Dynamic VRAM causes crash on AMD RX 7900 XTX (ROCm)

[iguchi333-dot]

Environment:

• GPU: AMD RX 7900 XTX
• ComfyUI: commit b353a7c (Integrate RAM cache with model RAM management)
• PyTorch: 2.10.0 (ROCm)
• OS: Windows

Issue:
After updating to commit b353a7c, ComfyUI crashes with
"Windows fatal exception: code 0xc000001d" when loading
SDXL model during FaceDetailer processing.

Reverting to 3696c5b fixed the issue.

This may be related to Dynamic VRAM not yet supporting AMD GPUs.

[iguchi333-dot]

Update: Reverting to 3696c5b did not fix the issue.
The crash still occurs with the same stack trace.

Workaround: Adding --disable-smart-memory to launch options
seems to prevent the crash (still testing).

This may be a ROCm/AMD-specific issue with the new memory
management system.

[iguchi333-dot]

Further update: Still crashes with access violation at
module_mmap_residency in model_management.py.

Error changed from 0xc000001d to 0xc0000005 (access violation).
Crash occurs at mmap-related code, which may indicate
Dynamic VRAM's mmap feature is incompatible with ROCm/AMD GPUs.

[iguchi333-dot]

Update: Removing FaceDetailer from the workflow completely
prevents the crash. The issue appears to be specific to
FaceDetailer triggering model reloading with the new
memory management system on AMD/ROCm.

Workaround: Using an additional KSampler with denoise ~0.5
instead of FaceDetailer.