IssueTracker

Problem you have encountered: When using the vertexai interface for gemini from python, it is possible to specify response_logprobs (boolean indicating whether to return logprobs) and logprobs (number of logprobs to return per generated token, on [0,5]). gemini-1.5-flash does this, but gemini-2.0-flash returns only the first few tokens and then quits.

What you expected to happen: I expect the model to return one chosen_logprobs and top_logprobs entry per generated token in gemini-2.0-flash.

Steps to reproduce:

from vertexai.generative_models import GenerationConfig

generation_config = GenerationConfig(
    logprobs=5,
    response_logprobs=True,
)

client = GenerativeModel("gemini-2.0-flash")
response = await client.generate_content_async("What is the capital of Australia, and what territory is it in?", generation_config=generation_config)
print(len(response.candidates[0].logprobs_result.chosen_candidates))  # 6, not even up to Canberra
print(len(response.candidates[0].logprobs_result.top_candidates))  # 6

If I replace gemini-2.0-flash with gemini-1.5-flash, I instead get logprobs for the full response.

Other information (workarounds you have tried, documentation consulted, etc): I'm using the grpc transport method on Python 3.10. Here are some library versions:

google-ai-generativelanguage  0.6.15
google-api-core               2.24.1
google-api-python-client      2.161.0
google-auth                   2.38.0
google-auth-httplib2          0.2.0
google-cloud-aiplatform       1.81.0
google-cloud-bigquery         3.29.0
google-cloud-core             2.4.2
google-cloud-resource-manager 1.14.1
google-cloud-storage          2.19.0
google-crc32c                 1.6.0
google-generativeai           0.8.4
google-resumable-media        2.7.2
googleapis-common-protos      1.68.0
grpc-google-iam-v1            0.14.0
grpcio                        1.70.0
grpcio-status                 1.70.0
httpcore                      1.0.7
httplib2                      0.22.0
httpx                         0.28.1