[Bug] device-side assert triggered in contact/contact.cu at line 1633

[a-gostynski]

Summary

Error "device-side assert triggered in contact/contact.cu at line 1633" keeps coming up in the console.

The blender GUI says "Remote: Unrecoverable CUDA runtime or driver error: solver exited via fatal hook..."

Steps to Reproduce

1. Download the zip archive containing the .blend file (https://drive.google.com/file/d/1p0iAO-2FYjQj450KUa_qNZTqUuGm7ZyO/view?usp=sharing)
2. Open the provided .blend file
3. Capture body and head deformations if missing
4. Transfer and run

Expected Behavior

Simulation running

Actual Behavior

Error as stated above and failing to simulate, happens on frame 42.

Alternatively one time the following error happened, also on frame 42.

newton stalled: no acceptable step after 2048 iterations (last toi: 3.91e-05, toi_advanced: 1.24e-01)
an over-constrained configuration cannot be advanced: a prescribed pin driven into geometry that cannot yield has no way to resolve. Re-author the pin's path, or make it a soft pull pin.

Which is confusing, since there isn't a single pin in the scene.

Environment

Machine A:
Blender 5.2.0 + Add-on 1.0.16
Ubuntu 24.04.3 LTS
Local docker
Solver version from docker build #118
NVIDIA GeForce RTX 4060, Driver Version: 595.71.05, CUDA Version: 13.2

Machine B:
Ubuntu 24.04.2 LTS
Docker over ssh (Blender running on Machine A)
Solver version from docker build #118
NVIDIA GeForce RTX 4090, Driver Version: 580.95.05, CUDA Version: 13.0

Contribution Agreement

• I confirm that any code, patches, scripts, attachments, or other copyrightable material I include in this issue may be used, reproduced, modified, and distributed by this project under the Apache License 2.0 or as otherwise described in CONTRIBUTING.md.
• This is a bug report from the community (please provide a reference below).

Community Reference

No response

[ryichando]

Confirm the crash, and I have a hunch that it is closely related to #116 and #110.

I suspect the dress is partly getting pinched by the animated character's arms, leaving the solver with no way to avoid self-penetration. This is still just a suspicion, though I haven't verified it yet.

Currently, the animated character motion is absolute, meaning it is never pushed by dynamic sheets or solids. I will try bringing back an option for "soft constraints" and enabling self-collision even for static objects to see if that remedies this pitfall.

I will start working on this tomorrow morning. It's 8:45 PM here in Tokyo at the time of writing.

[ryichando]

Issue solved. I introduced an "Apply Soft Constraints" option for STATIC deformations, and the simulation now runs without crashing.

However, I noticed that when a 6% strain limit is enabled, the character animation's step stride is too large, and the legs become partly squished by the strain-limited skirt. See the video below:

weak-const-SL-on-50.mp4

If you turn off strain limiting, this does not happen, but the dress may look rubbery:

weak-const-SL-off-50.mp4

Either way, the simulator itself runs fine, and both outcomes above are expected.

The fix will be delivered in the next release, likely within the next 48 hours.

[a-gostynski]

Great stuff, thank you!

[ryichando]

This is taking longer than I expected, but we are in the final stretch. The Windows compiled binary is also working. The delivery should be soon.

windows-working-50.mp4

[ryichando]

Delivered:

https://github.com/st-tech/ppf-contact-solver/releases/tag/addon-2026-08-08-0540
https://github.com/st-tech/ppf-contact-solver/releases/tag/2026-08-08-03-55