high memory footprint: image encoding / decoding done twice in pydantic-ai + google-genai ?

[charlesmindee]

Question

Hello, when calling gemini through GoogleModel on big documents with a lot of pages, I have the memory usage which is diverging.
After profiling the memory, I found out that the operation which is taking most of the memory is the encoding / decoding of images (screenshot profiling report). The 2 first lines of the report have exactly the same memory usage, and when looking into the source code both lines are doing (almost) the same thing:

pydantic_ai.models.google.py l.524 in _map_user_prompt:

elif isinstance(item, BinaryContent):
        # NOTE: The type from Google GenAI is incorrect, it should be `str`, not `bytes`.
        base64_encoded = base64.b64encode(item.data).decode('utf-8')

google.genai._common.py l.520 in encode_unserializable_types:

if isinstance(value, bytes):
        processed_data[key] = base64.urlsafe_b64encode(value).decode('ascii')

Do you think there is a redundancy here ?
Do you have any suggestions to reduce memory usage ? I am sending the same images in the user prompt of many inferences but I couldn't find a way to share the encoding / decoding of images accross the calls since this operation is done at each inference.

Thank you !

Additional Context

No response

[DouweM]

@charlesmindee Yep looks like we're doing some redundant conversion there! Let me work on a fix.