[GH-ISSUE #120] [Windows][AMD][ROCm 7.13] Access violation in vrambuf_create() when DynamicVRAM allocates cast buffer (RX 9070 XT) #45

Open
opened 2026-10-26 22:24:01 +00:00 by bsakaguchi · 2 comments
Owner

Originally created by @frinking602 on GitHub (Sep 29, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-aimdo/issues/120

crash_log_excerpt_0929.log

GPU: AMD Radeon RX 9070 XT 16GB
OS: Windows
ROCm: 7.13
AIMDO: 0.4.15
Problem Description
AIMDO was working normally for a long period of time.
After installation I experienced several rare crashes (about 3 times in total), but the issue disappeared after restarting and workloads could continue normally.
Since last night the failure has become persistent:
All workflows fail when AIMDO is enabled
The same workflows succeed when AIMDO is disabled
No version changes were made to AIMDO, ROCm, ComfyUI or models before the issue started
Cleaning ComfyUI caches did not help
The crash now occurs consistently.
Expected Behavior
AIMDO should allocate the cast buffer and continue model execution.
This issue was initially rare (about 3 occurrences since installation), but since 2026-09-28 it has become reproducible and all AIMDO-enabled workflows now fail.

Originally created by @frinking602 on GitHub (Sep 29, 2026). Original GitHub issue: https://github.com/Comfy-Org/comfy-aimdo/issues/120 [crash_log_excerpt_0929.log](https://github.com/user-attachments/files/32788669/crash_log_excerpt_0929.log) GPU: AMD Radeon RX 9070 XT 16GB OS: Windows ROCm: 7.13 AIMDO: 0.4.15 Problem Description AIMDO was working normally for a long period of time. After installation I experienced several rare crashes (about 3 times in total), but the issue disappeared after restarting and workloads could continue normally. Since last night the failure has become persistent: All workflows fail when AIMDO is enabled The same workflows succeed when AIMDO is disabled No version changes were made to AIMDO, ROCm, ComfyUI or models before the issue started Cleaning ComfyUI caches did not help The crash now occurs consistently. Expected Behavior AIMDO should allocate the cast buffer and continue model execution. This issue was initially rare (about 3 occurrences since installation), but since 2026-09-28 it has become reproducible and all AIMDO-enabled workflows now fail.
Author
Owner

@frinking602 commented on GitHub (Sep 29, 2026):

Image
<!-- gh-comment-id:5885074300 --> @frinking602 commented on GitHub (Sep 29, 2026): <img width="1080" height="4480" alt="Image" src="https://github.com/user-attachments/assets/a784294f-3278-47cb-abd3-3d05bc30d3f3" />
Author
Owner

@0xDELUXA commented on GitHub (Oct 4, 2026):

ROCm 7.13 is quite old. Why are you using this specific release? Update to the latest stable with:

pip install --index-url https://nightly.repo.amd.com/rocm/whl-next/ torch[device-gfx1201] torchvision[device-gfx1201] torchaudio --upgrade

<!-- gh-comment-id:5983039653 --> @0xDELUXA commented on GitHub (Oct 4, 2026): ROCm 7.13 is quite old. Why are you using this specific release? Update to the latest stable with: `pip install --index-url https://nightly.repo.amd.com/rocm/whl-next/ torch[device-gfx1201] torchvision[device-gfx1201] torchaudio --upgrade`
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
Comfy-Org/comfy-aimdo#45
No description provided.