The Ascend 310 P chip is very slow in embedding for the first time, and fast in the second and later.

[zhaowenbin]

This problem does not occur on the 910B, but occurs on the 310P.
Some codes are as follows:

from FlagEmbedding import FlagModel, FlagReranker
......
model = FlagModel(base_model_path, use_fp16=True)
......
// That's the step. It's slow for the first time, 50 or 60 seconds for the first time, and then 0.0 seconds for the next.
query_embedding = model.encode(text_data) 

Is it because there's no adaptation?

[hanhainebula]

Hello, @zhaowenbin. This is normal since initializing multiple devices (refer to here) need some time. If the size of your text_data is small, you can add parameter devices="..." to use only one GPU to save the time for initializing multiple devices. The modified code:

from FlagEmbedding import FlagModel, FlagReranker
......
model = FlagModel(base_model_path, use_fp16=True, devices="...")  # `...` need to be replaced
......
// That's the step. It's slow for the first time, 50 or 60 seconds for the first time, and then 0.0 seconds for the next.
query_embedding = model.encode(text_data)