ComfyUI seems to ignore the --reserve-vram and/or --disable-smart-memory ? Is there anything going wrong ?
- Dominant language
- Python
- Stars
- 133k
- Forks
- 15.7k
- Avg merge
- 1d 6h
- Merged PRs (30d)
- 155
Description
### Your question
So I am using everything to reduce the vram usage amount.
`ComfyUI_windows_portable>.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --lowvram --force-fp16 --reserve-vram 2.4 --disable-smart-memory`
However I keep seeing that the memory usage of the loaded model is increased ->
and then I am getting memory errors in later generations as shown in the logs and the generation slows down significantly. Sometimes I get an OOM error and the generation dies.
I am basically using a XY generation of pictures with different guidance values and base_shifts. I am using the flux-q4-ks.gguf model (with clip models t5_v1.1_q8 and vit-l-text-detail ) and this [lora](https://civitai.com/models/730373/hyper-realism-lora-by-aidma-flux).
I can provide the workflow if needed for reproduction.
Seems like the issue is very close to #5958 , #5385 and #4318.
However I am using an SamplerCustomAdvanced and a nvidia card. So just wondering if this is a different problem. Or I am doing something wrong.
### Logs
```powershell
[2025-01-01 20:36:33.946]
[ARequested to load Flux
[2025-01-01 20:36:34.009] 0 models unloaded.
[2025-01-01 20:36:34.118] loaded partially 6320.525634765625 6320.427978515625 0
[2025-01-01 20:36:34.123] Attempting to release mmap (44)
[2025-01-01 20:36:34.259]
[2025-01-01 20:36:34.260] [A
[2025-01-01 20:36:34.261] [AERROR lora diffusion_model.single_blocks.0.linear2.weight Allocation on device
[2025-01-01 20:38:46.045]
[2025-01-01 20:38:46.045] [A
[2025-01-01 20:40:26.966] [A
[2025-01-01 20:42:00.008] [A
[2025-01-01 20:43:26.563] [A
[2025-01-01 20:45:06.902] [A
[2025-01-01 20:46:33.755] [A
[2025-01-01 20:48:04.298] [A
[2025-01-01 20:49:35.220] [A
[2025-01-01 20:52:40.229] [A
[2025-01-01 20:55:56.443] [A
[2025-01-01 20:58:42.092] [A
[2025-01-01 21:01:50.056] [A
[2025-01-01 21:02:17.609] [A
bosh3: 100%|█████████████████████████████████████████████████████████████████| 1/1 [25:43<00:00, 1543.35s/it]
bosh3: 100%|█████████████████████████████████████████████████████████████████| 1/1 [25:43<00:00, 1543.35s/it]
[2025-01-01 21:02:17.612]
[2025-01-01 21:02:17.613]
[ARequested to load Flux
[2025-01-01 21:02:17.679] 0 models unloaded.
[2025-01-01 21:02:17.759] loaded partially 6384.427978515625 6383.381103515625 0
[2025-01-01 21:02:17.765] Attempting to release mmap (40)
[2025-01-01 21:02:17.869]
[2025-01-01 21:02:17.869] [A
[2025-01-01 21:02:17.870] [AERROR lora diffusion_model.single_blocks.0.linear2.weight Allocation on device
[2025-01-01 21:04:24.271]
```
### Other
Total VRAM 8192 MB, total RAM 40352 MB
pytorch version: 2.5.1+cu124
Forcing FP16.
Set vram state to: LOW_VRAM
Disabling smart memory management
Device: cuda:0 NVIDIA GeForce RTX 3070 Laptop GPU : cudaMallocAsync
Using pytorch cross attention
### Loading: ComfyUI-Manager (V2.55.5)
### ComfyUI Version: v0.3.7-13-g44db978 | Released on '2024-12-10'
Contributor guide
Assessment
This issue has not been assessed yet.