Self hosted OpenAI compatible model does not call tools

[magaton]

Hello, I am trying to migrate langchain agent to a langgraph and hitting a problem with locally hosted qwen-2.5 (Tabby wrapper through ChatOpenAI). The tools are not called!
The tools are executed when I use either Ollama through ChatOllama or OpenAi model through ChatOpenAI.

I have tried three options:

1. with compiling graph of nodes and ToolNode

 tool = TavilySearchResults(max_results=2)
 tools = [tool]
 llm = ChatOpenAI(model="Qwen2.5-14B-Instruct-6.0bpw-h6-exl2", api_key="asdf", base_url="http://....", verbose=True)
 llm_with_tools = llm.bind_tools(tools)
 workflow = StateGraph(State)
 workflow.add_node("chatbot", chatbot)
 tool_node = ToolNode(tools=[tool])
 workflow.add_node("tools", tool_node)
  workflow.add_conditional_edges(
            "chatbot",
            tools_condition,
   )
   workflow.add_edge("tools", "chatbot")
   workflow.set_entry_point("chatbot")
   self.graph : CompiledStateGraph =workflow.compile()       

2. with langgraph builtin: create-react-agent without state_modifier
3. with langgraph builtin: create-react-agent with state_modifier where I set a specific System prompt to instruct how to call tools

 system = """
            Answer the following questions as best you can. You have access to the following tools:
            {tool_names}

            TOOLS
            ------
            Assistant can ask the user to use tools to look up information that may be helpful in answering the users original question. The tools the human can use are:
 
            {tool_descriptions}
            
            RESPONSE FORMAT INSTRUCTIONS
            ----------------------------

            When responding to me, please output a response in one of two formats:

            **Option 1:**
            Use this if you want the human to use a tool.
            Markdown code snippet formatted in the following schema:

            ```json
            {{
                "action": string, \\ The action to take. Must be one of {tool_names}
                "action_input": string \\ The input to the action
            }}
            ```

            **Option #2:**
            Use this if you want to respond directly to the human. Markdown code snippet formatted in the following schema:

            ```json
            {{
                "action": "Final Answer",
                "action_input": string \\ You should put what you want to return to use here
            }}
            ```

        """.format(tools=tools, tool_names=tool_names, tool_descriptions=tool_descriptions)


        
        self.graph : CompiledStateGraph = create_react_agent(llm, tools, state_modifier=system, debug=True)

None of the above works and I would conclude that this particular model or tabby itself cannot call tools, but this proves not to be the case.

Old langchain agents just work, meaning the tool is properly called.

search = DuckDuckGoSearchRun()
tools = [search]
memory = ConversationBufferWindowMemory(
    memory_key="chat_history", k=5, return_messages=True, output_key="output"
)
agent = initialize_agent(
    agent="chat-conversational-react-description",
    tools=tools,
    llm=llm,
    verbose=True,
    early_stopping_method="generate",
    memory=memory,
    handle_parsing_errors=True,
)

The question I use in all of the scenarios is: "Who won US Open 2024?"

What should I do in langgraph (both in create-react-agent and classic graph with Nodes and ToolNodes) to get the tool_call working?