Open jhyeom1545 opened 3 months ago
We also don't know the specific reason. But the difference is very small and it does not influence performance.
@staoxiao Thanks for your comment. I don't know exactly why, but the result of calculating the embeddings on different servers using the same GPU model seems to be the same.
I am using flagbedding well. Among them, I found something I don't understand. If the same sentence is inferred from different GPUs on different servers, the value of the embedding vector is different.
example
The value changes from the 8th digit of the prime number, so if you know why, I would appreciate it if you could let me know.