microsoft / microsoft/onnxruntime-genai

0.5.2 DML 2x to 4x Slower than 0.4.0 (Big regression)

Open
#1,114 1 comment 0 reactions 0 assignees View on GitHub
ep:DML performance
Dominant language
C++
Stars
1.1k
Forks
354
Avg merge
2d 16h
Merged PRs (30d)
85

Description

genai 0.5.2 appears to be between 2x and 4x slower than genai 0.4.0, in fact it is only about 50% faster than CPU mode.

If have just tested both 0.4.0 and 0.5.2 and it definitely is a vast difference in DML mode.

GPU Quadro P5000.

model: microsoft/Phi-3-mini-4k-instruct-onnx
c#
BTW, I checked it is not the new DirectML.dll library causing this.

Contributor guide

No contributing guide indexed for this repository

Research direction

No source file, test, or entry point is named. First reproduce the comparison between genai 0.4.0 and 0.5.2 in DML mode with the Phi-3-mini-4k-instruct-onnx model on a Quadro P5000, then compare against CPU mode. Done means identifying the regression and validating improved DML performance without attributing it to the new DirectML.dll library.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp, csharp
Domain
machine-learning, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.