microsoft / microsoft/onnxruntime-genai

.Net After updating to .5, Phi3.5Mini outputs some meaningless characters

Open
#1,109 0 comments 0 reactions 0 assignees View on GitHub
model quality
Dominant language
C++
Stars
1.1k
Forks
354
Avg merge
2d 16h
Merged PRs (30d)
85

Description

Window11 24H2Ltsc .netCore 8 cuda Model:Phi-3.5-mini-instruct-onnx\cuda\cuda-int4-awq-block-128

We have a sample dialog to test the output.

![Image](https://github.com/user-attachments/assets/247b8765-6bdb-490c-86ff-9df24aaf64da)

in 0.4
Phi3.5 output
"I dont know"
![Image](https://github.com/user-attachments/assets/82edf74c-9499-4a5f-a829-59a0f7f8cefb)

In 0.5.2
![Image](https://github.com/user-attachments/assets/ff3565cf-5860-46c6-9521-63f6d2133d2b)

Code
![Image](https://github.com/user-attachments/assets/151fce15-eef7-48b9-bfd3-2219de93c081)

In 0.5.2 Qwen2 It will also output some meaningless exclamation points,
Qwen2 Model Qwen2-1.5B

We made the conversion with reference to the [link](https://github.com/microsoft/onnxruntime-genai/tree/main/src/python/py/models#full-usage) parameter Int4 Cuda
Qwen2 output

![Image](https://github.com/user-attachments/assets/f32a2649-1d45-4126-8c66-7152663c86e6)

Contributor guide

No contributing guide indexed for this repository

Research direction

Start by reproducing the sample dialog with Phi-3.5-mini-instruct-onnx and Qwen2-1.5B on version 0.5.2, comparing the output with 0.4. Review the conversion parameters in the linked model-conversion documentation. Done means both models produce meaningful output rather than meaningless characters or exclamation points.

Written by the indexing model from the issue text.

Assessment

Domain
ai
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.