Got Wan2.2 I2V running 2.5x faster on 8xH100 using Sequence Parallelism + Magcache
https://preview.redd.it/07lwyvl5zryf1.png?width=1200&format=png&auto=webp&s=ad22c52c861c18c94c54f27bbe71a6e120a8f3e7
Hey everyone,
I was curious how much faster we can get with Magcache on 8xH100 instead of 1xH100 for Wan 2.2 I2V. Currently, the original repositories of Magcache and Teacache only support 1GPU inference for Wan2.2 because of FSDP, as shown in this GitHub issue.
I managed to scale Magcache on 8XH100 with FSDP and sequence parallelism. Also experimented with several techniques: Flash-Attention-3, TF32 tensor cores, int8 quantization, Magcache, and torch.compile.
The fastest combo I got was FA3+TF32+Magcache+torch.compile that runs a 1280x720 video (81 frames, 40 steps) in 109s, down from 250s baseline (8xH100 sequence parallelism and FA2 only) without noticeable loss of quality. We can also play with the Magcache parameters for a quality tradeoff, for example, E024K2R10 (Error threshold =0.24, Skip K=2, Retention ratio = 0.1) to get 2.5x + speed boost.
Full breakdown, commands, and comparisons are here:
👉 Blog post with full benchmarks and configs
👉 Github repo with code
Curious if anyone else here is exploring sequence parallelism or similar caching methods on FSDP-based video diffusion models? Would love to compare notes.
Disclosure: I worked on and co-wrote this technical breakdown as part of the Morphic team
https://redd.it/1om8sr9
@rStableDiffusion
https://preview.redd.it/07lwyvl5zryf1.png?width=1200&format=png&auto=webp&s=ad22c52c861c18c94c54f27bbe71a6e120a8f3e7
Hey everyone,
I was curious how much faster we can get with Magcache on 8xH100 instead of 1xH100 for Wan 2.2 I2V. Currently, the original repositories of Magcache and Teacache only support 1GPU inference for Wan2.2 because of FSDP, as shown in this GitHub issue.
I managed to scale Magcache on 8XH100 with FSDP and sequence parallelism. Also experimented with several techniques: Flash-Attention-3, TF32 tensor cores, int8 quantization, Magcache, and torch.compile.
The fastest combo I got was FA3+TF32+Magcache+torch.compile that runs a 1280x720 video (81 frames, 40 steps) in 109s, down from 250s baseline (8xH100 sequence parallelism and FA2 only) without noticeable loss of quality. We can also play with the Magcache parameters for a quality tradeoff, for example, E024K2R10 (Error threshold =0.24, Skip K=2, Retention ratio = 0.1) to get 2.5x + speed boost.
Full breakdown, commands, and comparisons are here:
👉 Blog post with full benchmarks and configs
👉 Github repo with code
Curious if anyone else here is exploring sequence parallelism or similar caching methods on FSDP-based video diffusion models? Would love to compare notes.
Disclosure: I worked on and co-wrote this technical breakdown as part of the Morphic team
https://redd.it/1om8sr9
@rStableDiffusion
ComfyUI Tutorial: Take Your Prompt To The Next Level With Qwen 3 VL
https://youtu.be/cfgtvXeYYb0
https://redd.it/1omavip
@rStableDiffusion
https://youtu.be/cfgtvXeYYb0
https://redd.it/1omavip
@rStableDiffusion
YouTube
ComfyUI Tutorial: Take Your Prompt To The Next Level With Qwen 3 VL #comfyui #comfyuitutorial
On this tutorial I will show you how to generate prompt by analyzing an image input using qwen vl3 new model dedicated for prompt extraction using input image, it allows you to extract all the needed data like poses, outfit, colors, environment to create…
Dataset tool to organize images by quality (sharp / blurry, jpeg artifacts, compression, etc).
I have rolled some of my own image quality tools before but I'll try asking. Any tool that allows for grouping / sorting / filtering images by different quality criteria like sharpness, blurriness, jpeg artifacts (even imperceptible), compression, out-of-focus depth of field, etc - basically by overall quality?
I am looking to root out outliers out of larger datasets that could negatively affect training quality.
https://redd.it/1omac5p
@rStableDiffusion
I have rolled some of my own image quality tools before but I'll try asking. Any tool that allows for grouping / sorting / filtering images by different quality criteria like sharpness, blurriness, jpeg artifacts (even imperceptible), compression, out-of-focus depth of field, etc - basically by overall quality?
I am looking to root out outliers out of larger datasets that could negatively affect training quality.
https://redd.it/1omac5p
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit
Explore this post and more from the StableDiffusion community
Warping Inception Style Effect – with WAN ATI
https://www.youtube.com/watch?v=AI9-1G7niXY
https://redd.it/1omcu7n
@rStableDiffusion
https://www.youtube.com/watch?v=AI9-1G7niXY
https://redd.it/1omcu7n
@rStableDiffusion
YouTube
Wan ATI with ComfyUI
We're diving into an old wan model called ATI that lets you control trajectories.
download link to everything
https://huggingface.co/lividtm/Wan_ATI/tree/main
download link to everything
https://huggingface.co/lividtm/Wan_ATI/tree/main