Why do GPU and CPU embedding outputs differ for the same input? Is normal?

[jygmysoul]

I am using the embedding example, the execution parameters are as follows
embedding.exe -ngl 200000 -m I:\JYGAIBIN\MetaLlamaModel\Llama2-13b-chat\ggml-model-f32_q4_1.gguf --log-disable -p "Hello World!"

The first three embedding values ​​are output when the CPU executes the embedding
-4.67528416e-08
-1.07059577e-06
1.76811977e-06

The first three embedding values ​​are output when the GPU (-ngl 200000) executes the embedding
5.86615059e-08
-1.02221782e-06
1.78800110e-06

Why are the same "Hello World!" inputs different? Does llama.cpp currently correctly support GPU and CPU embedding?

Also, does llama.cpp have specific instructions for underlying API functions, or usage precautions? In addition to those on github, is there any interface documentation website? Thank you

Possible Answer
I think for the same input content, the GPU and CPU output embedding values ​​should be the same

Which of these two results is correct? Do GPUs currently support computational embedding and fine-tuning?

[jygmysoul]

UP

[sprappcom]

@jygmysoul did u use the same seed?

[jygmysoul]

用的是默认的种子,应该是不一样的,这个也有影响吗?

[sprappcom]

@jygmysoul of course. different seed, diff answer