HunyuanVideo-1.5 120 frames 12G oom

[zwukong]

81 frames is ok, while 120 frames can not run properly. Vram can sometimes over max vram and then speed becomes very slow in 121 frams. wan2.2 is ok,but hunyuan not . 3.12,2.8

[rattus128]

Vram can sometimes over max vram and then speed becomes very slow in 121 frams. wan2.2 is ok,but hunyuan not . 3.12,2.8

Hi. Can we get you hardware specs, and the video resolution of what you are generating. If you can attach a workflow that's ideal.

[zwukong]

4070 12G, 480p, your example workflow ,hunyuanvideo1.5_480p_i2v_cfg_distilled_fp8_scaled ,8 steps

[rattus128]

4070 12G, 480p, your example workflow ,hunyuanvideo1.5_480p_i2v_cfg_distilled_fp8_scaled ,8 steps

Thanks. I'll give it a go on my 3060.

[rattus128]

@zwukong no issues for me at 480p121f on the 3060 12G. Few things to try from here though.

I notice the VAE uses more VRAM than the Model and I got forced to go back to the tiled VAE. Can you confirm whether the KSampler or VAE Decode is the problem?

If its a VAE problem, try manually tiling. The workflow has an example VAE Decode tiled node, hook that up instead of the default VAE Decode.

If you are on Windows, can you change this setting to "Prefer no Sysmem Fallback":

Run again, and if it is OOMing the way you describe, this hopefully forces it to crash with OOM report instead of run really slow. Then paste me the log and we can see where the OOM is happening.

This setting can also help the automatically tiler figure out it needs to tile.

For any OOMing run, can we get your load statistics? In the UI, expand the console and paste the content. It should read something like this:

got prompt
Requested to load CLIPVisionModelProjection
loaded completely; 3196.40 MB usable, 787.72 MB loaded, full load: True
Requested to load AutoencodingEngine
loaded partially; 1351.83 MB usable, 1351.83 MB loaded, 1052.61 MB offloaded, lowvram patches: 0
Requested to load HunyuanVideo15
loaded partially; 7301.02 MB usable, 7301.02 MB loaded, 639.97 MB offloaded, lowvram patches: 0
  0%|          | 0/20 [00:00<?, ?it/s]Interrupting prompt 030cdc54-0921-4c6f-a7b2-9bf18528090b
  0%|          | 0/20 [01:41<?, ?it/s]
Processing interrupted
Prompt executed in 106.10 seconds
got prompt
0 models unloaded.
loaded partially: 7300.95 MB loaded, lowvram patches: 0
 75%|███████▌  | 6/8 [06:15<02:05, 62.66s/it]

[zwukong]

in my experience，if the total size over the max vram ,it will become slow.

[zwukong]

4 steps cost 2 minutes, very slow

[rattus128]

4 steps cost 2 minutes, very slow

You have lowvram patches on load. Do you have Loras?

You might have to send me the complete workflow as I have similar load numbers to you but I don't go over 11GB of VRAM.

[kanarch66]

vae decode is the problem

loaded partially; 128.00 MB usable, 127.84 MB loaded, 2276.62 MB offloaded, lowvram patches: 0 Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.

[zwukong]

hunyuan_video .json what i mean is shared vram ,11G vram + 2G shared vram >12G, speed will become slow

[zwukong]

@rattus128 i think it is the same problem as flux2

[rwfsmith]

I hit the same problem during VAE decode on my 5090 if I try to run the native 1080p upscale workflow from the default templates.