Unable to test gemini service

[psydok]

Describe the bug
gemini does not support endpoints with /v1/..., and there is no way to override it.

Expected behavior
is it possible to add the ability to redefine endpoints for certain services/models?

Environment
Include all relevant environment information:

1. OS [e.g. Ubuntu 20.04]: -
2. Python version [e.g. 3.12.2]: 3.10

To Reproduce
I've tried using the gemini documentation for load testing via the guideline: https://ai.google.dev/gemini-api/docs/openai

export GUIDELLM__PREFERRED_ROUTE="chat_completions" &&
      export GUIDELLM__OPENAI__MAX_OUTPUT_TOKENS=512 &&
      export GUIDELLM__MAX_CONCURRENCY=233 &&
      export GUIDELLM__REQUEST_TIMEOUT=300 &&
      export GUIDELLM__PREFERRED_PROMPT_TOKENS_SOURCE=local &&
      export GUIDELLM__PREFERRED_OUTPUT_TOKENS_SOURCE=local &&
      guidellm benchmark
      --target https://generativelanguage.googleapis.com/v1beta/openai/
      --rate-type sweep --rate 5
      --model gemini-2.5-flash --processor ${processor}
      --random-seed 1234
      --max-requests=100
      --data "prompt_tokens=4096,output_tokens=512"
      --backend-args '{"extra_body":{"chat_template_kwargs":{"enable_thinking":false}}, "headers":{"Authorization": "Bearer GEMINI_API_KEY"}}'
      --output-path "${output_dir}/benchmarks.json"

Errors
The error is that guidellm could not get a list of available models by endpoint /v1/models

Additional context
Add any other context about the problem here. Also include any relevant files.

[psydok]

@sjmonson
I wrote adapter to gemini to test the service through guidellm. But I noticed that gemini/openai-gpt do not support ignore_eos.
Can you tell me if there is any way to configure the generation of prompta in guidellm so that there is some kind of instruction at the beginning of the text (for example, do not generate eos, as in llmperf)?

[sjmonson]

If the model server does not support ignore_eos or an equivalent like min_tokens then there is no way to get an exact number of tokens from the server in a single request. I have never heard of llmperf but a model cannot just decide to not send eos tokens based on user prompting.

[markurtz]

Noting here that the 0.4 release will have fully customizable pathways for the request formats that are sent. The default will still run as is, but you'll have workarounds with a bit of work. Adding to 0.4 milestone for when that is released