NVIDIA / NVIDIA/TensorRT-LLM

[PERF] MM models. Transfer images main->worker in low precision

Open
#6,068 2 comments 0 reactions 1 assignee View on GitHub

Nobody has claimed this yet.

Investigating Performance triaged
Dominant language
Python
Stars
14.7k
Forks
2.8k
Avg merge
2d 23h
Merged PRs (30d)
489

Description

Problem

After #6004 we still has sufficient overhead for image preparation.
Profile looks like

Image

send is around 15% from image preparation time.

Solution

I noticed that we transfer image in fp32 format. Every pixel takes 3*sizeof(fp32)=12 bytes. If we transfer(and maybe can process) images in int8 it might sufficiently reduce send overhead. (or at least bf16 - anyway the model run in bf16 and we anyway convert the image to bf16).

Contributor guide

Open the contributing guide

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.

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.