lllyasviel / lllyasviel/FramePack

RTX 4060 with SageAttention

Open
#205 8 comments 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
17.3k
Forks
1.7k
PR merge metrics
No merged PRs in 30d

Description

On my RTX 4060 Laptop, I am generating each second of video in about 4:09 (around 10s/it). It is taking around 16 mins for a 5 sec Video (with Sage Attention). First of all, great job, to the Developer, this is Amazing !

But With Sage Attention, there is a lot of variation going on throughout the 25 images being generated at each pass:
- It starts at about 20 s/it for the first 4 frames
- It keeps getting progressively quicker from Frame 4 onwards, until I'm down to 5s/it by Frame no 18 and it looks as if the whole pass is gonna finish in less than 3 mins.
- But then it suddenly slows down, with image 19 - 25 generating progressively slower, going from 5 to 6, 7, 9 and then 10it/s at Frame 25.

The VRAM and RAM stay at the same usage throughout the pass but the speed nosedives at the very end. If it did not I could be getting whole 5 sec video in around 10 mins but due to this I end up at 16 mins for the 5 seconds. Is this the normal behaviour for Sage Attn ? Or am I getting bottlenecked somewhere at the end of each pass ?

Contributor guide

No contributing guide indexed for this repository

Research direction

The report contains RTX 4060 Laptop timing observations for SageAttention but names no source files or tests. Start by reproducing the 25-frame pass with the reported hardware and settings, then trace the per-frame timing and memory behavior. Done means identifying whether the late-pass slowdown is expected or a reproducible bottleneck.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.