modelscope / modelscope/DiffSynth-Studio

H100 GPU - FP8 inference acceleration

Open
#1,055 1 comment 1 reaction 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
13.1k
Forks
1.3k
Avg merge
13h 12m
Merged PRs (30d)
45

Description

Thanks for the great work !

I followed the Qwen-Image inference example in below link (recommended for faster inference) for comparing FP8 vs bfloat16:
./accelerate/Qwen-Image-FP8.py

But for FP8 (torch.float8_e4m3fn), the inference is slower than bfloat16.

FP8 takes 42s per image with a single-gpu instance and num_inference_steps=40
bfloat16 takes 33s under the same setting.

Setup:

  1. H100 GPU which supports FP8
  2. Flash Attention 3 is available

How to improve FP8's inference speed?

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with examples/qwen_image/accelerate/Qwen-Image-FP8.py and run the reported FP8 and bfloat16 configurations on an H100. Compare the per-image timings and inspect the inference path and available Flash Attention 3 setup. Done means identifying and addressing the cause of FP8 being slower, with measurements showing improved FP8 performance under the same settings.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.