[Issue][rocdxg]: VBAR allocation fails in hipMemAddressReserve under ROCm/ROCDXG on WSL

[yzhang19-amd]

Problem Description

Description

In a WSL2 + ROCm/ROCDXG environment, running ComfyUI Flux Schnell FP16 inference fails when DynamicVRAM uses comfy-aimdo to create a VBAR.

The issue has been reproduced on both:

• APU: GFX1151
• Discrete GPU: Navi48 XTX

The failure occurs while reserving GPU virtual address space:

[ERROR] aimdo: /project/src/model-vbar.c:236:ERROR:
Could not reseve Virtual Address space for VBAR

MemoryError: VBAR allocation failed

Environment

• OS: Windows WSL2 Ubuntu-24.04
• ROCm: 10.1 0816 Nightly
• GPU 1: APU, GFX1151
• GPU 2: Discrete GPU, Navi48 XTX
• PyTorch: 2.13.0 0817 Nightly
• ComfyUI commit: cc0fc21fea7a6a82f568362b15b7fbd713b419c1
• comfyui-aimdo: 0.4.13

Operating System

WSL2/Ubuntu-24.04

CPU

AI MAX395/Ryzen 9 7900X

GPU

8060S/Navi48XTX

ROCm Version

10.1 0816 Nightly Build

ROCm Component

No response

Steps to Reproduce

Reproduction

1. Clone ComfyUI

git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
git checkout cc0fc21fea7a6a82f568362b15b7fbd713b419c1

python3 -m pip install -r requirements.txt

2. Install a ROCm-compatible PyTorch build matching the installed ROCm version, then install the ComfyUI requirements:

3. Prepare the Flux Schnell model

Download the standard Flux Schnell model files and place them in the following directories:

models/text_encoders/clip_l.safetensors
models/text_encoders/t5xxl_fp16.safetensors
models/vae/ae.safetensors
models/diffusion_models/flux1-schnell.safetensors

Model and workflow details are available in the official ComfyUI documentation:
https://docs.comfy.org/tutorials/flux/flux-1-text-to-image

4. Start ComfyUI with DynamicVRAM enabled

From the root of the ComfyUI repository, run:

python3 main.py \
    --listen 127.0.0.1 \
    --port 8188 \
    --enable-dynamic-vram

Open the ComfyUI web interface, load the standard Flux Schnell text-to-image workflow, and queue one prompt.

Additional Information

The issue can not reproduce on Native Linux environment(ubuntu24.04)