Comfy-Org/comfy-aimdo

Fork 0

mirror of https://github.com/Comfy-Org/comfy-aimdo.git synced 2026-11-04 23:37:21 +00:00

UI stalling during VAE Decode with DynamicVRAM enabled #37

New issue

Closed

opened 2026-10-26 22:23:58 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 22:23:58 +00:00

Owner

Originally created by @ktla672-creator on GitHub (Sep 3, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-aimdo/issues/104

Title

[ROCm/Windows][RX 9070 XT] Severe GPU/UI stalling during VAE Decode with DynamicVRAM enabled

Description

I am experiencing severe GPU/Windows UI stalls during image generation with DynamicVRAM enabled on an AMD Radeon RX 9070 XT.

The issue is reproducible with a relatively simple SDXL workflow using the normal VAE Decode node.

Interestingly, disabling DynamicVRAM with --disable-dynamic-vram makes the problem disappear under otherwise essentially identical conditions.

Hardware / Software

OS: Windows
GPU: AMD Radeon RX 9070 XT
VRAM: 16 GB (16304 MB reported)
GPU architecture: gfx1201
ROCm: 7.14
PyTorch: 2.12.0+rocm7.14.0
ComfyUI: 0.34.3
comfy-aimdo: 0.4.15
comfy-kitchen: 0.2.31
Python: 3.12.12
VAE dtype: torch.bfloat16
VAE attention: split attention
Attention backend: PyTorch attention

Problem

With DynamicVRAM enabled, the first generation may cause several seconds of noticeable Windows UI/GPU stuttering around the VAE Decode stage.

On subsequent generations, the problem can become much worse. In my latest test:

First generation: Prompt executed in 63.43 seconds
Second generation: Prompt executed in 121.41 seconds

The actual SDXL sampling itself remained normal:

30 sampling steps
approximately 17 seconds for each generation

During the VAE/output stage, Windows becomes severely unresponsive. The mouse and desktop UI can become very sluggish or appear frozen.

Interestingly, turning the monitor off and then back on immediately restores normal Windows responsiveness.

There is no obvious Python exception, CUDA/HIP out-of-memory error, or ComfyUI crash in the log.

Relevant log output

With DynamicVRAM enabled:

aimdo: ... installing 6 hooks
comfy-aimdo WDDM adapter match: AMD Radeon RX 9070 XT
comfy-aimdo inited for GPU: AMD Radeon RX 9070 XT (VRAM: 16304 MB)
DynamicVRAM support detected and enabled

During generation:

Using split attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16

Requested to load SDXL
Model SDXL prepared for dynamic VRAM loading. 4896MB Staged.

100%|██████████| 30/30 [00:17<00:00, 1.75it/s]

Requested to load AutoencoderKL
0 models unloaded.
Model AutoencoderKL prepared for dynamic VRAM loading. 159MB Staged.

Prompt executed in 63.43 seconds

Second generation:

100%|██████████| 30/30 [00:17<00:00, 1.73it/s]

0 models unloaded.
Model AutoencoderKL prepared for dynamic VRAM loading. 159MB Staged.

Prompt executed in 121.41 seconds

A/B test

I also tested the same setup with:

--disable-dynamic-vram

With DynamicVRAM disabled:

ComfyUI 0.34.3
comfy-aimdo 0.4.15
same RX 9070 XT
same ROCm 7.14
same PyTorch 2.12.0+rocm7.14.0
same SDXL workflow
normal (non-tiled) VAE Decode

I was able to generate two images consecutively without the severe GPU/UI stall.

This makes DynamicVRAM/comfy-aimdo the main suspect, although I cannot completely rule out an interaction between ComfyUI, comfy-aimdo, ROCm/HIP and the Windows WDDM driver.

Additional observation

Using VAE Decode Tiled also avoids the problem, but I would prefer to use the normal VAE Decode node because the same workflow works normally when DynamicVRAM is disabled.

The fact that:

SDXL sampling itself remains fast and stable (~17 seconds),
the slowdown happens around the VAE/output stage,
Windows itself becomes unresponsive,
toggling the monitor off/on restores the desktop immediately, and
disabling DynamicVRAM makes the problem disappear,

makes me suspect a DynamicVRAM/comfy-aimdo memory residency, VRAM eviction, or WDDM/HIP interaction issue rather than a simple VRAM exhaustion problem.

Expected behavior

VAE Decode should complete normally without causing the Windows desktop/GPU to stall.

Actual behavior

With DynamicVRAM enabled, VAE Decode can cause severe GPU/Windows UI stalls, with the problem becoming significantly worse on the second consecutive generation.

Reproduction

Start ComfyUI 0.34.3 on Windows with an RX 9070 XT and ROCm 7.14.
Keep DynamicVRAM enabled.
Use an SDXL workflow with the normal VAE Decode node.
Generate an image.
Generate another image immediately afterward.
Observe the VAE/output stage and Windows desktop responsiveness.

For comparison, restart ComfyUI with:

--disable-dynamic-vram

and repeat the same workflow. In my testing, the severe stalling does not occur.

Logs

I can provide the complete ComfyUI logs for both cases:

DynamicVRAM enabled — problematic
DynamicVRAM disabled — normal

Please let me know if any additional diagnostics would help, such as specific environment variables, debug logging, GPU/VRAM telemetry, or a minimal reproduction workflow.

https://chatgpt.com/share/6a996e24-65c0-83e9-b729-15852c2aacc3

Originally created by @ktla672-creator on GitHub (Sep 3, 2026). Original GitHub issue: https://github.com/Comfy-Org/comfy-aimdo/issues/104 ## Title [ROCm/Windows][RX 9070 XT] Severe GPU/UI stalling during VAE Decode with DynamicVRAM enabled ## Description I am experiencing severe GPU/Windows UI stalls during image generation with **DynamicVRAM enabled** on an AMD Radeon RX 9070 XT. The issue is reproducible with a relatively simple SDXL workflow using the normal `VAE Decode` node. Interestingly, disabling DynamicVRAM with `--disable-dynamic-vram` makes the problem disappear under otherwise essentially identical conditions. ### Hardware / Software * OS: Windows * GPU: AMD Radeon RX 9070 XT * VRAM: 16 GB (16304 MB reported) * GPU architecture: gfx1201 * ROCm: 7.14 * PyTorch: 2.12.0+rocm7.14.0 * ComfyUI: 0.34.3 * comfy-aimdo: 0.4.15 * comfy-kitchen: 0.2.31 * Python: 3.12.12 * VAE dtype: `torch.bfloat16` * VAE attention: split attention * Attention backend: PyTorch attention ### Problem With DynamicVRAM enabled, the first generation may cause several seconds of noticeable Windows UI/GPU stuttering around the VAE Decode stage. On subsequent generations, the problem can become much worse. In my latest test: * First generation: `Prompt executed in 63.43 seconds` * Second generation: `Prompt executed in 121.41 seconds` The actual SDXL sampling itself remained normal: * 30 sampling steps * approximately 17 seconds for each generation During the VAE/output stage, Windows becomes severely unresponsive. The mouse and desktop UI can become very sluggish or appear frozen. Interestingly, turning the monitor off and then back on immediately restores normal Windows responsiveness. There is no obvious Python exception, CUDA/HIP out-of-memory error, or ComfyUI crash in the log. ### Relevant log output With DynamicVRAM enabled: ```text aimdo: ... installing 6 hooks comfy-aimdo WDDM adapter match: AMD Radeon RX 9070 XT comfy-aimdo inited for GPU: AMD Radeon RX 9070 XT (VRAM: 16304 MB) DynamicVRAM support detected and enabled ``` During generation: ```text Using split attention in VAE VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16 Requested to load SDXL Model SDXL prepared for dynamic VRAM loading. 4896MB Staged. 100%|██████████| 30/30 [00:17<00:00, 1.75it/s] Requested to load AutoencoderKL 0 models unloaded. Model AutoencoderKL prepared for dynamic VRAM loading. 159MB Staged. Prompt executed in 63.43 seconds ``` Second generation: ```text 100%|██████████| 30/30 [00:17<00:00, 1.73it/s] 0 models unloaded. Model AutoencoderKL prepared for dynamic VRAM loading. 159MB Staged. Prompt executed in 121.41 seconds ``` ### A/B test I also tested the same setup with: ```text --disable-dynamic-vram ``` With DynamicVRAM disabled: * ComfyUI 0.34.3 * comfy-aimdo 0.4.15 * same RX 9070 XT * same ROCm 7.14 * same PyTorch 2.12.0+rocm7.14.0 * same SDXL workflow * normal (non-tiled) VAE Decode I was able to generate two images consecutively without the severe GPU/UI stall. This makes DynamicVRAM/comfy-aimdo the main suspect, although I cannot completely rule out an interaction between ComfyUI, comfy-aimdo, ROCm/HIP and the Windows WDDM driver. ### Additional observation Using `VAE Decode Tiled` also avoids the problem, but I would prefer to use the normal VAE Decode node because the same workflow works normally when DynamicVRAM is disabled. The fact that: 1. SDXL sampling itself remains fast and stable (~17 seconds), 2. the slowdown happens around the VAE/output stage, 3. Windows itself becomes unresponsive, 4. toggling the monitor off/on restores the desktop immediately, and 5. disabling DynamicVRAM makes the problem disappear, makes me suspect a DynamicVRAM/comfy-aimdo memory residency, VRAM eviction, or WDDM/HIP interaction issue rather than a simple VRAM exhaustion problem. ### Expected behavior VAE Decode should complete normally without causing the Windows desktop/GPU to stall. ### Actual behavior With DynamicVRAM enabled, VAE Decode can cause severe GPU/Windows UI stalls, with the problem becoming significantly worse on the second consecutive generation. ### Reproduction 1. Start ComfyUI 0.34.3 on Windows with an RX 9070 XT and ROCm 7.14. 2. Keep DynamicVRAM enabled. 3. Use an SDXL workflow with the normal `VAE Decode` node. 4. Generate an image. 5. Generate another image immediately afterward. 6. Observe the VAE/output stage and Windows desktop responsiveness. For comparison, restart ComfyUI with: ```text --disable-dynamic-vram ``` and repeat the same workflow. In my testing, the severe stalling does not occur. ### Logs I can provide the complete ComfyUI logs for both cases: * DynamicVRAM enabled — problematic * DynamicVRAM disabled — normal Please let me know if any additional diagnostics would help, such as specific environment variables, debug logging, GPU/VRAM telemetry, or a minimal reproduction workflow. https://chatgpt.com/share/6a996e24-65c0-83e9-b729-15852c2aacc3

bsakaguchi closed this issue

2026-10-26 22:23:58 +00:00

bsakaguchi commented

2026-10-26 22:23:58 +00:00

Author

Owner

@ktla672-creator commented on GitHub (Sep 5, 2026):

https://github.com/Comfy-Org/ComfyUI/issues/16062

@ktla672-creator commented on GitHub (Sep 5, 2026): https://github.com/Comfy-Org/ComfyUI/issues/16062

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#37

No description provided.

Rows
Columns