 using the command

pip install torchdata==0.7.1 --no-deps

This ensures that my torch version won't be affected. Unfortunately, after the change a mismatch in the C++ underlying dynamic library versions appeared:

/home/kk203/Downloads/Soft/Test-RF/RFdiffusion/rfdiffusion/util.py:253: UserWarning: Using torch.cross without specifying the 
dim arg is deprecated.
Please either pass the dim explicitly or simply use torch.linalg.cross.
The default value of dim will change to agree with that of linalg.cross in a future release. (Triggered internally at /pytorch
/aten/src/ATen/native/Cross.cpp:63.)
Z = torch.cross(Xn, Yn)
Traceback (most recent call last):
File "/home/kk203/Downloads/Soft/Test-RF/RFdiffusion/scripts/run_inference.py", line 25, in
from rfdiffusion.inference import utils as iu
File "/home/kk203/Downloads/Soft/Test-RF/RFdiffusion/rfdiffusion/inference/utils.py", line 6, in
from rfdiffusion.diffusion import get_beta_schedule
File "/home/kk203/Downloads/Soft/Test-RF/RFdiffusion/rfdiffusion/diffusion.py", line 12, in
from rfdiffusion.util_module import ComputeAllAtomCoords
File "/home/kk203/Downloads/Soft/Test-RF/RFdiffusion/rfdiffusion/util_module.py", line 7, in
import dgl
File "/home/kk203/.conda/envs/RFdiffusion/lib/python3.10/site-packages/dgl/init.py", line 16, in
from . import (
File "/home/kk203/.conda/envs/RFdiffusion/lib/python3.10/site-packages/dgl/dataloading/init.py", line 13, in
from .dataloader import *
File "/home/kk203/.conda/envs/RFdiffusion/lib/python3.10/site-packages/dgl/dataloading/dataloader.py", line 27, in
from ..distributed import DistGraph
File "/home/kk203/.conda/envs/RFdiffusion/lib/python3.10/site-packages/dgl/distributed/init.py", line 5, in
from .dist_graph import DistGraph, DistGraphServer, edge_split, node_split
File "/home/kk203/.conda/envs/RFdiffusion/lib/python3.10/site-packages/dgl/distributed/dist_graph.py", line 11, in
from .. import backend as F, graphbolt as gb, heterograph_index
File "/home/kk203/.conda/envs/RFdiffusion/lib/python3.10/site-packages/dgl/graphbolt/init.py", line 55, in
load_graphbolt()
File "/home/kk203/.conda/envs/RFdiffusion/lib/python3.10/site-packages/dgl/graphbolt/init.py", line 45, in load_graphbolt
raise FileNotFoundError(
FileNotFoundError: Cannot find DGL C++ graphbolt library at /home/kk203/.conda/envs/RFdiffusion/lib/python3.10/site-packages/d
gl/graphbolt/libgraphbolt_pytorch_2.10.0.so

I would like to ask if there are any good ways to resolve version conflicts between these libraries, or if anyone can recommend versions of torch, DGL, and other libraries to install for the 5070 Ti GPU so that GPU acceleration can be used.

[rclune]

Is there a particular functionality you need from RFdiffusion?

RFdiffusion3 (in foundry) is able to do much of what the original RFdiffusion can do and works on more modern systems.

However if you do need functionality only included in RFdiffusion, have you tried any of the suggestions in this issue? #349

[Lucky-xf]

Is there a particular functionality you need from RFdiffusion?

RFdiffusion3 (in foundry) is able to do much of what the original RFdiffusion can do and works on more modern systems.

However if you do need functionality only included in RFdiffusion, have you tried any of the suggestions in this issue? #349

This was also a small oversight on my part; I didn't see issue #349. I just took a look at it, and I think it should be helpful to me, so I will give it another try. At the same time, thank you very much for reminding me to use Foundry.

[Lucky-xf]

Is there a particular functionality you need from RFdiffusion?

RFdiffusion3 (in foundry) is able to do much of what the original RFdiffusion can do and works on more modern systems.

However if you do need functionality only included in RFdiffusion, have you tried any of the suggestions in this issue? #349

I just followed the guidance in issue 349 to recompile and install the relevant configuration files, and I reran my job script. As can be seen in the Slurm logs, my GPU is now being detected：

[2026-03-18 08:42:56,243][main][INFO] - Found GPU with device_name NVIDIA GeForce RTX 5070 Ti. Will run RFdiffusion on NVI
DIA GeForce RTX 5070 Ti

but unfortunately the following error still occurs:

Error executing job with overrides: ['inference.output_prefix=./diffusion_outputs/chimera', 'inference.input_pdb=moti
f_fixed.pdb', 'inference.num_designs=100', 'contigmap.contigs=[A2-434/15-30/A1129-1131/15-30/A1463-1465/5-15]', 'inference.ckp
t_override_path=/home/kk203/Downloads/Soft/GPU-Rfdiffusion/RFdiffusion/models/ActiveSite_ckpt.pt', 'diffuser.T=50']
Traceback (most recent call last):
File "/home/kk203/Downloads/Soft/GPU-Rfdiffusion/RFdiffusion/scripts/run_inference.py", line 84, in main
x_init, seq_init = sampler.sample_init()
^^^^^^^^^^^^^^^^^^^^^
File "/home/kk203/Downloads/Soft/GPU-Rfdiffusion/RFdiffusion/rfdiffusion/inference/model_runners.py", line 326, in sample_in
it
self._init_cyclic_reses(self.mask_str, self.contig_map)
File "/home/kk203/Downloads/Soft/GPU-Rfdiffusion/RFdiffusion/rfdiffusion/inference/model_runners.py", line 564, in _init_cyc
lic_reses
self.cyclic_reses = torch.zeros_like(mask).bool().to(self.device).squeeze()
^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

[rclune]

That likely means that there's still a mismatch between your CUDA version and some of the dependencies RFdiffusion is using, most likely PyTorch. Have you checked to make sure that everything torch-related is now pointing to CUDA 12.8?