Eval bug: Embedding output differs significantly between b4712 and b4713

[lingchingteng]

Name and Version

version: 5713 (4c9fdfb)
built with clang version 18.1.8 for x86_64-pc-windows-msvc

Operating systems

Windows

GGML backends

CUDA

Hardware

CPU Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
GPU NVIDIA Quadro RTX 5000 with Max-Q Design

Models

bge-m3

Problem description & steps to reproduce

The embedding results are very different between commit b4712 and b4713.

Server command used:

.\llama-server.exe --hf-repo gpustack/bge-m3-GGUF --hf-file bge-m3-Q4_K_M.gguf --embedding -ngl 99

POST request:

curl.exe -d "{\"input\": \"Hello\"}" http://127.0.0.1:8080/v1/embeddings

Please let me know if this behavior is expected or if there was a change in the embedding logic between these versions.

First Bad Commit

#14217

Relevant log output

no relevant log