the gpu usage is more constant than before

JJaviEM the gpu usage is more constant than before

F

Furkan Gözükara SECourses•3/16/25, 2:06 PM

yes the lesser RAM usage the more GPU usage

J

JaviEMOP•3/16/25, 2:07 PM

and it's faster than ever

J

JaviEMOP•3/16/25, 2:07 PM

huge update

J

JaviEMOP•3/16/25, 2:10 PM

it's getting even faster. I have never seen less than 40s/it

JJaviEM it's getting even faster. I have never seen less than 40s/it

F

Furkan Gözükara SECourses•3/16/25, 2:13 PM

if teacache enabled it will

J

JaviEMOP•3/16/25, 2:47 PM

@Dr. Furkan Gözükara can you fix the profiles with FP8? It's even slower than bf16 now. It's only using 13GB VRAM and 64% GPU

J

JaviEMOP•3/16/25, 3:01 PM

FP8 videos are working fine now btw

JJaviEM FP8 videos are working fine now btw

F

Furkan Gözükara SECourses•3/16/25, 3:10 PM

for FP8

F

Furkan Gözükara SECourses•3/16/25, 3:10 PM

increase number of parameters

F

Furkan Gözükara SECourses•3/16/25, 3:12 PM

720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB
1 is : 18326 MB
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB
6,000,000,000 is : 28380 MB
6,500,000,000 is : 29330 MB
6,750,000,000 is : 29790 MB
16,000,000,000 is : 47030 MB
22,000,000,000 is : 47800 MB


720x1280 px, 81 frames, 14b 720p image to video model - FP8 precision (torch.float8_e4m3fn)

3,000,000,000 is : 19763 MB
6,750,000,000 is : 23350 MB

720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB
1 is : 18326 MB
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB
6,000,000,000 is : 28380 MB
6,500,000,000 is : 29330 MB
6,750,000,000 is : 29790 MB
16,000,000,000 is : 47030 MB
22,000,000,000 is : 47800 MB


720x1280 px, 81 frames, 14b 720p image to video model - FP8 precision (torch.float8_e4m3fn)

3,000,000,000 is : 19763 MB
6,750,000,000 is : 23350 MB

720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB
1 is : 18326 MB
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB
6,000,000,000 is : 28380 MB
6,500,000,000 is : 29330 MB
6,750,000,000 is : 29790 MB
16,000,000,000 is : 47030 MB
22,000,000,000 is : 47800 MB


720x1280 px, 81 frames, 14b 720p image to video model - FP8 precision (torch.float8_e4m3fn)

3,000,000,000 is : 19763 MB
6,750,000,000 is : 23350 MB

720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB
1 is : 18326 MB
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB
6,000,000,000 is : 28380 MB
6,500,000,000 is : 29330 MB
6,750,000,000 is : 29790 MB
16,000,000,000 is : 47030 MB
22,000,000,000 is : 47800 MB


720x1280 px, 81 frames, 14b 720p image to video model - FP8 precision (torch.float8_e4m3fn)

3,000,000,000 is : 19763 MB
6,750,000,000 is : 23350 MB

F

Furkan Gözükara SECourses•3/16/25, 3:12 PM

i am doing some testing

J

JaviEMOP•3/16/25, 3:12 PM

nice, thanks

F

Furkan Gözükara SECourses•3/16/25, 3:30 PM

480x832 px, 81 frames, 14b 480p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 8618 MB
1 is : 8618 MB
1,000,000,000 is : x MB
1,500,000,000 is : x MB
2,000,000,000 is : x MB
3,000,000,000 is : x MB
3,500,000,000 is : x MB
4,000,000,000 is : x MB
8,000,000,000 is : 15392 MB
12,000,000,000 is : x MB
22,000,000,000 is : 23080 MB

480x832 px, 81 frames, 14b 480p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 8618 MB
1 is : 8618 MB
1,000,000,000 is : x MB
1,500,000,000 is : x MB
2,000,000,000 is : x MB
3,000,000,000 is : x MB
3,500,000,000 is : x MB
4,000,000,000 is : x MB
8,000,000,000 is : 15392 MB
12,000,000,000 is : x MB
22,000,000,000 is : 23080 MB

480x832 px, 81 frames, 14b 480p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 8618 MB
1 is : 8618 MB
1,000,000,000 is : x MB
1,500,000,000 is : x MB
2,000,000,000 is : x MB
3,000,000,000 is : x MB
3,500,000,000 is : x MB
4,000,000,000 is : x MB
8,000,000,000 is : 15392 MB
12,000,000,000 is : x MB
22,000,000,000 is : 23080 MB

480x832 px, 81 frames, 14b 480p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 8618 MB
1 is : 8618 MB
1,000,000,000 is : x MB
1,500,000,000 is : x MB
2,000,000,000 is : x MB
3,000,000,000 is : x MB
3,500,000,000 is : x MB
4,000,000,000 is : x MB
8,000,000,000 is : 15392 MB
12,000,000,000 is : x MB
22,000,000,000 is : 23080 MB

F

Furkan Gözükara SECourses•3/16/25, 3:30 PM

i am gonna make presets to support fp8 too

F

Furkan Gözükara SECourses•3/16/25, 3:41 PM

all done

F

Furkan Gözükara SECourses•3/16/25, 3:41 PM

GPU VRAM Usages Table With Respected To Number of Persistent Params+

All tests made on secondary GPU thus they are 100% accurate

Therefore, make sure that you have such empty VRAM before starting the Wan 2.1 SECourses App 

1280x720 px, 81 frames, 14b text to video model - BF16 precision (torch.bfloat16)

0 is : 14168 MB - Below 24 GB GPUs
1,000,000,000 is : 15658 MB
2,000,000,000 is : 17664 MB
3,000,000,000 is : 20134 MB
3,750,000,000 is : 21134 MB
4,000,000,000 is : 21500 MB
4,250,000,000 is : 22000 MB - 24 GB GPUs
8,250,000,000 is : 29870 MB - 32 GB GPUs
22,000,000,000 is : 41700 MB - 48 GB GPUs

1280x720 px, 81 frames, 14b text to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 14168 MB - Below 24 GB GPUs
1,000,000,000 is : 14918 MB
8,000,000,000 is : 21260 MB
8,750,000,000 is : 21980 MB - 24 GB GPUs
22,000,000,000 is : x MB


480x832 px, 81 frames, 14b 480p image to video model - BF16 precision (torch.bfloat16)

0 is : 8618 MB - Below 12 GB GPUs
1,000,000,000 is : 9354 MB
1,500,000,000 is : 10090 MB - 12 GB GPUs
2,000,000,000 is : 11370 MB
3,000,000,000 is : 13080 MB
3,500,000,000 is : 14038 MB - 16 GB GPUs
4,000,000,000 is : 15140 MB
8,000,000,000 is : 22100 MB - 24 GB GPUs
12,000,000,000 is : 30060 MB - 32 GB GPUs
22,000,000,000 is : 38300 MB - 48 GB GPUs


480x832 px, 81 frames, 14b 480p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 8618 MB - Below 12 GB GPUs
2,500,000,000 is : 9794 MB - 12 GB GPUs
7,500,000,000 is : 14850 MB - 16 GB GPUs
8,000,000,000 is : 15392 MB
15,000,000,000 is : 21700 MB - 24 GB GPUs
22,000,000,000 is : 23080 MB - 32 GB GPUs


720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB - Below 24 GB GPUs
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB - 24 GB GPUs
6,000,000,000 is : 28380 MB
6,500,000,000 is : 29330 MB
6,750,000,000 is : 29790 MB - 32 GB GPUs
16,000,000,000 is : 47030 MB - 48 GB GPUs
22,000,000,000 is : 47800 MB - 80 GB GPUs


720x1280 px, 81 frames, 14b 720p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 18326 MB - Below 24 GB GPUs
3,000,000,000 is : 19763 MB
6,000,000,000 is : 22020 MB - 24 GB GPUs
14,000,000,000 is : 29840 MB
15,000,000,000 is : 30790 MB - 32 GB GPUs
22,000,000,000 is : 32013 MB - 48 GB GPUs

GPU VRAM Usages Table With Respected To Number of Persistent Params+

All tests made on secondary GPU thus they are 100% accurate

Therefore, make sure that you have such empty VRAM before starting the Wan 2.1 SECourses App 

1280x720 px, 81 frames, 14b text to video model - BF16 precision (torch.bfloat16)

0 is : 14168 MB - Below 24 GB GPUs
1,000,000,000 is : 15658 MB
2,000,000,000 is : 17664 MB
3,000,000,000 is : 20134 MB
3,750,000,000 is : 21134 MB
4,000,000,000 is : 21500 MB
4,250,000,000 is : 22000 MB - 24 GB GPUs
8,250,000,000 is : 29870 MB - 32 GB GPUs
22,000,000,000 is : 41700 MB - 48 GB GPUs

1280x720 px, 81 frames, 14b text to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 14168 MB - Below 24 GB GPUs
1,000,000,000 is : 14918 MB
8,000,000,000 is : 21260 MB
8,750,000,000 is : 21980 MB - 24 GB GPUs
22,000,000,000 is : x MB


480x832 px, 81 frames, 14b 480p image to video model - BF16 precision (torch.bfloat16)

0 is : 8618 MB - Below 12 GB GPUs
1,000,000,000 is : 9354 MB
1,500,000,000 is : 10090 MB - 12 GB GPUs
2,000,000,000 is : 11370 MB
3,000,000,000 is : 13080 MB
3,500,000,000 is : 14038 MB - 16 GB GPUs
4,000,000,000 is : 15140 MB
8,000,000,000 is : 22100 MB - 24 GB GPUs
12,000,000,000 is : 30060 MB - 32 GB GPUs
22,000,000,000 is : 38300 MB - 48 GB GPUs


480x832 px, 81 frames, 14b 480p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 8618 MB - Below 12 GB GPUs
2,500,000,000 is : 9794 MB - 12 GB GPUs
7,500,000,000 is : 14850 MB - 16 GB GPUs
8,000,000,000 is : 15392 MB
15,000,000,000 is : 21700 MB - 24 GB GPUs
22,000,000,000 is : 23080 MB - 32 GB GPUs


720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB - Below 24 GB GPUs
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB - 24 GB GPUs
6,000,000,000 is : 28380 MB
6,500,000,000 is : 29330 MB
6,750,000,000 is : 29790 MB - 32 GB GPUs
16,000,000,000 is : 47030 MB - 48 GB GPUs
22,000,000,000 is : 47800 MB - 80 GB GPUs


720x1280 px, 81 frames, 14b 720p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 18326 MB - Below 24 GB GPUs
3,000,000,000 is : 19763 MB
6,000,000,000 is : 22020 MB - 24 GB GPUs
14,000,000,000 is : 29840 MB
15,000,000,000 is : 30790 MB - 32 GB GPUs
22,000,000,000 is : 32013 MB - 48 GB GPUs

GPU VRAM Usages Table With Respected To Number of Persistent Params+

All tests made on secondary GPU thus they are 100% accurate

Therefore, make sure that you have such empty VRAM before starting the Wan 2.1 SECourses App 

1280x720 px, 81 frames, 14b text to video model - BF16 precision (torch.bfloat16)

0 is : 14168 MB - Below 24 GB GPUs
1,000,000,000 is : 15658 MB
2,000,000,000 is : 17664 MB
3,000,000,000 is : 20134 MB
3,750,000,000 is : 21134 MB
4,000,000,000 is : 21500 MB
4,250,000,000 is : 22000 MB - 24 GB GPUs
8,250,000,000 is : 29870 MB - 32 GB GPUs
22,000,000,000 is : 41700 MB - 48 GB GPUs

1280x720 px, 81 frames, 14b text to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 14168 MB - Below 24 GB GPUs
1,000,000,000 is : 14918 MB
8,000,000,000 is : 21260 MB
8,750,000,000 is : 21980 MB - 24 GB GPUs
22,000,000,000 is : x MB


480x832 px, 81 frames, 14b 480p image to video model - BF16 precision (torch.bfloat16)

0 is : 8618 MB - Below 12 GB GPUs
1,000,000,000 is : 9354 MB
1,500,000,000 is : 10090 MB - 12 GB GPUs
2,000,000,000 is : 11370 MB
3,000,000,000 is : 13080 MB
3,500,000,000 is : 14038 MB - 16 GB GPUs
4,000,000,000 is : 15140 MB
8,000,000,000 is : 22100 MB - 24 GB GPUs
12,000,000,000 is : 30060 MB - 32 GB GPUs
22,000,000,000 is : 38300 MB - 48 GB GPUs


480x832 px, 81 frames, 14b 480p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 8618 MB - Below 12 GB GPUs
2,500,000,000 is : 9794 MB - 12 GB GPUs
7,500,000,000 is : 14850 MB - 16 GB GPUs
8,000,000,000 is : 15392 MB
15,000,000,000 is : 21700 MB - 24 GB GPUs
22,000,000,000 is : 23080 MB - 32 GB GPUs


720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB - Below 24 GB GPUs
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB - 24 GB GPUs
6,000,000,000 is : 28380 MB
6,500,000,000 is : 29330 MB
6,750,000,000 is : 29790 MB - 32 GB GPUs
16,000,000,000 is : 47030 MB - 48 GB GPUs
22,000,000,000 is : 47800 MB - 80 GB GPUs


720x1280 px, 81 frames, 14b 720p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 18326 MB - Below 24 GB GPUs
3,000,000,000 is : 19763 MB
6,000,000,000 is : 22020 MB - 24 GB GPUs
14,000,000,000 is : 29840 MB
15,000,000,000 is : 30790 MB - 32 GB GPUs
22,000,000,000 is : 32013 MB - 48 GB GPUs

GPU VRAM Usages Table With Respected To Number of Persistent Params+

All tests made on secondary GPU thus they are 100% accurate

Therefore, make sure that you have such empty VRAM before starting the Wan 2.1 SECourses App 

1280x720 px, 81 frames, 14b text to video model - BF16 precision (torch.bfloat16)

0 is : 14168 MB - Below 24 GB GPUs
1,000,000,000 is : 15658 MB
2,000,000,000 is : 17664 MB
3,000,000,000 is : 20134 MB
3,750,000,000 is : 21134 MB
4,000,000,000 is : 21500 MB
4,250,000,000 is : 22000 MB - 24 GB GPUs
8,250,000,000 is : 29870 MB - 32 GB GPUs
22,000,000,000 is : 41700 MB - 48 GB GPUs

1280x720 px, 81 frames, 14b text to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 14168 MB - Below 24 GB GPUs
1,000,000,000 is : 14918 MB
8,000,000,000 is : 21260 MB
8,750,000,000 is : 21980 MB - 24 GB GPUs
22,000,000,000 is : x MB


480x832 px, 81 frames, 14b 480p image to video model - BF16 precision (torch.bfloat16)

0 is : 8618 MB - Below 12 GB GPUs
1,000,000,000 is : 9354 MB
1,500,000,000 is : 10090 MB - 12 GB GPUs
2,000,000,000 is : 11370 MB
3,000,000,000 is : 13080 MB
3,500,000,000 is : 14038 MB - 16 GB GPUs
4,000,000,000 is : 15140 MB
8,000,000,000 is : 22100 MB - 24 GB GPUs
12,000,000,000 is : 30060 MB - 32 GB GPUs
22,000,000,000 is : 38300 MB - 48 GB GPUs


480x832 px, 81 frames, 14b 480p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 8618 MB - Below 12 GB GPUs
2,500,000,000 is : 9794 MB - 12 GB GPUs
7,500,000,000 is : 14850 MB - 16 GB GPUs
8,000,000,000 is : 15392 MB
15,000,000,000 is : 21700 MB - 24 GB GPUs
22,000,000,000 is : 23080 MB - 32 GB GPUs


720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB - Below 24 GB GPUs
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB - 24 GB GPUs
6,000,000,000 is : 28380 MB
6,500,000,000 is : 29330 MB
6,750,000,000 is : 29790 MB - 32 GB GPUs
16,000,000,000 is : 47030 MB - 48 GB GPUs
22,000,000,000 is : 47800 MB - 80 GB GPUs


720x1280 px, 81 frames, 14b 720p image to video model - FP8 precision (torch.float8_e4m3fn)

0 is : 18326 MB - Below 24 GB GPUs
3,000,000,000 is : 19763 MB
6,000,000,000 is : 22020 MB - 24 GB GPUs
14,000,000,000 is : 29840 MB
15,000,000,000 is : 30790 MB - 32 GB GPUs
22,000,000,000 is : 32013 MB - 48 GB GPUs

F

Furkan Gözükara SECourses•3/16/25, 3:41 PM

time to update the app

J

JaviEMOP•3/16/25, 4:07 PM

I have updated but preset doesn't change for FP8. 16GB preset still use 3,500,000,000 on FP8

J

JaviEMOP•3/16/25, 4:21 PM

With 7,500,000,000 is using 0,4GB of shared VRAM on a 16GB VRAM GPU. I will try lowering it to 7,000,000,000 or 6,500,000,000

F

Furkan Gözükara SECourses•3/16/25, 4:33 PM

i havent updated app yet still few missing

J

JaviEMOP•3/16/25, 4:58 PM

FP8 7,500,000,000: 1278.54 seconds
FP8 7,000,000,000: 1137.68 seconds

J

JaviEMOP•3/16/25, 5:17 PM

BF16 3,500,000,000: 1002.77 seconds

J

JaviEMOP•3/16/25, 5:17 PM

FP8 has no sense on a 16GB VRAM GPU

J

JaviEMOP•3/16/25, 5:18 PM

Img to video 480p model*

J

JaviEMOP•3/16/25, 5:33 PM

10 videos generated this afternoon and no glitches or image corruption with teacache. I only had one big colour change with FP8. Very nice updates today.

JJaviEM 10 videos generated this afternoon and no glitches or image corruption with teac...

F

Furkan Gözükara SECourses•3/16/25, 5:46 PM

awesome so fixed finally

F

Furkan Gözükara SECourses•3/16/25, 6:44 PM

VRAM usages Wan 2.1, APP updated : https://www.reddit.com/r/SECourses/comments/1jcse0z/wan_21_vram_usages_very_detailed_chart_for_4_8_10/

From the SECourses community on Reddit: Wan 2.1 VRAM usages very de...

Explore this post and more from the SECourses community

F

Furkan Gözükara SECourses•3/16/25, 7:04 PM

sana 1.5 published but i will wait diffusers

J

JaviEMOP•3/16/25, 7:51 PM

@Dr. Furkan Gözükara are you using the new use_ret_steps in your implementation of teacache in wan 2.1? According to the developers it has more speed and quality in generation

J

JaviEMOP•3/16/25, 7:52 PM

The info is here, at the bottom https://github.com/ali-vilab/TeaCache/tree/main/TeaCache4Wan2.1

GitHub

TeaCache/TeaCache4Wan2.1 at main · ali-vilab/TeaCache

Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model - ali-vilab/TeaCache

JJaviEM @Dr. Furkan Gözükara are you using the new use_ret_steps in your implementation ...

F

Furkan Gözükara SECourses•3/16/25, 7:53 PM

i think ours matching

F

Furkan Gözükara SECourses•3/16/25, 7:53 PM

JJaviEM @Dr. Furkan Gözükara are you using the new use_ret_steps in your implementation ...

F

Furkan Gözükara SECourses•3/16/25, 7:54 PM

what we use here https://github.com/modelscope/DiffSynth-Studio/pull/448

GitHub

support teacache in wan by Artiprocher · Pull Request #448 · models...

J

JaviEMOP•3/16/25, 7:55 PM

nice

F

Furkan Gözükara SECourses•3/16/25, 9:16 PM

i did setup wan 2.1 for a client on runpod and it took freaking 65 min to install

F

Furkan Gözükara SECourses•3/16/25, 9:16 PM

shameless disk speed

F

Furkan Gözükara SECourses•3/16/25, 9:52 PM

just fixed FP8 + LoRA issue

F

Furkan Gözükara SECourses•3/16/25, 9:52 PM

whoever gets that run update

FFurkan Gözükara SECourses VRAM usages Wan 2.1, APP updated : https://www.reddit.com/r/SECourses/comments/1...

J

J2_Nedo•3/16/25, 10:25 PM

thank you Doc. but you know its Sunday, don't you have some time off, its important too

JJ2_Nedo thank you Doc. but you know its Sunday, don't you have some time off, its import...

F

Furkan Gözükara SECourses•3/16/25, 10:30 PM

i have been working since woken up

F

Furkan Gözükara SECourses•3/16/25, 10:30 PM

still working lol

J

J2_Nedo•3/16/25, 10:33 PM

you have to take care of you Doc.! stress kills

F

Furkan Gözükara SECourses•3/16/25, 10:34 PM

F

Furkan Gözükara SECourses•3/16/25, 10:34 PM

true

T

Tomfoolery•3/16/25, 10:40 PM

@Dr. Furkan Gözükara - you see skip-layer guidance has been implement for WAN in Comfy? Thoughts?

https://github.com/kijai/ComfyUI-KJNodes/commit/7c488a16ef420acf0276a4f8e31fc024a969d24b

GitHub

Add SkipLayerGuidanceWanVideo · kijai/ComfyUI-KJNodes@7c488a1

TTomfoolery @Dr. Furkan Gözükara - you see skip-layer guidance has been implement for WAN in...

F

Furkan Gözükara SECourses•3/16/25, 10:41 PM

some placebo effect

F

Furkan Gözükara SECourses•3/16/25, 10:41 PM

0 explanation :d

T

Tomfoolery•3/16/25, 10:42 PM

I know it was pretty well tested in Hunyuan video, that skipping layers 9 and 10 resulted in 'smoother' videos. Curious to see what the end results are for Wan

X

xlads•3/16/25, 10:47 PM

X

xlads•3/16/25, 10:48 PM

Going to try a 30 sec image to video 720p tomorrow.

Similar Threads

Similar Threads

Similar Threads