OOM on LTX Tiled Sampler Node
Running 5070ti 16GB VRAM + 64GB RAM. First pass works really well, but I always get OOM on the LTX Tiled Sampler Node after increasing the video length to 20 seconds. I tried reducing the properties down, but still get OOM. I have --lowvram set on comfyui.
Do you get OOM on a normal WF too? This should be more memory efficient since it is actually sampling less and it's set up to use comfy memory allocation. --lowvram might cause it try it without and see if it will just freeze. I also run LTX with --disable-pinned-memory for blackwell.
Removed --lowvram and added --disable-pinned-memory. LTX Tiled Sampler node OOM after 95 seconds. Is there a way to save off the latent to a file, and then load it to that node?
I would just choose bypass on it which is in the update. If that doesnt work replace it with a normal sampler, but if that OOM's too it was a size/resolution issue for the hardware.
I need to implement an actual Cuda pipeline on it to see what kind of gains that can get.
With a 5070ti 16GB VRAM + 128GB RAM I can get 28 seconds out of 10Eros_10SNodes_I2V_v3_TiledSampler.json. When trying 10Eros_10SNodes_LikenessGuideHelper_I2V_v3.2.json I OOM on the LTX Tiled Sampler if i try to do 12 seconds. Haven't tried fewer seconds than that yet.
With a 5070ti 16GB VRAM + 128GB RAM I can get 28 seconds out of 10Eros_10SNodes_I2V_v3_TiledSampler.json. When trying 10Eros_10SNodes_LikenessGuideHelper_I2V_v3.2.json I OOM on the LTX Tiled Sampler if i try to do 12 seconds. Haven't tried fewer seconds than that yet.
Same size and resolutions right?
Same here. On my 3090, stage 3 of the tiled sampler gets stuck at 99% VRAM. It works, but it takes anywhere from 30 minutes to forever to complete the third pass.
With a 5070ti 16GB VRAM + 128GB RAM I can get 28 seconds out of 10Eros_10SNodes_I2V_v3_TiledSampler.json. When trying 10Eros_10SNodes_LikenessGuideHelper_I2V_v3.2.json I OOM on the LTX Tiled Sampler if i try to do 12 seconds. Haven't tried fewer seconds than that yet.
Same size and resolutions right?
Yes, both at 736x1280.
This morning I had another look and noticed that 10Eros_10SNodes_I2V_v3_TiledSampler.json has 'bypass_tiling' checked by default. I checked it in 10Eros_10SNodes_LikenessGuideHelper_I2V_v3.2.json and it put out 28 seconds successfully. Incidentally, 10Eros_10SNodes_LikenessGuideHelper_I2V_v3.2.json output was quite blurry.