#H200 SXM produces black video

28 messages · Page 1 of 1 (latest)

burnt turtle
#

Hi everyone, I’m looking for some help.

I’m generating videos with InfiniteTalk using RTX 5090 / RTX 4090, running the ComfyUI SageAttention (CUDA 12.8) template.

On 5090 / 4090 everything works perfectly.
I sometimes use a Network Volume, and it works reliably there as well.

However, when I try to run the exact same setup on H200 SXM, I run into a serious issue:
• the output video contains only ~1 second (or even just 1 frame),
• and then the rest of the video is a black screen.

What I’ve already verified:
• I tried both the SageAttention template and a setup without it
• I clone the same GitHub repositories
• I use the exact same ComfyUI workflow
• Same models, same settings
• Same input image and the same MP3 audio file

Everything is literally 1:1 identical, except for the GPU.

On H200 there is no full crash — generation finishes — but the output video is broken.

Could this be related to:
• H200 / Hopper-specific behavior
• SageAttention or attention backend compatibility
• FP8 / precision differences
• CUDA 12.8 issues on Hopper GPUs

I’d really appreciate any ideas or similar experiences.
Thanks in advance 🙏

dusk bluffBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

torn sphinxBOT
still nexus
#

or maybe update your custom node?

#

sageattention might be also compiled for a specific nvidia capability, so it might notwork with other gpu architectures?

burnt turtle
#

The thumbnail gets generated, but after that the video is just a black screen while the audio continues to play.

still nexus
#

i think the generation fails, idk why tho, maybe to get more context share your logs also? and search somewhere else

#

i haven't experienced this problem before so can't help much😅

burnt turtle
still nexus
#

Maybe just upload the txt files here

burnt turtle
still nexus
#

Idk maybe it's just the model

#

Or your custom nodes

#

Or workflow

#

Try reporting the issue to the custom node that you're using, with the workflow file

#

And your model links

burnt turtle
still nexus
#

are you using the latest version of the custom nodes?

burnt turtle
burnt turtle
still nexus
#

Can you screenshot and share your workflow .json

#

maybe i can take a quick look

burnt turtle
#

this it with prefetch_blocks=1, but i tried prefetch_blocks=0

still nexus