[LiteLLM] Gemma models via Ollama enter infinite tool-calling loop due to wrong tool message role

[jfrometa-tlsi]

When using Gemma models (2B/4B) via Ollama through the LiteLLM adapter, agents with tools enter an infinite tool-calling loop and never produce a final response.

Root cause: _content_to_message_param serializes tool result messages with role="tool" (OpenAI-compatible default), but Gemma's chat template expects role="tool_responses" (according documentation: https://ai.google.dev/gemma/docs/core/prompt-formatting-gemma4).

This mismatch causes the model to misinterpret the tool result as a new turn instead of a response to its own tool call.

This is not a hardware or quantization issue — the same behavior occurs on high-end GPUs.

[sanketpatil06]

Hi @jfrometa-tlsi , To help us look into this, could you please share a minimal code example to reproduce the issue and let us know which ADK version you are using? Additionally, if you aren't already on the latest version (1.33.0), we highly recommend upgrading. Thanks.

[jfrometa-tlsi]

Hi, thanks for the quick response!

Here's the minimal reproduction info:

ADK version: [1.32.0]

---

Setup

from google.adk.models.lite_llm import LiteLlm

model = LiteLlm(
    model="ollama_chat/gemma4:e2b",
    api_base="http://localhost:11434",
    temperature=0.1,
    num_ctx=8192,
    num_predict=512,
    repeat_penalty=1.1,
)

agent = LlmAgent(
    name="orchestrator",
    model=model,
    instruction="...",
    tools=[my_tool],
)

runner = Runner(agent=agent, app_name=APP_NAME, session_service=session_service)

async for event in runner.run_async(
    user_id=user_id,
    session_id=session.id,
    new_message=query_content,
):
    ...

---

Root cause (confirmed via LiteLLM request logs)

When ADK serializes tool result messages, it uses role: "tool" (OpenAI format). However, Gemma 4's chat template expects role: "tool_responses" as documented here: https://ai.google.dev/gemma/docs/core/prompt-formatting-gemma4

This is visible in the raw request LiteLLM sends to Ollama:

{ "role": "tool", "content": "{...tool result...}", "images": [] }

Gemma's template does not recognize role: "tool", so it treats the tool result as a new user turn instead of a response to its own tool call. This causes the model to call the tool again indefinitely — an infinite loop.

This is not a hardware or quantization issue. Reproduced on high-end GPUs with multiple Gemma variants (2B, 4B).

---

Local fix applied (also submitted as a PR)

In _content_to_message_param, detecting when the target model is a Gemma variant served via Ollama and remapping role: "tool" → role: "tool_responses" resolves the loop entirely. Happy to discuss the approach in the PR if needed.

Let me know if you need anything else to reproduce it on your end!