Flux.2[Klein-4b] using extremely high VRAM (Overflow)

[Kranos555]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Your question

Updated comfyui(portable) to the latest version(0.9.2) to try out Flux.2 klein 4b but during inference it completely fills up my Vram (4gb) spilling into the shared gpu memory making the generations really sloow.
For record other models like Z-image-turbo, Qwen-image, Wan 2.2 and even LTX-2 runs perfectly fine without overflowing out from the available vram.

You guys are already doing a great job with memory management and it would be really helpful if a solution for this can be found.
Thanks..

Logs

Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
gguf qtypes: F32 (145), Q6_K (1), Q2_K (144), Q3_K (72), Q4_K (36)
Dequantizing token_embd.weight to prevent runtime OOM.
Requested to load Flux2TEModel_
loaded completely;  2024.07 MB loaded, full load: True
CLIP/text encoder model load device: cpu, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load Flux2TEModel_
gguf qtypes: F32 (60), Q6_K (30), Q5_K (50), BF16 (9)
model weight dtype torch.bfloat16, manual cast: torch.float32
model_type FLUX
Potential memory leak detected with model Flux2TEModel_, doing a full garbage collect, for maximum performance avoid circular references in the model code.
Requested to load Flux2
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 3038.95 MB offloaded, 822.45 MB buffer reserved, lowvram patches: 0
  0%|                                                                                            | 0/4 [00:00<?, ?it/s]

Other

No response

[Kranos555]

Added the log for more info. Potential memory leak.

[Kranos555]

I tried using it with a fresh new installation of comfyui and the problem still persists.

[Kranos555]

The memory usage keeps overflowing. No other model behave this way.

[zwukong]

4G LTX2? serious?😄 reserved vram can help you forever,there is a comfyui plugin too

[Kranos555]

4G LTX2? serious?😄 reserved vram can help you forever,there is a comfyui plugin too

Yeah though painfully slow, 4gb works with ltx2. And yeah i tried using --reserve-vram 4.0 alongwith --novram --cache-none and --disable-pinned-memory , these help with ltx2 and all other models but no luck with flux.2-klein. It runs slower than Ltx2 and even slower than wan 2.2 on my system.

[zwukong]

keep adding the reserve-vram params, until shared vram is 0

[Kranos555]

keep adding the reserve-vram params, until shared vram is 0

Tried it but no difference. Like in the screenshot above, the memory usage keeps increasing gradually which i don't think is typical behaviour.

[Kranos555]

Update: The gpu doesn't seem to be the problem because i tried running using diffusers and it works fine..

[Kranos555]

Update: It works now. Turns out i was using the wrong workflow, the empty latent image node was of sd3 and not flux2. Switched those up and now works fine.