ComfyUI

The scope of this repository is to document how to run Comfy UI with ROCm acceleration with a AMD 7900XTX under Windows.

Value Proposition

Nvidia and CUDA works better. But Nvidia makes you pay a significant premium for VRAM, and VRAM is critical to machine learning.

When choosing how to upgrade in 2025-01 my choices were

• AMD 7900XTX 24GB: 940€
• Nvidia 3090 24GB 4 years old used: 750€
• Nvidia 4090 24GB: 2500 €
• Nvidia 5090 32GB: 3500 €

In my region the 7900XTX still goes for 850€ to 950€ at the time of update 2025-12-09, and in my opinion this is amazing value to accelerate ComfyUI generation and local LLMs.

16GB cards are more affordable, but those 8GB extra go a long way in inference.

Hardware

• AMD RX 7900 XTX 24GB <--- The GOAT
• Intel 13700F
• DDR5 4x16GB 64GB 6400 CL32

ComfyUI vs AMD ROCm

The achille's heel of AMD card, is the stack. It's bad. As of 2026-01-25 ROCm has windows binaries for pytorch that works to an extent with decent performance with ComfyUI.

Now it works with the default driver 26.1.1 Make sure you install Pytorch option in the AI tab of adrenaline.

ComfyUI has a portable release ROCm 7.1.

Building it pip works better for me, as I can build with ROCm 7.2. Read the script and what it does.

EXTERNAL MODEL FOLDER

The environment can brick easily.

It is convenient to move the models outside the ComfyUI folder, so that when I rebuild, the models are all there. This also allows multiple local env to all access models without duplication.

extra_model_paths.yaml

comfyui:
    # Go up to the parent folder, and look for the model folder there
    base_path: ..\ComfyUI-Models
    # Model Folder
    checkpoints: checkpoints\
    clip: clip\
    clip_vision: clip_vision\
    text_encoders: text_encoders\
    configs: configs\
    controlnet: controlnet\
    diffusion_models: |
                diffusion_models
                unet
    embeddings: embeddings\
    loras: loras\
    upscale_models: upscale_models\
    vae: vae\
    # custom vibevoice node dumps here
    vibevoice: vibevoice\

INSTALLATION

AMD Driver,

Documentation listing ROCm Version History for each version

ROCm Wheels 7.2.1 P3.12

PREVIEW

Preview ROCm Wheels

As of 2026-04-06 Preview wheels use ROCm 7.10

RUN

uv run main.py

PORTABLE

It's very convenient, but it works somewhat different

ComfyUI manager

LAUNCH ARGUMENTS

ROCm cannot handle memory properly, there need to be mitigations 2026-07-26 testing

FLAGS

--windows-standalone-build core flag to make ROCm work

--disable-api-nodes remove the useless node that need cloud APIs to work

--enable-dynamic-vram weird interaction with ROCm, on big models cause timeouts

--enable-manager core flag to enable the manager. can omit it to load somewhat faster

--use-pytorch-cross-attention ??? I think it's not needed as it's already inside

--disable-smart-memory Important

Launch without Manager

Remove API nodes, but disables the manager

Uses ROCm workaround that prevent driver crashes

set COMFYUI_ENABLE_MIOPEN=1
set MIOPEN_FIND_MODE=2
.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-smart-memory --disable-api-nodes

Launch with Manager

The Manager needs the API to run

Uses ROCm workaround that prevent driver crashes

set COMFYUI_ENABLE_MIOPEN=1
set MIOPEN_FIND_MODE=2
.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-smart-memory --enable-manager

WORKFLOWS

Test VAE Decode

Zimage

Strong fast model, it works amazingly well

uv run main.py --windows-standalone-build --disable-smart-memory

• First execution: 29s
• Repeat with changed prompt: 19s
• Repeat: 16s

uv run main.py --windows-standalone-build --use-pytorch-cross-attention

• First execution: 32s
• Repeat: 18s

Performance

got prompt
Using split attention in VAE
Using split attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load ZImageTEModel_
loaded completely; 22392.36 MB usable, 7672.25 MB loaded, full load: True
gguf qtypes: F32 (245), F16 (24), Q4_K (120), Q6_K (30), BF16 (4), Q5_K (30)
model weight dtype torch.bfloat16, manual cast: None
model_type FLOW
Requested to load Lumina2
loaded completely; 22296.06 MB usable, 4834.06 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████| 9/9 [00:12<00:00,  1.39s/it]
Requested to load AutoencodingEngine
loaded completely; 11319.68 MB usable, 159.87 MB loaded, full load: True
Prompt executed in 28.92 seconds
got prompt
Requested to load ZImageTEModel_
loaded completely; 22392.36 MB usable, 7672.25 MB loaded, full load: True
Requested to load Lumina2
loaded completely; 22296.06 MB usable, 4834.06 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████| 9/9 [00:12<00:00,  1.40s/it]
Requested to load AutoencodingEngine
loaded completely; 11319.68 MB usable, 159.87 MB loaded, full load: True
Prompt executed in 19.32 seconds
got prompt
Requested to load Lumina2
loaded completely; 22296.06 MB usable, 4834.06 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████| 9/9 [00:12<00:00,  1.41s/it]
Requested to load AutoencodingEngine
loaded completely; 11319.68 MB usable, 159.87 MB loaded, full load: True
Prompt executed in 15.71 seconds
got prompt
Requested to load ZImageTEModel_
loaded completely; 22392.36 MB usable, 7672.25 MB loaded, full load: True
Requested to load Lumina2
loaded completely; 22296.06 MB usable, 4834.06 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████| 9/9 [00:12<00:00,  1.40s/it]
Requested to load AutoencodingEngine
loaded completely; 11319.68 MB usable, 159.87 MB loaded, full load: True
Prompt executed in 19.17 seconds

Qwen Edit

Strong model that is very good at executing edit instructions and taking multiple reference images.

uv run main.py --windows-standalone-build --disable-smart-memory

• First execution: 60s
• Repeat: 46s

Performance

got prompt
Using split attention in VAE
Using split attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
gguf qtypes: Q8_0 (198), F32 (141)
Dequantizing token_embd.weight to prevent runtime OOM.
Attenpting to find mmproj file for text encoder...
Using mmproj 'qwen2.5-vl-7b-instruct-q8_0-mmproj-fp16.gguf' for text encoder 'qwen2.5-vl-7b-instruct-q8_0.gguf'.
gguf qtypes: F32 (291), F16 (228)
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load WanVAE
loaded completely; 20161.56 MB usable, 242.03 MB loaded, full load: True
Requested to load QwenImageTEModel_
loaded completely; 22392.36 MB usable, 8946.75 MB loaded, full load: True
gguf qtypes: F32 (1088), BF16 (6), Q6_K (258), Q8_0 (2), Q5_K (20), Q4_K (560)
model weight dtype torch.bfloat16, manual cast: None
model_type FLUX
Requested to load QwenImage
loaded completely; 22033.91 MB usable, 12738.98 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:25<00:00,  6.31s/it]
Requested to load WanVAE
loaded completely; 18811.56 MB usable, 242.03 MB loaded, full load: True
Prompt executed in 58.36 seconds
got prompt
Requested to load QwenImage
loaded completely; 22033.91 MB usable, 12738.98 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:24<00:00,  6.24s/it]
Requested to load WanVAE
loaded completely; 18811.56 MB usable, 242.03 MB loaded, full load: True
Prompt executed in 45.39 seconds
got prompt
Requested to load QwenImage
loaded completely; 22033.91 MB usable, 12738.98 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:25<00:00,  6.29s/it]
Requested to load WanVAE
loaded completely; 18811.56 MB usable, 242.03 MB loaded, full load: True
Prompt executed in 43.96 seconds
got prompt
Requested to load QwenImage
loaded completely; 22033.91 MB usable, 12738.98 MB loaded, full load: True
100%|███████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:25<00:00,  6.29s/it]
Requested to load WanVAE
loaded completely; 18811.56 MB usable, 242.03 MB loaded, full load: True
Prompt executed in 44.29 seconds

uv run main.py --windows-standalone-build --use-pytorch-cross-attention

• First execution: 98s
• Repeat: 32s

Performance

got prompt
100%|███████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:35<00:00,  8.79s/it]
Requested to load WanVAE
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 242.00 MB offloaded, 22.78 MB buffer reserved, lowvram patches: 0
Prompt executed in 98.22 seconds
Requested to load QwenImage
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 12738.98 MB offloaded, 224.60 MB buffer reserved, lowvram patches: 0
100%|███████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:31<00:00,  7.82s/it]
Requested to load WanVAE
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 242.00 MB offloaded, 22.78 MB buffer reserved, lowvram patches: 0
Prompt executed in 31.98 seconds
Requested to load QwenImage
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 12738.98 MB offloaded, 224.60 MB buffer reserved, lowvram patches: 0
100%|███████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:31<00:00,  7.80s/it]
Requested to load WanVAE
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 242.00 MB offloaded, 22.78 MB buffer reserved, lowvram patches: 0
Prompt executed in 31.91 seconds
Requested to load QwenImage
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 12738.98 MB offloaded, 224.60 MB buffer reserved, lowvram patches: 0
100%|███████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:31<00:00,  7.84s/it]
Requested to load WanVAE
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 242.00 MB offloaded, 22.78 MB buffer reserved, lowvram patches: 0
Prompt executed in 32.06 seconds
Requested to load QwenImage
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 12738.98 MB offloaded, 224.60 MB buffer reserved, lowvram patches: 0
100%|███████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:31<00:00,  7.83s/it]
Requested to load WanVAE
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 242.00 MB offloaded, 22.78 MB buffer reserved, lowvram patches: 0
Prompt executed in 32.02 seconds

Hunyuan 3D 2.0 MV

This workflow uses A background removal model, followed by Qwen Edit Q4 to generate the back, followed by Hunyuan 2.0 multiview to generate the 3D model

Logs

LLM Image/Text safetensor

This uses ComfyUI LLM inference, with the same LLM Clip Krea2 uses. It's pragmatic as it's already loaded. Qwen 3 VL 4B FP8 safetensor

Performance is horrible. It does 1/4 of the speed that LM Studio does on Qwen 3.5 9B Q4 gguf.

Background Removal

It's a small model native to ComfyUI now, doesn't need third party packages

https://github.com/OrsoEric/comfyui-orso-character-sheet-generator

My first custom node for D&D 5E character sheet cards

EOL

Performance

xxx