ComfyUI-DN_AimdoMonitor

A ComfyUI custom node for diagnosing and evaluating comfy_aimdo's behavior at runtime. Rather than being a general VRAM monitor, it specifically validates whether aimdo's internal allocation tracking matches what the AMD driver actually reports via hipMemGetInfo - exposing hook failures, accounting drift, and - on Windows - silent VRAM paging to system RAM.

---

Requirements

Requirement	Notes
AMD GPU	Any ROCm-supported card.
ROCm	Must be installed and working with PyTorch.
PyTorch with ROCm	torch.cuda.mem_get_info() must be functional.
ComfyUI	Any recent version.
comfy_aimdo	Must be installed and loaded.

OS: Windows and Linux both supported. The WDDM paging detection is Windows-only by nature and simply won't trigger on Linux.

---

Installation

1. Clone or download this repository into your ComfyUI custom nodes folder:

ComfyUI/custom_nodes/

The folder must contain:

ComfyUI-DN_AimdoMonitor/
├── __init__.py
└── js/
    └── DN_AimdoMonitor.js

2. Restart ComfyUI normally:

python main.py

---

Usage

Floating button

A [Start VRAM Poll] button appears in the bottom-right corner of the ComfyUI interface. Click it to start polling; it turns red and becomes [Stop VRAM Poll]. The button is draggable and remembers its position across reloads.

While polling, stats are printed to the console every 0.5 seconds (only when values change or drift exceeds 256 MB):

[aimdo_poll] aimdo=4200MB  hip_used=4350MB  drift=+150MB

Workflow nodes

Two passthrough nodes are available for inline snapshots at specific points in a workflow:

• Aimdo VRAM Monitor (Latent) - connects inline on a LATENT wire
• Aimdo VRAM Monitor (Model) - connects inline on a MODEL wire

Both print a detailed snapshot to the console when the node executes:

[aimdo_monitor] after_model_load
[aimdo_monitor]   aimdo total_vram_usage :    4200 MB
[aimdo_monitor]   hipMemGetInfo used     :    4350 MB
[aimdo_monitor]   hipMemGetInfo free     :   11834 MB
[aimdo_monitor]   drift (aimdo - hip)    :    +150 MB
[aimdo_monitor]   OK: tracking within 512MB

Session summary

When polling is stopped - either via the button or on ComfyUI shutdown - a summary is printed to the console:

[aimdo_monitor] -- Session Summary --
[aimdo_monitor]   Samples collected  :     104
[aimdo_monitor]   Avg drift          :    +429 MB
[aimdo_monitor]   Max drift          :    +660 MB
[aimdo_monitor]   Drift direction    : Slightly negative - normal, driver overhead not tracked by aimdo
[aimdo_monitor]   Drift > 512MB      : 23 time(s)
[aimdo_monitor]   Drift > 1GB        : 0 time(s)
[aimdo_monitor]   WDDM paging events : 0
[aimdo_monitor]   Verdict            : Slightly degraded - drift > 512MB, aimdo mostly working but not fully accurate
[aimdo_monitor] ------------------------------------

Possible verdicts:

Verdict	Meaning
Ok - working as intended	Low drift throughout, hooks intercepting correctly.
Slightly degraded - drift > 512MB...	Drift exceeded 512MB at least occasionally, mostly working but not fully accurate.
Poor - drift exceeded 1GB frequently...	Drift exceeded 1GB in more than 30% of samples, or max drift spiked above 2GB. Accounting unreliable.
Fail - hooks partially failing...	aimdo was 0 for a significant portion of the session.
Fail - hooks not intercepting...	aimdo reported 0 for nearly the entire session while the GPU was in use.

Drift direction tells you which way aimdo is off. A slightly negative result is normal on Windows due to driver overhead that aimdo does not track. Negative (avg > 1GB) means aimdo is meaningfully undercounting. Strongly negative (avg > 2GB) means something is seriously wrong with hook interception.

---

Console warnings

Message	Meaning
*** AIMDO NOT ACTIVE: aimdo=0 but hip shows nMB used!	aimdo hooks are not intercepting allocations - check that comfy_aimdo is installed and loaded correctly.
*** WDDM PAGING TO RAM ***	(Windows only) VRAM is full and allocations are overflowing into system RAM. Generation will be slow.

---