You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
RTX 4060 with 16GB RAM. I know this is a low-end machine for video generation but over 40 mins for one sec? People's 3060s are faster than this. Is there a way to optimize?
What's your generation settings? Seems that you can decrease reserved VRAM to less than 6GB (default). Still I did not get the advertised speed on 3060 laptop. Not pretty sure how those guys got the number (5 second in 50 minutes compared to 1.5 hours and still not done).
What's your generation settings? Seems that you can decrease reserved VRAM to less than 6GB (default). Still I did not get the advertised speed on 3060 laptop. Not pretty sure how those guys got the number (5 second in 50 minutes compared to 1.5 hours and still not done).
All default settings rn, could it be due to outdated CUDA in the pack
All default settings rn, could it be due to outdated CUDA in the pack
Doesn't seems like the problem. If you'd like to optimize on VRAM usage, decrease the reserved VRAM size seems to be the only way.
BTW it seems that the current resource management system is home-made by the author and not the diffusers' inplementation. This can be the reason for the slower speed. I might try to implement diffusers' layer offload to see if it is different.
I also have ridiculously show speeds for some reason with a decently spec'd pc. Running a RX 3050, 12th gen i9-12900K, 64 gb memory
Been running for 1.4 hrs now (and going) and it only recently showed the 1 sec preview.
I also have ridiculously show speeds for some reason with a decently spec'd pc. Running a RX 3050, 12th gen i9-12900K, 64 gb memory Been running for 1.4 hrs now (and going) and it only recently showed the 1 sec preview.
๐ข
Reinstalled Python(even though supposedly this uses its own built-in environment) and sage, turned on teacache and now at approx. 44s/it. Still slow and nowhere near advertised speed but slightly better
I'm in the same boat. I'm on a RTX 3060 and still doing about 1 second every 45 mins. I've been messing around with the sliders but haven't noticed much change yet.
I guess the advertised speed is only for gpus with 16gb minimum, this assumption is based on the different posts I have come accross on reddit pages. Any vram below this, I think it takes more time than wan 2.1 safetensor file
Hi people - if you get unreasonably slow (if more than 2x slower than my tests) speed then it is possible that (1) your memory speed is bottlenecked, or (2) your swap memory is bottlenecked.
Usually (2) is more common than (1).
to solve (1), you take a look at your memory speed - if your system memory is too old, consider get a newer one. I believe my laptop is using like this. one can use something older but if too old or low end then it may cause bottleneck.
to solve (2), take the below steps:
look at this Image and find out whether the main swap is using HDD or SSD. If it is using HDD then you may want to change the setting to force the swap to use SSD.
if the swap is already using SSD, then make sure that the used disk has at least 100GB free space.
if this does not work then the bottleneck is in the SSD, maybe the SSD is too old or low end.
If the SSD is new and good but you still have problem, then it maybe bottlenecked by motherboard, but this is not very common.
also the speed reported in the first iteration is not accurate. it will drop after some warmup.
Reacted by Frankthetank02-rgb, gordy12gg, nitinmukesh, anw777, furana93, Othneil Drew, 10ANT, zjxdiu, Attashe, Hoseok Yoo and 8 more
Moved the program to another drive and the speed is now much bearable (~45sec/frame). Seems that hard drive speed plays a large factor if you do not have the VRAM to stuff all the things in. Would definitely suggest anyone with slower speeds to try other drives lying around.
However, I still believes that the current implementation does not use swap and it simply loads the required parameters from the file. I used to have the whole folder on my D drive and when it is running what I saw was surging D drive utilization and nearly nothing from C drive. I currently configured swap to only be on C drive.
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
The next step to initialize after this is really slow like around 10 minute. Is it because I set my Virtual Ram to 32 GB? Will it be better if I use the "Automatically manage paging file size for all drivers"?
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
The next step to initialize after this is really slow like around 10 minute. Is it because I set my Virtual Ram to 32 GB? Will it be better if I use the "Automatically manage paging file size for all drivers"?
I have a 3060 6gb vram, 32gb ram.
The current implementation seems to not be using tiled VAE decode, and with your setup it seems like it is just the norm. The 10-minute wait is just for decoding the diffused samples. Would be great if tiled VAE is implemented to speed up the process.
Hi people - if you get unreasonably slow (if more than 2x slower than my tests) speed then it is possible that (1) your memory speed is bottlenecked, or (2) your swap memory is bottlenecked.
Usually (2) is more common than (1).
to solve (1), you take a look at your memory speed - if your system memory is too old, consider get a newer one. I believe my laptop is using like this. one can use something older but if too old or low end then it may cause bottleneck.
to solve (2), take the below steps:
look at this Image and find out whether the main swap is using HDD or SSD. If it is using HDD then you may want to change the setting to force the swap to use SSD.
if the swap is already using SSD, then make sure that the used disk has at least 100GB free space.
if this does not work then the bottleneck is in the SSD, maybe the SSD is too old or low end.
If the SSD is new and good but you still have problem, then it maybe bottlenecked by motherboard, but this is not very common.
also the speed reported in the first iteration is not accurate. it will drop after some warmup.
This helped with debugging for me. Speeds are now comparable to what the readme says. It took about 9mins for each frame on a RTX 3050. I only have an SSD. My problem was that I was running on WSL2. This in itself shouldn't be a problem, but there's probably something wrong with how I enabled hardware acceleration because on previous tests the GPU wouldn't be under any significant load.
My RAM usage is about 50Gb
The GPU memory usage is about 6Gb at any case.
If you have less than 64Gb of RAM it is more probable you're using disk memory. If this is the case and you don't have a good SSD then you're in trouble.
So stop complaining. You already are at your setup limits.
Rule:
50Gb of RAM and 6Gb of VRAM = 5secs in 10min (worst case)
No RAM = DISK USAGE.
Disk Usage = Good SSD
Bad SSD = TROUBLE
No SSD = YOU ARE DOOMED
Even at worst case this is far the best tool you will find to create video with low requirements.
I said LOW ... not NONE.
So.. be patient or give up.
( or buy a better setup )
Just tried this out. I created a docker for it, running in ubuntu. Have 6xRTX3090, 32gb ram, running on old/slow sata drives. Probably took an hour or more to run the first time for a 5sec vid. CPU pegged at 100 for lots of it, lots of drive activity. Rebooted and ran everything fresh, this time cpu not pegging, but slowing system down. Not sure if should be running faster... may try something else since have more vram to play with.
$ docker logs framepack-ui
Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn']
Xformers is not installed!
Flash Attn is not installed!
Sage Attn is not installed!
Namespace(share=False, server='0.0.0.0', port=None, inbrowser=False)
Free VRAM 17.42950439453125 GB
High-VRAM Mode: False
Downloading shards: 100%|โโโโโโโโโโ| 4/4 [02:37<00:00, 39.43s/it]
Loading checkpoint shards: 100%|โโโโโโโโโโ| 4/4 [00:01<00:00, 3.39it/s]
Fetching 3 files: 100%|โโโโโโโโโโ| 3/3 [03:46<00:00, 75.40s/it]
Loading checkpoint shards: 100%|โโโโโโโโโโ| 3/3 [00:02<00:00, 1.28it/s]
transformer.high_quality_fp32_output_for_inference = True
To create a public link, set share=True in launch().
Unloaded DynamicSwap_LlamaModel as complete.
Unloaded CLIPTextModel as complete.
Unloaded SiglipVisionModel as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete.
Loaded CLIPTextModel to cuda:0 as complete.
Unloaded CLIPTextModel as complete.
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Loaded SiglipVisionModel to cuda:0 as complete.
latent_padding_size = 27, is_last_section = False
Unloaded SiglipVisionModel as complete.
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [15:45<00:00, 37.84s/it]
/usr/local/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Decoded. Current latent shape torch.Size([1, 16, 9, 80, 76]); pixel shape torch.Size([1, 3, 33, 640, 608])
latent_padding_size = 18, is_last_section = False
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [10:26<00:00, 25.05s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Decoded. Current latent shape torch.Size([1, 16, 18, 80, 76]); pixel shape torch.Size([1, 3, 69, 640, 608])
latent_padding_size = 9, is_last_section = False
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [10:24<00:00, 25.00s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Decoded. Current latent shape torch.Size([1, 16, 27, 80, 76]); pixel shape torch.Size([1, 3, 105, 640, 608])
latent_padding_size = 0, is_last_section = True
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [10:26<00:00, 25.04s/it]
Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn']
Xformers is not installed!
Flash Attn is not installed!
Sage Attn is not installed!
Namespace(share=False, server='0.0.0.0', port=None, inbrowser=False)
Free VRAM 23.0460205078125 GB
High-VRAM Mode: False
Downloading shards: 100%|โโโโโโโโโโ| 4/4 [00:00<00:00, 473.76it/s]
Loading checkpoint shards: 100%|โโโโโโโโโโ| 4/4 [00:07<00:00, 1.94s/it]
Fetching 3 files: 100%|โโโโโโโโโโ| 3/3 [00:00<00:00, 2080.51it/s]
Loading checkpoint shards: 100%|โโโโโโโโโโ| 3/3 [00:06<00:00, 2.28s/it]
transformer.high_quality_fp32_output_for_inference = True
To create a public link, set share=True in launch().
Unloaded DynamicSwap_LlamaModel as complete.
Unloaded CLIPTextModel as complete.
Unloaded SiglipVisionModel as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete.
Loaded CLIPTextModel to cuda:0 as complete.
Unloaded CLIPTextModel as complete.
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Loaded SiglipVisionModel to cuda:0 as complete.
latent_padding_size = 27, is_last_section = False
Unloaded SiglipVisionModel as complete.
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [14:46<00:00, 35.46s/it]
Just tried this out. I created a docker for it, running in ubuntu. Have 6xRTX3090, 32gb ram, running on old/slow sata drives. Probably took an hour or more to run the first time for a 5sec vid. CPU pegged at 100 for lots of it, lots of drive activity. Rebooted and ran everything fresh, this time cpu not pegging, but slowing system down. Not sure if should be running faster... may try something else since have more vram to play with.
$ docker logs framepack-ui Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is not installed! Flash Attn is not installed! Sage Attn is not installed! Namespace(share=False, server='0.0.0.0', port=None, inbrowser=False) Free VRAM 17.42950439453125 GB High-VRAM Mode: False Downloading shards: 100%|โโโโโโโโโโ| 4/4 [02:37<00:00, 39.43s/it] Loading checkpoint shards: 100%|โโโโโโโโโโ| 4/4 [00:01<00:00, 3.39it/s] Fetching 3 files: 100%|โโโโโโโโโโ| 3/3 [03:46<00:00, 75.40s/it] Loading checkpoint shards: 100%|โโโโโโโโโโ| 3/3 [00:02<00:00, 1.28it/s] transformer.high_quality_fp32_output_for_inference = True
To create a public link, set share=True in launch(). Unloaded DynamicSwap_LlamaModel as complete. Unloaded CLIPTextModel as complete. Unloaded SiglipVisionModel as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete. Loaded CLIPTextModel to cuda:0 as complete. Unloaded CLIPTextModel as complete. Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Loaded SiglipVisionModel to cuda:0 as complete. latent_padding_size = 27, is_last_section = False Unloaded SiglipVisionModel as complete. Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [15:45<00:00, 37.84s/it] /usr/local/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec warnings.warn( Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Decoded. Current latent shape torch.Size([1, 16, 9, 80, 76]); pixel shape torch.Size([1, 3, 33, 640, 608]) latent_padding_size = 18, is_last_section = False Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [10:26<00:00, 25.05s/it] Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Decoded. Current latent shape torch.Size([1, 16, 18, 80, 76]); pixel shape torch.Size([1, 3, 69, 640, 608]) latent_padding_size = 9, is_last_section = False Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [10:24<00:00, 25.00s/it] Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Decoded. Current latent shape torch.Size([1, 16, 27, 80, 76]); pixel shape torch.Size([1, 3, 105, 640, 608]) latent_padding_size = 0, is_last_section = True Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [10:26<00:00, 25.04s/it] Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is not installed! Flash Attn is not installed! Sage Attn is not installed! Namespace(share=False, server='0.0.0.0', port=None, inbrowser=False) Free VRAM 23.0460205078125 GB High-VRAM Mode: False Downloading shards: 100%|โโโโโโโโโโ| 4/4 [00:00<00:00, 473.76it/s] Loading checkpoint shards: 100%|โโโโโโโโโโ| 4/4 [00:07<00:00, 1.94s/it] Fetching 3 files: 100%|โโโโโโโโโโ| 3/3 [00:00<00:00, 2080.51it/s] Loading checkpoint shards: 100%|โโโโโโโโโโ| 3/3 [00:06<00:00, 2.28s/it] transformer.high_quality_fp32_output_for_inference = True
To create a public link, set share=True in launch(). Unloaded DynamicSwap_LlamaModel as complete. Unloaded CLIPTextModel as complete. Unloaded SiglipVisionModel as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete. Loaded CLIPTextModel to cuda:0 as complete. Unloaded CLIPTextModel as complete. Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Loaded SiglipVisionModel to cuda:0 as complete. latent_padding_size = 27, is_last_section = False Unloaded SiglipVisionModel as complete. Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [14:46<00:00, 35.46s/it]
Are you running it on Docker?????
Damn....
I quit.
Just tried this out. I created a docker for it, running in ubuntu. Have 6xRTX3090, 32gb ram, running on old/slow sata drives. Probably took an hour or more to run the first time for a 5sec vid. CPU pegged at 100 for lots of it, lots of drive activity. Rebooted and ran everything fresh, this time cpu not pegging, but slowing system down. Not sure if should be running faster... may try something else since have more vram to play with.
$ docker logs framepack-ui Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is not installed! Flash Attn is not installed! Sage Attn is not installed! Namespace(share=False, server='0.0.0.0', port=None, inbrowser=False) Free VRAM 17.42950439453125 GB High-VRAM Mode: False Downloading shards: 100%|โโโโโโโโโโ| 4/4 [02:37<00:00, 39.43s/it] Loading checkpoint shards: 100%|โโโโโโโโโโ| 4/4 [00:01<00:00, 3.39it/s] Fetching 3 files: 100%|โโโโโโโโโโ| 3/3 [03:46<00:00, 75.40s/it] Loading checkpoint shards: 100%|โโโโโโโโโโ| 3/3 [00:02<00:00, 1.28it/s] transformer.high_quality_fp32_output_for_inference = True
To create a public link, set share=True in launch(). Unloaded DynamicSwap_LlamaModel as complete. Unloaded CLIPTextModel as complete. Unloaded SiglipVisionModel as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete. Loaded CLIPTextModel to cuda:0 as complete. Unloaded CLIPTextModel as complete. Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Loaded SiglipVisionModel to cuda:0 as complete. latent_padding_size = 27, is_last_section = False Unloaded SiglipVisionModel as complete. Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [15:45<00:00, 37.84s/it] /usr/local/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec warnings.warn( Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Decoded. Current latent shape torch.Size([1, 16, 9, 80, 76]); pixel shape torch.Size([1, 3, 33, 640, 608]) latent_padding_size = 18, is_last_section = False Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [10:26<00:00, 25.05s/it] Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Decoded. Current latent shape torch.Size([1, 16, 18, 80, 76]); pixel shape torch.Size([1, 3, 69, 640, 608]) latent_padding_size = 9, is_last_section = False Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [10:24<00:00, 25.00s/it] Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Decoded. Current latent shape torch.Size([1, 16, 27, 80, 76]); pixel shape torch.Size([1, 3, 105, 640, 608]) latent_padding_size = 0, is_last_section = True Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [10:26<00:00, 25.04s/it] Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn'] Xformers is not installed! Flash Attn is not installed! Sage Attn is not installed! Namespace(share=False, server='0.0.0.0', port=None, inbrowser=False) Free VRAM 23.0460205078125 GB High-VRAM Mode: False Downloading shards: 100%|โโโโโโโโโโ| 4/4 [00:00<00:00, 473.76it/s] Loading checkpoint shards: 100%|โโโโโโโโโโ| 4/4 [00:07<00:00, 1.94s/it] Fetching 3 files: 100%|โโโโโโโโโโ| 3/3 [00:00<00:00, 2080.51it/s] Loading checkpoint shards: 100%|โโโโโโโโโโ| 3/3 [00:06<00:00, 2.28s/it] transformer.high_quality_fp32_output_for_inference = True
To create a public link, set share=True in launch(). Unloaded DynamicSwap_LlamaModel as complete. Unloaded CLIPTextModel as complete. Unloaded SiglipVisionModel as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete. Loaded CLIPTextModel to cuda:0 as complete. Unloaded CLIPTextModel as complete. Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. Unloaded AutoencoderKLHunyuanVideo as complete. Loaded SiglipVisionModel to cuda:0 as complete. latent_padding_size = 27, is_last_section = False Unloaded SiglipVisionModel as complete. Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB 100%|โโโโโโโโโโ| 25/25 [14:46<00:00, 35.46s/it]
Are you running it on Docker????? Damn.... I quit.
yah, took a few times building it to get the python dependencies, but ended up working. kinda new to this, so lots of experimenting. edit: what made the docker build process easier was having an llm explaining things along the way.
5090
linux
sage 2.1
teacache on
10 sec video 6-7 mins
image resolution 1344x1728
gradio version (not comfyui version)
(but i have a different issue: action does not start until 05 secs for a 10 secs video / 09 secs until a 20 sec video )
Every of four steps now takes about 4:15 / ~10 s/it = ~ 17 min total for 5 sec video
So increasing RAM size (up to certain limit) doubled the speed!
From these 80 GB of memory used about 47 GB, free 31 GB, so it seems 64 GB must be enough. (Though I'd prefer to utilize this extra memory for speed or quality.)
I'm also using 4060 6gb with 16gb ram. on Windows I tried to run this and I was able to get 80/s . which was really horrible so I though why shouldn't give a try on linux so i did install ubuntu (with same environment like nvidia, cuda and cudnn drivers installed and same swap size) I was able to get around 26/s . I have a question is there still room to increase it? or its max what i can get? because its taking like avg 12 min for 1 sec clip only.
I'm also using 4060 6gb with 16gb ram. on Windows I tried to run this and I was able to get 80/s . which was really horrible so I though why shouldn't give a try on linux so i did install ubuntu (with same environment like nvidia, cuda and cudnn drivers installed and same swap size) I was able to get around 26/s . I have a question is there still room to increase it? or its max what i can get? because its taking like avg 12 min for 1 sec clip only.
Hello everyone,
I am facing a similar problem than @furana93 with my configuration:
-RTX5090
-Intel(R) Core(TM) Ultra 9 285K (3.70 GHz)
-96go ram (4800MT/s)
-motherboad:ASUS ROG STRIX Z890-E GAMING WIFI
-Framepack folder on SSD (around 4000mo/s)
I do not think bottleneck is problem but maybe I am wrong.
As you can see below I have no problem to generate the 25 steps (roughly it takes around 30sec). The problem appeared after, at the steps : Loaded/Unloaded AutoencoderKLHunyuanVideo.
It took a very very long time to complete/start and I don't know why. Any ideas?
I also noticed that the time between each 25 steps generation increase at each iteration (for instance: 2min/4min/20min).
Finally, i need roughly 30min to generate 5sec video that is not a standard with this configuration I guess?
To create a public link, set share=True in launch().
Unloaded DynamicSwap_LlamaModel as complete.
Unloaded CLIPTextModel as complete.
Unloaded SiglipVisionModel as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete.
Loaded CLIPTextModel to cuda:0 as complete.
Unloaded CLIPTextModel as complete.
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Loaded SiglipVisionModel to cuda:0 as complete.
latent_padding_size = 27, is_last_section = False
Unloaded SiglipVisionModel as complete.
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
28%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 7/25 [01:14<02:12, 7.38s/it]
92%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 23/25 [03:39<00:24, 12.10s/it]
100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 25/25 [03:58<00:00, 9.53s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
/home/dk/pinokio/api/Frame-Pack.git/app/env/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Decoded. Current latent shape torch.Size([1, 16, 9, 80, 76]); pixel shape torch.Size([1, 3, 33, 640, 608])
latent_padding_size = 18, is_last_section = False
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
4%|โโโโ | 1/25 [00:18<07:14, 18.09s/it]
8%|โโโโโโโโ | 2/25 [00:36<06:59, 18.22s/it]
76%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 19/25 [02:51<00:53, 8.92s/it]
100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 25/25 [04:11<00:00, 10.06s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
/home/dk/pinokio/api/Frame-Pack.git/app/env/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Decoded. Current latent shape torch.Size([1, 16, 18, 80, 76]); pixel shape torch.Size([1, 3, 69, 640, 608])
latent_padding_size = 9, is_last_section = False
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
0%| | 0/25 [00:00<?, ?it/s]
4%|โโโโ | 1/25 [00:18<07:19, 18.30s/it]
52%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 13/25 [01:54<01:06, 5.55s/it]
76%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 19/25 [02:53<00:53, 8.96s/it]
100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 25/25 [04:11<00:00, 10.07s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
/home/dk/pinokio/api/Frame-Pack.git/app/env/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Decoded. Current latent shape torch.Size([1, 16, 27, 80, 76]); pixel shape torch.Size([1, 3, 105, 640, 608])
latent_padding_size = 0, is_last_section = True
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
12%|โโโโโโโโโโโโ | 3/25 [00:56<06:51, 18.71s/it]
100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 25/25 [04:04<00:00, 9.78s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
/home/dk/pinokio/api/Frame-Pack.git/app/env/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Decoded. Current latent shape torch.Size([1, 16, 37, 80, 76]); pixel shape torch.Size([1, 3, 145, 640, 608])
Activity
Frankthetank02-rgb commentedon Apr 20, 2025
GPU usage also not ideal.
able2608 commentedon Apr 20, 2025
What's your generation settings? Seems that you can decrease reserved VRAM to less than 6GB (default). Still I did not get the advertised speed on 3060 laptop. Not pretty sure how those guys got the number (5 second in 50 minutes compared to 1.5 hours and still not done).
Frankthetank02-rgb commentedon Apr 20, 2025
All default settings rn, could it be due to outdated CUDA in the pack
able2608 commentedon Apr 20, 2025
Doesn't seems like the problem. If you'd like to optimize on VRAM usage, decrease the reserved VRAM size seems to be the only way.
BTW it seems that the current resource management system is home-made by the author and not the diffusers' inplementation. This can be the reason for the slower speed. I might try to implement diffusers' layer offload to see if it is different.
othneildrew commentedon Apr 20, 2025
I also have ridiculously show speeds for some reason with a decently spec'd pc. Running a RX 3050, 12th gen i9-12900K, 64 gb memory
Been running for 1.4 hrs now (and going) and it only recently showed the 1 sec preview.
๐ข
Frankthetank02-rgb commentedon Apr 20, 2025
Reinstalled Python(even though supposedly this uses its own built-in environment) and sage, turned on teacache and now at approx. 44s/it. Still slow and nowhere near advertised speed but slightly better
thorsil commentedon Apr 20, 2025
I'm in the same boat. I'm on a RTX 3060 and still doing about 1 second every 45 mins. I've been messing around with the sliders but haven't noticed much change yet.
oscarchuncha commentedon Apr 20, 2025
I guess the advertised speed is only for gpus with 16gb minimum, this assumption is based on the different posts I have come accross on reddit pages. Any vram below this, I think it takes more time than wan 2.1 safetensor file
lllyasviel commentedon Apr 20, 2025
Hi people - if you get unreasonably slow (if more than 2x slower than my tests) speed then it is possible that (1) your memory speed is bottlenecked, or (2) your swap memory is bottlenecked.
Usually (2) is more common than (1).
to solve (1), you take a look at your memory speed - if your system memory is too old, consider get a newer one. I believe my laptop is using like this. one can use something older but if too old or low end then it may cause bottleneck.
to solve (2), take the below steps:
look at this Image and find out whether the main swap is using HDD or SSD. If it is using HDD then you may want to change the setting to force the swap to use SSD.
if the swap is already using SSD, then make sure that the used disk has at least 100GB free space.
if this does not work then the bottleneck is in the SSD, maybe the SSD is too old or low end.
If the SSD is new and good but you still have problem, then it maybe bottlenecked by motherboard, but this is not very common.
also the speed reported in the first iteration is not accurate. it will drop after some warmup.
MarklarMarklar commentedon Apr 20, 2025
I can get 48sec/it on a 3060 laptop GPU with 6GB VRAM. I'm just using the default parameters with teacache turned on.
able2608 commentedon Apr 20, 2025
Moved the program to another drive and the speed is now much bearable (~45sec/frame). Seems that hard drive speed plays a large factor if you do not have the VRAM to stuff all the things in. Would definitely suggest anyone with slower speeds to try other drives lying around.
However, I still believes that the current implementation does not use swap and it simply loads the required parameters from the file. I used to have the whole folder on my D drive and when it is running what I saw was surging D drive utilization and nearly nothing from C drive. I currently configured swap to only be on C drive.
furana93 commentedon Apr 20, 2025
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
The next step to initialize after this is really slow like around 10 minute. Is it because I set my Virtual Ram to 32 GB? Will it be better if I use the "Automatically manage paging file size for all drivers"?
I have a 3060 6gb vram, 32gb ram.
able2608 commentedon Apr 20, 2025
The current implementation seems to not be using tiled VAE decode, and with your setup it seems like it is just the norm. The 10-minute wait is just for decoding the diffused samples. Would be great if tiled VAE is implemented to speed up the process.
othneildrew commentedon Apr 20, 2025
This helped with debugging for me. Speeds are now comparable to what the readme says. It took about 9mins for each frame on a RTX 3050. I only have an SSD. My problem was that I was running on WSL2. This in itself shouldn't be a problem, but there's probably something wrong with how I enabled hardware acceleration because on previous tests the GPU wouldn't be under any significant load.
Hope this helps
othneildrew commentedon Apr 20, 2025
Alr, last message lol
Out of curiousity I setup a cloud GPU server with the following specs:
1 GPU / 48 GB VRAM / 8 vCPU / 64 GB RAMI was able to get around 1 frame every 1:29:00. So a 10 sec video took about 14 mins. This has a lot of potential!
25 remaining items
icemagno commentedon Apr 28, 2025
My RAM usage is about 50Gb
The GPU memory usage is about 6Gb at any case.
If you have less than 64Gb of RAM it is more probable you're using disk memory. If this is the case and you don't have a good SSD then you're in trouble.
So stop complaining. You already are at your setup limits.
Rule:
50Gb of RAM and 6Gb of VRAM = 5secs in 10min (worst case)
No RAM = DISK USAGE.
Disk Usage = Good SSD
Bad SSD = TROUBLE
No SSD = YOU ARE DOOMED
Even at worst case this is far the best tool you will find to create video with low requirements.
I said LOW ... not NONE.
So.. be patient or give up.
( or buy a better setup )
Netshadows commentedon Apr 28, 2025
Just tried this out. I created a docker for it, running in ubuntu. Have 6xRTX3090, 32gb ram, running on old/slow sata drives. Probably took an hour or more to run the first time for a 5sec vid. CPU pegged at 100 for lots of it, lots of drive activity. Rebooted and ran everything fresh, this time cpu not pegging, but slowing system down. Not sure if should be running faster... may try something else since have more vram to play with.
$ docker logs framepack-ui
Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn']
Xformers is not installed!
Flash Attn is not installed!
Sage Attn is not installed!
Namespace(share=False, server='0.0.0.0', port=None, inbrowser=False)
Free VRAM 17.42950439453125 GB
High-VRAM Mode: False
Downloading shards: 100%|โโโโโโโโโโ| 4/4 [02:37<00:00, 39.43s/it]
Loading checkpoint shards: 100%|โโโโโโโโโโ| 4/4 [00:01<00:00, 3.39it/s]
Fetching 3 files: 100%|โโโโโโโโโโ| 3/3 [03:46<00:00, 75.40s/it]
Loading checkpoint shards: 100%|โโโโโโโโโโ| 3/3 [00:02<00:00, 1.28it/s]
transformer.high_quality_fp32_output_for_inference = True
To create a public link, set
share=Trueinlaunch().Unloaded DynamicSwap_LlamaModel as complete.
Unloaded CLIPTextModel as complete.
Unloaded SiglipVisionModel as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete.
Loaded CLIPTextModel to cuda:0 as complete.
Unloaded CLIPTextModel as complete.
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Loaded SiglipVisionModel to cuda:0 as complete.
latent_padding_size = 27, is_last_section = False
Unloaded SiglipVisionModel as complete.
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [15:45<00:00, 37.84s/it]
/usr/local/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Decoded. Current latent shape torch.Size([1, 16, 9, 80, 76]); pixel shape torch.Size([1, 3, 33, 640, 608])
latent_padding_size = 18, is_last_section = False
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [10:26<00:00, 25.05s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Decoded. Current latent shape torch.Size([1, 16, 18, 80, 76]); pixel shape torch.Size([1, 3, 69, 640, 608])
latent_padding_size = 9, is_last_section = False
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [10:24<00:00, 25.00s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Decoded. Current latent shape torch.Size([1, 16, 27, 80, 76]); pixel shape torch.Size([1, 3, 105, 640, 608])
latent_padding_size = 0, is_last_section = True
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [10:26<00:00, 25.04s/it]
Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn']
Xformers is not installed!
Flash Attn is not installed!
Sage Attn is not installed!
Namespace(share=False, server='0.0.0.0', port=None, inbrowser=False)
Free VRAM 23.0460205078125 GB
High-VRAM Mode: False
Downloading shards: 100%|โโโโโโโโโโ| 4/4 [00:00<00:00, 473.76it/s]
Loading checkpoint shards: 100%|โโโโโโโโโโ| 4/4 [00:07<00:00, 1.94s/it]
Fetching 3 files: 100%|โโโโโโโโโโ| 3/3 [00:00<00:00, 2080.51it/s]
Loading checkpoint shards: 100%|โโโโโโโโโโ| 3/3 [00:06<00:00, 2.28s/it]
transformer.high_quality_fp32_output_for_inference = True
To create a public link, set
share=Trueinlaunch().Unloaded DynamicSwap_LlamaModel as complete.
Unloaded CLIPTextModel as complete.
Unloaded SiglipVisionModel as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete.
Loaded CLIPTextModel to cuda:0 as complete.
Unloaded CLIPTextModel as complete.
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Loaded SiglipVisionModel to cuda:0 as complete.
latent_padding_size = 27, is_last_section = False
Unloaded SiglipVisionModel as complete.
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
100%|โโโโโโโโโโ| 25/25 [14:46<00:00, 35.46s/it]
icemagno commentedon Apr 28, 2025
Are you running it on Docker?????
Damn....
I quit.
Netshadows commentedon Apr 28, 2025
yah, took a few times building it to get the python dependencies, but ended up working. kinda new to this, so lots of experimenting. edit: what made the docker build process easier was having an llm explaining things along the way.
alumbs commentedon Apr 28, 2025
For some reason, its using my first GPU which is an Intel GPU not my second GPU which is my Nvidia GPU
How can I change it?
Zarrac commentedon May 3, 2025
5090
linux
sage 2.1
teacache on
10 sec video 6-7 mins
image resolution 1344x1728
gradio version (not comfyui version)
(but i have a different issue: action does not start until 05 secs for a 10 secs video / 09 secs until a 20 sec video )
ccwrapper commentedon May 10, 2025
My experience:
Intel Core i5-12600 / GeForce RTX 3070 Ti 8 GB VRAM
FramePack installed on a nvme drive, default settings,
SETTINGS 1:
16 GB RAM, 50 GB pagefile (manual settings) on a different (system) nvme drive.
Every of four steps takes about 11 min / ~26 s/it = ~44 min total for 5 sec video
+++ now adding extra RAM +++
SETTINGS 2:
80 GB RAM, page file - automatic system settings
Every of four steps now takes about 4:15 / ~10 s/it = ~ 17 min total for 5 sec video
So increasing RAM size (up to certain limit) doubled the speed!
From these 80 GB of memory used about 47 GB, free 31 GB, so it seems 64 GB must be enough. (Though I'd prefer to utilize this extra memory for speed or quality.)
mHamzaIqbal1998 commentedon May 13, 2025
I'm also using 4060 6gb with 16gb ram. on Windows I tried to run this and I was able to get 80/s . which was really horrible so I though why shouldn't give a try on linux so i did install ubuntu (with same environment like nvidia, cuda and cudnn drivers installed and same swap size) I was able to get around 26/s . I have a question is there still room to increase it? or its max what i can get? because its taking like avg 12 min for 1 sec clip only.
llinlluo commentedon May 20, 2025
Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn']
Xformers is not installed!
Flash Attn is not installed!
Sage Attn is not installed!
Namespace(share=False, server='127.0.0.1', port=None, inbrowser=True)
Free VRAM 22.455078125 GB
High-VRAM Mode: False
Downloading shards: 100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 4/4 [00:00<00:00, 2442.45it/s]
Loading checkpoint shards: 100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 4/4 [00:00<00:00, 4.44it/s]
Fetching 3 files: 100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 3/3 [00:00<?, ?it/s]
Loading checkpoint shards: 100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 3/3 [00:00<00:00, 10.66it/s]
transformer.high_quality_fp32_output_for_inference = True
To create a public link, set
share=Trueinlaunch().How to put Free VRAM 22.455078125 GB to use?
gaming09 commentedon May 20, 2025
are these speeds to be expected gpu is a 4070ti super 32gig ddr4 4000
Habbybbo commentedon Aug 9, 2025
It seems RAM is the key, I'm runnig with 32gb ddr5 3200mhz, due to stability issues and 3060 12gb. So if I add more RAM it should be way better.
eric-gitta-moore commentedon Aug 24, 2025
CPU RAM too small ๐
Mou-Mou-L commentedon Oct 18, 2025
clevesim commentedon Nov 3, 2025
Hello everyone,
I am facing a similar problem than @furana93 with my configuration:
-RTX5090
-Intel(R) Core(TM) Ultra 9 285K (3.70 GHz)
-96go ram (4800MT/s)
-motherboad:ASUS ROG STRIX Z890-E GAMING WIFI
-Framepack folder on SSD (around 4000mo/s)
I do not think bottleneck is problem but maybe I am wrong.
As you can see below I have no problem to generate the 25 steps (roughly it takes around 30sec). The problem appeared after, at the steps : Loaded/Unloaded AutoencoderKLHunyuanVideo.

It took a very very long time to complete/start and I don't know why. Any ideas?
I also noticed that the time between each 25 steps generation increase at each iteration (for instance: 2min/4min/20min).
Finally, i need roughly 30min to generate 5sec video that is not a standard with this configuration I guess?
tekken001 commentedon Nov 21, 2025
To create a public link, set
share=Trueinlaunch().Unloaded DynamicSwap_LlamaModel as complete.
Unloaded CLIPTextModel as complete.
Unloaded SiglipVisionModel as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete.
Loaded CLIPTextModel to cuda:0 as complete.
Unloaded CLIPTextModel as complete.
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
Loaded SiglipVisionModel to cuda:0 as complete.
latent_padding_size = 27, is_last_section = False
Unloaded SiglipVisionModel as complete.
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
28%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 7/25 [01:14<02:12, 7.38s/it]
92%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 23/25 [03:39<00:24, 12.10s/it]
100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 25/25 [03:58<00:00, 9.53s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
/home/dk/pinokio/api/Frame-Pack.git/app/env/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Decoded. Current latent shape torch.Size([1, 16, 9, 80, 76]); pixel shape torch.Size([1, 3, 33, 640, 608])
latent_padding_size = 18, is_last_section = False
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
4%|โโโโ | 1/25 [00:18<07:14, 18.09s/it]
8%|โโโโโโโโ | 2/25 [00:36<06:59, 18.22s/it]
76%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 19/25 [02:51<00:53, 8.92s/it]
100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 25/25 [04:11<00:00, 10.06s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
/home/dk/pinokio/api/Frame-Pack.git/app/env/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Decoded. Current latent shape torch.Size([1, 16, 18, 80, 76]); pixel shape torch.Size([1, 3, 69, 640, 608])
latent_padding_size = 9, is_last_section = False
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
0%| | 0/25 [00:00<?, ?it/s]
4%|โโโโ | 1/25 [00:18<07:19, 18.30s/it]
52%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 13/25 [01:54<01:06, 5.55s/it]
76%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ | 19/25 [02:53<00:53, 8.96s/it]
100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 25/25 [04:11<00:00, 10.07s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
/home/dk/pinokio/api/Frame-Pack.git/app/env/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Decoded. Current latent shape torch.Size([1, 16, 27, 80, 76]); pixel shape torch.Size([1, 3, 105, 640, 608])
latent_padding_size = 0, is_last_section = True
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
12%|โโโโโโโโโโโโ | 3/25 [00:56<06:51, 18.71s/it]
100%|โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ| 25/25 [04:04<00:00, 9.78s/it]
Offloading DynamicSwap_HunyuanVideoTransformer3DModelPacked from cuda:0 to preserve memory: 8 GB
Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete.
Unloaded AutoencoderKLHunyuanVideo as complete.
/home/dk/pinokio/api/Frame-Pack.git/app/env/lib/python3.10/site-packages/torchvision/io/_video_deprecation_warning.py:5: UserWarning: The video decoding and encoding capabilities of torchvision are deprecated from version 0.22 and will be removed in version 0.24. We recommend that you migrate to TorchCodec, where we'll consolidate the future decoding/encoding capabilities of PyTorch: https://github.com/pytorch/torchcodec
warnings.warn(
Decoded. Current latent shape torch.Size([1, 16, 37, 80, 76]); pixel shape torch.Size([1, 3, 145, 640, 608])
Is this a normal speed?
OS : Ubuntu 24.04.3 LTS
CPU : amd 5600x
MB : ASRock X570 Phantom Gaming 4 WiFi ax
RAM : 16*4= 64gb
GPU : 4060ti 16gb
SSD : 1tb
pinokio ai Framepack