can this model run on Hopper GPU

#8
by simonlindelta - opened

can this model run on Hopper GPU

You can but not so meaningful since Hopper doesn't have HW-supported FP4. It would be slow.

You can but not so meaningful since Hopper doesn't have HW-supported FP4. It would be slow.

Is 5070ti with torch2.7+cu128 okay now?

You can but not so meaningful since Hopper doesn't have HW-supported FP4. It would be slow.

Is 5070ti with torch2.7+cu118 okay now?

cu118 doesn't support blackwell.

This comment has been hidden
This comment has been hidden (marked as Resolved)

@simonlindelta Please try this build for H100 (80G) x 8, supporting inference with FP4:

https://hub.docker.com/r/tutelgroup/deepseek-671b

Sign up or log in to comment